Static Voltage Stability Assessment Based On Support Vector Machine

Voltage collapse has become one of the most important problems which have threatened the operation safety of electric power systems. It is necessary to evaluate the distance between the operation state and the voltage critical point in order to escape from the voltage collapse.By means of calculating the critical point, the loading margin to voltage collapse can be determined. But at the critical point, the Jacobian matrix of conventional power flow equations becomes singular. The continuation power flow method of getting the critical point by tracing the PV curve has been applied to overcome this difficulty. The calculation speed of this method is slow for power systems with high dimension, so it is difficult to realize real-time voltage stability assessment. The application of a faster and more reliable evaluation technique is very important to shorten the evaluation time.Support vector machine(SVM), a method based on statistics learning theory, is a machine learning algorithm of the new era. It equals to solving a quadratic programming problem in the principle of minimum structural risk. This algorithm is featured with strong forecasting ability, global optimization and fast speed of approaching, etc. Hence a method of model construction which based on SVM is presented for the power system static voltage stability assessment, and a SVM model for voltage stability assessment is established in this dissertation. Basing on the operation state, the critical point can be estimated by the model being trained according to the test results. This method takes full advantage of SVM's ability to solve the problems with high dimension, nonlinear and small sample. Hence, with quicker assessment speed and higher forecast precision, better generalization ability is guaranteed. Compared with ANN model, it can be seen that SVM model has higher precision.In this dissertation, the features of the data set are extracted by using principle component analysis to reduce the input dimension. Then the principle component containing the information of sample data is sent to support vector machine for training. This proposal combines the feature extraction ability of principle component analysis together with the excellent nonlinear function approaching ability of support vector machine. The empirical results show that with high forecast precision, the input dimension can be reduced by this method.