Sei sulla pagina 1di 39

Energy Conversion and Management 156 (2018) 459–497

Contents lists available at ScienceDirect

Energy Conversion and Management


journal homepage: www.elsevier.com/locate/enconman

Review

Solar photovoltaic generation forecasting methods: A review T


a a,⁎ a,b
Sobrina Sobri , Sam Koohi-Kamali , Nasrudin Abd. Rahim
a
UM Power Energy Dedicated Advanced Centre (UMPEDAC), Level 4, Wisma, R&D UM, Jalan Pantai Baharu, 59990 Kuala Lumpur, Malaysia
b
Renewable Energy Research Group, King Abdulaziz University, Jeddah 21589, Saudi Arabia

A R T I C L E I N F O A B S T R A C T

Keywords: Solar photovoltaic plants are widely integrated into most countries worldwide. Due to the ever-growing utili-
Solar photovoltaic zation of solar photovoltaic plants, either via grid-connection or stand-alone networks, dramatic changes can be
Renewable energy power plant anticipated in both power system planning and operating stages. Solar photovoltaic integration requires the
Modelling and planning capability of handling the uncertainty and fluctuations of power output. In this case, solar photovoltaic power
Spatial and temporal horizons
forecasting is a crucial aspect to ensure optimum planning and modelling of the solar photovoltaic plants.
Smart grid forecasting
Accurate forecasting provides the grid operators and power system designers with significant information to
design an optimal solar photovoltaic plant as well as managing the power of demand and supply. This paper
presents an extensive review on recent advancements in the field of solar photovoltaic power forecasting. This
paper aims to analyze and compare various methods of solar photovoltaic power forecasting in terms of char-
acteristics and performance. This work classifies solar photovoltaic power forecasting methods into three major
categories i.e., time-series statistical methods, physical methods, and ensemble methods. To date, Artificial
Intelligence approaches are widely used due to their capability in solving the non-linear and complex structure of
data. The performance analysis shows that these methods outperform the traditional methods. Recently, the
ensemble methods were also developed by researchers to extract the unique features of single models to enhance
the forecast model performances. This combination produces accurate results compared to individual models.
This paper also elaborates on the metrics assessment which was implemented to evaluate the forecast model
performances. This work provides information which is beneficial for researchers and engineers who are in-
volved in the modelling and planning of the solar photovoltaic plant.

1. Introduction International Energy Agency (IEA), are supposed to generate 196 GW


(in most grid-connected plants) by the end of 2015. An additional 40
Global warming and the critical depletion of fossil fuel over the past nations that were excluded from the IEA Photovoltaic Power System
decades have encouraged the use and development of renewable energy Programme (IEA PVPS) produced about 31 GW of solar power. Fig. 1
sources (RES). Renewable energy sources e.g., solar, wind, hydropower, illustrates the evolution of global solar PV installation from 2000 to
and geothermal energy have not only been acknowledged as novel so- 2015. Solar PV installation for both IEA PVPS and other countries has
lutions to the issues listed above but also reflect the future of energy increased dramatically from 2007 to 2015. About 70% of solar PV in-
advancement. In substituting conventional sources, solar energy has stallation came from IEA PVPS countries [4]. In early 2016, 120 solar
emerged as a most popular approach and is implemented in many na- PV plants with a capacity of more than 50 MW operated in at least 23
tions worldwide compared to others. Solar energy becomes the most countries i.e., Philippines, Uruguay, Pakistan, Kazakhstan, Honduras,
promising source for generating power for residential, commercial, and Guatemala, Denmark, and Australia [3]. The Global Future Report 2013
industrial applications. Solar photovoltaic (PV) systems use PV cells Renewable Energy Policy Network for the 21st Century (REN21) has
that convert solar irradiation into electric power [1]. Solar PV is used in projected that the capacity of global solar PV has the potential to reach
stand-alone and grid-connected systems to supply power for home ap- 400–800 GW by 2020 [2].
pliances, lighting, and commercial and industrial equipment [2]. Malaysia is one of the Asian countries that receives high solar ra-
In fact, the number and the size of the solar PV plants have grown diation. The availability of solar energy is highly influenced by climatic
rapidly at a worldwide level, due to their essential role in generating conditions over the year. Malaysia is located on the South China Sea
electricity [3]. Several nations that are in collaboration with the and lies between 1° and 7° on the North latitude and 100° and 120° on


Corresponding author.
E-mail addresses: sam.koohi@gmail.com, sam.koohi@ieee.org (S. Koohi-Kamali).

https://doi.org/10.1016/j.enconman.2017.11.019
Received 26 May 2017; Received in revised form 1 November 2017; Accepted 6 November 2017
Available online 01 December 2017
0196-8904/ © 2017 Elsevier Ltd. All rights reserved.
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

Nomenclature GCPV Grid-connected Photovoltaic


GEM Global Environmental Multi-scale
ACE Average Coverage Error GFM Generalized Fuzzy Model
AGC Automatic Gain Control GFS Global Forecast System
AI Artificial Intelligence GHI Global Horizontal Irradiance
AnEn Analog Ensemble GIcsk Clear Sky POA irradiance
ANFIS Adaptive Neuron Fuzzy Inference System GLSSVM Group Least Square Support Vector Machine
ANN Artificial Neural Network GMDH Group Method of Data Handling
AR Autoregressive GMM Gaussian Mixture Method
ARMA Autoregressive-Moving Average GP Genetic Programming
ARIMA Autoregressive Integrated Moving Average GRNN General Regression Neural Network
ARIMA-BP Autoregressive Integrated Moving Average-Back-propa- GSO Genetical Swarm Optimization
gation GW Gigawatt
ARX Autoregressive with Exogenous Input H Horizontal surface
AWS Automated Weather Stations HH Time in Hours
BCRF Bias Compensation Random Forest HMC Higher-order Markov Chain
BLUE Statistical model of Blue Sky HPANN Hybrid Physical Artificial Neural Network
BP Back-Propagation HSV Hue-Saturation-Value
BPNN Back-propagation Neural Network Ho extra-terrestrial global solar radiation on a horizontal
Bagging-BP Bagging-Back-propagation surface
BSRN Baseline Surface Radiation Network I Irradiance
BS Brier Score ICP Internal Coverage Probability
BSS Brier Skill Score Ics Clear Sky Irradiance
Cap cloud-advection-versus-persistence, cap IEA International Energy Agency
CARDS Coupled Autoregressive and Dynamical System IS Interval sharpness
CC Coefficient Correlation Io Solar constant
CDF Cumulative Distribution Function KM Kernel Methods
CFB Cascaded Feed-forward Back-propagation kNN k-Nearest Neighbour
CDHMM Continuous Density Hidden Markov Model KSI Kolmogorov-Smirnov test Integral
CMV Cloud Motion Vector kt∗ Forecast clear sky index
CRPS Continuous Ranked Probability Score L Loss function
CS Cuckoo Search LES Linear Exponential Smoothing
CSKY-Glo Clear Sky Global Horizontal Radiation LM-BP Levenberg-Marquardt Back-Propagation
CS-OP-ELM Cuckoo Search-Optimally Pruned-Extreme Learning LOO-CV Leave-Out-Outcross Validation
Machine LSS Large Scale Solar
CSRM Clear Sky Solar Radiation Model LS-SVM Least Square Support Vector Machine
CVRMSE Coefficient of Variance based on Root Mean Square Error LVQ Learning Vector Quantization
C1 Custom Network-1 MA Moving Average
C2 Custom Network-2 MAE Mean Absolute Error
DAS Data Acquisition System MAD Mean Absolute Deviation
DBN Deep Belief Network MAID Mean Absolute Internal Deviation
DCNN Deep Convolutional Neural Network MAPE Mean Absolute Percentage Error
DD Number of Day MARE Mean Absolute Relative Error
DDM Data-Driven Model MARS Multivariate Adaptive Regression Splines
DH Diffuse Horizontal MASE Mean Absolute Scaled Error
DHI Diffuse Horizontal Irradiance MaxAE Maximum Absolute Error
DHR Dynamic Harmonic Regression MBE Mean Bias Error
DN Direct Normal Radiation MdAPE Median Absolute Percentage Error
DNI Direct Normal Irradiance ME Systematic Error
DNN Deep Neural Network MeAPE Median Absolute Percentage Error
DP BP due point temperature MEF Mean Error Function
EA Evolutionary Algorithm MLP Multi-layer Perceptron
EBP Evidence-Based Practise MM Number of months
ECMWF European Centre for Medium-Range Weather Forecasts MM-MOS MOS by the weather company Meteomedia GmBH
Elman-BPNN Elman Back-propagation Neural Network MOS Model Output Statistical
ELM Extreme Learning Machine MPE Mean Percentage Error
ESS Exponential Smoothing State MRE Mean Relative Error/Magnitude Relative Error
ETS Exponential Trend Smoothing MRSR Multi-Response Sparse Regression
FA Fuzzy ARTMAP MSE Mean Square Error
FF Firefly MW Megawatt
FFB Feed-forward back propagation NAM North American Mesoscale Forecast System
FFN Feed-forward Network NAR Non-Linear Auto-regressive
FFNN Feed-forward Neural Network NARX Non-Linear Auto-regressive Network with exogenous
GA Genetic Algorithm input
GBRT Gradient Boosted Regression Trees NCAR National Center for Atmospheric Research

460
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

NCEP National Centres for Environmental Prediction RRMSE Relative Root Mean Square Error
NMAE Normalized Mean Absolute Error RSKY-Glo Real Sky Global Horizontal Radiation
NMBE Normalized Mean Bias Error RW Random Walk
NN Neural Network Rh relative humidity
NOAA National Oceanic and Atmospheric Administration SAE Stacked auto-encoder
NOCT Nominal Operating Cell Temperature SARIMA Seasonal Autoregressive Integrated Moving Average
NRMSE Normalized Root Mean Square Error SDE Standard Deviation Error
NST New Straits Time SEDA Sustainable Energy Development Authority
NWP Numerical Weather Prediction SES Simple Exponential Smoothing
ǹ Sunshine duration SMA Seasonal Moving Average
nE Normalized Error SOM Self-Organizing Map
n/N relative sunshine duration SPP Solar Power Prediction
nRMSE Relative Root Mean Square Error SREF Short-Range Ensemble Forecast
OBS Optimal Brain Surgeon SS Skill Score
OLSs Ordinary Least Squares ST Statistical Model
OP Operating Point STI Successful Tracking Index
PCA Principal Component Analysis SURFRAD Seven Surface Radiation Budget Network
PDF Probability Distribution Function SVM Support Vector Machine
PDM Pattern Discovery Method SVM-FFA Support Vector Machine-Firefly Algorithm
PFLRM Partial Functional Linear Regression Model SVR Support Vector Regression
PHANN Physical Hybrid Artificial Neural Network SZA solar zenith angle
PM Persistence Model TB K-Means Transformation Based K-Means
PNN Probabilistic neural network TCWB Taiwan Central Weather Bureau
POA Plane-of-array TNB Tenaga Nasional Berhad
PR Pressure TSI Total Sky Imager
PSO Particle Swarm Optimization Tavg average ambient temperature
PSS Persistence Tmax maximum temperature
PV Photovoltaic Tmax,ambient maximum ambient temperature
PvPP Photovoltaic Power Production Tmin minimum temperature
PVPS Photovoltaic Power System Programme Tmin,ambient minimum ambient temperature
QR Quantile Regression UCSD University of California, San Diego
R/r statistical correlation coefficient USI University of California San Diego Sky Imager
R2 coefficient of determination V Solar Variability
RBF Radial Basis Function Vp water vapour pressure
RBFNN Radial Basic Function Neural Network WD Wavelet Decomposition
RBN Radial Basis Neural Network WkNN Weighted k-Nearest Neighbour
RBR red-blue-ratio WMAE Weight Mean Absolute Error
RE Relative Error WNN Wavelet Neural Network
REN21 Renewable Energy Policy Network for the 21st Century WRF Weather Research and Forecasting model
RF Random Forests WRF-MT Weather Research and Forecasting-Meteosat
RGB red-green-blue WS Wind Speed
RMAE Relative Mean Absolute Value WSI Whole Sky Imager
rMBE Relative Mean Biased Error WT Wavelet Transform
RME Mean Relative Error WT-BP Wavelet Transform-Back-Propagation
RMSE Root Mean Square Error XGB Extreme Gradient Boosting Machines
RNN Recurrent Neural Network ρ correlation coefficient
ROC Relative Operating Characteristic σ2 error variance
ROCSS Relative Operating Characteristic Skill Score

the East longitude with an average solar radiation of 15 MJ/m2 per day. Solar energy is a sustainable resource amongst other RES [8]. The in-
The solar energy resources in Malaysia are four times higher than the tegration of solar PV plants into power grids has received much at-
world’s fossil fuel as reported by [5,6]. Nowadays, about 90% of Ma- tention due to its ability in generating electric power. Solar plants are
laysia’s electricity generation relies on fossil fuels. This dependency for widely used in smart grids. The implementation of large scale grid-
a long-term period is not a secure option. This mandates the govern- connected solar PV plants has shown significant issues to the power
ment of Malaysia to encourage the use of RES to replace the fossil fuel networks such as system stability, reliability, electric power balance,
by solar energy. The government of Malaysia has promoted the use of reactive power compensation and frequency response [1,9]. Solar PV
renewable energy by implementing several policies and acts i.e., Na- power forecasting has emerged as a brilliant way to address these is-
tional Green Technology Policy 2009, Sustainable Energy Development sues.
Authority Act 2011, and Renewable Energy Act 2011. Malaysia targets Modelling stage includes sizing, designing, and simulating the solar
a capacity of approximately 18,700 MW in terms of solar PV integration PV plants. Many researchers have focused on the optimization of solar
by 2050 [7]. PV power generation in terms of the number of PV modules, storage
Renewable Energy Sources (RES) have dynamic nature and highly and inverter capacity, and controller types [10]. This can improve the
depend on geographical locations and weather conditions. Integrating operation of renewable energy based power grids by proper energy
large scale RES into existing energy systems becomes a major challenge. storage scheduling [11]. In solar PV plant, sizing is a crucial part of the

461
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

Fig. 1. Evolution of solar PV installations (GW) from 2000 until 2015 [4].

system design. The modelling of solar PV plants does not only focus on possess the ability to combine both linear and non-linear approaches.
sizing but also it is important for optimal planning of PV power plants Therefore, the ensemble methods outperform individual approaches.
under variable meteorological conditions [8]. The intermittency of the This paper can provide the researchers and engineers who work in
solar energy resources could result in dramatic problems in balancing the area of solar PV plants modelling and planning with useful in-
between the power generation and load demand. It is necessary to formation about forecasting techniques. The readers can easily gain
schedule solar PV plants in the planning stage to closely match gen- information about the methods since each method together with their
eration with the predicted load profile. The reliable forecast informa- characteristics, merits and demerits are discussed and summarized. The
tion on the solar sources assists grid operators to anticipate a case of rest of this paper is organized as follows: Section 2 describes the
shortage or abundance of solar power. Accurate solar forecasting methods used in solar PV power forecasting, factors influencing solar
eliminates the impact of solar PV output uncertainty, improves the PV power forecasting, and the table of summarization (see Table 2).
stability of the system, increases the penetration level of the PV system, Section 3 discusses the metrics assessment of solar PV power forecasting
and reduces maintenance cost of ancillary devices [12]. In order to to compute and analyze the performance of the forecast models. Fi-
meet requirements of the decision-making process of modelling, the nally, Section 4 concludes the work.
forecasting models strongly rely on the spatial and temporal resolu-
tions, meteorological variability, selection of input parameters, and 2. Solar photovoltaic power forecasting techniques
training algorithms. The training modes will give effect to the fore-
casting performance. Two types of training methodologies i.e., online This section briefly describes the various solar forecasting methods.
and offline are applied to eleven-data driven models in order to eval- The solar forecasting techniques were classified into three major models
uate the fitness and flexibility of the forecast models performances as i.e., time-series statistical, physical and ensemble methods.
presented in [11]. The solar PV power forecasting method could be
deployed to optimize the usage of solar energy. Thus, solar PV power
2.1. Time series statistical methods
forecasting is a strong tool that helps the system operators and power
system designers to model and manage the solar PV plants efficiently
Time-series is a sequence of observations on a parameter measured
[13].
at successive points in time [15]. The statistical methods were used to
The state-of-the-art of this paper is to review, classify, and compare
reconstruct the relations between the past meteorological parameters
the solar PV power forecasting techniques. The purpose of this review
and hourly irradiance [16]. The statistical model does not require the
article is to analyze and compare different solar PV power generation
internal states information of the system to model it [17].
forecasting techniques. There are three major methods i.e., statistical-
time series methods, physical methods, and ensemble methods. The
statistical approach is comprised of five sub-models i.e., (i) Artificial 2.1.1. Artificial Neural Network
Neural Network (ANN), (ii) Support Vector Machine (SVM), (iii) Artificial Neural Network (ANN) is known as Artificial Intelligence
Markov Chain, (iv) Autoregressive, and (v) Regression models. These (AI) system that reflects a human brain. The ANN modelling tools can
statistical methods are highly dependent on the historical data, ease of solve the complex of non-linear data, data arranging, pattern recogni-
implementation, and the ability to extract the past data to forecast time- tion, simulation, and optimization. The accuracy of a forecast using
series. Artificial Intelligence (AI) approaches such as Artificial Neural ANN models relies on input parameters, training algorithm, and
Networks (ANNs) are also considered as statistical methods. The phy- structure configuration [18]. The complex nature of most practical
sical methods consist of three sub-models i.e., (i) Numerical Weather problems can be effectively solved via ANN in solving non-linear and
Prediction (NWP), (ii) Sky Imagery, and (iii) Satellite-Imaging models. complex data without a prior assumption of the nature of these corre-
These physical models depend on the interaction between the physical lations. The most critical phase for the implementation of ANN is the
state and the dynamic motion of solar radiation that takes place in the training phase [19]. Fig. 2 shows the basic configuration of ANN that
atmosphere. The ensemble method refers to any combination of sta- consists three layers known as input layers, hidden layers, and output
tistical or physical methods. This concept is to mix different models layers [20]. The modelling of ANN can be classified into three stages:
with unique features to address the limitation of an individual model (1) The design stage consists of ANN type, number of neurons in each
thus enhancing the forecast performance [14]. These techniques also layer, number of layers, input and output parameters, training, and
validation of selected datasets, (2) training phase takes place when the

462
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

August. The RMSE values show a range between of 33–55 W and


37–63 W for measurement and testing, respectively [24].
Kardakos et al., proposed a time-series of Seasonal Autoregressive
Integrated Moving Average (SARIMA) implemented with solar radia-
tion prediction derived from Numerical Weather Prediction (NWP)
model, and ANN with multiple inputs, to predict GCPV power genera-
tion plant for both intraday and day-ahead on an hourly basis. The
Normalized Root Mean Square Error (NRMSE) was used to evaluate the
proposed model performance. The findings show that the average va-
lues for NRMSE at an annual basis were 11.12% (Modified SARIMA),
11.26–11.46% (ANN), 12.89% (Pure SARIMA), and 13.71%
(Persistence model) [25].
Chen et al., developed the Fuzzy and Neural Network (NN) method
to predict solar radiation using weather and solar radiation data prior to
a month. The historical data such as hourly solar irradiation, sky con-
ditions, and average hourly temperature were chosen as input for the
NN. The fuzzy logic classifies various sets of temperature, sky condi-
tions, and data which were integrated into NN to predict solar radia-
tion. The proposed model was validated under various sky conditions
i.e., sunny, cloudy, and rainy days, and a combination of these three sky
Fig. 2. Basic ANN structures [20]. conditions. The simulation results show that the Mean Absolute
Percentage Error (MAPE) values obtained for the proposed model were
data were presented to ANN and the weight was modified until a pre- 6.03–9.65% for each case, respectively, which was demonstrated better
determined condition was satisfied and finally, (3) validation phase, accuracy than the other methods [26].
where the ANN model was built using unseen data, excluding those Mellit et al., proposed two ANN models i.e., ANN-1 and ANN-2, to
used in the training phase. Upon successful validation test, the model is forecast solar power generated by 50 Wp Si-Polycrystalline PV modules
ready to perform its designed function or further modification was re- with solar irradiance and air temperature as the input variables. Model
quired in the previous stage as indicated in Fig. 3. ANN-1 predicts solar PV power in cloudy cases with an average daily
The Multilayer Perceptron (MLP) structure has emerged as the fa- solar irradiation less or equivalent to 400 W m−2/day. Model ANN-2
vourite form of ANN. This MLP structure contains an input layer, one or predicts sunny-day cases with its average solar irradiation exceeding
more hidden layer, and an output layer. The input layer and output 400 W m−2/day. The developed models were simulated using the
layer represent vectors x and y , respectively. Meanwhile, the hidden MATLAB software program. As a result, ANN-2 shows a performance
layer was characterized as a number of neurons. The non-linear func- enhancement with MBE values ranging from 0.94% to 0.98% and Root
tion of the tangent hyperbolic function f (x ) , also known as activation Mean Square Error (RMESE) less than 0.2%. In comparison with other
functions as shown in Eq. (1): approaches i.e., one-diode, analytical, polynomial regression, and
multiple linear regressions methods, the ANN models demonstrate the
e x −e−x
f (x ) = lowest MRE of 2.5% and 2.3% for the cloudy day and sunny-day, re-
e x + e−x (1)
spectively [27].
Therefore, the Neural Network (NN) is a non-linear mapping be- Notton et al., proposed three ANN models to forecast solar radiation
tween input (vector x ) and output (vector y ) with the given inputs (d ) , on a tilted plane. The solar irradiation prediction was carried out by
hidden neurons (h) , and single linear output correlations as shown in using 5 years of 10-min solar radiation data obtained from the
Eq. (2):
h d
⎡ ⎛ ⎞⎤
y = y (x ;w ) = ∑ ⎢wj f ⎜∑ wji. x i⎟ ⎥
j=0 ⎣ ⎝ i=0 ⎠⎦ (2)

The parameters of NN model were denoted as weights while the


connection between the layers was termed biases i.e., designated by
subscripts i = 0 and j = 0 . The NN parameters were designated by
vector w that supervises the non-linear mapping [22].
Mellit et al., proposed an ANN-MLP architecture by using the
Levenbeg-Marquardt Back-Propagation (LM-BP) algorithm to predict
solar irradiance based on 24-h ahead for GCPV plant in Trieste, Italy.
The input parameters applied were air temperature, daily solar irra-
diance, and day of the months. The simulation results demonstrated
that the metric performances of correlation coefficient were 98% and
94% for sunny and cloudy days, respectively. Meanwhile, the Mean
Absolute Error (MAE) and Mean Bias Error (MBE) values were 3.21%
and 8.54% for the forecast and actual Grid-Connected Photovoltaic
(GCPV) plant power output, respectively [23].
Izgi et al., developed an ANN to forecast solar power derived from a
750 W solar PV panel. A comparison between measurement and pre-
diction values of ANN was carried out with correlation coefficient and
RMSE. As a result, the best forecast of solar power for short-term and
mid-term forecast horizons was 5 min and 35 min, respectively, in
Fig. 3. The modelling process flowchart for ANN [21].
April. Nonetheless, the results were extended to 3 min and 40 min in

