Genetic Neural Networks - Based Strategy for Fast Voltage Control in Power Systems

Transcription

1 Genetic Neural Networks - Based Strategy for Fast Voltage Control in Power Systems M. S. Kandil, A. Elmitwally, Member, IEEE, and G. Elnaggar The authors are with the Electrical Eng. Dept., Mansoura university, Egypt ( kelmitwally@yahoo.co.uk). Abstract-In this paper, a new technique for voltage and reactive power control based on genetic-ann is proposed. Genetic algorithm is employed to give more proper initial weights for the decision-making neural network. Then, a fast complete training of the ANN is accomplished using backpropagation algorithm. The training data is obtained by running an adjusting power flow program at different operating conditions of the power system. The ANN has a three-layer feedforward structure. The input vector includes the load level of the power system and the required voltage values at the load buses. The output vector is the setting values for the various reactive power control devices incorporated into the power system. The integrated scheme is applied to two example systems which are the IEEE- 30 bus and middle delta zone network in Egypt. The results obtained clearly show that, the ANN approach is capable of alleviating voltage violations in power system. Furthermore, a comparison with adjusted power flow technique shows the clear superiority of the proposed ANN to achieve the proper control action in a shorter computational time. Keyword: Voltage control, neural network, genetic algorithm, adjusted power flow, fast control. I. INTRODUCTION One of the main requirements in power system is to keep the load bus voltage within limits specified for the proper operation of equipment. Any change to the system configuration or in power demands can result in higher or lower voltages in the system. The operator can improve this situation by reallocating reactive power generations in the system by control devices such as transformer tap settings, generator voltage regulator and switching reactive power sources. In this respect, an efficient control technique is needed [1]. In the past, several techniques have been employed using sensitivity relationships and gradient search approaches to overcome this complex problem [2-3]. This techniques give the approximate changes in bus voltages for a given control action. In these approaches, the bus voltage violations are alleviated one by one. So these methods can be used in small number of violations. In case of many violations, the method may run into an infinite number of iteration. To avoid these difficulties linear programming (LP) approach [4-5] has been proposed to yield the control actions. In most of these studies, the LP problem has been formulated using real valued control variables in order to reduce the computational effort. However, these methods are complex and require significant computational effort to determine the required adjustments to control variables. Artificial intelligence (AI) methods have also been applied to control voltage to be within acceptable limits. The expert system (ES) techniques [6] are applied to identify the system operating conditions, detect the bus or buses at which certain constraints have been violated, and select the appropriate control actions to alleviate the voltage violations. Therefore, ES decides and gives proper signals to perform the control actions of the power system. However, in the case of large power system, the knowledge base is larger and consequently the search time will be increased. References [7-9] presented fuzzy logic theory to optimal control of reactive power. Using fuzzy sets operators the coefficients of the objective function are calculated for each bus and membership functions are defined for bus voltages. Reference [10] presents Artificial Neural Network (ANN) to control the voltage and reactive power in power systems. A three Layer Feed-forward ANN with Back Propagation training algorithm is trained to give the proper control action required to achieve reactive power and voltage control [11-12]. In this paper, a Genetic-Based ANN (GBANN) technique is proposed to control the voltage in power system. The genetic algorithm is embedded to intelligently decide the initial weights for the ANN learning process. Thus, the computation time for the ANN learning can be largely reduced. In addition, the genetic algorithm can prevent the ANN learning from being trapped in a local minimum. The training data is generated by running the modified power flow program (MPF) at several conditions [13-16]. The proposed technique is tested by computer simulation on IEEE 30-bus system and 220 kv Delta Network, Egypt. The obtained results are satisfactory and show superior performance compared to other techniques. II. FRAMEWORK OF GENETIC-BASED ANN In the proposed genetic-based ANN approach, three-layer feed forward ANN is used [17]. Genetic algorithm is employed to generate the best initial weights of the ANN and the back-propagation technique is applied to train the ANN. A. Genetic Algorithm Genetic Algorithm (GA) is a search algorithm based on the natural selection and genetics [18]. The features of genetic algorithm are different from other search techniques in several aspects. First, the algorithm works with a population strings. Thus searching many peaks in parallel, and hence reducing the possibility of local minimum trapping. Secondly, GA works with a coding of parameters instead of the parameters themselves. The coding of parameters will help the genetic operator to evolve the current state into the next state with Reference Number: W

