THE PURPOSE of dynamic security assessment (DSA)

Transcription

1 942 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 8, NO. 4, JULY 1997 Dynamic Security Contingency Screening and Ranking Using Neural Networks Yakout Mansour, Fellow, IEEE, Ebrahim Vaahedi, Senior Member, IEEE, and Mohammed A. El-Sharkawi, Fellow, IEEE Abstract This paper summarizes B.C. Hydro s experience in applying neural networks to dynamic security contingency screening and ranking. The idea is to use the information on the prevailing operating condition and directly provide contingency screening and ranking using a trained neural network. To train the two neural networks for the large scale systems of B.C. Hydro and Hydro Quebec, in total 1691 detailed transient stability simulation were conducted, 1158 for B.C. Hydro system and 533 for the Hydro Quebec system. The simulation program was equipped with the energy margin calculation module (second kick) to measure the energy margin in each run. The first set of results showed poor performance for the neural networks in assessing the dynamic security. However a number of corrective measures improved the results significantly. These corrective measures included: 1) the effectiveness of output; 2) the number of outputs; 3) the type of features (static versus dynamic); 4) the number of features; 5) system partitioning; and 6) the ratio of training samples to features. The final results obtained using the large scale systems of B.C. Hydro and Hydro Quebec demonstrates a good potential for neural network in dynamic security assessment contingency screening and ranking. Index Terms Contingency, dynamic security, neural network, ranking, screening. I. INTRODUCTION THE PURPOSE of dynamic security assessment (DSA) is to determine which contingencies may cause power system limit violations or system instability. The ultimate objective in this process is to derive operating guidelines for defining the areas of secure operation. In simple terms a DSA process includes the following steps: selection of critical contingencies; performing a detailed stability analysis for each critical contingency; analysis and checking of results for violations; repeat steps 2) and 3) until a barely secure case is obtained. In the past two decades a number of researchers have focused on enhancing step 2) which is the analysis part [9]. The subject of this project is step 1) which is intended to reduce the number of contingencies for which detailed stability analysis has to be performed. In other words one can view step 1) as the screening and ranking stage of the DSA. Historically, the dynamic security contingency screening and ranking methods unlike their static security counter parts Manuscript received December 12, 1996, revised March 5, Y. Mansour and E. Vaahedi are with with B.C. Hydro, Burnaby, Canada. M. A. El-Sharkawi is with the Department of Electrical Engineering, University of Washington, Seattle, WA USA Publisher Item Identifier S (97) have not received a great deal of attention. With the recent progress in analytical techniques [6] and major advancements in computer hardware, contingency screening for DSA seems to be more feasible and justified. So far, the only methods suggested for dynamic security CS&R are the analytical methods using the transient energy function (TEF) [6] and the use of artificial intelligence including expert systems and neural networks [4]. B.C. Hydro recently completed a CEA project with the objective of exploring the application of neural network to online dynamic security assessment [2]. In that project, various areas of dynamic security assessment were analyzed and potential areas for neural-network application were identified and ranked. In total, nine applications were identified with four being ranked highly. The report recommended that the applications ranked highly should be considered for implementation. This paper summarizes B.C. Hydro s experience [1], [2], [3] in applying neural networks to dynamic security contingency screening and ranking. The idea is to use the information on the prevailing operating condition and directly provide contingency screening and ranking using a trained neural network. Obviously the preference in this work was only to use features that can be easily measured or calculated from measurements such as steady state pre- and postcontingency features and avoid dynamic ones which could require time consuming simulations. The idea behind the application proposed here is similar to that EPRI project [4] except that the implementation including the neural-network design, the choice of features and outputs are considerably different. The final product presented here serves as the Beta test version of B.C. Hydro s DSA contingency screening and ranking module. II. STUDY SYSTEMS The systems of B.C. Hydro and Hydro Quebec were used to test the neural-network applications. These systems are described below. A. B.C. Hydro System The bulk system is composed of two main generation subsystems, the Peace River and the Columbia River. The Peace River system has a generating capacity of 3400 MW the majority of which is transmitted over about 1000 km of 500- kv transmission system to the main load center. The Columbia system has a generating capacity of 4730 MW located between 200 to 500 km from the main load center. The relatively large transfers over such long distances characterizes the B.C. Hydro /97$ IEEE

