MATLAB/GUI Simulation Tool for Power System Fault Analysis with Neural Network Fault Classifier

Transcription

1 MATLAB/GUI Simulation Tool for Power System Fault Analysis with Neural Network Fault Classifier Ph Chitaranjan Sharma, Ishaan Pandiya, Dipak Swargari, Kusum Dangi * Department of Electrical Engineering, Malaviya National Institute of Technology, Jaipur Abstract This paper proposes a MATLAB based Graphical User Interface (GUI) tool which can serves as a user friendly visual tool for power system fault analysis. This GUI calculates fault level voltages and currents for all the different types of faults and displays them along with their waveforms accordingly. The GUI will serve as an educational tool to help the students understand the intricacies involved in fault analysis. Different Artificial Neural Network (ANN) architectures are proposed for designing an appropriate classifier for classification of the different types of faults that occurs in a real system. Finally, in order to test the classifier in test system conditions it is integrated with the developed GUI. Introduction A huge amount of capital investment is made at present to generate electric power and transmit it over a long distance to consumers in a stable, viable and optimal fashion. To attain stable supply of electrical power, the power system must be reliable. It is also important to run the power system at high or peak efficiencies and protect it from unavoidable accidents or faults. Faults usually occur due to insulation failure, flashover, physical damage or human error. It can be broadly classified as symmetrical fault and unsymmetrical fault. Power system fault analysis provides the necessary information for the proper selection of switchgear, setting of relay and stability of system operation. The problem consists of determining bus voltages, line current fault level MVA during various types of faults. The three phase balanced fault is used to select and set phase relays, while line-to-ground fault is used for ground relays. Fault study thus helps in improving the stability and reliability of the whole system. The proposed GUI tool will serves as a fast and accurate visual tool for fault analysis. It calculates the various fault levels and presents them in the GUI screen along with the waveforms of the current and voltages. Authors in [1,4] have also presented different GUI tools for fault analysis but they seem to lack in one or the other important aspects such as the line diagram of the system in consideration, the waveforms etc. This paper rather presents a complete tool for the analysis of fault. Power system transmission line fault identification is very important to ensure quality performance of the power system. Since the restoration of power requires extremely quick judgement, it is important to classify the fault in a very short period of

2 time. Several techniques have been implemented for analysis of power system faults. Conventional approaches have some difficulties in achieving the desired speed, selectivity and accuracy. Neural networks posses powerful characteristics such as fast learning, fault tolerance and ability to produce correct output when fed with partial input. Hence it can be used for fault classification at high speed and accuracy. The classification has been done using Multi Layer Perceptron (MLP) and Probabilistic Neural Network (PNN) architectures. A comparative study has been made between the two architectures to determine the better classifier. After the selection of the better classifier has been made, it is integrated with the developed GUI for further testing. Here GUI is used as a simulator which acts a test system and provides required input to the ANN classifier. The various voltages and currents at the different buses during a fault condition are generated by the GUI and are given to the ANN classifier. The classifier classifies the type of fault that has occurred and displays on in the GUI screen. GUI tool for power system fault analysis A MATLAB GUI tool has been developed using the Graphical User Interface Development Environment (GUIDE) toolbox to calculate the short-circuit fault currents and fault voltages in power transmission lines. Presented program provides the calculation of three-phase, single line-to-ground, line-to-line, and double line-to-ground faults on transmission lines. Balanced three phase fault are analysed using a single phase equivalent circuit and unbalanced faults are analysed using the symmetrical components method. The process of creation of a GUI using GUIDE consists of two steps: creation of a layout and programming of the GUI. Firstly the type of components needed, interaction required and the technique that is to be used is determined. Then using the Layout Editor the components such as tables, text box, axes, push buttons, panels etc. are added to the GUI. After laying out work is done the GUI is saved which creates a corresponding.fig file and an.m file. Then the GUI is programmed by adding the corresponding callbacks to each component. The flowchart below gives the stepwise explanation of the programming procedure: Read positive /negative and zero sequence impedances Read Base MVA and the Faulted Bus No. Calculation of various Fault Level Currents and Voltages Determine the fault Currents and Voltages and plot their corresponding waveforms Display the Voltages and Currents in the corresponding boxes Figure 1: Flowchart showing the procedure for working of GUI. Firstly the positive/negative and zero sequence impedances are taken in the form of two tables namely Zdata1 and Zdata0. This includes all the impedances of the generators, transformers and the lines connecting the different buses. Then