2 minimum computations. Thirdly, GA evaluates the fitness of each string to guide its search instead of the optimization function. Finally, GA explores the search space where the probability of finding improved performance is high. In genetic terms, each parameter is represented by a string structure. This is similar to the chromosome structure in natural genes. A group of strings is called population. Strings in a given population when mutated or crossovered produce a new generation. For each generation, all the populations are evaluated. The individual (string) with high fitness will have more chances of evolving to the next generation (or next state). The process of GA is surprisingly similar to the biological process. It involves selection of strings, mutually swapping of strings (crossover), and changing partial strings (mutation) [19]. B. Estimation of initial weights for ANN The main computation blocks of the GA are described below [20]. Step 1: Generate Initial Population Pool A random number generator is used to generate individuals. The chromosome length is the number of weights in the ANN topology. Step 2: Conversion From Phenotype and Genotype The phenotype is the physical expression of a chromosome where the genotype is the genetic structure of a chromosome. In neural network terms, the genotype is the binary code of the neural network weights. The actual person characteristics (slim, tall, blue eyes, etc) are the phenotype. In neural networks, the phenotype is the weights in decimal form. In the proposed genetic-based neural network approach, all initially generated individuals need to be evaluated. For convenience of processing, these initially generated individuals are converted from genotype (chromosome) to phenotype (weights). Step 3: Conversion of Error to Fitness The main issue of training is to minimize the objective function, which is the difference between the neural network output and the target output. With GA, the minimization is achieved in such a way that smaller functional values are mapped to larger fitness values. Although the best conversion function can be somewhat problem-dependent, one function that has been found to be generally useful is the exponential function [4]: ( ) kv v = e F (1) Where k is a negative number. Since the neural network error is within the range from zero to one after normalization, using k = -20 is sufficiently effective. This would map an error of to a raw fitness of 0.98, 0.01 to 0.82 and 0.1 to The transformed fitness is then normalized by the average fitness. Therefore, each fitness can be interpreted as the expected number of times it would be chosen from the population pool. Step 4: Parent Selection Two arrays (strings) are selected to be parents. The arrays are selected from the strings with higher fitness. Parents with high fitness can be selected more than once. This is called the frequency of selection. Hence, a higher fitness is correlated with a higher probability of parent selection. Step 5: Crossover A crossover point along the chromosome is randomly selected. Both parents are split at this point. Part of the genes of one parent, and one side of this point, will be given to the child. The rest of the child s genes will come from the other side of the crossover point in the other parent. Step 6: Mutation Mutation plays a vital role in genetic optimization as it introduces new genetic material that was lost by chance through poor selection of mates. However, the overuse of mutation may destroy good chromosome in the next generation. In this paper, the probability of mutation is selected equal to Step 7: Decision The number of randomly generated individuals in the population is set equal to 50, and the number of generations is set equal to 3. This number is arbitrary selected. However, a big number may lead to a longer computation time. All the individuals in each generation must be evaluated with respect to the network error. If the network error is improved, the genes of this individual will be decoded to the network weights. These weights can then be used as initial weights for the neural network learning. C. Back-propagation learning algorithm The algorithm gives a prescription for changing the weights in any feed-forward ANN to learn a training set of input-output pairs. The back-propagation learning algorithm (BPLA) is essentially an optimization method that uses an iterative gradient algorithm that is designed to minimize the mean square error between the actual output of the neural network and the desired output [21]. The algorithm is performed in two successive steps: forward propagation and back propagation. In the forward propagation, a pattern vector at the input layer with its desired output pattern at the output layer is simultaneously applied to the network. The error detected at the output layer is then back propagated through the network to update the connection weights according to the generalized delta rule [21]. III. GENERATION OF THE TRAINING DATA The well-trained neural network should give the right decision for different operating conditions. To achieve this important goal, the training data should be selected carefully. In this case, the training data cover a wide range of steady state operating conditions. Once ANN is trained, it is tested for validation by considering operating conditions not Reference Number: W