463
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

Mediterranean region of Ajaccio located in France. For the first and isotropic model due to the investigation of isotropic and anisotropic sky
second models, five input parameters were selected i.e., declination, conditions for total radiation incident on the tilted surface [32].
time, zenith angle, 10-min extra-terrestrial horizontal irradiation, and Giorgi et al., predicted power from a 960 kWp GCPV plant im-
10-min horizontal global irradiation. Meanwhile, an extra parameter plemented in Italy using statistical methods based on multi-regression
was added to the third model i.e., inclination angle parameter. The analysis and Elman ANN for 1–24-h ahead. Several time-series data
results indicate that elimination of one parameter improves the accu- were applied as inputs i.e., PV power, module temperature, ambient
racy of RMSE and Relative Mean Absolute Error (RMAE) values at an temperature, and solar irradiance for tilt angles of 3° and 15°, respec-
approximate of 9% and 5.5%, respectively [28]. tively. Three different input vectors were developed for Elman ANN as
Yona et al., developed the PV power output forecasting based on input for the training dataset. As a result, the probability of prediction
insolation prediction for 1-day ahead by utilizing weather reported error (Ei) in the range of −10% to 10%, declined drastically as the time
data, fuzzy theory, and Recurrent Neural Network (RNN). The insola- horizon was increased. By taking into account all the weather para-
tion forecast data generated by the fuzzy model functions to train the meters as input vectors, it provides the best prediction of PV power
RNN smoothly. From the simulation results, the forecast error of MAE measurement [12].
shows an average value of 0.1327 kW. This forecasting error indicates Teo et al., proposed an ANN model based on Extreme Learning
the lowest error in comparison with other methods i.e., persistence Machine (ELM) training algorithms to forecast solar PV power. Three
method, only using fuzzy theory, a combination of fuzzy theory and datasets were tested with a set of 100 hidden nodes. The activation
correction method, and the combination of fuzzy theory, correction function was used as a logistic function due to its best performance, in
method, and feed forward neural network [29]. comparison with a hyperbolic function. This model was evaluated by
Amrouche et al., integrated a new methodology for the daily pre- using the RMSE metric. From the simulation 1 analysis, the values of
diction of local Global Horizontal Irradiance (GHI) using a combination RMSE and standard deviation were 3.8564 and 0.0798, respectively,
method of spatial modelling and ANN. The new method was tested at compared to the initial simulation which resulted in RMSE and stan-
two locations i.e., French Alps and South France. The ANN-based dard deviation values of 4.1096 and 0.0946, respectively. Modifying
method was integrated to estimate the local GHI based on daily weather input variables and increasing training data led to great improvements.
forecasts derived from US National Oceanic and Atmospheric (US The sequence of variables also gives impact on model performance
NOAA) for four adjacent locations. The new model was compared to [33].
geometric and statistical models via Mean Square Error (MSE) and Kashyap et al., compared eight various NN models i.e., Feed-for-
RMSE assessments. The findings represent that the sixth validation in- ward Network (FFN), Feed-forward Back-propagation (FFB), Radial
dicates the highest MSE and RMSE values with 53.9693 W/m2 and Basis Neural Network (RBN), Probabilistic Neural Network (PNN),
92.5523 W/m2, respectively. The lowest MSE and RMSE values were Elman Back-propagation Neural Network (Elman-BPNN), Cascaded
recorded at the second validation with 16.4593 W/m2 and 33.10 W/m2, Feed-forward Back-propagation (CFB), Custom Network-1 (C1), and
respectively. The developed ANN-based model shows better accuracy Custom Network-2 (C2) to evaluate the most relevant and independent
compared to statistical and geometric models [21]. parameters of hourly global radiation in India. The selected training
Almonacid et al., proposed a new methodology to predict 1-h ahead parameters were direct normal radiation, diffuse horizontal irradiance,
of the power PV generator output based on the dynamic ANN model. wind direction, wind speed, pressure, bulb point temperature, due point
Two major atmospheric parameters i.e., solar irradiance and air tem- temperature, time in hours, number of days, and number of months.
perature were applied due to their impact on the performances of These parameters show the advancement of ANN structure, delay,
structures and PV generator losses. This technique was measured by neuron, and transfer function. The comparison of accuracy was per-
using the linear regression analysis, in which the actual and forecast formed by using the results obtained from the various parameters and
data were compared. The findings from the analysis show that all global models selected. As a result, it was found that delay, neuron, transfer
irradiance and air temperature level ranges perform effectively with function, model parameters, and error RMSE values were 15 or 30, 10
Correlation Coefficient close to 1, RMSE equal to 3.38%, and Relative or 20, tansig, ELM, BP or DN, and 9–10 with 25–35% of training in the
Error (RE) around 0 with a margin of error ± 8% [30]. tested cases [34].
Dahmani et al., applied an MLP-ANN model to forecast the tilted Cornaro et al., proposed two ANN-based models to forecast 24-h
global solar irradiation derived from the horizontal data gathered from ahead hourly solar irradiance. The first model was Statistical Model
Bouzareah in Algeria. Several parameters were used as inputs such as (ST) based on ANN that applied ground measurement data for fore-
horizontal global, extra-terrestrial 5-min irradiation, declination, zenith casting. The second model was Model Output Statistics (MOS) based
angle, and azimuth angle to determine the accuracy of ANN. This model ANN that corrected NWP. These two models were based on ANN and
was integrated with 2 years of solar radiation data. It can be concluded were generated by a master optimization process to optimize the neu-
that, with 4 input parameters i.e., declination, extraterrestrial, zenith rons number found in the hidden layer and to choose a suitable en-
angle, and global horizontal solar radiation is sufficient to forecast 5- semble of ANN. Data gathered for a year were analysed in ESTER
min global solar irradiation on a 36.8 tilted plane. It was discovered Laboratory located at the University of ROME for model training and
that the relative RMSE value was approximately 8.81% for optimum validation. The performances of ST, MOS, and NWP models were
configuration which presented an accurate performance within a short- compared to the Persistence Model (PM) model by using ANN MATLAB
term period [31]. Toolbox software. The results show an improvement of RMSE with
Kaushika et al., forecasted the direct, diffuse, and global solar ra- values 10% for ST, 15–25% for NWP, and 30% for MOS with respect to
diations on both horizontal and inclined surfaces by using ANN PM model. The MOS model presents the greatest RMSE improvement
methods. Eight variables i.e., latitude, longitude, the altitude of loca- compared to the PM model [16].
tion, the month of the year, local mean time, monthly mean hourly total Zhang et al., analysed the solar power prediction challenges and
rainfall, monthly mean hourly relative humidity, and monthly mean then presents a similar day-based prediction tool to develop 1-day
durations of the sunshine per hour, were chosen as input parameters. ahead estimation for small-scale solar power output prediction. The
The mapping between climatic measurement data functions as numeric data is obtained from solar PV systems in three different locations
inputs while atmospheric clearness indices for the beam, diffuse, and namely San Diego (USA), Braedstrup (Denmark), and Catania (Italy).
global radiations act as numeric target variables in the MATLAB ana- The proposed technique consists of the SDD engine and forecasting
lysis. The developed ANN model exhibits an excellent compatibility engine. The SDD engine determines historical days that are similar to
with an overall RMSE and MBE value of 5.19% and −0.914%, re- the target day. The recorded power values of those days were used to
spectively, for global radiation. The results indicate a favourable predict the power output for the target forecast day. The Radial Basis

464
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

Function Neural Network (RBFNN), Least Square Support Vector f (x ) = wφ (x ) + b (3)


Machine (LS-SVM), k-Nearest Neighbour (kNN), and Weighted k- n
Nearest Neighbour (WkNN) were integrated as the forecasting engine. 1 1
RSVMs =
2
‖w‖2 + C
N
∑ L (x i,di )
From the numerical results analysis, the proposed approach effectively i=1 (4)
generates the next-day forecast. The forecasting engines of kNN and
where φ (x ) is high-dimensional space characteristics that was mapped
WkNN are combined with the SDD engine, resulted in the lowest error
to input space vector (x ), parameters w and b are normal vector and
[35]. 1 n
scalar, respectively and C n ∑i = 1 L (x i,di ) represents the empirical risk.
Lima et al., developed a methodology based on a combination of
Variables w and b can be predicted by minimizing the regularized risk
Weather Research and Forecasting model (WRF) with statistical ANN
function after initiating positive slack parameters ξi and ξi∗ that re-
post-processing method for estimation of solar radiation in the Brazilian
present upper and lower excessive deviations.
North-Eastern region. The weather forecast outputs supplied by WRF
n
model with observational ground data were used as ANN input data. 1 1
Minimize RSVMs (w,ξ ∗,ξ ) = ‖w‖2 + C ∑ L (ξi,ξi∗)
The developed output model was compared to the data of surface solar 2 N (5)
i=1
irradiance obtained from Automated Weather Stations (AWS). The
RMSE, MSE, systematic error (ME), and statistical correlation coeffi- ⎧ di−wφ (x i ) + bi ≪ ε + ξi
cient (r) metrics were utilized to evaluate the model performances. ⎪
Subject to wφ (x i ) + bi−di ≪ ε + ξi
From the comparison analysis, the proposed model shows a reduction of ⎨ ∗
⎪ ξi,ξi ≫ 0, i = 1,…,l (6)
bias error at around 50 W/m2 in average while the RMSE (%) was re- ⎩
duced by 30% for the ANN model in relation to the WRF model [36]. 1
where 2
‖w‖2
represents the regularization term, C is the error penalty
Sharma et al., proposed a 15-min and hour-ahead solar irradiance factor utilized to determine the trade-off between the regularization
prediction approach by introducing a mixed Wavelet Neural Network term and empirical risk, ε is the loss function equivalent to approx-
(WNN) in a tropical region in Singapore. The Morlet wavelet and imation precision of the training data point, and l is the elements
Mexican hat wavelet were used in the WNN architecture with LM back- number in the training dataset. Eq. (3) can be solved by introducing the
propagation approach. The performance of the developed WNN model Lagrange multiplier and optimality constraints, thus the generic func-
was validated by assessing MBE and NRMSE. The proposed WNN model tion is given below:
gives the lowest MBE and NRMSE values. As for the 15-min ahead of
l
prediction, the lowest MBE and NRMSE was 4.01% and 9.42%, re- (βi−βi∗) K (x i,x j ) + b
f (x ) = ∑
spectively. Meanwhile, for the 1-h ahead of prediction, the lowest MBE i=1 (7)
and NRMSE was 5.56% and 15.41%, respectively [37].
where K (x i,x j ) = φ (x i ) φ (x j ) . Term K (x i,x j ) refers to kernel function,
Graditi et al., developed a comparative analysis of three different
which is the internal product of vector x i and vector x j , in the feature
models i.e., phenomenological model (Sandia model), MLPNN model,
space φ (x i ) and φ (x j ), respectively. The SVMs functions are to execute
and regression model. The analysis was carried out by using dataset
data correlation through non-linear mapping. The kernel methods must
consists of electrical, thermal, and meteorological data obtained from
be able to perform in a high-dimensional and implicit feature space
the year 2006–2012. From the analysis performances, the statistical
without compute the data coordinates in the space, but rather by simply
learning model of regression and MLP model shows an accurate fore-
computes the internal products between the images of all pairs of data
casting compared to the classical reference model. The nRMSE%_AVG
in the feature space. This method was known as the direct computation
and R2_AVG values of ANN model were 6.66 and 0.982, respectively.
approach to the kernel function, denoted by K. There were four types of
The further analysis represents an identification of minimum dataset
basic kernel functions i.e., linear, sigmoid, polynomial, and radial basis
required to train the ANN and regression model by using GA algorithm.
functions. The radial basis function (RBF) has emerged as the best
The effectiveness of optimization strategy is validated by comparing the
kernel function because of its computational effectiveness, con-
forecasting performances of the optimal dataset with randomly chosen
venience, reliability, ease of adaptation to optimize other adaptive
dataset employed [38].
techniques, and its adaptability in handling complex parameters. Fig. 4
Hossain et al., proposed an Extreme Learning Machine (ELM) model
illustrates the SVM architecture. It consists of input layers of dimension
to forecast PV output power of three grid-connected PV plant installed
p , hidden layer of j kernel inner products, and linear output layer. The
on the rooftop of PEARL laboratory in University of Malaya for 1-h and
RBF kernel function only requires the set solution of linear equations,
1-day ahead. Based on the simulation analysis, the proposed ELM model
rather than of lengthy and computationally demanding quadratic pro-
shows more reliable results than ANN and SVR models in terms of the
gramming problem for its training [42]. The definition of the nonlinear
forecasting capability and accuracy. For the day ahead PV output power
RBF kernel function was given below:
forecasting, the average computational time of training and validation
2
process was 0.22 s and 0.0335 s, respectively. For the hourly ahead PV K = (x i,x j ) = e γ ‖ xi − x j ‖ (8)
output power forecasting, the average computational time of training
and validation process was 0.296 s and 0.0153 s, respectively, con-
sidering all underlying PV types. The proposed model exhibits faster
processing time than other models [39].

2.1.2. Support Vector Machine


Support Vector Machines (SVMs) were categorized as a machine
learning technique. It was applied to many significant areas i.e., com-
puting, hydrology, and environmental fields. This approach was widely
used in prediction, classification, pattern recognition, and regression
analysis. The SVM has exhibited great performances compared to other
methods such as NN and other traditional statistical models [40]. These
models can prevent over-fitting of training data, dismiss iterative tuning
of model parameters, require a few kernels, make faster computations,
and have good generalization and convergence. The SVM equations
Fig. 4. SVM architecture [43].
based on Vapnik’s theory approximates are as listed below [41]:

465
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

where x i and x j are vectors in the input space. Parameter γ = −1/(2σ 2) Mohammadi et al., integrated a hybrid Support Vector Machine
where σ denotes Gaussian noise level of standard deviation. The three with Wavelet Transform (SVM-WT) model to forecast horizontal global
variables associated to RBF Kernels are γ ,ε, and C . The precision of SVM solar radiation for an Iranian Coastal City. The various combinations of
model depends on the selection of model parameter(s). input parameters i.e., daily global radiation on a horizontal surface,
Shi et al., developed the PV power output prediction for 1-day relative sunshine duration, minimum ambient temperature, maximum
ahead based on SVM and the features of weather categorization at a PV ambient temperature, average ambient temperature, relative humidity,
station in China. The historical model data were categorized into four water vapour pressure, and extra-terrestrial global solar radiation on a
classes of historical weather conditions i.e., cloudy, sunny, rainy, and horizontal surface were used to predict the global solar radiation. The
foggy. The input dataset for each model includes 15-min interval his- performance of this new model was compared to other models i.e.,
torical PV output power from the nearest day with the similar day type, Autoregressive Moving Average (ARMA), ANN, and Genetic
average temperature, minimum temperature, and maximum tempera- Programming (GP). The statistical analysis of SVM-WT model shows
ture of the next day from a local weather report. As a result, the sunny performance enhancement as it outperforms other models [40].
model was emerged as the best among the others in PV power fore- Olatomiwa et al., predicted the mean horizontal global solar ra-
casting. For the proposed model, the average forecasting errors of RMSE diation by integrating a new hybrid machine learning technique known
and MRE were 2.10 MW and 8.64%, respectively. Thus, the proposed as Support Vector Machine Firefly Algorithm (SVM-FFA). Three me-
GCPV systems in the developed model show an effectiveness in the teorological variables i.e., sunshine duration, maximum temperature,
performance [44]. and minimum temperature are defined as inputs. The performance of
Mellit et al., proposed a short-term forecast of meteorological time the proposed model was determined by RMSE, r, MAPE, and R2. The
series parameters by using LS-SVM at Medina City. The input para- proposed model was compared to GP and ANN models as well. The
meters were wind speed, wind direction, air temperature, relative hu- results obtained shows that the statistical indicator obtained the highest
midity, atmospheric pressure, and solar irradiation. The effectiveness of values for R2 and r, whereas the lowest values for RMSE and MAPE
the proposed model was determined via coefficient of determination [41].
(R2) and MAPE. This model was compared to four ANN architectures Ramli et al., presented a solar radiation forecast on PV panel sur-
i.e., MLP, RBF, RNN, and PNN. Two meteorological time series para- faces with specific tilt angles by using SVM and ANN methods at Jeddah
meters were selected i.e., air temperature and pressure to test the and Qassim in Saudi Arabia. The direct, diffuse, and global solar ra-
capability of LS-SVM in forecasting the future based on prior values. diation data on the horizontal surface were utilized in solar radiation
From the results analysis, the LS-SVM represents the best performance prediction. The performance and the comparison of the developed
compared to the other ANN architecture [43]. models were evaluated using RMSE, Coefficient Correlation (CC), Mean
Zeng et al., developed a short-term solar power forecast by using an Relative Error (MRE), and computation speed. As a result, the MRE
LS-SVM model with the following input parameters i.e., historical data values of SVM were 0.33 and 0.51 in comparison with ANN that ob-
of atmospheric transmissivity (2-dimensional form), wind speed, re- tained 0.19 and 1.16, respectively. SVM offers higher precision, more
lative humidity, and sky-cover. The proposed model performance was robust computation, and quickest in forecasting on the tilted surface at
determined by MAE, MAPE, and correlation coefficient (ρ). The vali- a speed of 2.15 s compared to 4.56 s for ANN during the training phase
dation was carried out by comparing Autoregressive (AR) and RBFNN [48].
models. The proposed model shows a good performance and outper- Wolff et al., developed PV power forecasting for 15-min to 5-h
forms other models. This is because SVM does have good generalization ahead via statistical learning model Support Vector Regression (SVR).
and is able to solve non-linear and time-varying issues of solar radia- This developed model was evaluated and served as an alternative to
tion. The simulation results for 1-h solar power prediction shows that complement other physical prediction models via PV power measure-
the transmissivity function produces small prediction errors with MAE ment, NWP, and Cloud Motion Vector (CMV) irradiances. The RMSE
values of 34.230 W/m2, 33.773 W/m2, and 62.864 W/m2 compared to and BIAS metrics were determined between measurement and forecasts
34.372 W/m2, 46.839 W/m2, and 64.504 W/m2 for sigmoid function, of PV power. The prediction of PV measurement for SVR indicates good
respectively [45]. results in prediction up to 1-h ahead, while NWP-based predictions
Chen et al., developed the SVM model to forecast daily solar ra- produce better period forecast starting at 3-h ahead, and CMVs emerged
diation. Seven SVM-models (SVM1-SVM7) with various input para- as the best amongst them. The combination of these three input sources
meters and five empirical sunshine-based models (linear, exponential, has provided the best result. The error reduction between single and
linear exponential, quadratic, and cubic) were evaluated by utilizing regional average forecasts resulted in over 75% for the combination of
meteorological data gathered from 3 stations in China. In assessing the forecast model and high-quality regional PV power predictions in short-
performances of the models, evaluation using RMSE, Relative Root term forecast horizons up to 1-h ahead [49].
Mean Square Error (RRMSE), and R2 were conducted. The simulation
results indicate that the SVM models produce average RMSE values of 2.1.3. Markov chain
2.094 MJ m−2 i.e., 10% lower than 2.323 MJ−2 generated by empirical Markov chain models have a sequence number of random para-
models. Therefore, SVM exhibits better performance than the empirical meters. The Markov property needs serial dependence only between its
models. Due to such remarkable improvement, the SVM method has adjacent states. It can be defined as the systems consists of a chain of
gained superiority over other empirical and conventional methodolo- connected states. The next state of Markov process only depends on the
gies for daily solar radiation prediction by utilizing sunshine duration current state. Transition refers to the changes in the system states. Fig. 5
[46]. illustrates transition probabilities which are known as probabilities
Ekici proposed the next day’s forecast of solar insolation by in- related to various state changes. The process is identified by initial
corporating LS-SVM with RBF Kernel Function in Elazig, Turkey. The state, state space, and transition matrix which is describing the prob-
daily mean temperature, maximum temperature, sunshine duration, abilities of transitions. The Markov process considers all possible states
and insolation of the day before used as input parameters to predict the and transitions that are included in the process definition. Therefore,
daily solar insolation of LS-SVM. The simulation of the proposed model the process is kept in the next state and never terminated [50].
was conducted with SVM Toolbox of MATLAB and its performance was From the mathematical equations, Markov chain is a sequence of
assessed with several statistical parameters i.e., RMSE, RME, MEF, random parameters X 1,X 2,X 3… with the Markov property, as shown in
Absolute Fraction of Variance, and CVRMSE. The proposed model was the following:
proven to be very effective and feasible for estimating solar insolation
Pr (Xn + 1 = x|Xn = x n,…,X1 = x1) = Pr (Xn + 1 = x|Xn = x n ) (9)
values using prior meteorological data [47].

466
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

Fig. 5. A simple two-state Markov Chain [50].

The probable values of Xi generate a countable set S which is known


as the state space of chain. Markov chain is defined by a sequence of
directed graphs where the edges of graph n are labelled by the possi-
bilities of passing from a single state at time n into another state at time
n + 1.
Pr (Xn + 1 = x|Xn = x n ) (10)

Similar information is presented by transition matrix from time n to


n + 1. These Markov chain are always assumed to be time-homogenous
where the matrix and the graph are not n-dependent [50]. Fig. 6. General schematic diagram for ARMA-based predictions [55].

