Chapter 1 INTRODUCTION

1.1 Need of Prediction of overloading in transmission lines

Today the Generation system has expanded enormously but the expansion of transmission system has not kept pace with it. As a result the transmission system is overloaded, resulting in high losses. The problem of monitoring the power flows and bus voltages in a power system is very important in maintaining system security and fast prediction is essential for controlling these quantities. As power systems have become more stressed due to increased loading and large interconnections, there will be an increase in cases of voltage limit violation and line loading limit violation, particularly in contingency conditions like line outage, generator outage etc. The alleviation of emergency transmission line overload is a critical problem in power system operation. An efficient, reliable and direct method is always desirable. The concept of local optimisation is introduced, and a method is developed for the same. This gives the proper sequence of control actions and adjustment in control variables to alleviate line overloads [2]. This is essential that power flows in all the branches respect their specified limits not only in base case condition but also in stressed/line outage conditions. The most severe violation in a line flow limit can be due to different contingencies. Therefore an immediate need arises to take corrective action for alleviation of line overloads in the overloaded lines of the system. There are several methods based on optimal power flow (OPF) for the corrective and preventive control actions along with economy and security function. Power system is a dynamic system as the operating state of it continually changes with respect to time. The emergency state of line overloading may occur as a result of sudden increase in system demand, outages of generator or transmission line or failure in any of the system components.

1.2 Literature survey

In reference [3], a corrective switching algorithm has been proposed which relieves overloads and voltage violations as well. The traditional form of load control (shedding) is quite disruptive to consumers and so often avoided. In reference [4], a non-disruptive load control method has been developed to switch small pieces of load, so that interruptions are effectively unnoticed by consumers. Several power flow methods are available to compute line flows in a power system like Gauss Seidel iterative method, Newton-Raphson method, fast decoupled power flow method and dc power flow method but these are either approximate or too slow for on-line implementation [5,8].
1

In [6] an approach based on radial basis function neural network (RBFN) is presented for corrective action planning to alleviate line overloading in an efficient manner. Effectiveness of the proposed method is demonstrated for overloading alleviation under different loading/contingency conditions in 6-bus system and 24-bus RTS system.

1.2.1 Role of Artificial neural networks in Power Systems

Due to the uncertainty in the results now a days methods have been developed that use artificial intelligence (ANN, fuzzy logic). These methods are accurate and efficient in discovering similarities among large bodies of data and synthesizing fault tolerant model for nonlinear, partly unknown and noisy/corrupted system. Artificial neural network is the functional imitation of a human brain which simulates the human intuition in making decisions and drawing conclusions even when presented with complex, noisy, irrelevant/partial information. The information going to the input layer neurons (units) of artificial neural network is recoded into an internal representation and the outputs are generated by the internal representation rather than by the input pattern. It can model any non-linear function without knowledge of the actual model structure and during testing phase it gives the result in very short time. A neural network consists of a number of neurons, which are the elementary processing units that are connected together according to some pattern of connectivity. The development of artificial neural network involves into two phases, training or learning phase and testing phase. Developing a neural network is unlike developing software, because the network is trained, not programmed. Most of the published work in power system area utilizes multi-layer perceptron (MLP) model based on back propagation (BP) algorithm, which usually suffers from local minima and over fitting problems [6].

1.3 Organisation of this report

In this report, a cascade neural network based approach is proposed for fast identification and prediction of transmission line overloading. The developed cascade neural network is a combination of an Identification module (ANN1) and a Prediction module (ANN2). All the training patterns are applied to the identification module, which is trained to classify them either in overloaded class or in under-loaded class using a modified BP algorithm. The identified overloaded cases are then passed to the prediction module (which is a feed-forward counter propagation neural network) for prediction of line overloading. The input features for CNN are taken from the set of real power injections at generation (PV) and load (PQ) buses, and reactive power injections at PQ type buses as these independent and known variables (prior to power flow analysis) influence the line flows most. Due to large
2