3 included in the training set and evaluating the errors between the desired and actual values of the outputs. The training data consists of two vectors. The first is called the input vector (independent variables) and the second is called the output vector (control variables). The input vector includes the load multiplier (LM) and the load bus voltage (V L ). The output vector includes the generation reactive power (Q G ), the reactive power source (Q C ) and the transformer tapping (T). The training data is generated by running the adjusted load flow program at different operating conditions. The synchronous condenser (Q S ) is included as a generator with zero real power output. The following procedure is used to generate the required training data: 1. Define the load level condition (LM). 2. Define the state of control devices (ON or Off). 3. In case of Off state (without control), run the modified power flow program to obtain the state of input vector (LM, V L ) 4. Save the above data for ANN training as input vector. 5. In case of ON state (with control), run the modified power flow program to obtain the control actions ( Q C, Q G, Q S, T) required to keep the voltages within limits. 6. Save the above data for ANN training as output vector and modify the state of input vector. 7..Save the modified data as input data to the next load level condition. 8. Go to step 3 9. Repeat the steps from 4 to 7 in order to obtain a suitable number of training examples. IV. DESIGN OF GENETIC-BASED ANN Fig. 1 describes the framework of genetic based neural network. It is summarized as follows: 1. Determine the best initial weights of the ANN using GA. 2. Generate the training and testing data. 3. Train the ANN using Back-Propagation method. 4. Test the trained ANN. 5. If the error is acceptable, then save the results. 6. Else go to step 3 to repeat training. A. Case study 1 V. APPLICATION The IEEE 30-bus system is used in this study to test the proposed approach. [7] - [9]. To improve the efficacy of the GBANN approach, it has been trained with different operating conditions. These different operating conditions were selected at loading level in the range from 10% to 105% with respect to the nominal operating point. Each pattern contains18 input elements and 11 output parameters. NO GENERATE THE BEST INITIAL WEIGHTS USING GENETIC ALGORITHM GENERATE THE TRAINING AND TEST DATA TRAIN ANN USING BPA VALUE TEST THE TRAINED ANN YES IF ERROR IS ACCEPTABLE? PRINT THE RESULTS STOP Figure (1) Framework of genetic-based neural network For estimating the initial weights, the GA is applied with three generations. In each generation, 50 individuals are generated in the population pool. Table1 shows the parameters and results of the GA. It also indicates the minimum and maximum error of the generated individuals after each generation. Only the individual with minimum error will be qualified to be decoded to the initial weights for the ANN training. The ANN has been trained until the final error becomes % of maximum error with 10 iterations. The structure of the trained ANN is as follows: 18 neurons / nodes on the input layer, 6 neurons on the hidden layer, and 11 neurons on the output layer. To assess the effectiveness of the trained ANN, it is tested with an unforeseen pattern, i.e. 100% loading conditions. The test results are given in Table 2 that compares the control actions taken by MPF technique versus those obtained by the proposed GBANN technique. The MPF technique suggests changes on the control actions that are listed in the third column of Table 2. The GBANN technique suggests to adjust the control variables as shown in the second column of Table 2. It is clear that the test results are within acceptable error. Then, both techniques MPF and GBANN give proper control actions to keep the load voltages within limits. Executing the control actions of GBANN give the results of load voltages as in Fig.2. The same figure compares the resultant load voltages obtained by MPF and GBANN techniques. It is clear that error between load voltages obtained by the two techniques is acceptable. Then, the GBANN is capable of suggesting proper control action to keep voltages at load buses within limits. Reference Number: W

4 Table 1 Parameters and results of the GA Population size 50 Number of generation 3 Crossover rate 0.8 Mutation rate Initial Min. error population Max. error Generation Min. error # 1 Max. error Generation Min. error # 2 Max. error Generation Min. error # 3 Max. error B. Case Study 2 The 220 kv delta network of delta zone, Egypt is used to test the proposed approach [22]. Table 2 Test results at 100% load level Control Absolute ANN MPF Variables error T T T T Q S Q S Q S Q S Q C Q C Q G Total number of retries = 10 Number of input nodes = 18 Number of output nodes = 11 Numberof hidden layers = 1 Number of hidden nodes = 6 operating point. Each pattern contains10 input elements and 5 output parameters. Table 3 Parameters and results of the GA for case study 2 Population size 50 Number of generation 3 Crossover rate 0.8 Mutation rate Initial Min. error population Max. error Generation Min. error # 1 Max. error Generation Min. error # 2 Max. error Generation Min. error # 3 Max. error Tables 3 and 4 are similar to Tables 1 and 2, respectively but for case study 2. The ANN has been trained until the final error becomes % of maximum error with 10 iterations. The structure of the trained GBANN is as follows: 10 neurons / nodes on the input layer, 7 neurons on the