Bhardwaj et al., developed a solar radiation prediction technique by


using a combination of Continuous Density Hidden Markov Model where Non-negative integer (p) , coefficients (∅i ) , and et is assumed to
(CDHMM) with Generalized Fuzzy Model (GFM). CDHMM with Pearson be white noise with a mean zero mean and constant variance σe2 .
R was applied to extract the shape-based clusters from the meteor- The ARMA model is reliant on two primary models i.e., MA and AR
ological variables. Then, GFM was used to predict solar radiation ac- models as presented in the following [22,54]:
curately. The meteorological variables used in this estimation process p q
are wind speed, relative humidity, temperature, sunshine hour, atmo- S (t ) = ∑ ∝i S (t −1) + ∑ βj e (t −j )
spheric pressure, and solar radiation data. The model performance was i=1 j=1 (12)
compared to ANFIS and ANN. The 15 sets with various combinations of
input meteorological variables have been generated to determine the S (t ) represents solar irradiance prediction at time t . In AR model, p
denotes the order of AR process, while αi is the ith AR coefficient. In MA
response of each meteorological variable in solar radiation prediction.
According to the simulation results, sunshine duration was considered model, q is the order of MA error term, βj is the jth MA coefficient while
as the main variable because sets that excluded this variable display the e (t ) is the white noise that refers to uncorrelated random parameters
insignificant outcomes. The values of RMSE, MAPE, and correlation with zero mean and constant variance. A general schematic diagram of
coefficients for the developed models were 7.9124, 3.0083, and 0.9921, the ARMA-based prediction model is indicated in Fig. 6 [55].
respectively. Hence, the proposed model offers good forecasts com- The ARMA model was applied widely for autocorrelated time-series
pared to the common model used [51]. data. This model is an excellent tool to forecast future values of speci-
Sanjari et al., developed a 15-min ahead to estimate the Probability fied time-series. Hence, this model is commonly known as ARMA (p,q)
Distribution Function (PDF) for the generated power of PV systems model where p and q are AR and MA orders, respectively. The ARMA
based on Higher-order Markov Chain (HMC) model. Solar irradiance has emerged as the favourite model because it is able to extract valuable
and ambient temperature were used as main parameters to classify the statistical properties and the adoption of well-known Box-Jenkins
various operating conditions in the PV system due to its strong impact method. The main requirement of ARMA model, it must be a stationary
upon PV output power system. The Pattern Discovery Method (PDM) time-series.
was applied as a classification procedure on historical data. Next, HMC Autoregressive Integrated Moving Average (ARIMA) is an extension
was built based on the classified historical data of PV power in each of ARMA model. The ARIMA models form a general class of linear
operating point (OP). The proposed model performance was assessed by model that is widely applied for predicting and modelling time-series.
determining the MAE forecast error. This model successfully outper- The ARIMA (p,d,q) model of time-series {X 1,X 2,…} is represented in the
forms other approaches i.e., Support Vector Regression, Chronological following [22]:
Probability, and Pair Copula Construction with an average MAE value Φp (B )ΔdXt = Θq (B ) at (13)
of 2.18% [52].
where
2.1.4. Autoregressive Φp (B ) = 1−∅1 B−∅2 B2−⋯−∅p B p (14)
An autoregressive model measures the correlations between de-
pendent and independent parameters. Further categorization depends Φp (B ) = 1−θ1 B−θ2 B2−⋯−θq B q (15)
on how time series are treated i.e., stationary or non-stationary and
linear or non-linear. The stationary time-series refers to the time-series where B is backward shift operator, BXy = Xy − 1,Δ = 1−B is backward
fluctuate around the static mean [17]. difference, Φp and Θq are polynomials of orders p and q , respectively.
The equation for Autoregressive, AR (ρ) is presented as follows [53]: The ARIMA (p,d,q) model is a product of AR part ( p ) Eq. (14), an in-
tegrated part of I (d ) = Δ−d , and a moving average of MA (q) part Eq.
Rt = ∅0 + ∅1 Rt − 1 + ∅2 Rt − 2 + ⋯+∅p tt − p + et (11) (15). The variables in Φ and Θ are selected so that the zero values in

467
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

both polynomials could lie outside the unit circle to avoid the creation forecasting error, an improvement as compared to another case in
of unbounded processes. which raw inputs are used. A significant reduction in prediction errors
The ARIMA approaches are reference estimators in the global irra- obtained when the proposed model was applied to the stationarized
diation prediction area. It shows the stochastic process combines the AR series, as opposed to the use of regular forecasting. The dimensionality
element to MA element, after differentiating at proper time steps to issue can be solved by implementing the Lasso variable selection pro-
eliminate any trend. This is the method used in the ARIMA models to cedure to the spatio-temporal model. An additional reduction of the
allow treatment of non-stationary series. forecasting error level of PV production was shown when meteor-
Dazhi et al., developed an hourly solar irradiance prediction using ological variables such as wind speed were included into the spatio-
the time-series analysis, which included the cloud cover effects. Three temporal model. The nRMSE was improved by 20% or more compared
prediction methods were generated by using several meteorological to the use of state-of-the-art forecasting techniques [59].
data that are defined as input variables. The first technique measured
hourly GHI data by decomposing three additive elements i.e., seasonal, 2.1.5. Regression
irregular, and trend. Trend and irregular elements acted as input for the Regression analysis refers to the method that determines a func-
next hour GHI prediction in the ARIMA model. The ARIMA model is tional relationship of the model between response (dependent) and
more flexible and capable than other classical models as it can effec- predictor (independent) parameters. Regression analysis is called as
tively deal with seasonal variations, random errors, and outliers. All ‘univariate regression’ when only one response parameter is involved
autocorrelations can be removed from the model residuals [56]. The and called ‘multivariate regression’ where two or more parameters are
second technique measures the DNI and DHI separately by using the considered. The regression analysis represents a repetitive process so
decomposition method to predict GHI whereas the third technique that the outputs are utilized to analyze, verify, criticize, and modify the
learns on cloud cover effect by using the ARIMA model for prediction of input. The univariate linear regression approach identifies the corre-
cloud transients. These three techniques were developed by using two lation between parameters by fitting the linear equation to the data. For
stations in Miami and Orland located in the USA. As a result, the third the multiple linear regression, the linear fitting retains the constant all
model presents the lowest values for MBE with overall values of 0.39 except one of the predictor parameters. Two regression analysis
and −0.27 and overall values of RMSE are 29.73 and 32.80 for Miami methods i.e., simple linear regression and multiple linear regression
and Orland, respectively. Hence, it has been proven that the application were widely used with the complexity of correlation between para-
of cloud cover information in prediction as implemented in model 3, meters. Linear regression techniques were described based on the ca-
enhanced accuracy especially when solar irradiation measurements tegorization as indicated in Table 1 [60].
were scarce [57].
Huang et al., developed a global solar radiation forecast method for i Simple linear regression:
1-h ahead by combining the techniques of AR and dynamical system
model in Mildura Town located in Australia. The forecasting ability of The mathematical equation of simple linear regression is:
AR (2) model, Luncheroni model, and a mixture of both models were
identified. The proposed model was analysed using MeAPE, MBE, KSI, Y = β0 + β1 X + ε (16)
and NRMSE. From the results analysis, the combined model outper-
where Response parameter (Y ) , predictor parameter (X ), regression
forms the other models, hence selected as the model that was upgraded
parameters (β0 and β1), and ε refers to the error to calculate the dif-
by including added components, which led to the development of
ference between forecasted and observed data. The forecasted value of
CARDS models. The error analysis shows that the CARDS model has
Eq. (16) can be expressed as:
successfully decreased forecasting error in the combined model by
33.4% for MeAPE [53]. Y ̂ = β0̂ + β1̂ X (17)
Yang et al., proposed an Autoregressive with Exogenous Input based
Statistical model (ARX-ST) to enhance the accuracy of solar PV power where Forecasted value is known as Y  and regression coefficients
production prediction. The developed models consider the both local evaluation is denoted as β . The forecasted values are acquired for any
and geographically correlated information of solar PV production from set of values of the predictor parameters that differ from the observed
other sites. The proposed model performance was assessed and com- data.
pared to PSS model for multiple forecast horizons. From simulation
results, the suggested time-scales for the ST forecast is at 1-h and 2-h ii Multiple linear regression:
ahead by using the real solar data. For 1-h ahead, MAE of the ST model
is 50.79%, 41.8%, and 5.15% lower than the PSS, BPNN, and AR model, Multiple linear regression refers to the generalization of a simple
respectively. For 2-h ahead, MAE of the ST model is 60.2%, 47.27%, linear regression model. Multiple linear regression models consist more
and 8.09% lower than the PSS, BPNN, and AR model, respectively [58]. than one predictor parameter:
Agoua et al., developed a statistical spatio-temporal methodology to
Y = β0 + β1 X1 + β2 X2 + ⋯+βp Xp + ε (18)
forecast the power output of a specific plant for a few minutes up to 6 h
ahead. Two datasets were considered to accommodate different cli- where Response parameter (Y ) , predictor parameters (X1,X2 …Xp ) with p
matic conditions and spatial densities of installed PV plants. A new being the number of parameters, regression coefficients (βo,β1…βp), and ε
stationarization process was introduced to overcome the issue of non- is the error to estimate the difference between the forecasted and ob-
stationarity of the production series. This result shows the reduction of served data. The forecasting equation derived by Eq. (18) is given by:

Table 1
Linear regression techniques classification.

Regression type Response parameters Predictor parameters Regression equation

Univariate Simple 1 1 Y ̂ = β0̂ + β1̂ X


Multiple 1 ≥2 Y ̂ = β0̂ + β1̂ X1 + β2̂ X2 + ⋯+βp̂ Xp

Multivariate ≥2 ≥1 Yi ̂ = βi,0̂ + βi,1̂ X1 + βi,2̂ X2 + ⋯+βi,̂ p Xp

468
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

Y ̂ = β0̂ + β1̂ X1 + β2̂ X2 + ⋯+βp̂ Xp (19) determine the multisite prediction of solar power generation for 1–6-h
ahead. An analysis and optimization of the application of Gradient
where Forecasted value is known as Y  and regression coefficients Boosted Regression Trees (GBRT) for prediction of future power gen-
evaluation is known as β . The forecasted values are defined same as eration from PV rooftop installations were conducted. The multi-site
mentioned in the simple linear regression. In evaluating and measuring model shows competitive results in terms of RMSE on all forecast
the objectives of regression analysis, the observations derived from the horizons when compared to single site linear autoregressive and var-
observed data are η > p + 1. iations of GBRT models. The GBRT model shows simpler and more
reliable than the conventional forecasting techniques in terms of pre-
iii Multivariate linear regression dictive performance and the simplicity of the model setup. All the
presented benchmark models i.e., persistence model, climatology, and
Multivariate linear regression refers to the generalization of mul- recursive autoregressive model were outperformed by the GBRT models
tiple linear regression methods. Multivariate linear regression examines for each forecast horizon of 1–6-h ahead. The best results in terms of
the probability linear relationships between multiple response and NRMSE was 0.100 and 0.137 for a lead time of 1 and 6, respectively
multiple predictor parameters. Eq. (19) can be expressed as Eq. (20) to [65].
demonstrate that two or more response parameters are being analyzed:
2.2. Physical methods
Yi ̂ = βi,0̂ + βi,1̂ X1 + βi,2̂ X2 + ⋯+βi,̂p Xp (20)
The interactions between solar radiations and atmospheric compo-
In the multivariate regression, both predictor and response para- nents such as aerosols and gases (CO2, H2O, N2, O2, O3, etc.) are known
meters are associated with each other. as the physical satellite model for solar irradiance prediction.
Trapero et al., proposed solar irradiation prediction for 1–24-h Atmospheric conditions are measured by using local meteorological
ahead by applying the Dynamic Harmonic Regression (DHR) model. data and remove the necessity of solar irradiance data at the surface.
Hourly Direct Normal Irradiance (DNI) and GHI were selected to assess However, these models are required to convert digital counts from sa-
the proposed model. The error metrics of relative Mean Bias Error tellite-based radiometers into corresponding flux density. Thus a device
(rMBE) and rRMSE were evaluated to measure the performances of the with appropriate and frequent calibration is necessary [66].
proposed model and compared to PM, Seasonal PM, Exponential Trend
Smoothing (ETS) in a state framework, and ARIMA. From the results 2.2.1. Numerical weather prediction
analysis, DHR model gives precise and consistent rRMSE values of NWP tools were helpful in the advancement of power prediction
29.66% and 46.79% for both GHI and DNI, respectively. Due to the models for electrical plants based on renewable sources. The objectives
dissimilar magnitude of both rMBE and rRMSE, deriven from DNI and of these NWP tools are to provide information about atmosphere con-
GHI, there is difficulty in prediction for DNI even for the short-term ditions for a given time-scale. These NWP models can be categorized
forecast. Hence, the proposed DHR can be considered as a competitive into two categories i.e., global models and mesoscale models. Global
prediction method as it outperforms other models [61]. models simulate the features of the atmosphere to a worldwide scale
Wang et al., proposed a 24-h ahead PV power output forecasting while mesoscale models simulate the features of the atmosphere for a
based on the Partial Functional Linear Regression Model (PFLRM) limited area i.e., continents, countries, and regions [67]. NWP models
method. The PFLRM was integrated by using both functional and were used widely to predict atmospheric state up to 15-days ahead
multiple linear regression models. The accuracy of the PFLRM model [68]. There were 15 weather services that are actively operated by
was determined with RMSE, Mean Absolute Deviation (MAD), and these Global NWP models i.e., GFS run by US US NOAA and European
MAPE. The results were compared to NN models to assess the perfor- Centre for Medium-Range Weather Forecasts (ECMWF). Basically, these
mances in prediction. Based on the simulation analysis, the RMSE, global models have coarse resolution and disallow detailed mapping of
MAPE, and MAD values of the proposed model were 59.17, 11.34%, small-scale features. It happens even the resolution has increased
and 40.277, respectively. The regularized PFLRM has been more ef- drastically over the last few years until the present moment in the range
fective for estimation than NN methods [62]. of 16–50 km depending on the model [22]. These NWP models are
Li et al., developed a simple non-linear regression model of the based on numerical integration of coupled differential equations that
Multivariate Adaptive Regression Splines (MARS) to forecast solar explain the transportation of radiation mechanism and the dynamics of
power output for 24-h ahead. The MARS model was applied on the the atmosphere. The benefit of this forecast model is that they are based
daily output of grid-connected 2.1 kW PV system. The MARS model is a on deterministic physical models [16].
data-driven approach without assuming the relationship between the Mathiesen et al., analysed the MOS correction of GHI by using three
power output and predictors. It processes the capability of handling operational NWP models at the Continental of the USA. This analysis
nonlinearity while maintaining the simplicity of the classical multiple validates North American Mesoscale Forecast System (NAM), GFS, and
linear regression models. Based on the result analysis, the MARS models ECMWF by using the Seven Surface Radiation Budget Network
show the best results in terms of RMSE, followed by ARMAX, MLR, (SURFAD) ground measuring data. The forecast performance was
kNN, SVR, CART, ANN, and ARIMA models. From the error assessment, evaluated with RMSE and MBE where MOS was applied to decrease the
the MARS values of RMSE, MAD, and MAPE are 119.0, 89.8, and MBE metrics in estimating GHI. The values for MBE were acquired for
69.2%, respectively. The MARS model is able to provide more reliable every single NWP model as a function of solar zenith angle (SZA) and
forecast performance than other various forecasting models [63]. forecast clear sky index (k t∗) to gain a bias correction function via MOS.
Massidda et al., forecasted the power output of a 1.3 MW PV plant in The local MOS function based on the SZA and k t∗ estimation shows a
Borkum, Germany by integrating the Multilinear Adaptive Regression decrease in RMSE by 20.1 W m−2, 17 W m−2, and 25.6 W m−2 for
Splines and NWP. The proposed model forecasts the power production NAM, ECMWF, and GFS, respectively, thus efficiently eliminating MBE.
of a PV plant for one day ahead. Past historical power output data and The MOS-correlated NWP prediction based on SZA and k t∗ offers an
the available weather forecasts of the GFS NWP model were used by the important baseline precision to evaluate other prediction methods.
developed regression model. Based on the simulation analysis, the re- MOS can reduce MBE for all NWP models. Thus, the best solar forecast
sults were favourable even though only a relatively low number of was presented by the MOS-corrected GFS with an RMSE value of
samples and features were considered. The forecasts errors of RMSE and 85 W m−2 followed by ECMWF, and NAM. The most precise estimation
MAE were 177.8 kW and 125.9 kW, respectively [64]. for cloudy and clear sky conditions are ECMWF and GFS, respectively
Persson et al., proposed a non-parametric machine learning to [69].

469
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

Fernandez et al., developed a short-term prediction system for persistence model. The matching forecast errors increase with two as-
hourly electrical energy generation in real GCPV plant with a combi- pects i.e., cloud speed and forecast horizon [76].
nation of two cascade structures of NWP and an AI-based model. The Gohari et al., developed a solar power output prediction for 15-min
first module is the NWP model corresponded to GFS world coverage ahead. The proposed model was applied to clear sky index by making
was developed based on National Centres for Environmental Prediction comparisons with the Total Sky Imager (TSI) and the University of
(NCEP). The output parameters of the first model are defined as inputs California San Diego Sky Imager (USI). The persistence forecast was
in the second model which represents another NWP model but with required as a reference to compare the performances of the instruments.
larger spatial and temporal resolutions compared to GFS. This second The persistence forecast was developed for each forecast point by
model corresponded to the MM5 model that offers prediction values for averaging 60 s of ground-measured clear sky index. The performances
the chosen weather parameters to the third model. This model refers to of the developed TSI and USI were evaluated by using statistical metrics
the hourly energy generation prediction model. The outcomes are de- of Euclidean dispersion, Pearson Correlation Coefficient, and Spearman
livered into other forecast methods i.e., ARIMA, k-NN, ANN, and Correlation Coefficient. The results show that the clear sky index cor-
Adaptive Neuron Fuzzy Inference System (ANFIS). The MLP-ANN relation of USI was between 0.7 and 0.82 while TSI was between 0.55
model emerged as the best forecast model amongst others. The RMSE and 0.69. Hence, USI indicates better consistency compared to TSI on
value for the rated power PV plant is 11.79% for both 16- and 39-h the days evaluated [77].
forecast time-scales [70]. Peng et al., developed a short-term solar irradiance estimation for
Monteiro et al., developed the mesoscale NWP to construct a data- novel 3D cloud detection and tracking system based on multiple TSIs.
base of weather parameters. A transition model was developed using Firstly, a supervised classifier was developed to recognize clouds at the
prior data of power production. A historical similar mining mechanism pixel level and the output cloud mask. Secondly, an intelligent algo-
was applied to designate dissimilar weight cases based on their simi- rithm was calculated to measure the block-wise base height and the
larity with the current case. Once the probability matrices were de- motion of every cloud layer based on the images captured from the
termined, the power transitions were identified by generating prob- multiple TSIs. This information was then posted out to stitch pictures
abilities for each future prediction interval. The simulation results show together into larger views that were used for solar prediction. This
that the forecast skill, NRMSE, and average improvement were 36.3%, system can robustly recognize clouds and track layers. The statistical
10.14%, and 0.8%, respectively, on the second benchmark model with metrics of MAE and RMSE were evaluated to measure the forecasting
MLP for 1 to 24-h ahead forecast [67]. performance by the whole dataset. The proposed model shows at least
26% improvement for all irradiance prediction between 1 and 15-min
2.2.2. Sky imagery in comparison with the persistent model [78].
Sky imager is a digital camera that provides a good quality image of Montesinos et al., developed a short-term and mid-term (1–180-
the sky from horizon-to-horizon and it is applied for clouds detection, min) solar irradiance forecasting approach for all sky conditions i.e.,
cloud height measurement above ground, and cloud motion determi- partially-cloudy, cloudless, and overcast skies by integrating a total sky
nation [71]. Most existent sky imagers were designed to image a camera of TSI 880 model. The digital image levels were transformed
hemispheric view of the sky. Their built-in cameras are either oriented into red-greenblue (RGB) and hue-saturation-value (HSV) colour
upwards or downwards in obtaining a direct image or in capturing spaces. Afterwards, the cloud movement was implemented for future
reflections from a spherical mirror, respectively. The efficiency of cloud irradiance pixel position and to determine the future irradiance values
tracking functions and detection techniques may difference depends on by integrating the maximum correlation technique. The results indicate
sky imagers type [72,73]. Total Sky Imager (TSI) is an example of the the average relative Root Mean Square Error (nRMSE) values of beam
sky imager which is produced by the Yankee Environmental Systems as irradiance, diffuse irradiance, and global irradiance were 25.44%,
illustrated in Fig. 7. The equipment has a hemispherical mirror with a 11.6%, and 11.17%, respectively. These values remain constant
downwards position pointing towards the CCD camera located above it.
The mirror is equipped with a Sun tracking shadow band that shields
the optical sensor from the consequences of solar reflection. Fig. 8
shows a TSI image for Solar Forecasting Initiative captured at the
University of California Merced. The spinning shadow band is seen
nearby the bottom left side of the image and the support for the CCD
camera is seen at the top of the image [66]. The solar predictions based
on sky imagery analysis has four elements i.e., (1) obtaining the sky
image near the forecast site by using Sky Imager device, (2) sky image
data are analyzed to recognize clouds, (3) cloud motion vector is esti-
mated by utilizing consecutive images, and (4) cloud location and
motion vector data are used for short-term probabilistic and determi-
nistic cloud cover, irradiance, and power predictions. The sky imagery
has benefits with complete information of structure, extent, and motion
of existing clouds at the time of estimation is made. These data were
applied to develop very short-term forecast (minutes ahead) of future
cloud patterns in solar generation facility area [74].
Chow et al., developed a method of sub-kilometre cloud shadow
now-casting and estimated using a ground-based sky imager for several
selected days at the University of California, San Diego Solar Energy
Test-Bed. The sky imager was used to evaluate the solar source in the
solar energy application to enhance both spatial and temporal resolu-
tions through satellite and numerical prediction methods. The matching
error between two cloud maps and cloud-advection-versus-persistence
(cap) error were evaluated to predict cloud cover. The results show that
cap error is a universal metric compared to forecast performance. In the
Fig. 7. TSI-880 from Yankee Environmental Systems [66,75].
30 s forecast, the forecast error was reduced to 50–60% compared to the