number of such variable in any practical power system, it is not possible to consider all these variables as inputs to an ANN, as it will increase the input dimension, the size of the neural network i.e. the number of interconnection weights and ultimately the training time. So to overcome this difficulty, the variables are grouped into different clusters and from each cluster one representative variable is selected as the input feature to the proposed neural network. In this paper, angular distance based clustering [7] is applied for the selection of input features. Optimal training of the CNN has been achieved by iteration wise monitoring the error patterns for the training set and the validation set. The proposed neural network based technique is applied for overloading identification and prediction at different loading/ generation conditions in IEEE-14 bus system. Chapter 2 presents the methodology used for the prediction of overloading in transmission lines with the help of CNN. Chapter 3 presents the Simulation results and graphs representing the results.

Chapter 2 PROPOSED METHODOLOGY

The block diagram of the proposed Cascade neural network is shown in Fig. 2.1. By perturbing the load at all the buses randomly in wide range of system operating conditions, a large number of load patterns have been generated and full ac power flow analysis has been carried out for each case to calculate real line-flow in each line of the power system. Angular distance based clustering [7] is applied for selection of input features from the set of real power injections at generation and load buses, and reactive power injections at load buses.

Input

Identification Module

Output (0.1/0.9) 0.1 (Under-loaded) 0.9 (Overloading) Prediction (Overloaded) module Input

Fig. 2.1 Cascade neural network for identification and prediction of overloading

An identification module which is a three-layered feed-forward neural network with single clamped output as shown in Fig. 2.2, is developed to classify the transmission lines condition either in overloaded or in under-loaded class taking into consideration the real power flows and the maximum power flow limits in different lines.

Fig. 2.2 Identification module (three-layered ANN).

The identification module is trained using modified BP algorithm [8] such that its target output is high (0.9) when presented with a sample from overloaded class (C1), and low (0.1) when presented with a sample from under-loaded class (C2). During training, the outputs of the identification module greater than 0.9 are clamped to 0.9, similarly outputs smaller than 0.1 are clamped to 0.1 to reduce the likelihood of the network getting stuck in local minima. The training set for identification module contains many more exemplars for under-loaded class C2 than for class C1. When training a multilayer feed-forward neural network with standard back-propagation for such a two-class problem, in which one dominant class contains far more exemplars than a subordinate class (i.e., the training set is imbalanced), the rate of convergence of net output error is very low.

Fig.2.3 Relationship between gradient vectors

The reason for this is that negative gradient vector computed by back-propagation for an imbalanced training set does not initially decrease the error for the subordinate class as shown in Fig. 2.3. Consequently in the initial iteration, the net error for the exemplars in the subordinate class increases significantly. The subsequent rate of convergence for the exemplars of the subordinate class is very low. To solve this problem, in place of standard BP algorithm, a modified BP algorithm [8] is used that calculates a direction of the descent vector v in weight space, which is downhill for class C1 as well as for class C2 in each iteration as shown in Fig. 2.4 i.e. v satisfies ( ) ( ) (2.1)

( ) and ( ) refer to the errors (sum of squared errors) for the patterns where belonging to class C1 and class C2 respectively at Kth iteration.

Fig. 2.4 Direction of gradient vectors in modified algorithm

Direction of v is set so that it bisects the angle between

( ( ) ) ( ( ) )

( ) and

( ): (2.2)

and thus the rate of learning can be accelerated by one order of magnitude for such two class problems [8]. The prediction module is a feed forward counter propagation neural network (CPNN), which uses a different mapping strategy namely counter propagation. The counter propagation network provides a practical approach for implementing a pattern-mapping task, since learning is fast in this network [10]. The counter propagation neural network is shown in Fig. 2.5. This neural network model is a combination of Kohonen network and a Grossberg outstar. The CPNN model (Fig. 2.5) involves both supervised and unsupervised learning. The Kohonen network implements the winner-take-all (competitive) strategy for the weights from the units in the input layer to the units in the hidden layer, and the Grossberg outstar maps the winning neuron into the desired output. The unsupervised and supervised training are applied to train the CPNN model.