2 MANSOUR et al.: DYNAMIC SECURITY CONTINGENCY SCREENING 943 system as one of the systems limited by transient stability. In this system, generation shedding has proved to be one of the most effective discrete supplementary controls for maintaining stability. For these studies the B.C. Hydro base case used had 1393 buses and 209 machines. B. Hydro Quebec System The Hydro Quebec system is composed of two main generation complexes of Churchill Falls and James Bay in the north connected to the main load center through 735 kv heavily series compensated bulk transmission. In dynamic studies, the system is divided into two corridors, namely Eastern corridor and the Western corridor. The base case used for these studies, which was obtained from Hydro Quebec s System Planning Department, had 963 buses, 87 machines, and four two-terminal dc lines. The original base case obtained was extremely stable and it had to be heavily stressed to obtain more responsive dynamic behavior. III. NEURAL-NETWORK DESIGN The NN used in these studies was of the feedforward structure. It consisted of three layers, one for the input, one hidden, and one for the output. The size of the input layer was determined by the size of the input pattern. For an input pattern of, the number of input neurons was plus one neuron for biasing. Similarly, the size of the output layer was determined by the number of outputs each output neuron was assigned to one security index. The size of the hidden neurons was selected to provide the best test results for the given system. It was found that four six hidden neurons were adequate for all the studied cases. The input neurons of the feedforward structure are potential nodes. They pass the information to the neurons in the hidden layer in a weighted form. The neurons in the hidden layer process this information with a nonlinear function. The hidden neurons pass the information to the output neurons which are linear current nodes. The data of each study were normalized and shuffled a few hundred times before training the NN. This was done to enhance the randomness of the data and to eliminate the sequential bias effect. After the data were shuffled, they were divided into two groups, one was used for training and the other for testing. For proper NN performance, test data and training data must be different although belonging to the same statistical source. The property of being able to correctly classify previously unseen inputs is referred to as generalization. During the training, the NN was closely monitored to prevent network memorization (over training), or saturation. The presence of any of these problems would have resulted in unreliable results. When a NN memorizes the training data, it reproduces acceptable results for patterns that have used during the training, but unacceptable results with high errors when tested on unseen patterns. Such a NN is useless for the security applications but may find applications in areas where memorization is the goal. There are different techniques to ensure that an NN has learned and not memorized. They are all based on the fact that a properly trained NN should respond with equal error measures to both training and testing patterns. The training patterns in this study were first divided into two subsets where one subset was almost four times the second. After the first few iterations, the NN was trained on the larger subset and tested on the smaller subset. If the errors of the larger subset and the smaller subset were comparable, the training patterns were reshuffled and divided again into two subsets and so on. The error function typically takes the form shown in Fig. 1. The training error decreases along with the number of iterations while the testing error decreases at first, bounces around, and then starts to increase. The optimal learning and generalization are achieved close to the global minimum of the testing error indicated by the arrow. This technique has been used throughout this work with successful results. Another problem with NN training is network saturation. This problem can have a disabling effect if the nonlinear functions of the hidden neurons reach the upper or lower saturation limits. In this case any wide change in the input would produce no or minimal change in the output, and the neurons in this case are paralyzed. It is not uncommon to have some neurons in the saturation region, but too many would render the NN ineffective. To prevent network saturation, the hidden neurons were closely monitored. When one third of the neurons reach the saturation limit and stay there, the neurons were randomly disturbed and the learning process continued. IV. FEATURE SELECTION Features are system attributes which contain information that can be used to reconstruct the internal structure and pattern within the neural network. The task of deciding which system quantities to choose as features is a difficult task and there are formal feature selection algorithms which with some degree of success can be used for this purpose. In dynamic security assessment applications, it was shown [2] that for proper reconstruction of the system model, three types of information are required: 1) initial conditions; 2) precontingency network connectivity; and 3) contingency. Based on this argument and practical experience in developing offline dynamic security look-up tables, the key system attributes given in Table I were suggested to be used as features for dynamic security assessment [2]. This table includes two type of features, namely static and dynamic features. Static features are those attributes representing steady state preand postcontingency conditions. Dynamic features are those which reflect dynamic system conditions following the contingency. The above features were also augmented with the aggregate dynamic features suggested by ABB in their EPRI work [4]. The original list of features for the B.C. Hydro system contained 57 elements compared to 61 for that of the Hydro Quebec system. As studies progressed, the feature sets were refined to maximize the sample to feature ratio while main-