470
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

back-propagation MLP model. Firstly, the satellite images were identi-


fied and Self-Organizing Map (SOM) was utilized to perform the char-
acterization of cloud cover index. Next, the derived cloud cover index
was considered by utilizing the ESS model to predict the next time step
cloud cover index in Singapore. The solar irradiance was evaluated
based on the cloud cover index prediction with back-propagation MLP.
The proposed model performances were analyzed by measuring the
error of nRMSE, R2, and Normalized Mean Bias Error (NMBE) and
compared to other forecasting models i.e., ARIMA, Linear Exponential
Smoothing (LES), Simple Exponential Smoothing (SES), and Random
Walk (RW). The simulation results show enhancement in the fore-
casting model which improved by 6% than the best forecast accuracy of
other statistical models. The proposed hybrid model outperforms other
forecasting models [81].
Aguiar et al., developed a satellite-derived, ground data model in-
corporated with solar radiation and total cloud cover forecasted by
ECMWF to improve intra-day solar prediction. The clear sky index was
used as a solar radiation parameter with statistical models in this work.
The results obtained from the developed ANN which incorporated
ground and exogenous data were used as inputs. The simulation results
Fig. 8. Image captured with TSI by Coimbra and several collaborators at the University of
show that the combination of Neural Network, ECMWF, and Satellite
California Merced on 1st June 2011 [66].
derived model was suggested for the Gran Canaria Island. This model
improves the NN model (only with ground data) with an increased
throughout the study period under all sky conditions. The proposed percentage in RMSE with forecast time-scales from 15.47% to 22.17%
technique represents that sky-cams can become more accurate than for Co-Pozo Izquierdo station and 25.15–34.09% for C1-Las Palmas
satellite images for days with clouds, hence showing good results for station model [82].
beam and diffuse radiation elements [79].
2.3. Ensemble methods
2.2.3. Satellite imaging
The concepts of satellite imaging are quite similar to sky imagery Ensemble methods were presented to solve the weakness of in-
models. The cloud pattern is determined from both visible and infrared dividual methods and to enhance their strengths and accuracy [17].
images taken from satellite-based sensors flying overhead. By con- These ensemble approaches can be integrated in three different ways
sidering a longer forecast time horizon, the geostationary satellite data i.e., linear, non-linear, and combination of both. These methods com-
obtained from several networks such as NOAA, Geostationary bine different models to enhance the forecasting accuracy [22]. The
Operational Environmental Satellite network for North and South ensemble model was used widely in both statistical and learning ma-
America, and Meteosat network for Europe, Africa, and central Asia chine techniques to obtain aggregated decisions by using multiple
provide cloud motion information and its properties. Firstly, a physical- predictors. These models can be classified into two categories i.e.,
driven model forecasts the clear-sky conditions at a particular site by competitive and cooperative ensemble forecasting [83].
utilizing inputs such as water vapour, elevation, aerosol content, and
ozone. Next, these clear-sky irradiance models are modulated by pre- 2.3.1. Cooperative ensemble forecasting
dicting the irradiance obtained from the satellite images. Consecutive To improve the performance of a forecasting technique, cooperative
satellite images are mixed to generate cloud motion vector fields which ensemble method can divide a forecasting task into a few sub-tasks and
can be utilized to forecast the future cloud locations. These methods hence each sub-task is solved individually. In [83] The overall fore-
were used effectively in irradiance predictions in the range of 1-min to casting results were acquired by aggregating the forecast values of all
5-h ahead. Nevertheless, it is less efficient when the clouds quickly predictors. There were two types i.e., pre-processing and post-proces-
forming and dissipating [71]. sing. Pre-processing can split the dataset of input into a few sub-data-
Marquez et al., developed the GHI predictions at temporal horizons sets whereby each sub-dataset can be modelled and estimated by the
of 30, 60, 90, and 120-min by using the hybrid technique of satellite predictor. Generally, all sub-datasets were comprised of the same pre-
images and ANN. The cloud fraction parameters were used as inputs to dictor. The final decision was obtained by summing up all outputs from
an ANN model. These variables referred to processes using cloud in- the predictor. Post-processing refers to time-series data that contain
dexing and velocimetry algorithms with a satellite image. The persis- more than one feature. Each feature is appropriate for only one specific
tence model was designated by utilizing the present clear-sky index for technique e.g., ARIMA models were suited for modelling linear time-
the future clear-sky index prediction which becomes a criterion to series.
compare the precision of the developed approaches. The forecast model
performance was determined using MBE, RMSE, and RMSE. As a result, 2.3.2. Competitive ensemble forecasting
the proposed methodology represents similar findings or improvement The competitive ensemble forecasting models use the multiple
from the other studies [80]. The proposed method outperforms the predictors with different parameters to build individual forecast models
persistence model about of 5–19% for 1 time-step forecasts and about of and to form an ensemble forecast model. Results of the forecasting
10–25% for multistep forecasts. Hence, the suggested hybrid ap- model from the selected models were aggregated by averaging.
proaches that applied on the combination of stochastic learning, Diversity is the main characteristic in competitive ensemble forecasting
ground, ground telemetry and/or remote sensing image processing and can be classified into kernel, parameter, and data diversity. If the
provides accuracy enhancement and robustness for the development of base predictors return similar decisions, it will result less improvement.
solar estimation [75]. In this competitive ensemble forecasting, if sub-tasks are similar, the
Dong et al., developed an hourly resolution solar irradiance forecast outputs of individual predictors will be same as well and the perfor-
by applying a satellite-image analysis and a hybrid forecasting ap- mance improvement of ensemble predictors will be marginal [83].
proach using the Exponential Smoothing State (ESS) space model and a Quan et al., developed a hybrid ANN evolutionary algorithm to

471
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

speed up the convergence during training stage, especially in the wind 2.7% and 6.13%, respectively, compared to the ANN model. The overall
case to minimize the overall forecasting error. The developed fore- efficiency is slightly dissimilar during cloudy days [88].
casting ANN model with a combination of Genetical Swarm Yang et al., proposed a 24-h ahead hourly prediction method for PV
Optimization (GSO) and Evidence-Based Practise (EBP) resulted in the generation by integrating the Weather-Based Hybrid Technique with
reliable and good in all the operative conditions. This developed model combination methods of SOM, SVR, Fuzzy Inference, and Learning
was tested with over one-year period of PV production. The proposed Vector Quantization (LVQ) approaches. The SOM and LVQ techniques
model was analysed and compared to stand-alone classical EBP for a six were used to categorize the PV power output historical data. During the
months period. The results show that the proposed model achieves good training phase, SVR was applied to train a set of input/output data for
performances and accuracy for clear weather and permanently covered solar irradiance, temperature, precipitation probability, and solar irra-
days. The NMAE error obtained for wind case over one year period was diance. The fuzzy inference approach was applied during the fore-
doubled compared to the NMAE error determined for PV plant for the casting phase to choose an appropriate training model for precise pre-
same time period [84]. diction based on weather information supplied by Taiwan Central
Zhang et al., proposed a solar radiation forecasting method using an Weather Bureau (TCWB). Two metrics criteria i.e., MRE and RMSE
RNN based predictive model with a combination of Particle Swarm were evaluated to validate the performances of the developed model
Optimization (PSO) and Evolutionary Algorithm (EA). An Elman-model and compared to the individual ANN and SVR approaches. The pro-
based on RNN was proposed as a new NN solution to monitor solar posed model indicates superior performance with MRE and RMSE va-
radiation. The combination of PSO and EA hybrid training algorithms lues of 3.29% and 350.2 W, respectively, for overall annual data in
were defined to determine the benefits of two global optimization al- comparison with ANN and SVR approaches [89].
gorithms. PSO stored valuable information together with local and Wu et al., developed a novel hybrid model to predict1-h ahead of PV
global memories throughout the evaluation process. The anomaly in EA output. There are five types of forecasting models developed to predict
stores a variety of PSO populations in expressing new search area. The the power output of three PV systems i.e., ANN, SVM, ARIMA, ANFIS,
simulation results show that the RMSE value for the best individual and the combination models by using the GA algorithm. The actual
algorithm was decreased to 0.028 when integrated with hybrid PSO-EA predictions were performed for these PV systems in Taiwan and
algorithm after 100 iterations. The proposed predictive model produces Malaysia. The proposed model uses the experimental database of solar
the precise solar radiation estimation. This model generates an ex- power output, solar irradiance, air, and module temperature data. It
cellent performance for solar radiation monitoring and efficient sizing. consists data from the Green Energy Office Building in Malaysia, the
Thus, stand-alone PV and hybrid power systems were enhanced and Taichung Thermal Plant of Taipower, and the National Penghu
improved, respectively [85]. University, Taiwan. Analytical results demonstrate that the hybrid
Bouzerdoum et al., developed a new hybrid method of Seasonal prediction model generates the most accurate predictions in most cases.
Auto-regressive Integrated Moving Average (SARIMA) and SVM to The proposed model performance was assessed by determining the
predict the solar power generated by a small-scale GCPV plant NRMSE. From the simulation results, the NRMSE values of hybrid
(208 kWp) in Trieste, Italy. The SARIMA model was developed to esti- model were 5.64%, 3.43%, and 6.57% for PTM Green Energy Office,
mate and analyze the linear component of the power generated. The Taichung Thermal Power Plant, and academic building at the National
SVM model was developed to determine non-linear patterns in residual Penghu University, respectively [90].
series derived from SARIMA model. The statistical errors of R, NRMSE, Filipe et al., integrated a hybrid solar power forecasting system with
NMBE, and MPE were evaluated to assess the performance of the de- a combination of both statistical and physical methods. The proposed
veloped model and compared to individual SARIMA and SVM models. model was comprised of two main features. Firstly, the combination of
As a result, the hybrid SARIMA-SVM shows the lowest error in com- an electrical model of the PV system with the statistical model that
parison with SARIMA and SVM models. The proposed model outper- converts NWP into solar power for short-term forecast horizon.
forms SARIMA and SVM models [86]. Secondly, the various NWP models combine with prior time-series data.
Haque et al., developed a short-term forecast for solar PV power The performance of this hybrid forecasting method was compared to
output by incorporating several intelligent algorithms i.e., a hybrid of two naive models i.e., persistence model and diurnal model. The pro-
the Wavelet Transform (WT), Fuzzy ARTMAP (FA), and Firefly (FF). posed methods show performance enhancement with the lowest RMSE
The WT extracts the ill-behaved characteristics of PV time-series while value in comparison with individual models. The suggested hybrid
FA was used to capture the non-linear fluctuation of PV power time- forecasting models represent an average improvement of 57.4% and
series. The FF algorithm was utilized to enhance the FA network to be 34.06% over the persistence and diurnal models, respectively [91].
more efficient and to improve the precision of the entire prediction. The Dong et al., developed an hourly solar irradiance forecasting
developed model performance was evaluated by assessing the metric method by integrating the hybrid of SOM, SVR, and PSO approaches.
accuracy of MAPE, NRMSE, and error variance (σ2). The results of the The SOM functions to divide the entire input space into assorted dis-
proposed model were compared to other prediction models such as jointed areas with various information on the correlation between input
BPNN, RBFNN, General Regression Neural Network (GRNN), FA, WT and output. The SVR was applied to model each disjointed area to re-
and BPNN, WT and RBFNN, WT and GRNN, and WT and FA. The cognize the attributes of the correlation. The PSO was applied to select
proposed model outperforms other methods [87]. the variables automatically in SVR modelling and to decrease the SVM
Gandelli et al., presented solar power PV plant prediction for the volatility with dissimilar variables. Two metrics criteria i.e., nRMSE and
short-term by implementing a new Physical Hybrid Artificial Neural nMBE were used to evaluate the developed model performance. The
Network (PHANN) based on the fundamental of Physical Model developed model was compared to other models i.e., ARIMA, LES, SES,
Constraints and ANN models. The atmospheric conditions were sup- and RW by using similar datasets. From the simulation results, the
plied by the Meteorological Service such as geographical coordinates of developed models show consistent performance for low irradiance
the site and the time and date to determine the sun position were used variability areas like Colorado (USA) and high irradiance variability
as input data. Several error metrics i.e., NMAE, Weight Mean Absolute areas like Singapore for 7 out of 10 months. Therefore, the proposed
Error (WMAE), and nRMSE were determined to evaluate the proposed hybrid model shows a good performance with an average prediction
model performance. The result from the simulation indicates the pro- accuracy of 4% for lower energy in comparison with other statistical
posed PHANN is more precise than ANN models. The accuracy of this models [92].
model highly depends on the historical weather prediction data which Wang et al., suggested a hybrid approach called Cuckoo Search-
was used during the training stage. The PHANN model represents the Optimally Pruned-Extreme Learning Machine (CS-OP-ELM) to predict
highest precision during sunny days with NMAE and WMAE values of clear sky global horizontal radiation (CSKY Glo) and real sky global

472
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

horizontal radiation (RSKY Glo). Firstly, Multi-response Sparse to determine its irregular patterns and anomalies. This approach is
Regression (MRSR) and Leave-One-Outcross-Validation (LOO-CV) were beneficial in discovering irregular patterns and outliers to give a better
applied to rank the neurons and prune the probably meaningless neu- characterization for data collection with the larger resolution to be fed
rons of Feed-forward Neural Network (FFNN) respectively. Next, the into MLPNN. The training phase was performed by incorporating
OP-ELM direct strategy and the direct recursive strategy were devel- MLPNN which was utilized for solar radiation prediction. The devel-
oped to generate several ensemble approaches. The CS optimized al- oped TB K-means performance was measured with several real datasets
gorithm was performed to calculate the weight coefficients. The de- and compared to K-Means++, K-Means∗, and SOM clustering algo-
veloped model performances were assessed by evaluating RMSE and rithm. A comparative analysis with six other methods i.e., Non-linear
MRE. Four models were selected for comparison i.e., ARMA, BP, direct- Autoregressive (NAR), RW, Simple SES, LES, ARIMA and Hybrid
strategy OP-ELM, and direct recursive strategy OP-ELM. The simulation models, inclusive of European Centre for Medium-Range Weather
results show that the CS-OP-ELM represented the lowest MRE and Forecasts (ECMWF-OL), Statistical model of Blue Sky (BLUE), MOS by
RMSE values. For the CSKY-Glo forecasting, the value of MRE and the weather company Meteomedia GmBH (MM-MOS), Weather
RMSE were 75% and 62.5%, respectively. The RSKY-Glo forecasting Research and Forecasting-Meteosat (WRF-MT), and Persistence
obtained 75% and 70.83% for MRE and RMSE, respectively. Based on methods were utilized as the baselines for solar prediction models. The
these two metrics criteria, the CS-OP-ELM exhibits better performance analysis result presents the proposed hybrid prediction model has ac-
compared to other models for both clear and real sky prediction [93]. curate performance [97].
Rana et al., proposed a half-hourly PV power outputs forecasting for Li et al., presented a hierarchical method by using machine learning
the next-day based on prior power data by using three various NN methods of ANN and SVR to predict energy productions from a solar PV
ensemble methods. Firstly, two A1 and A2 techniques were carried out system in Florida for 15-min, 1-h, and 24-h ahead. No exogenous data
to forecast all power outputs on next day concurrently. Secondly, an A3 was used in the forecasting model such as solar irradiance. The hier-
technique performs a repetition by using prior forecast data. Two me- archical approach forecast the power prediction from the inverters. In
trics criteria of MAE and MRE were determined. The proposed model the hierarchical forecasting method, the machine learning tools were
was compared to other models i.e., two persistence and SVR methods to used at a micro level for each inverter prediction. The performance was
evaluate the developed model performance. Four Australian PV power evaluated at a macro level of the entire plant by adding up the forecast.
half-hourly data of one year was defined as input datasets. From the From the results analysis, the hierarchical method outperforms tradi-
simulation results, the repetitive technique A3 show more precise than tional methods by using information of the power production from the
non-iterative techniques Al and A2. As a result, A3 outperforms A2 and micro level of the plant system [98].
A1 with average MRE values of 7.77% and 1.05%, respectively. Graditi et al., compared and analyzed two different models to pre-
Meanwhile, the average values of MRE for all cases were 9.43%, dict energy yield of a 1.05 kWp experimental PV plant located at ENEA
36.48%, and 7.18% for A1, A2, and A3, respectively. The hybrid NNs Portici Research Centre. The first model is physical modelization of the
provide greater accuracy compared to individual NN for all the three plant while the other model was correlated to various ANN topologies
models and A3 emerges as the most important approach in all practical i.e., MLP-ANN, n-Wavelet NN, and ANFIS. In the second case, a new
applications [94]. hybrid model known as Hybrid Physical ANN (HPANN) based on ANN
Dolara et al., integrated a new hybrid model known as PHANN to and clear sky solar radiation curves were developed and compared to
predict the PV power plants production for the day-ahead. This pro- the MLP-ANN method. The developed ANN model uses the data ob-
posed model combines an ANN technique with an analytical physical tained from Data Acquisition System (DAS) connected to an experi-
model of Clear Sky Solar Radiation Model (CSRM). The weather fore- mental plant consisting of micro-morph silicon modules. The error as-
casts were supplied by the MeteoService and CSRM curve was used as sessment shows that the AI-based method is better than the physical
the input. The results show that the PHANN method outperforms ANN method in estimating Pac. Moreover, the hybrid method outperforms
method for most days. This provides a significant result of NMAE% and ANN. In the case of HPANN model, the MBE/A was between 0.01% and
WMAE% which reaches almost 5% error reduction for many days [95]. 5.11%, the RMSE/A was between 5.91% and 10.09%, the CC was be-
Piazza et al., presented an hourly solar irradiation and wind speed tween 0.9824 and 0.9934 while the R2 was between 0.9644 and 0.9868
prediction model by integrating Non-Linear Autoregressive Network [8].
with exogenous input (NARX) based on multiple-step ahead. In this Antonanzas et al., presented a comparison study between extreme
study, the temperature was used as an exogenous parameter. The NARX gradient boosting (XGB) machines, deep neural networks (DNN),
model was combined with two other approaches i.e., GA-based opti- random forest (RF), and SVR to produce day ahead of photovoltaic
mization and pruning-based Optimal Brain Surgeon (OBS) techniques. power forecasting. The input set was added with the NWP of various
The optimization of GA shows an excellent network weight for com- meteorological variables to enhance predictions. A hybrid model was
putation while the OBS strategy identifies the optimal network struc- produced with a weighted combination of single-technique predictions.
ture. With that, an optimum NARX was developed. Two metrics indices Based on the simulation results, the SVR produces the best results
i.e., NRMSEr and Coefficient of Variance based on Root Mean Square compared to other single models with RMSE value of 22.49%. The
Error (CV (RMSE)r) were calculated to determine the performances of hybrid model outperforms all single models with RMSE value of
the models. The results of NRMSEr and CV (RMSE)r of solar radiation 22.24% [99].
forecasting were 6.1% and 32%, respectively, in the recall phase. As for Do et al., concentrated on the minimum period of training data
wind speed forecasting, NRMSEr and CV (RMSE)r were 7% and 47%, required for PV generation prediction for two climatic conditions i.e.,
respectively. The developed technique was used to fix the NARX model temperature and tropical weather conditions. This study was im-
to overcome several disadvantages of the repetitive test. It is especially plemented three different forecast models i.e., the scaled persistence
during the trial-and-error process where extra time was required to model, ANN, and the multivariate polynomial model in two PV in-
obtain the optimal network structure. Hence, the proposed model in- stallations i.e., Guadeloupe and Lille. From the results analysis, a longer
dicates the network improvement [96]. training duration shows a high accurate forecasting. At least 6 months
Azimi et al., proposed a hybrid prediction approach with a combi- training duration was required for PV generation forecasting in a
nation of the time-series analysis of MLPNNs and new cluster selection temperature climate such as Lille compared to only 3 months in a tro-
algorithm to forecast solar irradiance. Time-series was applied to pro- pical climate at Guadeloupe. Once the model was trained, the perfor-
cess historical solar radiation data and to change them into input- mance is better with the error of 10.69% at Lille compared to 11.97% at
output series. The developed Transformation Based K-Means (TB K- Guadeloupe. This research uses the scaled persistence model to predict
means) algorithm classifies the time-series data of solar into various sets the PV output power during the data collection as this model does not

473
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

require historical data. Once the data collection phase was completed, parameters perform slightly better than the five parameters with NMAE
the multivariate polynomial model was applied to provide a better values of 8.5% and 9.0%, respectively. The PHANN model shows higher
accuracy [100]. accuracy in the range of 6% [103].
Malvoni et al., investigated the performance accuracy of data-driven Oneto et al., proposed a data-driven approach to solve the problem
methods that were applied to the hybrid forecasting model Group Least of nowcasting and forecasting the photovoltaic power production
Square Support Vector Machine (GLSSVM). This hybrid model com- (PvPP). This developed model based on real data available for the
bines the Group Method of Data Handling (GMDH) and LS-SVM Savona Campus and data from the energy management systems (EMSs)
methods. The data pre-processing techniques of WD and the Principal of the smart poly-generation microgrid that feeds the buildings in the
Component Analysis (PCA) were proposed to decompose meteor- university area. The performance of three data-driven techniques i.e.,
ological data used as inputs for the forecasts. From the results, the Kernel Methods (KM), ELM, and RF was evaluated. The obtained results
forecast accuracy can be improved with the combination of the data- were compared to the forecasted EMS installed at the Savona Campus
driven forecasting method and the data pre-processing technique. The Site which uses physical-based models. The results show that RF was
Principal Component Analysis Wavelet Decomposition (PCAWD) out- the best method to nowcast and forecast the PvPP [104].
performs the PCA, WD, and WDPCA in most of the hour-ahead horizons Chiang et al., developed a data-driven framework that integrates
[101]. WD and the Bias Compensated Random Forest (BCRF) model to forecast
Cervone et al., developed an Analog Ensemble (AnEn), ANN, and a PV output using the current value of PV output and meteorological
combination of both models to assess the 72-h deterministic and sensor data. The proposed technique provides a better representation of
probabilistic forecasts of power generated by 3 PV stations located in the raw data by extracting the PV output time series and other me-
Italy. This methodology was used the weather and astronomical pre- teorological variables. The proposed model was compared to the BPNN,
dictions as the input. The National Centre for Atmospheric Research Wavelet-SVM, and RF models. The developed Wavelet-BCRF shows the
(NCAR) Yellowstone supercomputer was used in conducting experi- consistent performance as it was reduced by 32.12%, 20.64%, and
ments to test the parallel implementation of the proposed solution, 19.65% in terms of RMSE, MAE, and MAPE, respectively. The Wavelet-
ranging from one node to 450 nodes. The results show that the com- BCRF is the solution to the mismatch between the sampling frequency
bination of AnEn and ANN gives the best results and well-suited for and forecast horizons that were determined independently [105].
massive scale computation [102]. Wang et al., presented a novel hybrid method for deterministic PV
Ogliari et al., presented a comparison between the PV power output power forecasting based on WT and Deep Convolutional Neural
forecasting performed by physical-deterministic methods based on the Network (DCNN). To evaluate the probabilistic information in power
Nominal Operating Cell Temperature (NOCT) thermal model of the PV data, a probabilistic PV power forecasting model was developed. This
module and hybrid stochastic-deterministic. This hybrid model merges probabilistic method combines the proposed deterministic and spine
the CSRM and the ANN models. This proposed method functions to find quantile regression (QR). The deterministic and probabilistic fore-
the best performance conditions. For the deterministic model, three and casting models were verified on real PV power datasets collected from
five parameters of electric circuit models were considered. For the two PV farms in Belgium. Based on the results, the proposed WT with
hybrid PHANN, different approaches in the training set composition DCNN model exhibits a more stable forecasting ability for the de-
and dimension were analyzed. From the result analysis, three terministic performance evaluation. The proposed method based on WT

Fig. 9. Model classification based on temporal and spatial resolutions [22,108].