Fig. 2.5 Feed-forward counterpropagation network. 6

CPNN is trained in a two-phase process. In the first phase, the Kohonen layer neuron weights are adjusted to match the input. The second training phase helps to adjust the Grossberg weights in order to fit the desired neuron output. A large number of load patterns are generated randomly by perturbing the load at all the buses and generation in wide range. Newton-Raphson (NR) ac power flow program is developed to generate training/ testing patterns for different load scenarios such that some of the transmission lines get overloaded.

2.1 Line overloading calculations using Newton-Raphson power flow method

The objective of power flow or load flow study is to determine the voltage and its angle at each bus, real and reactive power flows in each line and line losses in the power system for specified power system conditions. The power flow studies are conducted for the purpose of planning (viz. short, medium and long range planning), operation and control. For the purpose of power flow studies, it is assumed that the three-phase power system is balanced and also mutual coupling between elements is neglected [12-13]. Variables associated with each bus of a power system include four quantities viz. voltage magnitude Vi, its phase angle , real power Pi and reactive power ; total 4m variables for m buses system. At every bus two variables are specified, the remaining two can be found by solving the 2m power flow equations. Depending upon which two are specified, the buses are classified as Swing Bus or Reference Bus, Generator Bus or PV Bus and Load Bus or PQ Bus. From the nodal current equations, the total current entering the ith bus of m bus system is given by Where is the admittance of the line between buses i and k and polar coordinates (2.3) is the voltage at bus k. In

(2.4) (2.5) (2.6)

Here, is the angle of the bus voltage and conjugate power will be Or
7

is bus admittance angle. At ith bus, complex

(2.7)

(2.8)

The real power at ith bus will be Or | | ( ( ) ) ( ) (2.10) (2.11) | |

( )

(2.9)

Similarly, the reactive power at ith bus will be Or | | ( ( ) ) ( ) (2.13) (2.14) | |

( )

(2.12)

Eqs. (2.10, 2.11) and (2.13, 2.14) are known as static power flow equations (SLPE). The power flow equations used in Newton-Raphson method for computation of voltage corrections are given as [ ] [ ][ ] (2.15)

where, H, N, J and L are the sub-matrices of the Jacobian in Eq. (2.15), having ikth elements as

(2.16)

Eq. (2.15) may be written as [ ] [ ] [ ] (2.17)

The solution of Eq. (2.17) provides the correction vector i.e. for all the PV and PQ type buses and s for all the PQ type buses, which are used to update the earlier estimates of s and Vs. This iterative process is continued till the convergence is obtained i.e. mismatch vector s for all the PV and PQ type buses and s for all the PQ buses become less than a pre assigned tolerance value . Once the solution of bus voltages (| | and for load buses and for generation buses) is found, the power flows in line between buses i and k can be calculated using nominal-pi representation of the line. Current flow from bus i towards bus k will be (2.18) where is line charging of the line between buses i and k. The power flow in the line i-k at the bus i is given by (2.19) In Newton-Raphson method, Jacobian elements are to be calculated and its inverse is also required using Eqs. (2.15) and (2.17) in each iteration. Due to this fact, the Newton-Raphson method requires more time per iteration. However, this method provides accurate results and is the most reliable ac power flow method. Hence, in this work the NR power flow program has been developed and run to calculate real power flows in different lines of a power system. Line overloading is evaluated as the amount by which the real power flow exceeds the maximum power flow limit for the line i.e. the line rating. Line overloading = real power flow in the line - rating of the line Or (2.20) Where is line overloading in any line i-k and is maximum power flow limit of line i-k. In case of overloading in any line i-k, the value of would be positive, otherwise it would be either negative or zero. The solution algorithm for prediction of line overloading can be summarized in following steps: (i) A large number of load patterns are generated randomly by perturbing the load at all the buses and real power generation at the generator buses. (ii) Full ac power flow programs are run for each case to compute real power flows. (iii) Input features are selected by using angular distance based clustering from the set of real power injections at PV and PQ buses, and reactive power injections at PQ type buses. (iv) The input data are normalised between 0.9 and 0.1. (v) The normalised input data along with the topology (normalised Gij and Bij corresponding to the line between buses i and j) are used for training of the cascade neural network.
9