3 using this data the corresponding impedance matrix is formed. After this the faulted bus number and the base MVA is also entered in the corresponding text boxes. Then the callbacks corresponding to each of the pushbuttons which represent the different types of the faults are programmed. Each of these callback functions comprises of codes and functions for the calculation of fault levels corresponding to particular the fault type. The inputs to these functions are the bus impedance matrices. Line-to-ground fault and the double line to ground fault requires the positive, negative and zero sequence bus impedance matrices. While the double-line fault function requires the positive and negative bus impedance matrices only. A single click to any of these pushbuttons will execute its corresponding callbacks which calculate various fault level currents and voltages. The per unit values of the fault voltages and currents and the fault level MVA are displayed. The GUI also determines the signals of Currents and Voltages and plots their corresponding waveforms in the axes. The line diagram of the system is also displayed which indicates the faulted bus. A sample test system [5] consisting of 11 buses and 3 generating units is taken to check the validity of the developed GUI tool Figure 2: Single line diagram of the test system. The developed GUI tool showing various results is given below Figure 3: GUI showing a Single Line to Ground fault at Bus no. 8

4 A Generalized GUI The developed GUI tool is so programmed that any new system can be loaded for which the analysis of faults can be performed. Loading of a new system to the GUI can be done in two ways. Firstly By directly changing the system impedance values in the tables (zdata1 and zdata0) of the INPUT panel. This procedure is limited to systems containing up to 40 buses and 60 lines. Moreover the picture of the line diagram of the system cannot be loaded using this procedure. When it comes to a larger system containing a large number of system components, direct data entry becomes practically unfeasible. So an attempt has been made to enable the loading files containing large system data. This can be done by clicking the pushbutton in the OPEN Panel. The data of a new system can be entered as three different files. The first two files (.m file format) contains the impedance data of the positive/negative and zero sequence. And the third file is the picture file (preferably.jpg file) of the line diagram of the new system. Clicking the push button in the open panel triggers the following processes: >> Enter the name of the file containing Zdata1 data in single quotes -> 'zdata1.m' >> Enter the name of the file containing Zdata0 data in single quotes ->'zdata0.m' >> Enter the name of the picture file (single line diagram) in single quotes ->'newsyspic.jpg' Power system fault classification using neural networks The main aim of power system fault classification is to classify the fault into one of the types according the current and voltage patterns. It is one of the key components in digital protection systems. Elements of inputs represent measurements of features selected to be useful for distinguishing between classes. Recent studies show that neural network based classifiers offer many advantages over their conventional counterparts. The design of neural network based fault classifiers for transmission lines involves four basic tasks: (i) collecting or producing sets of sample of faulted voltage and current waveforms (ii) pre-processing the data and extracting useful information (iii) choosing and training the most appropriate neural network (iv) testing of the trained neural network. In this paper MLP and PNN architectures are used to design the classifier. Design of MLP based Fault Classifier For performing fault classification for a particular system, the first task involves the creation of the network. Six input vectors consisting of the voltages and currents of all the phases are chosen. A target vector defining the type of the fault is also taken. The following are the inputs and targets of the neural network classifier. Inputs: 3 voltages of phases a, b and c. 3 currents of all phases a, b, c. Target: Type of the faults characterised by numbers from 1 to 10.

5 V a V b V c I a I b I c ANN Classifier SLG DL DLG 3LG Figure 4: Fault type detection using ANN classifier. A system with 11 buses which was considered earlier (Figure 2) is taken. A data table consisting of all the input vectors and targets for all the types of the fault in all the buses of the system are created. It consists of 110 samples. The data table is then normalised. An adequate normalization, not only for the network output variables but also for the input ones prior to the training process is very important to obtain good result and reduce significantly the calculation time. Training algorithm generally works best when the network inputs and targets are scaled, inputs and outputs are normalized into a specific range. Normalization is done in a range from 0 to 1.The formulae used for the normalization process is given by: Then the normalized data table is used in creating the feed forward network (MLP Neural Network). The network is then trained using 60-80% of the data samples. After the network is trained, it is being tested for the full data set that is with all the 110 samples. The performance of the network is then measured. The performances of the classifiers are gauged in terms of measures likeclassification accuracy, execution time and misclassification rate. The following plot shows the samples that are correctly classified or misclassified. The black circles represent the target values and the red dot represents the output of the MLP classifier. Misclassifi cation Number of Testing samples Figure 5: Performance plot for testing of MLP classifier. In the table shown below, parameters liked training algorithm, number of training samples etc. are varied in order to improve the performance of the classifier. It is observed that with the increase in the number of training samples the error reduces. The performance of Table 1: Performance of MLP classifier with 6 features.