474
Table 2
Summary of solar PV power forecasting methods.
S. Sobri et al.

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
(year) horizon e metrics measurement nal time variable methods
Advantages Disadvantages

Melit et al. Artificial Neural 24-h ahead – MBE Sunny day: The training – Mean daily solar Solar 1st July Trieste Italy Power – K-fold cross validation
[23], Network – Multi – RMSE r = 98–99% tine is 10- irradiance irradiance 2008–23rd generated method is carried out to
2010 Layer Perceptron –r Cloudy day: min – Mean daily air May 2009 & by 20 kWp check the generalization
architecture r = 94–96% temperature 23rd Nov GCPV plant capability of MLP-
2009–24th predictor
Jan 2010 – ANN-MLP architecture
has a lower computing
time

1zgi et al. Artificial Neural 0–60-min – RMSE Training phase: N/A – 750 Wp solar PV Solar power April and Istanbul, – – Levenberg-Marquardt
[24], Network – ahead –r 5 min panel August Turkey Back-propagation
2012 Levenberg RMSE = 19.9515 W – Ambient and cell algorithms provide the
Marquardt Back- r = 0.98188 temperature way to find the weights in
Propagation 35 min – Diffuse solar order to achieve the best
algorithm RMSE = 25.3816 W irradiation mapping amongst input
r = 0.96979 and output
Testing phase:
5 min
RMSE = 35.4374 W
r = 0.93849
35 min
RMSE = 54.1102 W
r = 0.83766

475
Kardakos – Seasonal- Day and NRMSE Lower NRMSE SARIMA N/A Solar radiation PV power 1st Jan Greek – Pure – The pure SARIMA – The pure SARIMA
et al. Autoregressive Intra-day with exogenous: 2011–31st territory SARIMA model improves the model cannot
[25], Integrated Moving ahead Factor = 11.12% Dec 2012 model prediction closely to the forecast accurately
2013 Average ANN = 11.26–11.42% – real PV generation for the day-ahead PV
– Artificial Neural Persistence intra-day prediction generation when
Network with model previous days show
multiple inputs irregular patterns
with respect to the
forecast day

Chen et al. Fuzzy and Neural Hourly MAPE MAPE = 6.03–9.65% N/A – Past solar radiation Solar Past one- N/A – Statistical – Fuzzy logic is used for – The forecast
[26], Networks methods starting data radiation month data – Fuzzy an optimal clustering accuracy for fuzzy
2013 from – Sky conditions algorithm process, where fuzzy logic is not good as
8.00 a.m. to (sunny, rainy and – Neural logic functions to the NNs for a single
18.00 p.m cloudy) Network differentiate different sets weather condition
– Temperature of sky situations and
temperature

Mellit et al. Artificial Neural Hourly – RMSE Cloudy day: N/A – Solar irradiance Power of PV 1st Jan Marmara – – ANN model does not
[27], Network – Multi – MBE – R is between 93% and – Air temperature modules 2011–24th University Polynomial require more variables or
2013 Layer Perceptron –R 97% Feb 2012 Istanbul, and multiple complex calculations,
architecture – MBE in the range of Turkey linear unlike implicit models
−0.7% and −1.1% regression – The categorization of
– RMSE is less than 0.2% – Analytical the days into cloudy and
Sunny day: approach sunny provides a
– R is between 96% and – One-diode significant improvement
97% model of ANN-based models
– MBE in the range of compared to the single
−0.94% and −0.98% database which consists
– RMSE is less than 0.2% of all day
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

– The faults based on the


power loss in the PV
module can be detected
by using the proposed
ANN

Notton et al. Artificial Neural 10-min – MAE For optimal N/A – Declination Solar global 5 years France Traditional – This ANN model allows
[28], Network – Multi ahead – RRMSE configuration: – Time irradiation method getting the reliable input
2013 Layer Perceptron – MBE RMSE is around 9% – Zenith angle on tilted data for optimizing and
architecture with – RMAE RMAE is around 5.5% – 10-min of extra- plane modelling solar energy
Feed-Forward Back- – RMSE terrestrial horizontal systems for both
Propagation – R2 irradiance and photovoltaic and thermal
horizontal global applications
irradiation – From the sensitivity
– Inclination angle analysis, by minimizing
(450 and 600) the input data number
couldn’t change the
optimal network
structure. This sensitivity
study is not time
consuming and simple in
its application

Yona et al. Based on insolation 24-h ahead – MAE Average of MAE: N/A – Weather reported PV power Actual Japan – Fuzzy – RNN with feedback
[29], forecasting by – RMSE By using combination of data (Humidity, cloud output insolation theory loop was selected due to

476
2013 utilizing weather fuzzy theory, the amount, and data: 1st Jan – Fuzzy its effectiveness. This is
reported data, fuzzy correction method, and temperature) – 31st Dec theory and because by using simply
theory, and Radial RNN provide the lowest – Insolation forecast 2005 correction NN it is hard to keep the
Basis Neural values of MAE data with fuzzy method past information in time-
Network average = 0.1327 kW – series data forecasting
Persistence – The insolation forecast
forecasting data assists to train the
method RNN smoothly
– Feed-
forward NN

Amrouche Spatial modelling 1-day ahead – MSE During 2nd period, N/A Daily weather Local Global June – French – Statistical – The proposed model
et al. and Artificial Neural – RMSE lowest MSE = 16.4593 forecasts provided by horizontal 2008–May Alps model can forecast the GHI by
[21], Networks – Multi and RMSE = 33.10 for 2 the US NOAA: solar 2009 – Cadarache – Geometric only using NOAA
2014 Layer Perceptron different location – Solar radiation only irradiation model forecasted information at
architecture respectively – Solar radiation with the desired location
ambient temperature – The model can learn the
GHI profile continuously.
Therefore, this model is
suitable for locations,
which do not have a
measured data

Almonacid Dynamic Artificial 1-h ahead – RE – R2 close to 1 N/A – Global solar PV power 2011–2012 Spain Linear – The developed model is
et al. Neural Network: – RMSE – RMSE = 3.38% irradiance output regression a new tool, that
[30], Non-linear – R2 – RE around 0 with a – Air temperature generator analysis effectively employing
2014 Autoregressive margin of error – PV power output solar energy sources into
Neural Network around ± 8% module the electrical grid

5-min ahead N/A


(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

Dahmani Artificial Neural – MAE and The relative RMSE value – Horizontal Solar tilted 2 years of 5- Bouzareah, – 5 input – This model allows the
et al. Network – Multi nMAE was approximately extraterrestrial global min data City of data completion of some tilted
[31], Layer Perceptron – MBE and 8.81% for optimum irradiance irradiation Algiers – 4 input solar radiation missing
2014 architecture nMBE configuration – Solar declination data data from a database
– RMSE and – Zenith angle (without consisting of the
nRMSE – Azimuth angle azimuth) horizontal data
– Horizontal – This model allows
irradiation dispensing of an inclined
pyranometer usage once
the developed ANN has
sufficient accuracy

Kaushika Artificial Neural Hourly – RMSE RMSE = 5.19% N/A – Latitude Clearness January–De- New Delhi, NASA SSE – The model becomes
et al. Network based on – MBE MBE = −0.194% – Longitude index: cember India data sets useful to forecast the
[32], explicit approach – Altitude – Global direct, diffuse, and global
2014 – Local mean time solar solar radiation
– Month of the year radiation components at an
– Monthly mean – Direct arbitrary location
hourly relative solar
humidity and total radiation
rainfall – Diffuse
– Mean duration of the solar
sunshine per hour radiation

Giorgi et al. – Artificial Neural 1, 3, 6, 12, – NMBE Prediction Error, Ei = in N/A – PV power PV power 5th March GCPV plant, Different – The first layer of Elman

477
[12], Network -Elman and 24-h – NMAE the range of −10% to – Module temperature 2012–5th Italy input network, consists of
2013 algorithm ahead – SD +10% – Ambient March 2013 vector- I, II recurrent connection that
– Statistical methods – NRMSE temperature and III allows generation and
based on multi- – SDE – Solar irradiance on detection of time-varying
regression – DISP plane with a tilt angle patterns
– SKEW of 3° and 15°
– KURT

Teo et al. Artificial Neural N/A RMSE Training phase: N/A – Ambience PV power 8th June N/A – Activation – ELM training speed is
[33], Network –ELM – Standard temperature 2014–6th function generally faster, avoids
2015 algorithm deviation = 0.0798 – Panel temperature July 2014 used choosing additional
– RMSE = 3.5366 – Aggregated – Input parameters such as the
Testing phase: Daily Energy variables learning rate and
– Standard – Irradiance sequences stopping criterion, good
deviation = 0.0881 – Power – Number of generalization, and
– RMSE = 3.8574 input universal approximation
variables capabilities
used

Kashyap Neural Network – Last 50-h – RMSE – RMSE error = 9–10% N/A – Global horizontal Global Solar 5000-h of India – FFNN – The proposed model – Timing
et al. Levenberg error and 25–30% irradiance Radiation the year – FFBP has excellent results for – No guarantee of
[34], Marquardt – Delays – Delays = 15/30 – Diffuse horizontal 2008 – RBNN the different neurons the neuron setting
2015 algorithm with – Neuron – Neuron = 10/20 irradiance – PNN application
multiple parameters – Transfer – Transfer – Direct normal – Elman-
function function = Tansig radiation BPNN
– Model – Model = Elman back – Bulb point – Cascaded
– propagation temperature FFBP
Parameters – Parameters = Bulb – One-point – Custom
point temperature/ temperature network-1
Direct Normal Radiation – Pressure
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

– Wind direction – Custom


– Wind speed network-2
– Hours
– Number of days
– Number of months

Cornaro et al. Statistical Model 24-h ahead – RMSE Improvement RMSE N/A Statistical model: Global Jan ESTER Persistence – The master – The ANN
[16], (ST) and Model – MAE metrics of Persistence STNNS horizontal 2008–Dec outdoor Model optimization process generalization
2015 Output Statistics – MBE Model: – Ordinal day number irradiance 2011 Laboratory functions to optimize the capability will lose
(MOS) based on an STNN = 10% – Daily irradiation University neuron number in the when the neurons
ensemble Artificial NWP = 15–25% – Average daily of Rome hidden layer number is
Neural Networks by MOSNN = 30% temperature “Tor – ST model used only increasing
master optimization – Clearness index Vergata” local measured data,
process MOS model: Site while MOS corrects the
ECMWF-MOSNNS NWP data
– Ordinal day number
– Daily irradiation
– Average daily
temperature
– Clearness index
– Daily irradiation
prediction provided by
the ECMWF database

Zhang et al. SDD engine and 24-h ahead – nMAE The lowest overall errors N/A – Site 1 = Ambient Solar power – Site – Site – Naïve The kNN and WkNN The RBFNN and LS-

478
[35], Forecasting Engine – nRMSE obtained when the kNN temperature, cell 1 = 1st July 1 = USA model forecasting engines are SVM need a model
2015 of RBFNN, LS-SVM, and WkNN forecasting temperature, and wind 2010– 31st – Site simple and can compute building and
kNN, and WkNN engines combined with speed December 2 = Denma- effectively compared to training step. This
the SDD engine – Site 2 = Climate 2011 rk RBFNN and LS-SVM step is not
Site 1: forecast data including – Site – Site straightforward due
nMAE = 7.4% (WkNN) of GHI, high, medium, 2 = 1st 3 = Italy to the requirements
nRMSE = 9.82% (WkNN) low, and total cloud January – of model
Site 2: cover, fog, total 31st identification,
nMAE = 6.38% (kNN) ambient temperature December prediction, and
nRMSE = 8.72% (kNN) and wind speed 2006 testing stages that
Site 3: – Site 3 = Solar – Site are sometimes
nMAE = 7.74% (WkNN) altitude, GHI, DNI, 3 = 1st computationally
nRMSE = 10.37% total cloud cover, and January – burdensome
(WkNN) ambient temperature 31st
December
2011

Lima et al. NWP-WRF model by 24-h ahead – RMSE ANN: N/A – Weather forecast Surface 2009 and Brazilian The surface – The WRF model is a – The NWP model
[36], using Artificial – ME – MSE = 20.95 W/m2 outputs supplied by solar 2011 for North- solar mesoscale NWP model forecast presents a
2016 Neural Networks – MSE – RMSE = 29.48 W/m2 WRF model irradiance rainy and Eastern irradiance which functions in both large bias, thus
– R – R = 0.88 – Observational dry seasonal Region data forecasts operation and reducing their
WRF: ground data obtained by atmospheric research. It reliability and
– MSE = 29.51 W/m2 automated is an on-hydrostatic confidence
– RMSE = 42.87 W/m2 weather model, fully
– R = 0.78 stations compressible, and could
satisfy the demand

Mixed Wavelet 15-min and – MBE WNN outperforms other Clear sky index Solar 2014 of 1- Singapore – – The high signal
Neural Network 1-h ahead – NRMSE benchmark models in irradiance year data Persistence compression ability of
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

Sharma et al. terms of all performance The WNN model wavelets is suitable for – The optimal WNN
[37], metrics takes 11 min – ARIMA modelling of the non- architecture varies
2016 15-min: for training model stationary environmental for different seasons
NRMSE = 9.42% – EST model variables with high
1-h: – ANN-MLP information content
NRMSE = 15.41% – Unnecessary for online
monitoring and real-time
prediction of weather
parameters such as
irradiance

Graditi et al. Comparative N/A – nRMSE Sandia model: N/A Meteorological, PV power 2006–2012 Italy Schokley- Advantages of Gas:
[38], analysis between a – R2 nRMSE%_AVG = 10.09 thermal, and electrical Sandia – Identify the minimum
2016 Phenomenological – R2_AVG = 0.978 data consists of: model and meaningful datasets
Sandia model and Regression model: – GIpoa (Global – Useful in optimization
two statistical nRMSE%_AVG = 7.01 radiation on plane of process where the
learning techniques R2_AVG = 0.980 array) objective functions are
– Artificial Neural ANN model: – Tamb (Ambient discontinuous, non-
Network and nRMSE%_AVG = 6.66 temperature) differentiable, and non-
Regression models R2_AVG = 0.982 – Wspeed (Wind speed) linear
– Tm (PV module
temperature)

Hossain et al. Extreme Learning – Day ahead – RMSE Day ahead: Day ahead: – Daily average solar PV power 1st January Malaysia – SVR – ELM has very high
[39], Machine (ELM) – 1-h ahead – R2 – RMSE = 13.83–21.84% – 0.22 s irradiance 2015–30th – ANN learning speed with

479
(2017) – MAPE (training) (training) – Ambient September higher performance
– RMSE = 17.89–35.39% – 0.0335 s temperature 2016 (12
(testing) (testing) – Module temperature months)
1-h ahead: 1-h ahead: – Wind speed
– RMSE = 55.32–89.55% – 0.2496 s – PV output data
(training) (training)
– RMSE = 54.96–90.1% – 0.0153 s
(testing) (testing)

Shi et al. Support Vector 1-day ahead – MRE Average N/A – Historical data in the PV power 13th China – Cloudy – Has good calculating – Requires a big
[44], Machine – RMSE – MRE = 8.64% previous movement output January model speed and convergence amount of data for
2012 – RMSE = 2.10 MW – Weather report of 2010–29th – Foggy results compared to ANN fitting and
next day October model regression
2010 – Rainy
model
– Sunny
model

Mellit et al. Least Square – 1-h ahead – R2 LS-SVM outperforms 4-min for – Solar irradiance – Solar 1st Jan – Medina ANN – The reliability of LS-
[43], Support Vector – RMSE ANN models by choosing training – Wind speed irradiance 31st Dec City, Saudi models: SVM method can be
2013 Machine – MBE only 2 parameters time – Wind direction – Wind 2011 Arabia – MLP checked by using K-fold
– MAPE Pressure: computation – Air temperature speed – RBF cross validation instead
– R2 = 0.971 – Relative humidity – Wind – RNN of the holdout technique
– MAPE = 0.48% – Pressure direction – PNN
Air temperature: – Air Only use 2
– R2 = 0.974 temperature parameters
– MAPE = 1.2% – Relative (air
humidity temperature
– Pressure and
pressure)
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

Zeng et al. Least Square – 1-h, 2-h, and – MAE Minimum of MAE = 33.7 N/A – Historical data of Solar power 1991–2005 US -Seattle, – AR – Able to capture the
[45], Support Vector 3-h ahead – MAPE w/m2 by using atmospheric Deriver – RBFNN time-varying and non-
2013 Machine –ρ transmissivity transmissivity in 2-D Miami linear patterns of the
normalization methods form solar irradiation data
– Meteorological – Has a good
parameters i.e. sky generalization ability
cover, relative – Hourly solar radiation
humidity, and wind in 2-D representation
speed gives more insight into
the solar radiation
pattern

Chen et al. Support Vector Daily -RMSE Average: N/A Sunshine duration Solar 1976–2010 China – Empirical – The SVM tries to reduce
[46], Machine – RRMSE RMSE = 2.094 mJm−2 radiation models the generalization error
2013 – R2 (10% lower than 2.323 – SVM 1–7 upper bound rather than
mJm−2 produced by with local training error
empirical models) different – Since the optimality
input problem is complex, it
attributes delivers a unique solution

Ekici [47], Least Square – Next-day – RMSE R2 = 99.294 N/A Daily solar 2000–2003 Elazig city AI – The advantages of RBF
2014 Support Vector – MRE RMSE = 0.004384 – Daily mean insolation Turkey techniques kernel function are:
Machine – MEF MRE = 9.9617 temperature And 1. Can compete
– R2 MEF = 3.3188 – Daily max empirical effectively
– CVRMSE CVRMES = 0.094611 temperature techniques 2. Reliable

480
– Sunshine duration in literature 3. Simple
– Solar insolation of 4. Ease of adaption for
the day before optimization due to its
capability in handling the
more complicated factors

Mohammadi Support Vector Daily and – MAPE Daily: N/A Horizontal Jan Iranian – ANN – By using wavelet
et al. Machine – monthly – MABE MAPE = 6.9996% – Daily global solar global solar 1992–Dec coastal city – GP analysis, the resulting
[40], Wavelet algorithm – RMSE MABE = 0.8405 mJ/m2 radiation radiation 2005 – ARMA decomposed components
2015 – RRMSE RMSE = 1.4295 mJ/m2 – Sunshine duration with proper scaled
– R2 RRMSE = 7.94679% – Average, max, and resolution show the
R2 = 0.9086 min ambient perfect analysis
Monthly: temperature – The wavelet analysis is
MAPE = 3.2601% – Relative humidity suitable to analyze the
MABE = 0.5104 mJ/m2 – Water vapour frequency and time
RMSE = 0.6618 Mj/m2 pressure domain data due to its
RRMSE = 3.6935% ability to extract non-
R2 = 0.9742 periodic and transient
signal data

Olatomiwa Support Vector Monthly – RMSE Training: N/A – Sunshine duration Global solar 1987–2007 Nigeria- – ANN – The proposed SVM-FFA
et al. Machine – Firefly – R2 RMSE = 0.6988 – Max and min radiation Iseyin, – GP developed model shows
[41], algorithm – r R2 = 0.8024 temperature Maidu guri, an attractive approach to
2015 – MAPE r = 0.8956 and Jos forecast global solar
MAPE = 6.1768 radiation, which is
Testing: applicable for all case
RMSE = 1.8661 studies even for different
R2 = 0.7280 climate conditions
– The FFA is applied to
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

r = 0.8532 determine optimal SVM


MAPE = 11.5192 parameters

Ramli et al. Support Vector N/A – RMSE – In training and testing Maximum – Global radiation Solar 1 year Jeddah and ANN – The SVM shows an
[48], Machine – CC phase, the values of MRE training – Direct normal radiation Qassim, excellent performance
2015 – MRE for SVM are 0.33 and time during radiation Saudi due to it is the ability in
– Speed of 0.51 respectively training =- – Diffuse radiation Arabia maintaining the
computation – Faster in solar radiation 2.15 s robustness even for the
data forecast with the Maximum very small dataset
lowest speed of 2.15 s time during – The SVM has the faster
compared to 4.56 s for testing =- speed of computation
ANN during training 0.0468 s – The SVM has the
– SVM has lower RMSE capability to reduce
values errors during
computations

Wolff et al. Support Vector 15-min and – RMSE The error minimization is N/A – PV power PV power March–Nov- Germany Physical – SVR can generate better – The learning stage
[49], Regression 5-h ahead – Bias over 75% for the measurements ember 2013 modelling prediction at clear sky is working well on
2016 combination of model – Irradiance forecasts approaches and overcast conditions single systems, but
prediction and high from NWP on average this
quality of regional PV – Irradiance forecasts advantage is lost
power forecasts from cloud motion especially for very
vectors (CMV) high irradiance
input cases

Bhardwaj Density Hidden Long-term – RMSE – RMSE = 7.9124 N/A – Sunshine hour Solar 2009–2011 India – ANFIS Advantages of HMM: Disadvantages of

481
et al. Markov Model, forecasting – MAPE – MAPE = 3.0083 – Ambient radiation – ANN – It can deal with HMM:
[51], (HMM) with – R-value – R-value = 0.9921 temperature – sequential structures and – There is the
2013 Generalized Fuzzy – Relative humidity Combinatio- a fair amount of inadequacy of
Model (GFM) – Wind speed n of inputs uncertainty global context in
– Atmospheric meteorolo- – Have the powerful modelling
pressure gical optimization algorithms – The initial
parameters such as Expectation assumption is
Maximization required for the
Advantages of soft- statistical
computing: distributions
– GFMs have very Disadvantages of
efficient modelling soft-computing:
structures – There is no
– No initial assumptions specific way to deal
are required for the with time sequential
statistical distributions structures
– The computation
of learning
algorithms is costly
and time consuming

Sanjari et al. Higher-order 15-min MAE Proposed method has the N/A Historical data: Probabilistic Years 2014 Australia – Pair – PDF of the PV generated
[52], Markov Chain ahead lowest MAE with average – PV output power distribution and 2015 Copula power can be predicted
2016 value of MAE is 2.18% – Ambient function Constructio- precisely irrespective of
temperature (PDF) of the n the PV system operating
– Solar irradiance generated – point
PV power Chronologi- – This model can be
cal updated with the
Probability sequential data
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

– Support
Vector
Regression

Dazhi et al. Autoregressive Next hour – MBE Minimum RMSE N/A – Global horizontal Solar Orlando: USA – Advantages by
[57], Integrated Moving – RMSE = 29.73% (Miami) irradiance irradiance October Persistence considering cloud cover
2012 Average model = 32.80% (Orlando) – Direct normal 2005 model effects:
irradiance Miami: – Model 1 – Only low resolution is
– Diffuse horizontal December forecasted needed
irradiance 2004 – Model 2 – Only ground-based
– Cloud cover forecasted cloud cover data is
– Model 3 required to predict the
forecasted solar irradiance for the
next-hour
– Very useful during the
lack of solar irradiance
measurements in
determining its accuracy

Huang et al. Coupled Auto- 1-h ahead – MeAPE CARDS model has N/A Solar N/A Mildura, Literature – Improve the ability of
[53], Regressive and – MBE reduced prediction error radiation Australia review prediction in difficult
2013 Dynamical System – KSI of combination model (other weather conditions
(CARDS) model – NRMSE (AR (2) and Luncheroni methods) – The dynamic system
model): model is suitable for
MeAPE = 33.4% high-frequency analysis

482
(11.31–7.53%) of data such as solar
irradiance

Yang et al. Spatial-Temporal 5-min, 15- – MAE ST model outperforms N/A Real historical data Solar PV The whole California – – The proposed model
[58], Autoregressive with min, 1-h, – RMSE the other models at 1-h obtained from year 2010 Persistence was already
2015 Exogenous Input and 2-h ahead and 2-h ahead California Irrigation (PSS) model implemented, and it was
(ST-ARX) model ahead Management – AR expanded by adding the
Information System – BPNN meteorological
(CIMIS) website information to give more
accurate prediction
without creating an
additional computational
burden
– Quality-controlled-real
data are used to improve
the forecast accuracy in
training and validation
stages

Agoua et al. Statistical spatio – Few min up nRMSE The nRMSE can be N/A First dataset: PV power First France – – The forecasting error – High dimension
[59], temporal Model to 6 hours improved by 20% or – Time series of the dataset: Autoregress- can be reduced by data amplifies the
(2017) more as compared to the measured PV – Starting ive including the effects of dimensionality
state-of-the-art generation of 9 plants from July – Random meteorological problem
forecasting techniques sets 2013 (20 forest parameters such as wind
Second dataset: months) speed
– Different spatial Second
densities of installed dataset:
PV plants – November
2014–March
2016
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

Trapero et al. Dynamic 1–24-h – rRMSE Global Horizontal N/A Solar irradiation data -GHI Jan 2009 Spain – Naive – The DHR provides fast
[61], Harmonic ahead – rMBE Irradiation – DNI and Dec – Seasonal automatic identification
2015 Regression – rMBE = 0.21 2011 naive – This developed model is
– rRMSE = 29.66 – ETS in a the first used in solar
Direct Normal Irradiation state space irradiation prediction
– rMBE = 3.82 framework
– rRMSE = 46.79 – ARIMA