2.2 Input Feature selection

In any ANN application, if a large number of inputs are used, the size of a neural network and the number of interconnection weights will increase and the training of neural network will be extremely slow. To avoid this problem, some feature selection technique is to be applied so that only those inputs could be selected which strongly affect the output of a neural network. Several feature selection methods like entropy reduction method, principal component method, correlation coefficient based method, angular distance based method etc. are available in the literature. However, in this paper, angular distance based clustering of real and reactive power injections has been applied for input feature selection. The basic purpose of clustering [7] is to group the total M system variables (SV1, SV2, . . ., SVM) into G clusters such that the variables in a cluster have similar characteristics. One representative variable from each cluster is picked out as a feature for the cluster. Thus the number of variables will be reduced from M to G. The system variables considered here are real power injections at all the buses except slack bus and reactive power injections at PQ buses only. A large number of patterns (say N) are generated by perturbing the load at all the buses and real power generation at the PV buses randomly in wide range of operating conditions and Newton-Raphson power flow as discussed in Section 2.1 has been run to obtain line flows for each case. Each operating point i.e. the power injections can be described by state vector, (i = 1, 2, . . ., N). All the N state vectors can be represented in the form of a matrix X as-

The ith row of matrix X contains the values for M system variable (SV1, SV2, . . ., SVM) at ith operating point while jth column of matrix X consists of SVj variables in the N training patterns. In matrix X, if we define a column vector, (2.21) Then the M system variables SVj (j = 1,2, . . .,M) can be clustered based on these column vectors CYj. These system variables with similar vectors CYj will be grouped in one cluster. In the angular distance based clustering [7] technique of feature selection, the system variables are clustered on the basis of angular distance between them. The cosine value of the angle between two vectors CYj and CYk is defined as (2.22)
10

||

Two vectors CYj and CYk, which are similar to each other, will have a small angle between them. Hence, the cosine value as obtained by Eq. (2.22) can be used to evaluate the degree of similarity between two vectors. If cos is greater than a specified threshold, the two vectors CYj and CYk are considered as two similar vectors and are put in the same group (cluster). Otherwise a new cluster is formed for the vector CYk. The clustering process is carried out for all the system variables. Once all the variables are processed, the algorithm is repeated until stable clusters are formed. For any cluster g, the representative variable is selected by finding out a system variable SVg whose vector is closest to the cluster vector Cg. The system variable SVg will be feature variable for the cluster g. Thus, G features can be selected out of M system variables. In this way, the input features i.e. the real and reactive power injections are selected for training of the cascade neural network.

11

Chapter 3 SIMULATION RESULTS

The cascade neural network based method has been tested for prediction of overloading in IEEE 14-bus system [12], which is composed of 14 buses and 20 lines. Changing the load at each bus and generation at PV buses randomly several load patterns were generated. The power factor of loads at different buses was maintained constant. The NR power flow method was used to compute power flows for each loading scenario. The input as well as output data were normalised between 0.9 and 0.1. Since only one neural network is trained for a number of line-flows, a topology number is required along with the input data to train the neural network. The topology number in the form of Gij and Bij corresponding to the line between buses i and j are used for training of the identification module. To reduce the training time the adaptive learning rate was also used along with modified BP algorithm in identification module [8]. The data for IEEE-14-bus system were taken from [12] with buses renumbered to make bus-1 as slack bus, buses 25 as PV buses and buses 6 14 as load (PQ) buses. As many as 120 load scenarios were generated by changing the load at each bus and generation randomly in wide range ( 50% of base case) and the full ac power flow was run for each load pattern to compute the real line flows in different lines of the sample system. Using angular distance based clustering, 10 no. of power injections (P2, P3, P4, P6, P9, P13, P14, Q9, Q10, Q11) were selected with the threshold as 0.927. In addition to these features the line parameters G and B are also included as inputs, making the total 12 inputs for CNN. All these input data were normalised between 0.1 and 0.9 to overcome the problem of data suppression by large valued inputs. In this paper, each input parameter X is normalized as Xn according to the following equation: ( Where ) is maximum value of X and is minimum value of X. (3.1)

