Sei sulla pagina 1di 6

Demand Forecasting in Residential Distribution

Feeders in the Context of Smart Grids


Bruno Pires Campos
Milton Rafael da Silva
Federal University of Itajuba
Itajub, Minas Gerais, Brazil
brunopcampos10@gmail.com; miltonrafa@yahoo.com.br

AbstractInvestments on smarter solutions for power systems safer, avoiding the need of standby capacity of power plants
are currently growing and many large Smart Grid projects are [3]. So, the peak demands are important points for Smart grid
underway throughout the world. With photovoltaic and wind regulation, and forecasting that moment is indispensable.
power generation, consumers will also be able to produce and According to [4] forecasting algorithms could save up to
sell energy. This distributed generation will require a very 0.5M$/MW per year in renewable energy sector and about
efficient management of the power system. In this context, smart 10% of this value come from gains at operational efficiency.
meters will be crucial tools to measure and monitor the systems
In [13], forecasting is defined as making calculations or
performance. Each house will have a smart meter acquiring data
uninterruptedly, creating a large amount of data, which will predictions of a future event or condition based on analysis or
have to be processed into useful information for strategic studies of historical data, events or observations. One point
decisions. This paper analyses the forecasting of residential load needs careful attention when talking about forecasting
demand, using Auto Regressive with Exogenous Inputs (ARX), algorithms, the selection of input data that will compose the
Artificial Neural Networks (ANN) and Artificial Neural model; it needs to have high correlation with the model output.
Networks optimized by Genetic Algorithm (ANN-GA). The Short-term forecasting (one hour to several hours ahead) is
procedure that achieved best predictions was ANN working with useful for planning economic load dispatch and operation
GA. security in electricity market [5;6;7]. Forecasting of one week
ahead (medium-term) is proper for unit commitment decision,
Keywords Artificial neural networks, Autoregressive process;
Demand forecasting; Genetic algorithms; Smart grids; reserve and generator decisions [5].
However, according to [8], it is very difficult to forecast the
I. INTRODUCTION electricity load accurately, because the electric load is affected
by various variables, such as social factors, economic
With the advent of Smart grids [17], nowadays, electricity situations, irregular behaviors and other non-linear factors.
consumers can manage their consumption for many reasons, like It is known that load and energy production forecasting has
economic, reliability, or environmental purposes [1]. Consumers always been decisive for the effectiveness of power system
can react to the fluctuations in electricity price what influences planning and operation [9]. Many electric power companies
the demand side management (DSM) [6]. According to [2], forecasts power demand based on traditional prediction
Smart grids are intelligent power grids that execute autonomous methods. However, the relationship between power load and
adaptation of their components in order to optimize resource factors influencing power production is nonlinear, so it is
consumption (e.g., minimize energy transport; reshape demand difficult to identify its nonlinearity by using traditional
through the usage of automated devices in smart homes).In this prediction methods [4]. Advanced forecasting through
scenario, the use of smart meters is crucial. They are responsible evolutionary computation techniques help with reliable
for measuring the condition of the local residential grid and feed production predictions and the planning for additional power
the management system, which will store and process this supply, to act proactively. This aspect impacts on the economic
information. The main goal is to transform this data into useful balance of the systems especially in an integrated smart grid
information for strategic decisions such as network expansion, solution perspective [4]. Conventional methods in short-term
replacement or reinforcement of assets. On the distributed load forecasting are: regression models, ARIMA, ARX,
generation, each consumer will be able to supply energy too, ARIMAX or SARIMAX models, Box and Jenkins transfer
and they will be able to choose more profitable time to save and functions, optimization techniques, non-parametric regression,
store (battery charging) or to sell energy, because the price structural models, curve adjustment procedures. There are
fluctuates with load demand. It also promises to be more recent methods whose results are compared with the ones
advantageous for power generators, because at peak demands, mentioned such as Expert systems, Artificial neural networks,
energy price will be high and consumers will save and sell their Fuzzy inference and Neuro-fuzzy models [9].
energy, then the peak will decrease and power system can work Related with the size of the power grid and its load forecasting,
in a conventional power system, variations of different loads
inside the system may neutralize or attenuate oscillations effect
of each other on the total demand, resulting in fewer variations
and especially less outliers in system load [14]. On the other
hand, smaller system size of microgrids causes weaker
smoothing effect of load aggregation, leading to higher sharper
variations in a microgrid load. It is well known that volatility on the operation of human brain, namely large amount of
and predictability of a time series are tightly linked, the higher information processors elements highly connected to
the volatility, the lower the predictability [14]. So, modeling and themselves and capable of assimilating knowledge by training.