3 944 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 8, NO. 4, JULY 1997 Fig. 1. Variation of error with the number of iterations. TABLE I KEY FEATURES FOR DYNAMIC SECURITY ASSESSMENT In these studies both static and dynamic features were used to assess their benefits. The objective was to select features which require less computation and provide accurate security classification when using the neural network. Specific programs were written to extract the selected features from the cases. Fortran and PTI IPLAN programs were written for interaction with the power flow. User defined dynamic models were developed in Fortran to monitor dynamic features during the simulation. The features for each case were combined into one file for each run. These individual files were then formatted and merged together based on study, region, or contingency. The features in these merged files were then normalized for the use by the neural network. taining accurate neural-network results. This was achieved by eliminating those features not considered important for the considered contingency. V. IMPLEMENTATION The goal of this application was to be able to classify and rank different cases based on their dynamic security severity. It is crucial to find an output which reflects the ranking properly. In this application, the stability energy margin was selected as an output. This output reflects exactly the stability status of a case. This is one aspect of the work conducted here which sets it apart from previous efforts [2]. For comparison purposes, the largest angle swing was selected as another output. The largest angle swing, which was one of the outputs in the EPRI work [2], corresponds to the largest generator angle shift in the system.

4 MANSOUR et al.: DYNAMIC SECURITY CONTINGENCY SCREENING 945 To implement this application, detailed stability studies had to be conducted for different cases. For each case, the energy margin, the maximum swing angle, and all other system features had to be calculated. These values were then passed to the neural network for training and testing. To measure the energy margin for time domain simulations, B.C. Hydro s on-line DSA energy margin calculation module was used. VI. SAMPLE DEVELOPMENT To develop different samples, different base cases were first developed by changing network topology taking one component out of service at a time. New sets of base cases were then created by modifying the generation pattern of these cases. This allowed generation of more than 1000 base cases which then could be used with different contingencies to create more 1000 sample stability cases. The load level and generation patterns vary between cases. All major transmission lines remain in service in these base cases. Series compensation on these major lines was removed in a selected pattern to increase the stress on the system. For the B.C. Hydro System, initially 36 base cases were created for this CEA project. Modifying the generation pattern of these initial cases allowed another 72 cases to be created for a total of 108 base cases. Twenty-five B.C. Hydro system contingencies were selected for use in these studies. These contingencies were distributed on a regional basis and when combined with the base cases produced a total of 1158 simulations for the B.C. Hydro studies. For the Hydro Quebec system, the stressing of the base case was dependent on the corridor being studied. For the Western corridor, base cases were created by placing parallel 735 kv transmission lines out of service. As well, series compensation on some remaining in-service lines was removed. Generation patterns were modified to obtain different transfer levels. Through these various modifications 57 base cases were developed for Western corridor studies. The Eastern corridor proved more difficult to stress as this part of the system proved very stable transiently but would collapse due to voltage instability. Similar modifications to that of the Western corridor were performed on the Eastern corridor, however, only 32 base cases were developed. For the Western corridor a number of fault locations and contingencies were simulated. The fault locations were focused on the northern portion from Abitibi/Chibougamau to James Bay. These contingencies in combination with the base cases produced a total of 266 simulations for the Western corridor studies. The Eastern corridor fault locations were limited to the buses Arnaud, Montagnais, and Churchill Falls as the system proved highly stable at other locations. These contingencies in combination with the base cases produced a total of 267 simulations for the Eastern system. VII. RESULTS The systems of B.C. Hydro and Hydro Quebec were used to test the suitability of neural networks for contingency screening and ranking in