Each load scenario will generate 20 patterns corresponding to line flows in 20 lines and will provide 2400 patterns. Out of 120 load scenarios, 80 load scenarios (1600 patterns) were arbitrarily selected for training while, 20 load scenarios (400 patterns) were used as validation set and the remaining 20 load scenarios (400 patterns) were used for testing the performance of the identification and prediction modules. This validation set is used for checking the optimal training point of the developed CNN.

12

Outputs of the proposed identification module will identify the overloaded lines. If a line is overloaded under some loading condition, it would be classified in overloaded class and the corresponding output will be 0.9, otherwise it will be 0.1 for under-loaded class. The proposed identification module ANN1 (12-27-1) has the input layer of 12 neurons; one hidden layer of 27 neurons (optimum number of hidden nodes) and an output layer of 1 neuron. The identification module was trained using a modified BP algorithm to filter out (identify) the overloaded lines. The initial value of adaptive learning rate was selected as 0.5 which was changed during the neural network training according to the nature of training error in each iteration. These identified line overloaded cases were applied for training of the other ANN namely the prediction module for prediction of amount of overloading in those lines. Table I Identification of overloading for IEEE 14-bus system Line No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 From bus 1 1 2 2 2 3 4 4 4 6 6 6 7 7 8 9 9 10 12 13 To bus 2 8 3 6 8 6 11 12 13 7 8 9 5 9 4 10 14 11 13 14 Target T 0.9 0.1 0.9 0.9 0.9 0.1 0.1 0.1 0.1 0.9 0.1 0.1 0.1 0.1 0.9 0.1 0.1 0.1 0.1 0.1 ANN O 0.8949 0.1971 0.8957 0.8773 0.8585 0.1015 0.1042 0.1089 0.1049 0.8430 0.1033 0.1016 0.1011 0.1023 0.8645 0.1017 0.1034 0.1043 0.10407 0.1146 Class OL UL OL OL OL UL UL UL UL OL UL UL UL UL OL UL UL UL UL UL

In the above table-T: Desired output of identification module; O: actual output of identification module; OL: Overloaded Class; UL: Under-loaded class.

13

Out of 1600 training patterns, the identified overloading cases were 586 corresponding to 80 load scenarios. These 586 patterns were used to train the prediction module ANN2 i.e. the Counter propagation network (12-85-1). The initial values of as selected as 0.7 and consequently decreased to 0.2. The initial value of b was selected as 0.1 and consequently reduced to and 0.01. During testing phase, the 400 unseen patterns were tested for screening through the trained identification module. The trained identification module identified all the 147 overloading cases correctly. These 147 overloading cases corresponding to 20 operating scenarios were then passed to the trained prediction module for testing. Though the proposed cascade neural network identifies all the testing patterns correctly and predicts overloading amount accurately, the test results corresponding to only one load scenario are presented in Tables I and II. Table II Determination of overloading for IEEE 14-bus system Line No. 1 3 4 5 10 15 From bus 1 2 2 2 6 8 To bus 2 3 6 8 7 4 Target 0.6157 0.2695 0.3886 0.3408 0.1714 0.2313 ANN 0.6163 0.2712 0.3898 0.3391 0.1737 0.2352 Error(pu) -0.0006 -0.0017 -0.0012 -0.0017 -0.0023 -0.0039

Table I shows the performance of the identification module, which screens line flows in overloaded or under-overloaded class while Table II presents the overloading amount in overloaded lines screened by the identification module. As can be observed from Table I, the identification module screens all the overloading cases correctly. Table II shows that actual values of overloading amount and those obtained by the prediction module (ANN2) are very close to each other.