forecasting the load time series of a microgrid can be more The basic element of this network is called a neuron and it
complex, compared to the load of a conventional power system. operates as follows: inputs are multiplied by synaptic weights
In this paper, three approaches are applied to model the energy and are combined through a weighted sum. The result of this
consumption forecasting. First it is presented a classical combination then passes through a neuron activation function
solution, autoregressive with exogenous inputs (ARX) model. that will determine the output [12]. Two important parameters
Second, it is presented Artificial Neural Networks (ANN), used in the network are training rate and stop criteria of
which is an artificial intelligence technique and more recent than training. The first is a number between 0 and 1 that indicates
ARX. And, finally, it is presented Artificial Neural Networks the learning speed and must be carefully dimensioned to avoid
optimized by Genetic Algorithms (GA) which is proposed to over training that is nothing more than the network lose its
improve the results of ANN, selecting best data driven for ANN ability of generalization, when the net is so specialized that the
training [16]. These approaches were implemented and errors between the real data and the net output is too low, but
compared with one another for predicting a week-ahead this error can become too high for a different set of inputs. The
electricity consumption of residential customers from historical stop criteria may be maximum number of cycles. It can also be
smart meter data and weather information [13]. taken into account a minimum error as stop, but this method
can generate an increased training time and / or super training.
II. ARX MODELS There are basically two types of training: the supervised one
and the unsupervised one. The supervised training works with
Between the conventional methods in short-term load a tutor indicating the network a desired answer to the sets of
forecasting, there is ARX model, a linear regression method. In inputs, but in the unsupervised one there is no such statement,
this approach, there is no knowledge about that physical i.e., the network organizes itself to set an exit.
equations that describe the process or phenomenon to be studied MLP (Multi-layer Perceptron) net was made possible thanks to
[15]. In this approach, the goal is to find a mathematical model the development of the algorithm back propagation error. Such
that describes the relation between inputs and outputs through algorithm works in two steps: first, the outputs are processed
samples acquisition of the real process. Figure 1 shows the and the processing flows by layers until the output layer which
structure of the identification method. The four steps off occurs the comparison between the obtained and expected
identification are: 1) Input and output data acquisition of the output for that set of inputs, which generates an error, from this
system; 2) Choice of the model structure; 3) Parameters error it is obtained the gradient that is propagated on the
estimation; 4) Model Validation [11]. opposite direction to the initial flow in order to adjust network
A very important factor for a successful identification is the parameters aiming to minimize the error [12]. The activation
correct choice of the model structure (step 2), that will be used function used was sigmoid.
to compose the model [15]. At ARX models, it is required to
find the parameters of the mathematical relation between inputs IV. GENETIC ALGORITHMS
and outputs considering exogenous influences of factors that
have high correlation with system outputs. Among the approaches of the evolutionary computation
techniques, used in advanced forecasting is Genetic Algorithms
(GA). Genetic Algorithms (GA) are optimization and
stochastic searching techniques inspired in the theory of natural
selection of Charles Darwin, described in his famous book
The Origin of Species. The basic operation of the GA is to
generate an initial population constituted of individuals; each
individual is a solution of the problem at hand. From that,
individuals are selected by an evaluation function, which the
best evaluated individuals are selected for designing the next
generation, passing through some processes such as crossover
Figure 1. System Identification and mutation. This procedure is repeated until the defined stop
criterion is satisfied [10]. Fundamentally, Genetic Algorithms
The used method for parameters estimation (step 3) is the Least- are consisting of the following steps:
Square which minimizes the square error between the real and Selection: aims to perpetuate the individuals
estimated value through the equation of the model parameters: characteristics with the highest fitness levels. This index is
measured through a function called Fitness Function and
e = (XTX)-1XTY (1) individuals are classified from the fittest to the less fit. The
e is the vector of estimated parameters, X is the matrix of Fitness Function depends on the problem at hand to be
regressors and Y is the output vector. drawn up.
Crossover: It is the process in which individuals
III. ARTIFICIAL NEURAL NETWORKS exchange parts of their information generating new
individuals, which guarantees to the offspring
Another interesting method is the artificial neural networks. characteristics of interest and enables the search for new
Artificial neural networks (ANN) are computer systems based solutions and possibly more apt.
Mutation: It is a mechanism that randomly changes parts at same period of day, for example the peak of consumption
of individuals, for example, the state change of a particular are usually at the same hour of day (Hour of the load output
bit. This is of interest, because it generates diversity in the and Minute of the load output). Demand pattern is also