Wang et al. Partial Functional 1-day ahead -RMSE Regularized PFLRM: N/A – Intercept PV power 1st January Coloane – Neural – The regularized PFLRM
[62], Linear Regression – MAD K=2 – Pressure 2011–30th island of Network does not require previous
2016 Model – MAPE – RMSE = 59.1704 – Max, mean, min and June 2012 Macau models experience and provides
– Regularized – MAD = 40.6277 deviation temperature Special – Multiple reliable information,
algorithm – MAPE = 0.1134 – Humidity Administrat- Linear unlike NN models
K=7 – Insulation ive Region Regression – Simpler and involves
– RMSE = 59.3899 – Wind speed (SAR) models only a few parameters
– MAD = 40.4519 – Precipitation – Automatically chooses
– MAPE = 0.1204 – FPC score the parameters that
highly affect the response
parameters
– Able to model the non-
linear structure

Li et al. [63], Multivariate 24-h ahead -RMSE For testing data: N/A – Daily average, Solar Power 1st January Coloane Linear – Maintains the simplicity
(2016) Adaptive Regression – MAD – RMSE = 119.0 lowest, and highest 2011–20th Island of models: of the classical multiple
Splines (MARS) – MAPE – MAD = 89.8 temperature June 2016 Macau – MLR linear regression models

483
– MAPE = 69.2% – Dew temperature Special – ARIMA while possessing the
– Wind speed Administrat- – ARMAX capability of handling
– Precipitation ive Region Non-linear nonlinearity
amount (SAR) models: – Improved version of the
– Duration of – ANN traditional linear
insolation – SVR regression
– Humidity – kNN
– Air pressure – CART

Massidda Multilinear Adaptive 24-h ahead – RMSE Forecast errors: N/A – NWP data Solar PV Year 2014 Germany Persistence – Well-known for ‘data-
et al. Regression Splines – MAE – RMSE = 177.8 kW – Power production power model mining’ as it does not
[64], – MAE = 125.9 kW history data make any assumption or
(2017) The errors of daily energy set any particular class of
production of the plant: relationships between the
– RMSE = 1094.3 kWh input parameters and the
– MAE = 860.1 kW dependent parameter
– This model kept as
simple as possible
without the introduction
of any unnecessary
complexity

Persson et al. Gradient Boosted 1-h t 1–6-h ahead NRMSE was between 0.1 N/A – Historical power Solar Power 19th April Japan – – Able to catch highly – GBRT model has
[65], Regression Trees and 0.137 kW for lead observations Generation 2014–28th Persistence non-linear patterns no simple updating
(2017) (GBRT) time k = 1–k = 6 – Information about February model between input variables procedure
the location of the PV 2015 – and target response in a
installation Climatology fast and robust manner
– NWP data – Recursive
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

autoregres-
sive model

Mathiesen Analysis and MOS NAM: – MBE Reduction of RMSE is: N/A SURFRAD ground Global NAM: Continental – – MOS application to the
et al. correction from 3 – 1 h/36 h – RMSE – Nk = 20.1 W m−2 measurement data Horizontal 21st March US (CONUS) Persistence NWP irradiance output
[69], operational NWP – 3 h/84 h – E = 17 W m−2 Irradiance 2010–8th model can reduce the bias and
2011 models: GFS: – G = 25.6 W m−2 February – Clear sky RMSE
– North American – 3 h/180 h Effectively eliminated 2011 – The MOS model is
Model (NAM) ECMWF: MBE GFS: based on cloud cover
– Global Forecast – 3 h/144 h 23rd June variables and not
System (GFS) – 6 h/240 h 2010–8th irradiance
– European Centre February – MOS corrections lead to
for Medium-Range 2011 more precise solar
Weather Forecasts ECMWF: irradiance prediction
(ECMWF) 1st given by a location, time,
September and basis prediction
2009–31st model
August 2010

Fernandez Two cascaded 1–39-h RMSE ANN-MLP represents the N/A – Past values of hourly PV power 2nd June Spain – – The developed model is
et al. models of NWP- GFS ahead best prediction model energy production in 2007–27th Persistence able to provide direct
[70], (global) and MM5 with forecast horizons the PV plant May 2008 model energy hourly forecasts
2012 (mesoscale) with from 16-h to 39-h – Forecast values of – ARIMA that can be utilized as the
Artificial Neural The RMSE value is weather parameters – k-NN generation value for
Network – based 11.79% for the rated PV obtained from GFS – ANN- electricity markets

484
model power plan and MM5 modules based model – Offers the forecast
(MLP, ENN, values of the hourly
RBF, and electrical energy
TDNN) production for all the
– ANFIS hours of the following
day
– The forecast values can
be utilized for preparing
bid offers to the
electricity market

Monteiro Historical Similar 1–24-h – Forecast – Forecast skill = 36.3% N/A Historical values of PV power N/A Spain – Classical – HISIMI model gives a
et al. Mining (HISIMI) ahead skill – nRMSE = 10.14% variables persistence complete information
[67], model – nRMSE – Improvement to both corresponding to the model based on power
2013 benchmark models of PV plants – ANN transitions for each
MLP and Persistence based on forecasting time. The
model are 0.8% and MLP discrete probability
36.3% respectively functions were used to
completely spot forecasts
and uncertainty
prediction
– Easy to update the
database used by this
model

Chow et al. Total Sky Imager 30-s to 5- Cap errors Forecast error reduction N/A – Sky images – Cloud Since University Persistence – This single TSI is – The images are
[76], min ahead is 50–60% compared to – Sunshine parameter cover August 2009 of Model feasible for the short easily saturated near
2011 persistence model – Solar California, deterministic forecast the sun even with
irradiance San Diego horizons usage AGC of the CCD
camera due to the
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

low dynamic range


– Some part of the
data within each
image is lost
because of
obstruction the
shadow band and
camera arm
– The cloud
deformation,
evaporation,
condensation, and
uncertainty in cloud
base height can
cause the
inaccuracies of sky
conditions forecast

Gohari et al. Total Sky Imager 15-min – Pearson, USI outperforms than TSI N/A Global Irradiance Clear Sky TSI: USA Persistence Advantages of USI: Disadvantages of
[77], (TSI) and University ahead Spearman, USI: Index 1st–21st model – Does not have missing TSI:
2014 of California, San and – Correlations of clear sky July 2011 data caused by the – The shadow band
Diego Sky Imager Euclidean index 0.7 and 0.82 12th–17th shadow band on the TSI mirror
(USI) correlation TSI: July 2012 – Does not lose blocks the sky dome
coefficient – Correlations of clear sky USI: information to image view near the sun
index 0.55 and 0.69 1st–29th compression will give poor

485
Nov 2012, – Has both a higher nowcast
which is spatial and intensity
excluding resolution
12th Nov – No occultor blocking
2012 the sun
Advantages of TSI:
– Has a large spatial
average over the power
plant

Peng et al. Multiple Total Sky 1-min and – STI Achieve at least 26% N/A Weather and cloud – Solar 13th May US Persistent – The proposed model – Limitation in the
[78], Imagers 15-min – MAE improvement for all conditions: Irradiance 2013–3rd model has robustness in clouds observation FOVs
2015 ahead – RMSE irradiance predictions 1. Single layer clouds – 3-D Cloud June 2013 detection and layers – The stereo
compared to persistent 2. Multi-layer clouds Tracking tracking stitching of sky
model 3. Overcast or images from
extremely cloudy multiple TSI sources
4. Mixture thereof enlarge the visible
– Images features of range, but it can
sun-blocking pixel only expand the
– GHI prediction range to
a certain degree

Montesinos Sky Camera Images 15-min nRMSE Average nRMSE: N/A Digital Image Levels – Beam solar 2010–2014 Spain – Cloudless – The proposed model is
et al. ahead – Beam radiation skies only using digital image
[79], Irradiance = 25.44% – Diffuse – Partially- levels in forecasting from
2015 – Diffuse solar cloudy skies a terrestrial view
Irradiance = 11.60% radiation – Overcast
– Global – Global skies
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

Irradiance = 11.17% solar


(under all sky conditions) radiation

Marquez Combination of 30-min, 60- – MBE Improvement of RMSE N/A – Velocimetry – Global For GHI San Joaquin – – The combination of
et al. Satellite images and min, 90- – RMSE (lowest) – Cloud indexing horizontal data: Valley Persistence stochastic learning,
[75], Artificial Neural min, and 5–25% (information obtained irradiances 1st January models ground, and/or remote
2013 Network 120-min from satellite images) 2011–31st – Other sensing image processing,
ahead – Irradiances January methods in and ground telemetry
2012 Literature provide benefits in
Review robustness and accuracy
for the high-fidelity solar
forecasts development

Dong et al. Satellite image Hourly – nRMSE The developed model N/A Cloud cover index Solar For satellite Singapore – ARIMA – SOM is functioned to
[81], analysis and hybrid – R2 shows 6% improvement – Satellite images from Irradiance and SERIS – LES Model classify the cloud cover
2014 Exponential – nMBE of forecast accuracy the geostationary data: – SES Model index since it was the
Smoothing State compared to the best weather satellites of Sept – RW most effective NN
(ESS) Space with forecasting accuracy of Japan Meteorological 2010–July learning techniques
Artificial Neural all other statistical Agency (JMA). 2011 – The state space model is
Network model approaches – The solar irradiance a more sophisticated
time series dataset stochastic concept
from the because the state space
meteorological station model provides the
of the Solar Energy framework for computing
Research Institute of estimation interval and

486
Singapore (SERIS) other properties

Aguiar et al. Satellite-derived 1–6-h ahead – RMSE The best model is a N/A – Ground data set Global Data in the Island of: – Smart – By adding exogenous
[82], data and NWP – MAE combination of NN – Satellite-derived Horizontal year 2005 C0-Pozo Persistence data of satellite and
2016 model – Skill +ECMWF+SAT. RMSE: data set (Global Irradiance Izquierdo – ECMWF forecasts with
– C0-Pozo Izquierdo horizontal irradiation and C1-Las Climatology ground data on ANN
Station: and irradiation on top Palmas – NN models could improve its
1-h = 15.47% of the atmosphere) – NN+SAT accuracy in all cases
6-h = 22.17% – ECWMF data set – NN – The Bayesian
– C1-Las Palmas Station: (Total Cloud Cover +ECMWF framework of the
1-h = 24.15% and Surface Solar – NN learning process was used
6-h = 34.09% Radiation +SAT+EC- in NN to choose the most
Downwards) MWF relevant input, to avoid
the fitting problem, and
to optimize the ANN
architecture

Quan et al. Hybrid model: 1-day ahead – WMAE NMAE error for the wind N/A Physical variables – PV power 1 year Italy Classical – The proposed hybrid
[84], Hybrid Evolutionary – NMAE case (enma = 14.0%) is provided by the – Wind EBP alone learning process offers
2013 Algorithm to more than twice of NMAE weather forecasting power faster convergence
Artificial Neural calculated for the PV service compared to standard
Network plant for a period of 1- – Day EBP alone
year – Irradiance
– Environmental
temperature
– Humidity
– Atmospheric
pressure
– Wind speed
– Cloud coverage
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

Zhang et al. Recurrent Neural N/A RMSE Cumulative RMSE for the N/A – Solar radiation data Solar 27th June Washington N/A – PSO stores valuable
[85], Network based on best individual using – Real-time solar radiation 2011 and information with local
2013 hybrid learning hybrid PSO-EA algorithm energy data 15th July and global memory
algorithm dropped to 0.028 after 2011 during the whole of
incorporating 100 iterations evolution
Particle Swarm – The property of
Optimization (PSO) mutation in EA assists to
and Evolutionary retain the PSO diversity
Algorithm (EA) in flying into the new
search field

Bouzerdoum Hybrid model: Hourly – NMBE -NRMSE = 9.4010% N/A Climatic data: Solar PV 28th Trieste, Italy – SARIMA – The proposed model
et al. SARIMA and – NRMSE – NMBE = 0.1790% – Irradiance on the power January – model does not require any
[86], Support Vector – R – MPE = 2.7381% array plane 25th May – SVM forecasted meteorological
2013 Machine model – MPE – R = 0.9908 – Modules 2009 model parameters
temperature – The proposed model
– Ambient can be easily developed
temperature at array from well-known
side MATLAB function
Electrical data:
– String voltage
– Current
– Power
– AC power

487
Haque et al. Hybrid intelligent 1-h ahead – MAPE MAPE, NRMSE, and error N/A – Historical PV power PV power May 2011 Oregon – BPNN – The WT is developed to
[87], algorithms: – NRMSE variance of the proposed time-series – RBFNN extract the ill-behaved PV
2013 Combinations of. – Error model are lower – Solar radiation data – GRNN time-series
Wavelet Transform variance compared to other – FA characteristics. FA
(WT), Fuzzy (σ2) models in sunny, cloudy, – network is developed to
ARTMAP (FA) and rainy days for a WT+BPNN capture non-linear PV
network, and Firefly winter, spring, summer, – power fluctuation
(FF) optimization and fall seasons WT+RBFN- focussing on the plasticity
algorithm respectively N and stability. FF
– optimization algorithm is
WT+GRNN introduced to optimize
– WT+FA the FA network more
efficiently, thus enhances
the entire forecast
accuracy

Gandelli Hybrid method: Hourly – NMAE High accuracy and lowest N/A – Weather forecasts PV power 10 months N/A Standard – This proposed model is
et al. PHANN – Basic – WMAE of errors in Sunny days: – Geographical plant ANN model developed to forecast PV
[88], physical model and – nRMSE NMAE = 2.70% coordinate of site during: energy for optimal usage
2014 Artificial Neural WMAE = 6.13% – Date and time to – Sunny and RES management in
Network Partially cloudy day: determine the correct – Partially the future smart grids
NMAE = 5.7% sun position cloudy application
WMAE = 14.52% – Cloudy
Cloudy day 1: Day 1
NMAE = 23.61% – Cloudy
WMAE = 108.82% Day 2
Cloudy day 2:
NMAE = 11.66%
WMAE = 52.16%
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

Yang et al. A weather-based 1-day ahead – MRE The proposed method N/A Parameters of PV power 1st May Taiwan – – The proposed model is
[89], hybrid method: – RMSE outperforms than ANN temperature and 2012–30th Traditional used to improve the
2014 combination of and SVR models with the probability of April 2013 ANN efficiency of building
SOM, LVQ network, lowest error: precipitation – Simple energy, public energy
SVR method, and – MRE = 3.295% SVR method saving, and real-time
Fuzzy Inference – RMSE = 350.2 W control performance of
approach modularized EMS
including RES and
storage systems

Wu et al. Hybrid model: 1-h ahead NRMSE Hybrid prediction model N/A – Historical PV power N/A Taiwan and – ARIMA – The GA is used in this
[90], Combination of generates the most measurement data (PV Malaysia – LS-SVM model because it is
2014 ARIMA, SVM, ANN, accurate predictions in power output) – ANN routinely used to
and ANFIS methods most cases – NWP data (solar – ANFIS generate useful solutions
with GA algorithm NRMSE values of hybrid irradiance) for optimization,
model: prediction works, and
– PTM Green Energy search problems
Office (GEO) = 5.64% – Combination of several
– Taichung Thermal statistical and AI-based
Power Plant = 3.43% methods is utilized for
– An academic building at the optimum weight
the National Penghu between online
University measurements and
(NPU) = 6.57% meteorological forecasts

488
Filipe et al. Hybrid of different 1–10-h – MAE The average N/A – NWP Solar power May 2013 Portugal – – The electrical model
[91], models: ahead – RMSE improvement of 57.4% – Past observations of and Nov Persistence provides a good
2015 Combination of over persistence model the time series (solar 2014 model characterization and
physical and and 34.06% over the power) – Naïve interpretation of the
statistical methods – diurnal model model physical system. It does
AR model, electrical not require to know in
model, and advance the PV plants
statistical model features by metaheuristic
optimization algorithms
– The AR model is a more
accurate model for the
very short-term horizon

Dong et al. Hybrid approach Hourly – nRMSE Average: N/A Solar irradiance Solar Colorado, USA and – ARIMA – To overcome two major
[92], based on SOM, SVR, – nMBE – nRMSE = 4% lower historical data irradiance USA: Singapore – LES problems such as the non-
2015 and PSO error 1197–2013 – SES stationarity features of
– hybrid model well- Singapore: – RW the solar irradiance and
performs for 7 out of 10 2010–2013 noise level in the time-
months compared to series
other models

Wang et al. Hybrid model: CS- N/A – RMSE Clear sky: N/A – CSKY Glo data Solar 2008–2010 US – ARMA – For the input data, the
[93], OP-ELM model – MRE – Lowest MRE = 75% – RSKY Glo data radiation of – BP high-dimension data can
2015 – Lowest RMSE = 62.5% CSKY Glo – Direct be considered
Real sky: and RSKY strategy OP- – CS optimization
– Lowest MRE = 75% Glo ELM algorithm can be utilized
– Lowest – Direct- to validate the weight
RMSE = 70.83% recursive variables with properly
chosen variables, thus
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

strategy OP- eliminates redundant


ELM information and can
build a more effective
model
– The OP-ELM model is
able to build an NN with
a big neurons number,
rank neurons, and
determine the neurons
number
– Reduce the system loss
by using the optimized
hybrid strategy by CS
algorithms

Rana et al. Ensembles of Neural Next-day – MAE – A3 outperforms A2 with Training Previous power data PV power 1st January Brisbane, – Single NN – The proposed model is – The computational
[94], Networks – MRE 7.77% and A1 with 1.05% time is in 2013–31st Australia – Two suited for both online and cost of training an
2015 – The ensemble of NN range 3–10- December Persistence offline practical ensemble Neural
provides a higher min 2013 models as applications Network is higher
precision than individual the than a single Neural
NN. Average MRE for all benchmarks Network
cases is 9.43% for A1, – SVR based
36.48%, for A2, and iterative
7.18% forA3. approach

489
Dolara et al. Physical Hybridized Day-ahead – NMAE% The PHANN method N/A – Weather forecasts PV power 8 months North of Statistical – The proposed model is
[95], Artificial Neural – WMAE% outperforms traditional – Clear sky model plant data starting Italy ANN able to forecast the daily
2015 Network (PHANN): – nRMSE% methods. The error from 1st method PV power profile more
Combines Artificial reduction reaches 50% January accurately by using the
Neural Network during sunny days 2012 available real PV plant
with an analytical historical data in a
physical model specific place and
(Clear Sky Solar weather forecasting
Radiation Model, service historical data in
CSRM) the same location as well

Piazza et al. Nonlinear 8-h, 10-h, – NRMSE In recall phase: N/A – Temperature – Solar 2002–2008 Palermo, N/A – The proposed models
[96], Autoregressive 18-h, 24-h, – CV(RMSE) 1. Solar irradiation – Global solar irradiation Sicily Italy able to solve problems
2016 Network with and 48-h prediction irradiation – Wind which are involving the
Exogenous Input ahead – NRMSEr = 6.1% – Wind speed data speed modelling of non-linear
(NARX)- Neural – CV(RMSE)r = 32% dynamic systems
Network 2. Wind speed prediction
– NRMSEr = 7%
– CV(RMSE)r = 47%

Azimi et al. Hybrid method 1-h, 24-h, – MSE The proposed hybrid N/A Historical solar data Solar Between Colorado – – The developed
[97], based on TB K- and 48-h – MAE forecasting model shows radiation 2004 and states, US Persistence prediction model has a
2016 Means clusters ahead – RMSE more accurate 2013 method faster processing time
technique and – nRMSE performance with lowest – NAR – The TB K-Means
Multi-Layer – Forecast nRMSE (%) and higher – RW algorithm classifies the
Perceptron Neural skill forecast skill for months – SES solar time-series data into
Networks of January–October – LES multiple groups to better
– ARIMA recognize its anomalies
– Hybrid and irregular patterns
method
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

Li et al. [98], Hierarchical 15-min, 1-h, – MAE The lower error values of N/A – PV’s historical Power plant 1st January Florida Traditional – The machine learning
2016 approach based on and 24-h – MBE MBE, MAE, and RMSE for record output – 31st baseline (By algorithms study the non-
the machine ahead – RMSE hierarchical approach – Online December using whole linear relationship
learning algorithms: – rMBE compared to the meteorological 2014 plant between the clear sky
Artificial Neural – rRMSE traditional way services production solar properties and the
Network and – MPE informa- PV plant productions
Support Vector tion) directly rather than going
Regression through the solar
irradiance predictions
– The hierarchical
approach shows the
potential to have a
multiple-level forecast of
the PV system

Graditi et al. – Physical model N/A – MBE/A The hybrid method is N/A Data Acquisition AC energy 2006–2012 ENEA – Physical – The AI-based models – Requires a big
[8], – AI based models (%) more accurate than ANN System (DAS) yield of a Portici model only require information historical data
2016 (MLP-NN, n-Wavelet – RMSE/A In the case the HPANN connected to an thin-film PV Research – MLP-NN about irradiance, number to train
NN and ANFIS) (%) approach: experimental plant plant Centre – n-Wavelet temperature, and network
– HPANN – CC – MBE/A is between consisting of micro- NN geographic data to – By using a physical
– R2 0.01% and 5.11% morph – ANFIS acquire a better model, the dc
– RMSE/A is between silicon modules based – HPANN prediction of Pac power, the dc
5.91% and 10.09% on a combination of – The proposed models current, the main
– CC is between 0.9824 microcrystalline are insensitive to noise features data of the

490
and 0.9934 (u-Si:H) and included in the measured PV modules and
– R2 is between 0.9644 amorphous (a-Si:H) data inverter and a
and 0.9868 silicon technology detailed knowledge
of their prevailing
physical
relationships are
required to predict
Pac

Antonanzas Single and blended Day ahead – MAE For single model: N/A NWP data: PV power 2009–2010 North of – Two-day – Hybrid models appear
et al. model: (kWh) – SVR technique – GHI Spain persistence to foster the benefits of
[99], – Support Vector – RMSE outperforms the other – Air temperature – DNN individual techniques and
(2017) Regression (kWh) models with nRMSE of – Relative humidity – SVR improve model accuracy
Machines (SVR) – nRMSE 22.49%. – Wind speed – RF
– Deep Neural (%) For blended model: Deterministic – XGB
Network (DNN) – MBE – The blended model variables: – Blended
– Extreme Gradient (kWh) outperforms all the other – Sun position model
Boosting Machines models with nRMSE – Extraterrestrial
(XGB) value of 22.24%. irradiance
– Random Forests
(RF)

Do et al. – Scaled Persistence N/A – NMBE – The forecast errors were N/A – Data of PV PV 2 years (1st -Lille, – Scaled Multivariate Polynomial Multivariate
[100], Model – NMAE reduced from 12.71% to production at time (h) generation January France Persistence Model: Polynomial Model:
(2016) – ANN – NRMSE 10.69% starting from 6th – Two exogenous 2010–31st – Model – Simple, flexible, and – The forecast
– Multivariate months by using inputs (cloud cover December Guadeloupe, – ANN easy to compute accuracy is limited
Polynomial Model multivariate polynomial and air temperature) 2011) Caribbean – – Well-known and to one input
model Multivariate understand properties variable
Polynomial Persistence Model: Persistence Model:
Model – Does not need a series – The accuracy of
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

of historical data for the forecast is not


training high

Malvoni Data-driven 1-h, 3-h, 6- – Skewness For PCAWD (skewness): N/A – Ambient PV power 1st January Italy – Baseline – The wavelet is an
et al. methods: Hybrid h, 12-h, 18- – Kurtosis – 1-h = −0.49 temperature 2013–31st – WD efficient solution to
[101], model GLSSVM that h, and 24-h – 3-h = −0.81 – Solar irradiance on December – PCA reduce the noise in input
(2017) combines the GMDH ahead – 6-h = −1.22 tilted of array 2013 – WDPCA datasets before to
and LS-SVM – 12-h = −0.62 – Wind speed – PCAWD implement a prediction
methods merged – 18-h = −0.71 – PV power model to improve the
with data pre- – 2-h = −0.31 accuracy
processing For PCAWD (Kurtosis): – The PCA allows resizing
techniques of PCA – 1-h = 6.06 a dataset, considering
and WD – 3-h = 3.22 −6- uncorrelated and
h = 3.12 redundant information
– 12-h = 2.99
– 18-h = 2.80
– 2-h = 11.00

Cervone – ANN 72-h ahead – RMSE/NP The combination of ANN N/A – GHI PV power Station in Italy – ANN – AnEn methodology
et al. – AnEn – CORR and AnEn yielded the – Percent cloud cover Lombardy: (Lombardy, – AnEn presents a computational
[102], – ANN+AnEn – BIAS/NP best results (CC) – July Calabria, – ANN advantage because its
(2017) – Air temperature at 2 2010–Dece- and Sicily) +AnEn efficiency can be
meters above the mber 2011 enhanced through
surface (T2M) Station in parallelization
– Solar azimuth (AZ) Calabria: – The computational

491
– Elevation (EL) – April efficiency shown is suited
– Lead time (t) 2011–March for real-time applications
2013 of distributed PV power
Station in production when forecast
Sicily: must be quickly run for
– January thousands of stations
2010–Dece- – The AnEn methodology
mber 2011 is shown to scale
extremely well for
massively parallel
applications

Ogliari et al. Physical 1-day ahead NMAE 3-parameters: N/A – Ambient PV power 216 days Milano, –3 – Increasing the training
[103], -deterministic – NMAE = 8.5% temperature Italy parameters set size with any
(2017) methods based on 5-parameters: – GHI –5 composition approach,
the NOCT thermal – NMAE = 9.0% – Gpoa parameters forecasting errors are
model of the PV PHANN: – Wind speed – PHANN generally lowered
module and hybrid – NMAE = 6.0% – Wind direction – The capability of the
stochastic- – Pressure hybrid method to learn
deterministic – Precipitation both the inaccuracies of
models combining – Cloud cover the provided weather
the CSRM and ANN – Cloud type forecasts and the
peculiarities of the PV
plant

Oneto et al. Data-driven – 1-h ahead – MAE The actual nowcasting N/A – Actual PvPP PV power May Savona – KM – DDM performances are – DDM models
[104], approaches: – 1–7-day – MSE – MAE = 3.0 – Temperature of the 2014–Septe- Campus – ELM improved as having more slowly degrade its
(2017) – Kernel Methods ahead – NMSE Nowcasting DDM PvPP mber 2015 Smart – RF data performance if the
(KM) – REP – MAE = 0.6 (5 times – External Polygenerat- time horizon is
– Extreme Learning – PPMCC better) temperature ion
(continued on next page)
Energy Conversion and Management 156 (2018) 459–497
Table 2 (continued)

Author Forecast method Forecast Performanc- Forecast error Computatio- Input variables Forecast Data period Location Comparison Description
S. Sobri et al.