dynamic security assessment. A. B.C. Hydro System Results In these studies, initially, the complete set of the features (57) was used for the study including all the static and dynamic features. The outputs selected were energy margin and maximum swing angle. The first study, BC1, covered contingencies from the entire B.C. Hydro system. As was explained in Section III, the data for each study were normalized, shuffled and divided into two groups; one was used for training and the other for testing. In this case, the neural network was trained using 120 samples and then was tested using 24 samples. Figs. 2 and 3 show the performance of the trained neural network for this application. Fig. 2(a) gives a comparison of the actual values of the energy margin (target) with the neural-network predicted values (NN output). As a result of the normalization, these values are between 0.0 and 1.0. Fig. 2(b) gives the probability of error distribution. The error in this figure is defined as the actual value minus the predicted neural-network value. To develop this figure, the errors for different testing samples were first sorted in ascending order from maximum negative to maximum positive. The entire range was then divided into segments of 0.05 width. The probability of error associated with the testing samples to lie within each of these segments was then calculated. Fig. 2(b) shows that the average error is 14.66% with the errors spreading between 17.9% to 44.95%. The total number of misclassifications in this case was 13 with nine false alarms and four false dismissals. The maximum swing output results for this application are given in Fig. 3(a) and (b). The average error is 4.92% with an error range of 38.74% to 78.11%. Obviously, these results were unacceptable with very high maximum absolute errors. The summary of the results are provided in Table II. To investigate the accuracy improvement if the two security indices were decoupled, two more studies of BC2 and BC3 were conducted. The results obtained indicate that the accuracy does not improve noticably. To improve the accuracy, it was decided to break the system into three subsystems: peace (PEACE), south interior (SI), and lower mainland (LM) based on the applied contingency. This was justifiable since not all the features in one subsystem provide adequate information on the security status of another. The results of the PEACE1 and SI1 studies show no improvements compared to BC1 results (Table II). To further improve the accuracy, the neural-network outputs were decoupled to produce cases PEACE2 and SI2. The results obtained, as summarized in Table II, show that the accuracy of results are still too poor with absolute maximum errors of about 40% and 35% respectively. The biggest improvement was obtained when the number of features were reduced. The results obtained, as summarized in Table II, show that the average error for energy margin reduces to 4.22% with an error range of ( 18.72% to 5.36%). The average error for maximum swing output was 1.29% with an error range of ( 4.28% to 5.82%). The final set of results for the B.C. Hydro system was obtained when the number of features for the PEACE2 and SI2 cases were reduced deleting all the dynamic features to obtain

5 946 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 8, NO. 4, JULY 1997 (a) (b) Fig. 2. (a) Comparison of target and NN output for BC1 (energy margin); no. of false alarms = 9; no. of false dismissals = 4; no. of misclassifications = 13. (b) Probability distribution of error for BC1 (energy margin); average error = ; maximum positive error = ; maximum negative error = the new PEACE4 and SI-2 cases. The new results for the PEACE4 case, as shown in Fig. 4(a) and (b) and summarized in Table II, indicate a good improvement mainly because of the reduction in the number of features. B. Test Results of Hydro Quebec System Based on the experience obtained in the B.C. Hydro studies, the Hydro Quebec system was also divided into two subsystems: west (western corridor ) and east (eastern corridor). The first study was conducted on the West system (WEST3). In this study dynamic features were included with the neural network outputs of energy margin and maximum swing angle. From the original feature set with 61 elements, only 37 features relevant to the western corridor were chosen for this study. The results obtained for this case show poor accuracy. To improve the accuracy of the results, the steps taken in B.C. Hydro s studies were repeated here. The number of features were reduced and the neural-network output was reduced to the energy margin alone. The results obtained for the west system (Western) show that the accuracy improved over the previous case (WEST3). The HQ system characteristics were such that it was almost impossible to make it unstable. The amount of the data reflecting unstable performance was very limited and could not be relied on for classification. Therefore the HQ results should be taken as the neural-network estimate of the energy margin for stable cases only. VIII. DISCUSSION OF RESULTS The studies conducted indicate a path of progressive improvement in the accuracy of neural-network results driven by the following corrective measures. The fine-tuning and intelligent selection of key features improved the accuracy significantly. Dynamic features improve the accuracy marginally. Elimination of these features results in significant computational savings.