14

Fig. 3.1 Testing Performance of ANN 2

The performance of prediction module for all the unknown overloading patterns is shown in Fig. 3.1. The rms error for all the test patterns is 0.0043 p.u. The trained Cascade neural network is able to identify all the overloading cases correctly and at the same time able to predict the overloading accurately.

15

Chapter 4 CONCLUSIONS

The identification of overloaded lines and prediction of line overloading in different overloaded transmission lines are essential for secure and reliable operation of a power system. For this purpose analytical methods take a long time, as ac power flow analysis has to be carried out for any change in loading/generation condition. On the other hand, once the training of the cascade neural network is successfully accomplished, the identification of overloaded lines and prediction of overloading amount in those lines for unknown loading conditions is almost instantaneous. Once the overloading in different lines is predicted accurately, a fast and intelligent control action can be taken in the form of generation scheduling/load shedding to reduce the line overloads to the security limits. The proposed intelligent technique can be implemented for on-line MW security assessment in Energy Management Systems.

16

BIBLIOGRAPHY

[1] Savita Sharma, Laxmi Srivastava, Prediction of transmission line overloading using intelligent technique, Applied Soft Computing 8 (2008) pp-626633.(Base Paper) [2] Shandilya, H. Gupta, J. Sharma, A method for generation rescheduling and load shedding to alleviate line overloads using local optimisation, IEE Proceedings on Generation, Transmission and Distribution 140 (5) (1993) pp-337342. [3] S. Wei, V. Vittal, Corrective switching algorithm for relieving overloads and voltage violations, IEEE Transactions on Power Systems 20 (4) (2005) pp-18771885.\ [4] I. Hiskens, B. Gong, MPC-based load shedding for voltage stability enhancement, in: Proceeding of the 44th IEEE Conference on Decision and Control, Seville, Spain, December 1215, (2005), pp-44634468. [5] N. Kumar, R. Wangneo, P.K. Kalra, S.C. Srivastava, Application of artificial neural network to load flow, in Proceedings TENCON91, IEE Region 10 International Conference on EC3-Energy, Computer, Communication and Control System, vol. 1, 1991, pp-199203. [6] Daya Ram, Laxmi Srivastava, Manjaree Pandit, Jaydev Sharma, Corrective action planning using RBF neural network, Applied Soft Computing 7 (2007) pp-1055 1063. [7] Y.Y. Hsu, C.R. Chen, C.C. Su, Analysis of electromechanical modes using an artificial neural network, IEE Proceedings Generation, Transmission and Distribution 141 (1994) pp-198204. [8] A.N. Adupa, G.K. Purushothama, K. Parthasarathy, D. Thukaram, A fuzzy control for network overload alleviation, International Journal of Electrical Power & Energy Systems 23 (2) (2001) pp-119128. [9] R. Anand, Kishan G. Mehrotra, C.K. Mohan, S. Ranka, An improved algorithm for neural network classification of imbalanced training sets, IEEE Trans. Neural Networks 4 (1993) pp-962969. [10] K.L. Ho, Y.Y. Hsu, C. Yang, Short term load forecasting using a multilayer neural network with an adaptive learning algorithm, IEEE Trans. Power Systems 7 (1992) pp-141149.

17

[11] K.L. Lo, L.J.Peng, J.F. Macqueen, A.O. Ekwue, D.T.Y. Cheng, Fast real power contingency ranking using counter propagation neural network, IEEE Trans. Power Systems 13 (1998) pp-12591264. [12] L.L. Freris, A.M. Sasson, Investigation of the load-flow problem, Proc. IEE 115 (10) (1968) pp-14591470. [13] P.A. Chamorel, A.J. Germond, An efficient constrained power flow technique based on active reactive decoupling and the use of linear programming, IEEE Transactions on Power Apparatus and System 101 (1982) pp-158167.

18