population, it causes new possibilities to be tested, as well linked with seasons and its months, at warm countries, heat
as it can be helpful in the bypass some solution at a local system and fans are more used at summer months then at
minimum convergence. But if it is too high, the search can winter months (Month day of the load output).
become essentially random by not preserving features So, information about weather is usually high correlated with
already selected. energy consumption. Weather data is especially crucial for
photovoltaic systems (PV) and wind power forecasting, where
V. METHODOLOGY generation is intermittent. Cloudy day means low energy from
PVs. A day with draughts prevision means high power
The proposal is to get a model of load consumption, based on generation at wind farms Because of that, the variables of third
past patterns, that is able to forecast future loads. System column of Table 1 are related with weather (Temperature
identification techniques were used. This is based on the and Relative Humidity). The identification algorithm is fed
approach where the system is unknown but its inputs and with these inputs and then simulates correspondent future load
outputs are known. The goal is to get a relation between these measures (model outputs), minimizing the error between
inputs and outputs that describes the real system into a model. simulated and real data. When the identification process
The outputs of the model are future load data. An important (Figure 1) is done, the model can be used to forecast, and this
point is to properly select the model inputs. These inputs must prediction measure can feed the model to produce new
have high correlation with outputs to produce efficient model. predictions in advance, in close-loop iteration:
Talking about load forecasting, information about past load data, In this work different models obtained through three
time of the day, and weather conditions are usually correlated different algorithms are compared:
with load consumption. In this paper 11 kinds of inputs are used: 1) Autoregressive with exogenous inputs (ARX).
2) Neural Network
Table 1: Model inputs 3) Neural Network and Genetic Algorithm.
Past data Time Weather
Load measure of Hour of the load Temperature Load data are obtained from 190 houses at a city on the south
day before at same output of Netherlands with smart meters distributed between 5 feeders
time of load output that measure load consumption from 15 to 15 minutes. The
metering period was 2 months, from 24 December 2009 to 26
Load measure of Minute of the load February 2010. It was also taken into account data from air
last week at the output temperature and relative humidity from the studied region. One
same day and time important point, as first step, is to assure the data present the
of the load output minimum level of noise possible. The noise can be from many
Load measure of Week day of the Relative Humidity sources, such as uncertainties on metering or processing
lasts 15 minutes load output system, external events, etc. So, the data were filtered with a
moving average filter with window size of 3 samples to avoid
Load measure of training with wrong data.
lasts 30 minutes Month day of the
load output A. ARX
Mean of the last 24
hours of load As mentioned in section II, four steps are required for ARX
measures implementation. The first step of data acquisition is previously
described (data from smart meters). The second step, about the
The first column of Table 1 treats about past data of energy model structure, follows the choices of the input data (Table 1):
consumption. The data tends to have the same behavior at the
same hour of day (Load measure of day before at same time of ( ) = 1. ( 96) + 2. ( 672) + 3. ( 1) +
load output and Load measure of last week at the same day 4. ( 2) + 5. 1( ) + 6. 2( ) + + 11. 7( ) (2)
and time of the load output). Short past data can give an idea Where,
about the tendency of demand in that period (Load measure of Y(t): Load forecasting
lasts 15 minutes, Load measure of lasts 30 minutes), for Y(t-96): Load measure of day before at same time of load
example, if the energy consumption was low 15 minutes ago, output
the demand on next 15 minutes will probably remain low. Y(t-672): Load measure of last week at the same day and time
Following this idea, if the consumption was high at last day, of the load output
there is a good chance of this demand remains high at next day Y(t-1): Load measure o last 15 min
maybe because of its a cold period and the heat system is being Y(t-2): Load measure of last 30 min
more required (Mean of the last 24 hours of load measures). X1: Hour of the load output
The second column of Table 1 is linked with time information of X2: Minute of the load output
demand. These data are useful to track consumers pattern, as X3: Relative humidity
the load patterns differ among weekday, weekends and holydays X4: Temperature
(Week day of the load output). These data tends to be similar
X5: Mean of the last 24 hours of load measures outputs. For this reason, GA is used to select the best inputs for
X6: Month day of the load output ANN modeling.
X7: Week day of the load output The individual GA is an array with eleven elements that can
a1, a2 a11: model coefficients only assume 0 or 1. Each element is linked with one input.
When the element is 0, that input will be discarded at ANN
The third step needs data organization into a matrix of training. For example, if the individual is [1,1,1,1,1,1,1,1,1,1,1]
regressors with inputs (X) and the output vector (Y): all inputs will be considered, and result will be the same at
section above (V.B). But if the individual is
[1,0,1,0,0,0,0,0,1,0,0], only the 1st , 3rd and 9th inputs of
equation 2 will be used to train the net, so the input matrix
would be:

The targets vector (Y) is always the same. Initially, GA


function generates a set of eleven-element arrays with different
combinations of 0 and 1 (GA population). Each one is a GA
individual and represents a set of inputs for the model that is
Where, used as data driven to train a net. After the training step, GA
n = number of inputs. evaluation function measures the error between real and
simulated data from the net for each individual. The individual
The last step is to run the Least-Square equation (equation 1) to which produces the lowest error will be selected. Then the net
obtain the model coefficients. trained with the data driven from the chosen GA individual
will be used to produce prediction data. The GA process
B. ANN continues until the stop criterion is reached. For this paper, the
limit number of generations used is 15.
The kind of ANN used at this work was Multi-Layer
Perceptron (MLP) with supervised training function Levenberg- D. Error Analysis
Marquardt backpropagation and Gradient descent with
momentum backpropagation as learning function. The ANN Two methods are used to analyze the error between real and
pre-steps are very similar to the ARX. It uses the same simulated data produced by the models. One is the mean-square-
organization of the data into inputs (X) and targets (Y). The error (MSE):
neurons will be trained so when the first line of X is input, net
output will be the simulation of the first value at Y, and ( )
consecutively for the other inputs and correspondent targets. MSE(t) = (3)
ANN is trained to minimize the error between real targets and
simulated outputs. Training process continues until the stop MSE was also used at evaluation function on GA procedure and
criterion is reached, a minimum preset value of mean-square- at ANN training validation.
error between real data and net outputs at validation step. The other is mean-absolute-percentage-error (MAPE):
It was uses a Multi-Layer-Perceptron ANN with back
propagation error, the activation function is sigmoid. An %
important parameter is the number of neuron at hidden layer. MAPE(t) = (4)
This number cannot be too low; else the training process will be
too slow and even inefficient. Also, it cannot be too high, Where,
avoiding overspecialization. For this article, 8 neurons are At= Actual value
chosen at hidden layer because exploratory tests showed better Ft= Forecasting value
results with this number for that load data set. n= Number of measures