6 MANSOUR et al.: DYNAMIC SECURITY CONTINGENCY SCREENING 947 (a) (b) Fig. 3. (a) Comparison of target and NN output for BC1 (max swing). (b) Probability distribution of error for BC1 (max swing); average error = ; maximum positive error = ; maximum negative error = TABLE II SUMMARY OF ALL NEURAL NETWORK APPLICATION STUDIES

7 948 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 8, NO. 4, JULY 1997 (a) (b) Fig. 4. (a) Comparison of target and NN output for PEACE-4 (energy margin); no. of false alarms = 1; no. of false dismissals = 0; no. of misclassifications = 1. (b) Probability distribution of error for PEACE-4 (energy margin); average error = ; maximum positive error = ; maximum negative error = A neural network with two outputs of energy margin and maximum swing angle performs marginally worse than two single-output neural networks. The quality of results for both outputs of energy margin and maximum swing can be reasonably accurate. However, unlike the energy margin, the maximum swing angle may not provide an accurate security index. The partitioning of the study systems to smaller subsystems of interest based on applied contingencies had a major impact on the accuracy of the results. This was justifiabl,e since not all the features in one subsystem provide relevant information on the security status of another subsystem. An increase in the ratio of number of training samples to the number of features enhance the accuracy of the results. This ratio for the final results reported here are around ten. Figs. 5 and 6 summarize the result of this study. Fig. 5 shows the energy margin average error and the error range for different tests conducted. Fig. 6 gives the percentage of misclassifications in each case. Both of these parameters are important in assessing the performance of the neural network in this application. Misclassification indicates the number of cases which has been misclassified whereas the error range shows the maximum prediction error of the neural network which could happen in its classification. In our study, although, the neural network was designed with the focus on misclassification, the maximum prediction errors have also been given due to their importance. The final results of B.C. Hydro indicate that with a reasonable degree of accuracy the neural network has been able to predict the system stability margin. The total number of misclassifications for the combined B.C. Hydro system is three; two false alarms and one false dismissal. The maximum error range was ( 17.84% to 22.48%). The positive error gives the maximum possible error under-estimation (false alarms) whereas the negative one gives the maximum error for overestimation (false dismissals). Since the consequence of false

8 MANSOUR et al.: DYNAMIC SECURITY CONTINGENCY SCREENING 949 Fig. 5. Average error and error range for different cases (energy margin). Fig. 6. Misclassification percentage for different cases. dismissal is very serious, to ensure 100% accuracy so far as classification of false dismissals is concerned, an error margin of larger than 17.84% must be selected and any case with less energy margin should be classified as unstable. This way, contingencies can be screened with little danger of false dismissals at the expense of increasing the number of false alarms. The results obtained for Hydro Quebec are slightly worse than those of B.C. Hydro. For the total Hydro Quebec system combining the results of the Western and the Eastern systems, the total number of misclassifications were four false alarms with an error range of ( 10.59% to 34.55%). IX. CONCLUSIONS The neural network results obtained for the two systems of B.C. Hydro and Hydro Quebec show that with a reasonable degree of accuracy, the neural network was able to perform classification and predict the system margin of stability with an average error of about 5%, an average misclassification of 4% and an average error of about 10% for the misclassified cases. It is also noted that for each study case there were very few (two or three) outliers among the tested samples which increased the total classification error to ( 20% to 20%). The reason of for the high error for the outliers could be either lack of enough