6 the classifier is also affected by the number of hidden layer and the number of neurons in the hidden layer. As seen from the results tabulated above, for the MLP classifier the mean square error reduces as the number of neurons in the hidden layers is increased. It is also observed that the classifier gives approximately same results with two training algorithms takentrainlm and trainscg. Design of PNN Classifier The same procedure that has been adopted to create the MLP classifier is used in the creation of the PNN classifier. The data table created for the test system taken before is used to create the PNN classifier. Then the networked is trained for some percentage of the input samples (60-70%). The trained network is then tested for the whole 110 data samples. It is observed that the number of misclassification is zero i.e. the error in the classification is zero. The performance of the PNN classifier is plotted in figure 6. Number of testing samples Figure 6: Performance Plot for testing of the PNN classifier. Comparison of MLP and PNN Classifier The performances of the MLP and PNN based classifiers are compared for the determination of a better classifier. Table 6 given next page shows the comparison between the two types of Neural Network Classifiers. From the comparison table we can conclude that PNN classifier gives a better result giving zero misclassifications. The average Table 2: Performance of PNN classifier with 6 features. Table 2 shows the performance of the PNN classifier for the purpose of power system fault classification. The various performance parameters are calculated for different sets of architectures, No. of training/testing samples, error in classification and the elapsed time are noted for all the cases. training time is also lesser in case of PNN classifier. Hence, PNN architecture is more appropriate for designing the fault classifier. This classifier is further used to integrate with the GUI tool for further testing.

7 MLP Classifier PNN Classifier the fault that is created by the GUI. Average Error in Training = 2.9% Error in Testing = 3.8 % Average Training time = 2.7 sec Average No. of Misclassifications = 4 Average Error in training = 0% Average Error in Testing = 0% Average Training time = sec Average No. of Misclassifications= 0 Table 3: Comparison between MLP and PNN classifier. GUI (Simulates the condition of a real system) Provides the classifier with voltages and currents ANN Classifier (Identifies the type of fault) Takes its Input from GUI INTEGRATION OF ANN CLASSIFIER & GUI Integration of GUI Tool with ANN classifier An attempt has been made to integrate the ANN classifier in the previously developed GUI tool. This is to test the working of the classifier in test system conditions. Here the GUI acts as a simulator to simulate the conditions that prevail in a test system. Different types of fault can be created by the GUI tool. The GUI will provide the input of the ANN classifier i.e. the voltages and currents during the occurrence of the fault. Then the classifier can classify Figure 7: Integration of ANN classifier and GUI tool. For this integration a new panel called ANN CLASSIFIER OUTPUT. In this panel two edit box are added to display the type of the fault and the time taken for the classification. The input provided by the GUI is taken by the classifier. The classifier then normalises its input, loads its pre-trained network and then identifies the fault type that is occurring in the GUI. The time take in this process is also calculated. Thus by integrating the two we can Figure 8: Integration of GUI and ANN classifier.

8 test the performance of the classifier and the time taken for the classification. The new GUI with the ANN classifier is shown in the figure below. Conclusion The developed GUI tool can be used as an educational tool for power systems. Students tend to have difficulties in understanding the concept of faults and method of calculations. Using visual tools to calculate and present the variations, makes the concept of fault more understandable and lasting to the students. With GUI, anyone without the knowledge of programming can start applying this application software to solve the problem more efficiently. The GUI environment keeps most of the tedious and repetitive calculations in the background, allowing the user to spend more time in the analysis of the results obtained. ANN Fault Classifiers can be used to accurately find out the type of fault that occurs in a real system within a very short period of time. The comparison between MLP and PNN based classifier shows that PNN classifier is faster and more accurate. The integration of GUI and ANN classifier was tested on a computer having an Intel Core 2 Duo processor of 1.66 GHz and the average classification time was about 1.9 ms, which was quite fast. More components in the developed GUI tool can be added to make it a more powerful tool. For example the waveforms during the transient conditions can be imparted so as to enable a user to study the stability of the system. Digital meters can be employed to record the voltages and currents in the system. Using this data, the ANN classifier can be used with relays to design a high speed protective relay that can be used to support the conventional protective relaying system. The developed classifier can be integrated with a fault locator which can be designed by using pattern recognition in neural networks so that it can work as an independent relaying system which can detect, locate and classify the fault and give commands to clear out the fault. References [1] Savas Koç, Zafer Aydoğmus, A Matlab/GUI Based Fault Simulation Tool for Power System Education Mathematical and Computational application, Vol.14 no. 3 pp [2] Y. Fukuyama, Y.Ueki Fault analysis system using neural networks and artificial intelligence Fuji Electric Corporate Research and Development, Ltd. No. 1, Fujimachi Hino-city, Tokyo 191 Japan, /93, IEEE [3] Paul M. Anderson, Analysis of Faulted Power Systems IEEE Press Power System Engineering Series, New York, 1995 [4] M G Rabbanil MATLAB Based Fault Analysis Toolbox for Electrical Power System 4 th International Conference on Electrical and Computer Engineering ICECE 2006, December [5] Hadi Saadat, Power System Analysis, Tata McGraw-Hill 2002.