C. ANN and GA MAPE is better to compare results with different magnitudes


of real measures, because it is relative to error percentage from
GA is an optimization tool and it is used to improve the each value of load.
results of ANN implementation. As previously mentioned, a VI.RESULTS
relevant point at identification process is the correct selection of
inputs that must have high correlation with the system outputs. To test the methodology the data from a residential feeder was
At ANN procedure 11 inputs from Table 1 were chosen, but used. The load data used at ARX modeling and at training on
there is a possibility that not all of them might have influence on ANN and ANN with GA was from 24 December 2009 to 19
February 2010. The load data used to validate the model and get
the errors of forecasting compared to real data was from 20
February 2010 to 26 February 2010 (1 week).
At ARX identification the model coefficients (equation 2) were:

Table 2: ARX model coefficients


Coefficients Inputs
0,068 Y(t-96)
0,076 Y(t-672)
0,023 X1
0,021 X2
-0,003 X3
0,008 X4 Figure 3: ANN Forecasting and real data (Feeder 1).
0,051 X5
0 X6
0,008 X7
1,416 Y(t-1)
-0,617 Y(t-2)

The most important was Load measure of last 15 min and


Month day of load output was discarded from model. At
identification with ANN and GA, the selected individual was
[1,1,1,0,0,1,1,1,1,1,1]. Relative humidity and Minute of the
load output were excluded from the model. Then, the models
were used to forecast one week of load consumption. These
predictions are compared with real data in Figures 2, 3 and 4.

Figure 4: ANN with GA Forecasting and real data (Feeder 1).

These previsions can be evaluated on next table:

Table 3: Prevision errors (Feeder 1)


Prediction MSE MAPE
ARX 3.7859 22.3121
ANN 3.4378 20.1635
GA ANN 3.1134 16.6226

The best prevision was from the ANN with GA procedure.


The methodology was also tested for the sum of load from the
five feeders. The prevision errors were:

Table 4: Prevision errors (Sum of feeders)


Figure 2: ARX Forecasting and real data (Feeder 1).
Prediction MSE MAPE
ARX 101.207 9.7363
ANN 96.7715 9.29
GA ANN 97.9105 9.0402

In comparison with feeder 1 results, the sum of feeders had


worse results from MSE view, but it is because the magnitude of
the load measures is bigger than the feeder 1 data. When looking
at MAPE error, it can be seen that sum of feeder was easier to
forecast. Below, its shown the prevision by ANN and GA
technique:
each other when summed on bigger grid. It shows the worst
condition on load forecasting at residential Smart Grid view
with lower smoothing effect and evident outliers on load
demand.

VIII. ACKNOWLEDGEMENTS

The authors gratefully acknowledge the contributions of Dr.


Paulo F. Ribeiro, MSc. Roberto S. Netto and Dr. Ballard
Asare-Bediako to this paper.This work was funded by CAPES.