(year) horizon e metrics measurement nal time variable methods


Advantages Disadvantages

Machines (ELM) The actual forecasting – Relative humidity Microgrid longer during the
– Random Forests – MAE = 8.4 – Wind speed (SPM) PvPP prediction
(RF) Forecasting DDM – Wind direction
– MAE = 3.0 – Sea-level
atmospheric pressure
– Solar radiation
predicted PvPP by the
actual system

Chiang et al. Wavelet-coupled – 1-h ahead – RMSE For Wavelet-BCRF: N/A Dataset: Solar PV 95 days University – Multilayer – Wavelet-BCRF
[105], feature extraction – 3-h ahead – MAPE – RMSE = 1.12 – PV system output power of perceptron successfully solves the
(2017) and bias- – 6-h ahead – MAE – MAPE = 5.91% from the California, (MLP) mismatch between the
compensation – MAE = 0.50 Supercomputer on the San Diego – Wavelet- forecasting horizons and
random forest rooftop (UCSD) coupled sampling frequencies of
(BCRF) – Sensor data from a SVM input data, and allows the
nearby weather – RF sampling frequency and
station (records solar without forecast horizons to be
irradiance, UV index, wavelet- determined
temperature, and transform independently
humidity from indoor/
outdoor areas)

Wang et al. A deterministic and – 15-min For For deterministic N/A Historical PV power PV power January Belgium – For the – The proposed model Disadvantages of
[106], probabilistic ahead determi- performance evaluation: data 2015–Dece- overall exhibits high forecasting DCNN:
(2017) forecasting method – 30-min nistic – WT+DCNN model mber 2015 perfor- robustness irrespective of – Cannot be directly

492
– hybrid intelligent ahead method: exhibits a more stable mances of input training dataset, applied to fulfil the
approach based on – 1-h ahead – MAPE forecasting stability the seasons, and forecasting PV power
wavelet (WT), deep – 2-h ahead – RMSE – The performance of the proposed horizons forecasting tasks
convolutional neural – MAE RMSE has been averagely determi- – Advantages of DCNN: because the PV
network (DCNN), For improved by 52.25%, nistic 1. Exhibits a more concise power data is
and spine quantile probabilistic 48.79%, and 3.78%. forecasting form with fewer memory generally 1D data in
regression (QR) method: – The MAE index has model, four footprints and parameters time domain
– ACE been averagely improved cases over due to the use of weight
– IS by 56.64%, 50.69%, and different sharing technique
– CRPS 35.36%. forecasting 2. The pre-processing
For probabilistic horizons, process existed in SAE
performance evaluation: seasons, and and DBN is not required
– The proposed model’s PV data for DCNN training,
quantile sharpness was locations making DCNN preferable
averagely improved by – For for real-time application
49.10% in 15-min, performance
54.08% in 30-min, comparison,
61.42% in 1-h, and the BPNN,
59.94% in 2-h SVM, and
WT+SVM
are adopted
Energy Conversion and Management 156 (2018) 459–497
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

and DCNN exhibit the best forecasting performance for all data samples forecasting metrics review [17]. Generally, the accuracy error of solar
in Limburg as demonstrated by the statistical results. In terms of fore- prediction is referred to solar irradiance in terms of W/m2, while kW is
casting reliability, sharpness, and overall Continuous Ranked expressed for error of power output of the PV plant [109].
Probability Score (CRPS) skill, the developed probabilistic combination
with WT, DCNN, and QR model was shown to be very effective [106]. 3.1. Classical statistical indicators

2.4. Factors influencing solar photovoltaic power forecasting – Pearson’s correlation coefficient: Pearson’s correlation coefficient
measures the correlation between two parameters or two sets of data
The forecasting approaches highly depends on the geographical as given in the following:
locations, forecasting horizons, and selection of input parameters to
cov (ρ,ρ )̂
meet the requirements of the decision-making process. (ρ) =
σρ σ ρ ̂ (21)
2.4.1. Forecast horizons where ρ and ρ ̂ are indicates the actual and predicted solar output, re-
The forecast horizon is known as span of time into the future. The spectively. If the value of Pearson’s correlation coefficient is larger, it
forecasting can be classified into four types based on the forecasting indicates improvement in solar prediction skill.
period as given in the following [107]:
– Normalized Error (nE):
i Long-term prediction (1 year to 10 years ahead).
Ppred−Pmeas
ii Mid-term prediction (1 month to 1 year ahead). nE =
max(Ppred ) (22)
iii Short-term prediction (1 h or several hours ahead to 1 day or 1 week
ahead). where Pmeas and Ppred are known as measured and forecasted power
iv Very short-term prediction (1 min to several minutes ahead). outputs, respectively.

Fig. 9 illustrates the categorization of prediction models based on – Mean Absolute Error (MAE): MAE refers to the average distance
spatial and temporal resolutions. For the day-ahead and intra-days between the measured values and forecasting model. This analysis is
forecast horizons, ECMWF and GFS models represent the reliable out- appropriate to evaluate uniform prediction errors.
comes that are limited by their spatial resolution whereas WRF model N
was suggested because it permits for small temporal resolutions. Sta- 1
MAE =
N
∑ |Ppred−Pmeas |
tistical time-series models can be performed for intra-hour and intra- i=1 (23)
day time-scales. A big number of GHI measurement data on the ground
offers a big set of temporal series of irradiance which allows the de-
– Mean Bias Error (MBE): MBE metric is appropriate to assess the
velopment of statistical forecasting model [22,108].
forecast bias.
N
2.4.2. Parameters 1
Solar radiation is the main parameter in solar energy research, but it MBE =
N
∑ (Ppred−Pmeas )
i=1 (24)
is unavailable for most sites due to the non-availability of solar radia-
tion measuring devices at meteorological stations. It is significant to
predict the solar radiation for a location by using several climatic – Standard Deviation Error (SDE):
parameters i.e., sunshine duration, maximum ambient temperature, N
1
relative humidity, latitude, longitude, day of year, daily clear sky global SDE =
N
∑ (Ppred−Pmeas−MBE )2
radiation, total cloud cover, temperature, clearness index, altitude, i=1 (25)
months, average temperature, average cloudiness, average wind velo-
city, atmospheric pressure, reference clearness index, mean diffuse ra- – Root Mean Square Error (RMSE): RMSE measures global error
diation, mean beam radiation, month, extra-terrestrial radiation, eva- during the entire prediction period.
poration, and soil temperature. Sunshine duration is commonly used for
N
modelling of solar radiation due to its availability and measurement at 1
RMSE =
N
∑ (Ppred−Pmeas )2
most sites [18]. i=1 (26)

2.5. Table of summarization techniques


– Mean Absolute Percentage Error (MAPE): MAPE is used to assess
uniform prediction errors like MAE. Po is the capacity of analyzed PV
Three major techniques of solar PV power forecasting i.e., statistical
plants.
time-series, physical, and ensemble methods have been elaborated with
their main characteristics. Each method was compared together in N
100 Ppred−Pmeas
terms of forecast method, time horizon, metrics assessment, input and MAPE =
N
∑ Po
i=1 (27)
output parameters, computational time, benchmark model, advantages
and disadvantages. A summary of solar PV power generation fore-
casting techniques is described in Table 2. – Median Absolute Percentage Error (MdAPE):
Ppred−Pmeas ⎞
3. Metrics assessment of solar photovoltaic power forecasting MdAPE = median ⎜⎛ 100 ⎟

techniques ⎝ Pmeas ⎠ (28)

The accuracy of solar forecasting highly depends on two major – Coefficient of Determination (R2): It represents the correlations be-
factors i.e., geographic locations and forecast horizons. The perfor- tween real and forecasted values.
mance analysis of the developed methods can be measured by metrics
assessment. Some researchers have concentrated on the various metrics
which were used as the base in the complete evaluation, especially in

493
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

Var (Pmeas−Ppred ) good prediction [111].


R2 = 1−
Var (Ppred ) (29) MAE
MASE =
( )1
N−1
N
∑i = 2 |Pmeas,i−Pmeas,i − 1 | (35)
– Correlation Coefficient (ρ)
(Cov (Pmeas−Ppred ))2
ρ= – Kolmogorov–Smirnov Integral (KSI): KSI is a non-parametric test
Var (Ppred ) (30) that can be used to determine if two datasets are significantly dif-
ferent. Smaller KSI values represent both forecast and real values
– Skill Score (ss) or Forecasting Skill: were performed similarly which shows the better prediction. Zero
KSI value specifies that two datasets of CDFs are the same [112].
U
ss = 1−⎛ ⎞ Xmax
⎝V ⎠ (31) KSI = ∫Xmin Dn dx (36)
where
where Dn indicates the difference between two CDFs.
N 2
Uncertainty, U =
1
∑ ⎧ Ppred,t −Pmeas,t ⎫
N ⎨ Pcs,t ⎬ – OVER metric evaluates the statistical similarity on large prediction
t=1 ⎩ ⎭
errors between predicted and actual power curve.
N
1 Pmax
Solar variability, V =
N
∑ (Δk (t ))2 OVER = ∫Pmin tdp (37)
t=1

Δk is the power step-changes depending on irradiance types either clear where pmax and pmin are power generation maximum and minimum
sky irradiance (Ics ) or solar constant (Io). The ss measures the quality of values, respectively. t is defined as:
the performances by utilizing the persistence model as the benchmark. Dj−Vc if Dj > Vc
Better forecasting model is obtained when ss is closer to 1. t=⎧
⎨ 0 if Dj ⩽ Vc

3.2. Recently applied indicators where Vc and Dj represents the critical value and the difference between
two CDFs, respectively. Both KSI and OVER can be normalized by the
Classical statistical indicators are applied by most researchers. This term ac (ac = Vc (pmax −pmin )) to acquire KSIPer and OVERPer, respec-
classical indicator has failed in fulfilling the grid operator’s require- tively.
ments. For example, big prediction errors can affect the stability and
economics in real operations. Therefore, it is significant to perform a – KSD parameter is a combination of both KSI and OVER to allow
complete evaluation using appropriate metrics to assess power gen- continuity classification of the outcomes:
eration forecasting by considering various forecast time-scales and
geographical locations. These metrics assessments can be classified into KSD = w1 KSI + w2 OVER
four types i.e., statistical, ramp characterization, uncertainty quantifi-
cation, and economic metrics [17,110]. – w1 and w2 are weight parameters.

3.2.1. Statistical metrics Due to the combination of developed KSD and RMSE parameters, a
The most basic metrics applied to assess predictions are MAE and new metric known as RIO is generated. RIO metrics supplies informa-
RMSE. However, they are not able to differentiate between two sets of tion from CDFs and the distance among pairs.
values with equal mean and variance but with different skewness and
KSD + RMSE
kurtosis. Therefore, some distribution features which affects the system RIO =
2 (38)
operation have been overlooked. These two metrics cannot perform on
their own, they must be applied with a combination of other metrics
[110]. 3.2.2. Uncertainty quantification and propagation metrics
Rényi entropy of uncertainty quantification metrics was suggested
– Skew: by Zhang et al. [113]. As mentioned earlier, MAE and RMSE of classical
N 3 statistical metrics are only unbiased if they are based on a Gaussian
skew =
N
∑ ⎛ nE −nE ⎞
⎜ ⎟
distribution.
(N −1)(N −2) i=2 ⎝ SD ⎠ (32)
– Rényi entropy:
– Kurtosis: N
1
H∝ (X ) = log ∑ pi∝
N (N −1)
N 4
3(N −1)2 1−∝ 2
kurt =

∑ ⎛ nE −nE ⎞ ⎫ .
⎜ ⎟
i=1 (39)
⎨ (N −1)(N −2)(N −3) ⎝ SD ⎠ ⎬ (N −2)(N −3) (33)
⎩ i=1 ⎭ where a (a > 0 and a ≠ 1) is the Rényi entropy order and pi is the
probability density of ith discrete portion of the distribution. Basically,
– Maximum absolute error (MaxAE): MaxAE represents the biggest bigger values of Rényi entropy indicate larger uncertainties in the
prediction error. A larger value of MaxAE metric can cause a big predictions.
economic impact on grid operation.
3.2.3. Ramp characterization
MaxAE = max i = 1,2,… ,n |Ppred−Pmeas | (34)
In determining the ramp characterization metric, the swinging door
algorithm has been suggested [114]. This is because it is an easy and
– Mean Absolute Scaled Error (MASE): MASE claims to be scale free flexible technique where the threshold variable (∊) indicates the width
and has little awareness to outliers. Small values of MASEs represent of the ramp. The small values of ∊ represent many fluctuations whereas
the larger values represent the biggest changes.

494
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

3.2.4. Economic metrics lower bounds of interval for i−th example in the dataset. The Interval
Operating reserves have correlated with costs and the high pene- Coverage Probability (ICP) is also used by these researchers. This metric
tration of solar energy. Precise forecasting can minimize the operating determines the probability that the k values of time series Pt + 1…,Pt + k for
p
reserves amount and the operating system cost. When the forecasts the next k-length interval fall between the upper bound Umeas ,k ,i and the
p
ramp is modified, metrics of skewness, kurtosis, and Rényi entropy also lower bound Lmeas ,k ,i of the forecasted interval with the average of the
change as they are sensitive to any modification. Based on nonpara- entire observations dataset. The higher value of ICP indicates lower
metric statistical and sensitivity analysis, smaller set of metrics are prediction error.
chosen to evaluate the forecast accuracy. Therefore, the usage of MBE, N i+k
standard deviation, skewness, kurtosis, distribution of forecast errors, 1
ICP =
N. k
∑ ∑ Cj. 100%
Rényi entropy, RMSE, and OVERPer metrics are highly recommended. i=1 j=i+1 (44)

where N is sample number, k is interval length, and Cj :


3.2.5. Other metrics
p
Several researchers have proposed other metrics for error assess- 1 if Pj ∈ |Umeas,k,i,Lmeas
Cj = ⎧ ,k ,i
ment. Russo et al., introduced the error estimation as presented in Eq. ⎨
⎩ 0 otherwise
(40) [115]:
Almeida et al., used the modified MAE and MBE versions, desig-
Ni (Pmeas,i−Ppred,i )2 nated as MAE and MBE variation coefficient to evaluate market models
E = 100 ∑i −1
σi2 Ni (40) that penalize the hourly or daily energy error, respectively.

This metric has been proven to be more accurate compared to MAE


cvMAE =
RMSE, MAE or another similar type of metrics as it is proportional Pmeas (45)
correlates to the learning error with actual variability of the output. MBE
Alessandrini et al., introduced different metrics set from the basic cvMBE =
Pmeas (46)
ones. The statistical consistency was analysed by using the missing rate
error. It represents the fraction of observations at a lower/higher level, where Pmeas defines the mean of measured power. They evaluated the
where the lowest/highest forecasts are ranked above or below the performance of confidence interval output by using their model based
predicted lost rate of 1/(n + 1) (n is the number of ensemble members). on two aspects i.e., amplitude and accuracy. The amplitude of interval
Another metric that has been implemented is the Continuous Ranked was determined as its area could be normalized by the area of ob-
Probability Score (CRPS) that measures cumulative distribution func- servations. It presents an observation of energy forecasted in conjunc-
tions [116]. tion to the measured energy.
N Hong et al., discussed several metrics to assess the various models
1 ∞
CRPS =
N
∑ ∫−∞ (Fipred (x )−Fimeas (x ))2dx involved in the competition. After examining the suitability of MAE,
i=1 (41) KSI, CRPS, and pinball loss function, the second option was selected due
pred to its adequacy in probabilistic forecasts, simple execution, and com-
where Fi (x ) is the probabilistic prediction for CDF, is the Fimeas (x )
munication. The pinball loss function (L) was computed with all per-
observation of CDF for ith ensemble forecasts pair, and N is the number
centiles from 1st to 99th [120].
of available pairs. Lower CRPS value shows better forecasts. Zamo
et al., also integrated the CRPS and differentiated the reliability and the
potential term of CRPS. The statistical consistency is measured by the L (qa,y ) =
( a
⎧ 1− 100 )
(qa−y ) if y < qa

reliability term among the forecasted and the observed distribution. ⎨ a (y−q ) if y ≫ qa
⎩ 100 a (47)
The potential term is a perfect reliable CRPS model [117].
a
Alessandrini et al., developed the Brier Score (BS) for probabilistic where qa is quintile forecast, is target quintile, y is the observation
100
analysis that can be compared to the deterministic forecast of RMSE. used for prediction evaluation, and a = 1,2…99. Then, the score is
This approach measures the difference between the distribution of averaged over the target quintiles, time periods, and forecast horizons.
forecasted probability and its occurrences. The precise forecasting is obtained with a lower score as for the CRPS.
N
1 4. Conclusion
BS =
N
∑ (pn −on )2
i=1 (42)
This paper investigates solar PV power generation forecasting
where ρ is the predicted probability of categorical occurrence, on is the
techniques presented to date and describes the characteristics of various
categorical observation, and N refers to the total number of (ρn ,on )
forecasting techniques. These approaches are compared together in
pairs. Smaller value of BS represents better performance of the pre-
terms of forecast method, time horizon, measurement error, input and
diction model. Furthermore, the Brier Skill Score (BSS) represents the
output variables, computational time, and benchmark model. The
enhancement of a base model.
merits and demerits of different types of forecasting techniques are
Sperati et al., continuously applied this metric and presented the
discussed as well. This work classifies solar PV forecasting methods into
Relative Operating Characteristic Skill Score (ROCSS) based on ROC
three major categories i.e., time-series statistical, physical, and en-
curve that plots the false alarm rate. In ROCSS value approximation, 1
semble methods. Artificial Neural Network (ANN) and Space Vector
refers to precise forecasting [118].
Machine (SVM) are widely used due to their ability in solving complex
Rana et al., used Mean Absolute Interval Deviation (MAID) metrics
and non-linear forecasting models. The metrics assessment shows that
to assess their models that measured the deviation of forecasted interval
Artificial Intelligence (AI) models could decrease the error compared to
from the actual interval. A smaller value of MAID represents lower error
other statistical approaches. The ensemble method has been introduced
prediction [119].
recently for its ability to merge linear and non-linear techniques which
N enhances the accuracy and performance of models in comparison with
1 p p p p
MAID =
2N
∑ |Umeas ,k ,i−Upred,k ,i | + |Lmeas,k ,i −Lpred,k ,i | individual models. The metrics assessment that used for evaluating the
i=1 (43)
solar prediction accuracy is presented as well for specific applications
p p
where Umeas ,k ,iand Lmeas ,k ,i are the upper and lower bounds of the actual and hence the appropriate solar forecasting approaches can be selected
p p
k-length interval. Upred,k,i and Lpred ,k ,i are known as predicted upper and to ensure better performance. The comparative review and the most