training data or the performance of the used neural network. The training burden can be significantly reduced if similar to the practice of this paper, the transient stability program is equipped with an output analysis module such as second kick for early termination of simulations and margin calculations. This way each run will take a fraction (30% to 50%) of the time of a traditional simulation. REFERENCES [1] Y. Mansour et. al., Dynamic security assessment preprocessing, CEA 960 Project Final Rep., Sept [2] Y. Mansour, E. Vaahedi, A. Chang, M. A. El-Sharkawi, and S. Weerasooriya, Potential use of neural-network techniques for on-line dynamic security assessment of power systems, CEA Rep. ST-347C-P, Sept [3] Y. Mansour, E. Vaahedi, A. Chang, B. R. Corns, J. Tamby, M. A. El- Sharkawi, Large-scale dynamic security screening and ranking using neural network, Paper 96SM579-3, presented at the IEEE Summer Meet., Denver, CO, July [4] RP : Dynamic security analysis feasibility evaluation report, EPRI Rep., Apr [5] Y. Mansour, E. Vaahedi, A. Y. Chang, B. R. Corns, B. W. Garrett, K. Demaree, T. Athay, and K. Cheung, B.C. hydro s on-line transient

9 950 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 8, NO. 4, JULY 1997 stability assessment (TSA), IEEE Trans. Power Syst., vol. 10, Feb [6] G. D. Irisari, G. C. Ejebe, J. G. Waight, and W. F. Tinney, Efficient solution for equilibrium point in transient energy function analysis, IEEE Trans. Power Syst., pp , May [7] M. A. Pai, Energy Function Analysis for Power System Stability. Boston, MA: Kluwer, [8] A. A. Fouad and V. Vittal, Power System Transient Stability Analysis Using Transient Energy Function Method. Englewood Cliffs, NJ: Prentice-Hall, [9] W. W. Price, Rapid analysis of transient stability, IEEE Rep. 87TH PWR, Sept Yakout Mansour (S 77 M 77 SM 83 F 94) is presently the Manager of the Grid Operations and Inter-Utility Affairs Division of B.C. Hydro. Dr. Mansour is Chairman of the Power System Dynamic Performance Committee of IEEE. Ebrahim Vaahedi (S 78 M 79 SM 87) presently leads the Control Centre Technologies Department in the Grid Operations Division of B.C. Hydro, as well as being an Adjunct Professor at the Electrical Engineering Department of University of British Columbia, Vancouver. Mohammed A. El-Sharkawi (S 76 M 80 SM 83 F 95) received the B.Sc. degree in electrical engineering in 1971 from Cairo High Institute of Technology, Egypt, and the M.A.Sc and Ph.D. degrees in electrical engineering from the University of British Columbia, Vancouver, B.C., Canada, in 1977 and 1980, respectively. In 1980 he joined the University of Washington, Seattle, as a Faculty Member. He served as the Chairman of Graduate Studies and Research and is presently a Professor of Electrical Engineering and the Associate Chair. He co-edited an IEEE tutorial book on the applications of neural network to power systems. He organized and taught several international tutorials on intelligent systems applications, power quality and power systems, and he organized and chaired numerous panel and special sessions in IEEE and other international conferences. He published over 120 papers and book chapters in these areas and holds seven licensed patents. Dr. El-Sharkawi is the Founder of the international conference on the Application of Neural Networks to Power Systems (ANNPS) and the Cofounder of the international conference on Intelligent Systems Applications to Power (ISAP). He is a Member of the administrative committee of the IEEE Neural Networks Council representing the Power Engineering Society and Video Tutorial Chair of the IEEE Continuing Education Committee and the neural network council. He is the Founding Chairman of several IEEE task forces and working groups and subcommittees, including the task force on Application of Neural Networks to Power Systems, the working group on Advanced Control Strategies for dc-type Machines, and the task force on Intelligent Systems Application to Dynamic Security Assessment. He is the Cofounder of the IEEE Subcommittee on Intelligent Systems. He is a Member of the editorial board and Associate Editor of several journals, including the IEEE TRANSACTIONS ON NEURAL NETWORKS and Engineering Intelligent Systems.