XI. REFERENCES
[1] A. Motamedi, H. Zareipour, and W. D. Rosehart, Electricity Price
and Demand Forecasting in Smart Grids, IEEE Transactions on Smart
Grid, vol.3, no.2, pp664 674, June 2012.
Figure 5: ANN and GA Forecasting and real data (all feeders). [2] L. Hernndez, C. Baladrn, J. M. Aguiar, B. Carro, A. Snchez-
Esguevillas, J. Lloret, D. Chinarro, J. J. Gomez-Sanz and D. Cook, A
Multi-Agent System Architecture for Smart Grid Management and
For this case, GA selected individual [1,1,1,1,1,1,1,0,1,1,1], e.g., it Forecasting of Energy Demand in Virtual Power Plants, IEEE
discarded Month day of the load output from the model. Communications Magazine, pp106 113, January 2013.
These results were calculated through a computer with Intel(T) [3] Y. Simmhan, S. Aman, A. Kumbhare, R. Liu, S. Stevens, Q. Zhou
and V. Prasanna, Cloud-Based Software Platform for Big Data
Core(TM) i3-2350M CPU @ 2.30Ghz, 4.00 GB RAM, running Analytics in Smart Grids, Computing in Science & Engineering, pp 38
the MATLAB R2010b 64 bits. The processing time for each 47, July/August 2013.
approach is shown next: [4] D. Caputo, F. Grimaccia, M. Mussetta and R. E. Zich, Photovoltaic
Plants Predictive Model by means of ANN trained by a Hybrid
Evolutionary Algorithm, The 2010 International Joint Conference on
Table 5: Processing times Neural Networks (IJCNN), July 2010.
Technique Computational Time (s) [5] W. Y. Chang, A Literature Review of Wind Forecasting Methods,
Journal of Power and Energy Engineering, pp161 168, April 2014.
ARX 0.019 [6] I. N. Kakhki, H. Taherian and M. R. Aghaebrahimi, Short-Term
ANN 4.624 Price Forecasting Under High Penetration of Wind Generation Units in
Smart Grid Environment, 3rd International Conference on Computer and
ANN and GA 1441.374 Knowledge Engineering (ICCKE 2013), October 31 &November 1, 2013,
Ferdowsi University of Mashhad.
[7] Y. Hu, H. Ji and X. Song, To Forecast Short-Term Load in Electric
These times are needed to obtain the mathematical model (ARX), Power System Based on FNN, 2009 Sixth International Conference on
the net trained (ANN) and net trained with best inputs (ANN and Fuzzy Systems and Knowledge Discovery, pp 474 478, 2009.
GA). [8] J. Wang, L. Li, D. Niu, Z. Tan, An annual load forecasting model
based on support vector regression with differential evolution algorithm,
Journal of Applied Energy, pp 65 70, February 2013.
VII. CONCLUSION [9] C. Tranchita and A. Torres, Soft Computing Techniques for Short
Term Load Forecasting, Power Systems Conference and Exposition,
The three techniques properly forecasted the tendency of real 2004, October 2004.
data. It can be seen that the procedure that achieved best [10] T. Back, D.B. Fogel, Z. Michalewicz Evolutionary Computation 1:
Basic Algorithms and Operators, Ed.Taylor& Francis LTD, 2000.
predictions was ANN working with GA. For both applications, [11] L. Ljung System Identification Theory for the User, Ed. Prentice
load consumption of Feeder 1 and of sum of all five feeders, Hall, 1999.
ANN and GA produced best results and showed that discarding [12] S.S. Haykin Redes neurais-Princpios e prtica, Ed.Bookman,
some inputs can improve results and can bring savings with 2001.
[13] B. Asare-Bediako, W. L. Kling, and P. F. Ribeiro, Day-Ahead
sensors, data acquisition, data processing, computational cost Residential Load Forecasting with Artificial Neural Networks using
and time. Smart Meter Data, Power Tech (POWERTECH), 2013 IEEE Grenoble,
The three methods have different velocities at modeling. The June 2013.
quickest is ARX, because there is no training step, and it also [14] N. Amjady,F. Keynia,H. Zareipour, Short-Term Load Forecast of
Microgrids by a New Bilevel Prediction Strategy,IEEE transaction on
scales the most important inputs through its coefficients. The Smart Grid, vl. 1, no. 3, december 2010
ANN takes longer because it needs a training phase. The ANN [15] Shyh-Jier Huang ; Nat. Cheng Kung ; Kuang-Rong Shih Short-
and GA is the longest because for each tested data driven there term load forecasting via ARMA model identification including non-
is a training phase. But this modeling process is just needed Gaussian process considerations Power Systems, IEEE Transactions on
(Volume:18 , Issue: 2 ), May 2003
once to get the tendency of real data into a mathematical model [16] H. Niska, P. Koponen and A. Mutanen, Evolving Smart Meter Data
or trained ANN. Then the work is just to feed the model or net Driven Model for Short-Term Forecasting of Electric Loads, IEEE
with correct inputs and get forecasting data as outputs. Tenth International Conference on Intelligent Sensors, Sensor Networks
Other point was on the methodology application to sum of and Information Processing (ISSNIP),Singapore, 7-9 April 2015
[17] Takuro Sato,D. M. Kammen, B. Duan, M. Macuha, Z. Zhou, J. W.
feeders, which showed better MAPE results than on application Muhammad T., S. A. Asfaw Smart Grid Standards Future of the
to Feeder 1. It proves the smoothing effect on load aggregation Smart Grid 10 (391-392) - 2015
at bigger grids where variation effects off small grids cancel

Potrebbero piacerti anche