hidden layer, and 5 neurons on the output layer. Executing the control actions provided in Table 4, produces the load voltages revealed in Fig.3 which is equivalent to Fig.2 but for case study 2. It is clear that the deviations in load voltages obtained by the two techniques are acceptable. The GBANN is capable of suggesting proper control action to keep voltages at load buses within limits. Table 4 Test results at 100% load level for case study 2 Control Absolute Variables ANN MPF error QG QC QG QC QG Total number of retries = 10 Number of input nodes = 10 Number of output nodes = 5 Number of hidden layers =1 Number of hidden nodes =7 Figure (2) Voltages at load buses at full load To improve the efficacy of the GBANN approach, it has been trained with different operating conditions. These different operating conditions were selected at loading level in the range from 10% to 105% with respect to the nominal Figure (3) Voltages at load buses at full load for case study 2 Reference Number: W

5 A comparison between the computation time of the proposed GBANN technique and MPF is shown in Table 5. It is notable that GBANN requires very small computation time to satisfy voltage limits. Therefore, voltage control by GBANN is faster than voltage control by MPF technique. Table 5 Computation time comparison system GBANN MPF IEEE-30 bus 0.13 sec 6 sec Delta zone 15 bus 0.11 sec 3 sec VI. CONCLUSION A new technique for voltage and reactive power control based on genetic-ann is proposed in this paper. The training data is obtained by adjusting power flow program at different operating conditions. The results obtained clearly show that, the ANN approach is capable of alleviating voltage violations in power system. Furthermore, a comparison with MPF technique shows the clear superiority of the proposed ANN to achieve the proper control action in a shorter computational time. Therefore, it is possible to use the ANN technique on real time and implement it on a real power system. VII. REFERENCES [1] Gerald, B. Sheble, L., IEEE Tutorial Course, Reactive Power Basics, Problems and Solutions, 1987 [2] Kishore, A., Hill. E. F., Static Optimization Power Sources by use of sensitivity Parameters. IEEE Trans. On PAS, Vol. PAS90, 1971, pp [3] Tripathi, K. Martinez, C.A., Nirenlerg, S.A. Reactive Switching Simulation in Security Analysis at Florida Power and light System Control Center, IEEE Trans. On Power Apparatus and Systems. Vol.104, 1985, pp [4] Qiu, J., Shahidehpour, S.M. A New Approach for Minimizing Power Losses and Improving Voltage Profile, IEEE Trans. on Power Systems, Vol.1, PWRS- 2, No.2, [5] Hobson, E., Network Constrained Reactive Power Control Using Linear Programming, IEEE Trans. on Power Apparatus and Systems, Vol.1, PAS-99, No.3, 1980, pp [6] Liu, C.C, Tomsvick, K. An Expert System assisting Decision-Making of Reactive Power/Voltage Control, IEEE Trans. on Power Systems, Vol.1, No.3, [7] Abdul-Rahman, K.H., Shahidehpour,S.M. A Fuzzy- Based Optimal Reactive Power Control, IEEE Trans on Power Systems, Vol.8, No.2, May [8] Ching T.S., Chien, T.L. A New Fuzzy Control Approach to Voltage Profile Enhancement for Power Systems, IEEE Trans. on Power Systems, Vol.11, No.3, August [9] Yuan, Y.H., Kun, L.H., Chih. C.L., Lai, T.S., Chen, K.K., Change, B.S. Voltage Control Using a Combined Integer Programming and Rule-Based Approach, IEEE Trans. on Power Systems, [10] El-Sharkawi, M.A. Marks II, R.J., Weerasoritya, S. Neural Networks and their Application to Power Engineering, in control Dynamic System, Advances in theory and Applications, Vol.41, Part ¼ edited by C.T. Leondes, Academic press, San Diego, CA, [11] Hagan, M.H., Menhaj, M.B. Training Feedforward Network with the Marquardt Algorithm, IEEE Trans. on Neural Networks, Vol.5, No.6, Novamber [12] Demuch, H., Beale M. Neural Network Toolbox Manual for MATLAB, User s Guide, Version 6, [13] S. K. Chang and V. Brandwajn, Solving the Adjusted Interactions in Fast Decoupled Load Flow, IEEE Trans. On PWRS, pp , May [14] K. Chang and V. Brandwajn, Adjusted solutions in Fast Decoupled Load Flow, IEEE Trans. on PWRS, pp , May [15] G. A. Maria, A. H. Yuen, and J. A. Findlay, Control Variable Adjustment in Load Flow, IEEE Trans. On PWRS, pp , Aug [16] R. N. Allan and C. Arruda, LTC Transformers and MVAR Violations in the Fast Decoupled Load Flow, IEEE Trans. on PAS, PAS- 101, pp , Sept [17] Q. Zhou, J. Davidson and A. A. Fouad, Application of Artificial Neural Networks in Power System Security and Vulnerability Assessment, IEEE transactions on Power Systems, Vol. 9, No. 1, pp , February [18] M.Tarafdar Haque, and A. M. Kastiban, "Application of Neural Networks in power Systems; A Review", Proceedings Of World Academy Of Sience, Engneering And Tecnology, Volume 6 June 2005 Issn [19] M. Yoshimi, K. S. Swarup and Y. Izui, Optimal Economy Power Dispatch using Genetic Algorithms, Second International Forum on Applications of Neural Networks to Power Systems,Yokohama, Japan, April [20] S. J. Huang and C. Lien Huang, Application of Genetic- Based Neural Networks to Thermal Unit Commitment, IEEE transactions on Power Systems, Vol. 12, No. 2, PP , May [21] J. hertz, A. Krogh and R. G. Palmer, Introduction to the Theory of Neural Computations, Lecture notes Vol. I, Santo Fe Institute, [22] S. Kandil, A. Elmitwally, and G. Elnaggar, "New Expert System for Fault Diagnosis in Middle Delta Electricity Zone, Egypt," 11 th Middle East Power Engineering Conference MEPCON2006, Elmenya University, Egypt, Dec , Reference Number: W