495
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

recent information provided in this work is useful for researchers, output correction. IEEE Trans Sustain Energy 2013;4:527–33.
planners, and designers involved in solar power systems studies. [30] Almonacid F, Pérez-Higueras P, Fernández EF, Hontoria L. A methodology based
on dynamic artificial neural network for short-term forecasting of the power
output of a PV generator. Energy Convers Manage 2014;85:389–98.
Acknowledgement [31] Dahmani K, Dizene R, Notton G, Paoli C, Voyant C, Nivet ML. Estimation of 5-min
time-step data of tilted solar global irradiation using ANN (artificial neural net-
work) model. Energy 2014;70:374–81.
This work was supported by the Centre for Research Grant [32] Kaushika N, Tomar R, Kaushik S. Artificial neural network model based on in-
Management (PPGP) at University of Malaya [UMPEDAC-2016, MOHE terrelationship of direct, diffuse and global solar radiations. Sol Energy
HICOE – UMPEDAC]. 2014;103:327–42.
[33] Teo T, Logenthiran T, Woo W. Forecasting of photovoltaic power using extreme
learning machine. In: Smart grid technologies-Asia (ISGT ASIA), 2015 IEEE in-
References novative. IEEE; 2015. p. 1–6.
[34] Kashyap Y, Bansal A, Sao AK. Solar radiation forecasting with multiple parameters
neural networks. Renew Sustain Energy Rev 2015;49:825–35.
[1] Wan C, Zhao J, Song Y, Xu Z, Lin J, Hu Z. Photovoltaic and solar power forecasting
[35] Zhang Y, Beaudin M, Taheri R, Zareipour H, Wood D. Day-ahead power output
for smart grid energy management. CSEE J Power Energy Syst 2015;1:38–46.
forecasting for small-scale solar photovoltaic electricity generators. IEEE Trans
[2] Ismail AM, Ramirez-Iniguez R, Asif M, Munir AB, Muhammad-Sukki F. Progress of
Smart Grid 2015;6:2253–62.
solar photovoltaic in ASEAN countries: a review. Renew Sustain Energy Rev
[36] Lima FJ, Martins FR, Pereira EB, Lorenz E, Heinemann D. Forecast for surface solar
2015;48:399–412.
irradiance at the Brazilian Northeastern region using NWP model and artificial
[3] Fabiani Appavou AB, Barbel Epp (solrico) et al. In: Mastny L, editor. Renewables
neural networks. Renew Energy 2016;87:807–18.
2016 global status report.
[37] Sharma V, Yang D, Walsh W, Reindl T. Short term solar irradiance forecasting
[4] Trends in photovoltaic (PV) applications; 2016.
using a mixed wavelet neural network. Renew Energy 2016;90:481–92.
[5] Mekhilef S, Safari A, Mustaffa W, Saidur R, Omar R, Younis M. Solar energy in
[38] Graditi G, Ferlito S, Adinolfi G. Comparison of photovoltaic plant power produc-
Malaysia: current state and prospects. Renew Sustain Energy Rev 2012;16:386–96.
tion prediction methods using a large measured dataset. Renew Energy
[6] Bujang A, Bern C, Brumm T. Summary of energy demand and renewable energy
2016;90:513–9.
policies in Malaysia. Renew Sustain Energy Rev 2016;53:1459–67.
[39] Hossain M, Mekhilef S, Danesh M, Olatomiwa L, Shamshirband S. Application of
[7] Wong S, Ngadi N, Abdullah TAT, Inuwa I. Recent advances of feed-in tariff in
extreme learning machine for short term output power forecasting of three grid-
Malaysia. Renew Sustain Energy Rev 2015;41:42–52.
connected PV systems. J Clean Prod 2017;167:395–405.
[8] Graditi G, Ferlito S, Adinolfi G, Tina GM, Ventura C. Energy yield estimation of
[40] Mohammadi K, Shamshirband S, Tong CW, Arif M, Petković D, Ch S. A new hybrid
thin-film photovoltaic plants by using physical approach and artificial neural
support vector machine–wavelet transform approach for estimation of horizontal
networks. Sol Energy 2016;130:232–43.
global solar radiation. Energy Convers Manage 2015;92:162–71.
[9] Koohi-Kamali S, Rahim N, Mokhlis H, Tyagi V. Photovoltaic electricity generator
[41] Olatomiwa L, Mekhilef S, Shamshirband S, Mohammadi K, Petković D, Sudheer C.
dynamic modeling methods for smart grid applications: a review. Renew Sustain
A support vector machine–firefly algorithm-based model for global solar radiation
Energy Rev 2016;57:131–72.
prediction. Sol Energy 2015;115:632–44.
[10] Koohi-Kamali S, Rahim N, Mokhlis H. Smart power management algorithm in
[42] Shamshirband S, Petković D, Saboohi H, Anuar NB, Inayat I, Akib S, et al. Wind
microgrid consisting of photovoltaic, diesel, and battery storage plants considering
turbine power coefficient estimation by soft computing methodologies: compara-
variations in sunlight, temperature, and load. Energy Convers Manage
tive study. Energy Convers Manage 2014;81:520–6.
2014;84:562–82.
[43] Mellit A, Pavan AM, Benghanem M. Least squares support vector machine for
[11] Ferlito S, Adinolfi G, Graditi G. Comparative analysis of data-driven methods
short-term prediction of meteorological time series. Theoret Appl Climatol
online and offline trained to the forecasting of grid-connected photovoltaic plant
2013;111:297–307.
production. Appl Energy 2017;205:116–29.
[44] Shi J, Lee W-J, Liu Y, Yang Y, Wang P. Forecasting power output of photovoltaic
[12] De Giorgi MG, Congedo PM, Malvoni M. Photovoltaic power forecasting using
systems based on weather classification and support vector machines. IEEE Trans
statistical methods: impact of weather data. IET Sci Meas Technol 2014;8:90–7.
Ind Appl 2012;48:1064–9.
[13] Barbieri F, Rajakaruna S, Ghosh A. Very short-term photovoltaic power forecasting
[45] Zeng J, Qiao W. Short-term solar power prediction using a support vector machine.
with cloud modeling: a review. Renew Sustain Energy Rev 2016.
Renew Energy 2013;52:118–27.
[14] Leva S, Dolara A, Grimaccia F, Mussetta M, Ogliari E. Analysis and validation of 24
[46] Chen J-L, Li G-S, Wu S-J. Assessing the potential of support vector machine for
hours ahead neural network forecasting of photovoltaic output power. Math
estimating daily solar radiation using sunshine duration. Energy Convers Manage
Comput Simul 2017;131:88–100.
2013;75:311–8.
[15] Montgomery DC, Jennings CL, Kulahci M. Introduction to time series analysis and
[47] Ekici BB. A least squares support vector machine model for prediction of the next
forecasting. John Wiley & Sons; 2015.
day solar insolation for effective use of PV systems. Measurement 2014;50:255–62.
[16] Cornaro C, Pierro M, Bucci F. Master optimization process based on neural net-
[48] Ramli MA, Twaha S, Al-Turki YA. Investigating the performance of support vector
works ensemble for 24-h solar irradiance forecast. Sol Energy 2015;111:297–312.
machine and artificial neural networks in predicting solar radiation on a tilted
[17] Antonanzas J, Osorio N, Escobar R, Urraca R, Martinez-de-Pison F, Antonanzas-
surface: Saudi Arabia case study. Energy Convers Manage 2015;105:442–52.
Torres F. Review of photovoltaic power forecasting. Sol Energy 2016;136:78–111.
[49] Wolff B, Kühnert J, Lorenz E, Kramer O, Heinemann D. Comparing support vector
[18] Yadav AK, Chandel S. Solar radiation prediction using Artificial Neural Network
regression for PV power forecasting to a physical modeling approach using mea-
techniques: a review. Renew Sustain Energy Rev 2014;33:772–81.
surement, numerical weather prediction, and cloud motion data. Sol Energy
[19] Ogliari E, Grimaccia F, Leva S, Mussetta M. Hybrid predictive models for accurate
2016;135:197–208.
forecasting in PV systems. Energies 2013;6:1918–29.
[50] Hocaoglu FO, Serttas F. A novel hybrid (Mycielski-Markov) model for hourly solar
[20] Zhu H, Li X, Sun Q, Nie L, Yao J, Zhao G. A power prediction method for photo-
radiation forecasting. Renew Energy 2016.
voltaic power plant based on wavelet decomposition and artificial neural net-
[51] Bhardwaj S, Sharma V, Srivastava S, Sastry O, Bandyopadhyay B, Chandel S, et al.
works. Energies 2015;9:11.
Estimation of solar radiation using a combination of Hidden Markov Model and
[21] Amrouche B, Le Pivert X. Artificial neural network based daily local forecasting for
generalized Fuzzy model. Sol Energy 2013;93:43–54.
global solar radiation. Appl Energy 2014;130:333–41.
[52] Sanjari MJ, Gooi HB. Probabilistic forecast of PV power generation based on
[22] Diagne M, David M, Lauret P, Boland J, Schmutz N. Review of solar irradiance
higher-order Markov chain. IEEE Trans Power Syst 2016.
forecasting methods and a proposition for small-scale insular grids. Renew Sustain
[53] Huang J, Korolkiewicz M, Agrawal M, Boland J. Forecasting solar radiation on an
Energy Rev 2013;27:65–76.
hourly time scale using a Coupled AutoRegressive and Dynamical System (CARDS)
[23] Mellit A, Pavan AM. A 24-h forecast of solar irradiance using artificial neural
model. Sol Energy 2013;87:136–49.
network: application for performance prediction of a grid-connected PV plant at
[54] Box GE, Jenkins GM, Reinsel GC, Ljung GM. Time series analysis: forecasting and
Trieste, Italy. Sol Energy 2010;84:807–21.
control. John Wiley & Sons; 2015.
[24] Izgi E, Öztopal A, Yerli B, Kaymak MK, Şahin AD. Short–mid-term solar power
[55] Ssekulima EB, Anwar MB, Al Hinai A, El Moursi MS. Wind speed and solar irra-
prediction by using artificial neural networks. Sol Energy 2012;86:725–33.
diance forecasting techniques for enhanced renewable energy integration with the
[25] Kardakos E, Alexiadis M, Vagropoulos S, Simoglou C, Biskas P, Bakirtzis A.
grid: a review. IET Renew Power Gener; 2016.
Application of time series and artificial neural network models in short-term
[56] Phinikarides A, Makrides G, Kindyni N, Kyprianou A, Georghiou GE. ARIMA
forecasting of PV power generation. In: Power engineering conference (UPEC),
modeling of the performance of different photovoltaic technologies. In: 2013 IEEE
2013 48th international universities. IEEE; 2013. p. 1–6.
39th photovoltaic specialists conference (PVSC). IEEE; 2013. p. 0797-801.
[26] Chen S, Gooi H, Wang M. Solar radiation forecast based on fuzzy logic and neural
[57] Yang D, Jirutitijaroen P, Walsh WM. Hourly solar irradiance time series forecasting
networks. Renew Energy 2013;60:195–201.
using cloud cover index. Sol Energy 2012;86:3531–43.
[27] Mellit A, Sağlam S, Kalogirou S. Artificial neural network-based model for esti-
[58] Yang C, Thatte AA, Xie L. Multitime-scale data-driven spatio-temporal forecast of
mating the produced power of a photovoltaic module. Renew Energy
photovoltaic generation. IEEE Trans Sustain Energy 2015;6:104–12.
2013;60:71–8.
[59] Agoua XG, Girard R, Kariniotakis G. Short-term spatio-temporal forecasting of
[28] Notton G, Paoli C, Ivanova L, Vasileva S, Nivet ML. Neural network approach to
photovoltaic power production. IEEE Trans Sustain Energy 2017.
estimate 10-min solar global irradiation values on tilted planes. Renew Energy
[60] Fumo N, Biswas MR. Regression analysis for prediction of residential energy
2013;50:576–84.
consumption. Renew Sustain Energy Rev 2015;47:332–43.
[29] Yona A, Senjyu T, Funabashi T, Kim C-H. Determination method of insolation
[61] Trapero JR, Kourentzes N, Martin A. Short-term solar irradiation forecasting based
prediction with fuzzy and applying neural network for long-term ahead PV power
on dynamic harmonic regression. Energy 2015;84:289–95.

496
S. Sobri et al. Energy Conversion and Management 156 (2018) 459–497

[62] Wang G, Su Y, Shu L. One-day-ahead daily power forecasting of photovoltaic 18th international conference on. IEEE; 2015. p. 1–6.
systems based on partial functional linear regression models. Renew Energy [92] Dong Z, Yang D, Reindl T, Walsh WM. A novel hybrid approach based on self-
2016;96:469–78. organizing maps, support vector regression and particle swarm optimization to
[63] Li Y, He Y, Su Y, Shu L. Forecasting the daily power output of a grid-connected forecast solar irradiance. Energy 2015;82:570–7.
photovoltaic system based on multivariate adaptive regression splines. Appl [93] Wang J, Jiang H, Wu Y, Dong Y. Forecasting solar radiation using an optimized
Energy 2016;180:392–401. hybrid model by Cuckoo Search algorithm. Energy 2015;81:627–44.
[64] Massidda L, Marrocu M. Use of Multilinear Adaptive Regression Splines and nu- [94] Rana M, Koprinska I, Agelidis VG. Forecasting solar power generated by grid
merical weather prediction to forecast the power output of a PV plant in Borkum, connected PV systems using ensembles of neural networks. In: 2015 International
Germany. Sol Energy 2017;146:141–9. joint conference on neural networks (IJCNN). IEEE; 2015. p. 1–8.
[65] Persson C, Bacher P, Shiga T, Madsen H. Multi-site solar power forecasting using [95] Dolara A, Grimaccia F, Leva S, Mussetta M, Ogliari E. A physical hybrid artificial
gradient boosted regression trees. Sol Energy 2017;150:423–36. neural network for short term forecasting of PV plant power output. Energies
[66] Inman RH, Pedro HT, Coimbra CF. Solar forecasting methods for renewable energy 2015;8:1138–53.
integration. Prog Energy Combust Sci 2013;39:535–76. [96] Di Piazza A, Di Piazza MC, Vitale G. Solar and wind forecasting by NARX neural
[67] Monteiro C, Santos T, Fernandez-Jimenez LA, Ramirez-Rosado IJ, Terreros-Olarte networks. Renew Energy Environ Sustain 2016;1:39.
MS. Short-term power forecasting model for photovoltaic plants based on histor- [97] Azimi R, Ghayekhloo M, Ghofrani M. A hybrid method based on a new clustering
ical similarity. Energies 2013;6:2624–43. technique and multilayer perceptron neural networks for hourly solar radiation
[68] Lorenz E, Heinemann D. Prediction of solar irradiance and photovoltaic power. forecasting. Energy Convers Manage 2016;118:331–44.
Comprehen Renew Energy 2012;1:239–92. [98] Li Z, Rahman S, Vega R, Dong B. A hierarchical approach using machine learning
[69] Mathiesen P, Kleissl J. Evaluation of numerical weather prediction for intra-day methods in solar photovoltaic energy production forecasting. Energies 2016;9:55.
solar forecasting in the continental United States. Sol Energy 2011;85:967–77. [99] Antonanzas J, Urraca R, Aldama A, Fernández-Jiménez LA, Martínez-de-Pisón FJ.
[70] Fernandez-Jimenez LA, Muñoz-Jimenez A, Falces A, Mendoza-Villena M, Garcia- Single and blended models for day-ahead photovoltaic power forecasting. In:
Garrido E, Lara-Santillan PM, et al. Short-term power forecasting system for International conference on hybrid artificial intelligence systems. Springer; 2017.
photovoltaic plants. Renew Energy 2012;44:311–7. p. 427–34.
[71] Tuohy A, Zack J, Haupt SE, Sharp J, Ahlstrom M, Dise S, et al. Solar forecasting: [100] Do M-T, Soubdhan T, Robyns B. A study on the minimum duration of training data
methods, challenges, and performance. IEEE Power Energ Mag 2015;13:50–9. to provide a high accuracy forecast for PV generation between two different cli-
[72] Urquhart B, Kurtz B, Dahlin E, Ghonima M, Shields J, Kleissl J. Development of a matic zones. Renew Energy 2016;85:959–64.
sky imaging system for short-term solar power forecasting. Atmosph Meas Techn [101] Malvoni M, De Giorgi MG, Congedo PM. Forecasting of PV Power Generation using
Discuss 2014;7:4859–907. weather input data-preprocessing techniques. Energy Procedia 2017;126:651–8.
[73] Yang H, Kurtz B, Nguyen D, Urquhart B, Chow CW, Ghonima M, et al. Solar ir- [102] Cervone G, Clemente-Harding L, Alessandrini S, Delle Monache L. Short-term
radiance forecasting using a ground-based sky imager developed at UC San Diego. photovoltaic power forecasting using Artificial Neural Networks and an Analog
Sol Energy 2014;103:502–24. Ensemble. Renew Energy 2017;108:274–86.
[74] Pelland S, Remund J, Kleissl J, Oozeki T, De Brabandere K. Photovoltaic and solar [103] Ogliari E, Dolara A, Manzolini G, Leva S. Physical and hybrid methods comparison
forecasting: state of the art. IEA PVPS, Task 2013;14:1–36. for the day ahead PV output power forecast. Renew Energy 2017.
[75] Marquez R, Pedro HT, Coimbra CF. Hybrid solar forecasting method uses satellite [104] Oneto L, Laureri F, Robba M, Delfino F, Anguita D. Data-driven photovoltaic
imaging and ground telemetry as inputs to ANNs. Sol Energy 2013;92:176–88. power production nowcasting and forecasting for polygeneration microgrids. IEEE
[76] Chow CW, Urquhart B, Lave M, Dominguez A, Kleissl J, Shields J, et al. Intra-hour Syst J 2017.
forecasting with a total sky imager at the UC San Diego solar energy testbed. Sol [105] Chiang P-H, Chiluvuri SPV, Dey S, Nguyen TQ. Forecasting of solar photovoltaic
Energy 2011;85:2881–93. system power generation using wavelet decomposition and bias-compensated
[77] Gohari M, Urquhart B, Yang H, Kurtz B, Nguyen D, Chow C, et al. Comparison of random forest. In: Green technologies conference (GreenTech), 2017 ninth annual
solar power output forecasting performance of the total sky imager and the IEEE. IEEE; 2017. p. 260–6.
University of California, San Diego sky imager. Energy Procedia 2014;49:2340–50. [106] Wang H, Yi H, Peng J, Wang G, Liu Y, Jiang H, et al. Deterministic and prob-
[78] Peng Z, Yu D, Huang D, Heiser J, Yoo S, Kalb P. 3D cloud detection and tracking abilistic forecasting of photovoltaic power based on deep convolutional neural
system for solar forecast using multiple sky imagers. Sol Energy network. Energy Convers Manage 2017;153:409–22.
2015;118:496–519. [107] Raza MQ, Nadarajah M, Ekanayake C. On recent advances in PV output power
[79] Alonso-Montesinos J, Batlles F, Portillo C. Solar irradiance forecasting at one- forecast. Sol Energy 2016;136:125–44.
minute intervals for different sky conditions using sky camera images. Energy [108] Diagne HM, Lauret P, David M. Solar irradiation forecasting: state-of-the-art and
Convers Manage 2015;105:1166–77. proposition for future developments for small-scale insular grids. In: WREF 2012-
[80] Perez R, Kivalov S, Schlemmer J, Hemker K, Renné D, Hoff TE. Validation of short world renewable energy forum; 2012.
and medium term operational solar radiation forecasts in the US. Sol Energy [109] Widén J, Carpman N, Castellucci V, Lingfors D, Olauson J, Remouit F, et al.
2010;84:2161–72. Variability assessment and forecasting of renewables: a review for solar, wind,
[81] Dong Z, Yang D, Reindl T, Walsh WM. Satellite image analysis and a hybrid ESSS/ wave and tidal resources. Renew Sustain Energy Rev 2015;44:356–75.
ANN model to forecast solar irradiance in the tropics. Energy Convers Manage [110] Zhang J, Florita A, Hodge B-M, Lu S, Hamann HF, Banunarayanan V, et al. A suite
2014;79:66–73. of metrics for assessing the performance of solar power forecasting. Sol Energy
[82] Aguiar LM, Pereira B, Lauret P, Díaz F, David M. Combining solar irradiance 2015;111:157–75.
measurements, satellite-derived data and a numerical weather prediction model to [111] Hossain MR, Oo AM, Ali AS. Hybrid prediction method of solar power using dif-
improve intra-day solar forecasting. Renew Energy 2016;97:599–610. ferent computational intelligence algorithms. In: Universities power engineering
[83] Ren Y, Suganthan P, Srikanth N. Ensemble methods for wind and solar power conference (AUPEC), 2012 22nd Australasian. IEEE; 2012. p. 1–6.
forecasting—a state-of-the-art review. Renew Sustain Energy Rev 2015;50:82–91. [112] Espinar B, Ramírez L, Drews A, Beyer HG, Zarzalejo LF, Polo J, et al. Analysis of
[84] Quan DM, Ogliari E, Grimaccia F, Leva S, Mussetta M. Hybrid model for hourly different comparison parameters applied to solar radiation data from satellite and
forecast of photovoltaic and wind power. In: Fuzzy systems (FUZZ), 2013 IEEE German radiometric stations. Sol Energy 2009;83:118–25.
international conference on. IEEE; 2013. p. 1–6. [113] Zhang J, Zhang Y, Yu C-S. Rényi entropy uncertainty relation for successive pro-
[85] Zhang N, Behera PK, Williams C. Solar radiation prediction based on particle jective measurements. Quantum Inf Process 2015;14:2239–53.
swarm optimization and evolutionary algorithm using recurrent neural networks. [114] Florita A, Hodge B-M, Orwig K. Identifying wind and solar ramping events. In:
In: Systems conference (SysCon), 2013 IEEE international. IEEE; 2013. p. 280–6. Green technologies conference, 2013 IEEE. IEEE; 2013. p. 147–52.
[86] Bouzerdoum M, Mellit A, Pavan AM. A hybrid model (SARIMA–SVM) for short- [115] Russo M, Leotta G, Pugliatti P, Gigliucci G. Genetic programming for photovoltaic
term power forecasting of a small-scale grid-connected photovoltaic plant. Sol plant output forecasting. Sol Energy 2014;105:264–73.
Energy 2013;98:226–35. [116] Alessandrini S, Delle Monache L, Sperati S, Cervone G. An analog ensemble for
[87] Haque AU, Nehrir MH, Mandal P. Solar PV power generation forecast using a short-term probabilistic solar power forecast. Appl Energy 2015;157:95–110.
hybrid intelligent approach. In: 2013 IEEE power & energy society general [117] Zamo M, Mestre O, Arbogast P, Pannekoucke O. A benchmark of statistical re-
meeting. IEEE; 2013. p. 1–5. gression methods for short-term forecasting of photovoltaic electricity production.
[88] Gandelli A, Grimaccia F, Leva S, Mussetta M, Ogliari E. Hybrid model analysis and Part II: Probabilistic forecast of daily production. Sol Energy 2014;105:804–16.
validation for PV energy production forecasting. In: 2014 International joint [118] Sperati S, Alessandrini S, Delle Monache L. An application of the ECMWF
conference on neural networks (IJCNN). IEEE; 2014. p. 1957–62. Ensemble Prediction System for short-term solar power forecasting. Sol Energy
[89] Yang H-T, Huang C-M, Huang Y-C, Pai Y-S. A weather-based hybrid method for 1- 2016;133:437–50.
day ahead hourly forecasting of pv power output. IEEE Trans Sustain Energy [119] Rana M, Koprinska I, Agelidis VG. 2D-interval forecasts for solar power produc-
2014;5:917–26. tion. Sol Energy 2015;122:191–203.
[90] Wu Y-K, Chen C-R, Abdul Rahman H. A novel hybrid model for short-term fore- [120] Hong T, Pinson P, Fan S, Zareipour H, Troccoli A, Hyndman RJ. Probabilistic
casting in PV power generation. Int J Photoenergy 2014;2014. energy forecasting: global energy forecasting competition 2014 and beyond. Int J
[91] Filipe J, Bessa R, Sumaili J, Tomé R, Sousa J. A hybrid short-term solar power Forecast 2016;32:896–913.
forecasting tool. In: Intelligent system application to power systems (ISAP), 2015

497

Potrebbero piacerti anche