Power System Short-Term Load Forecasting Based On Support Vector Machine

Power system load forecast, which is the base of power market techniquesupporting system, is an essential part of power system operation, control andplanning.Support Vector Machines (SVM) is a new machine learning technology,which is based on Statistical Learning Theory. It can solve small-samplelearning problems better by using Experiential Risk Minimization in place ofStructural Risk Minimization. Because it has quite perfect theoretical propertiesand good learning performance, SVM theory becomes the new research hotspotafter the research of Artificial Neural Net.On the basis of analyzing the parameter performance of support vectormachine, a weight immune support vector machine method for short-term loadforecasting is presented in this paper, in which different sample is given adifferent weight and the parameters in SVM method are optimized by immunealgorithm. Through the simulation of interaction between antigens andantibodies the immune algorithm, which is designed according to mechanism ofthe immune systems of human and other mammals, can effectively surmount thepremature convergence and promote the diversity of colony. The simulationresults from short-term load forecasting example of actual power network showthat the presented weight immune SVM method can offer more accurateforecasting result than SVM method.A machine learning method so-called Wavelet kernel Support VectorMachine is given. We all know that mother wavelet function can create waveletframe, and the kernel function with wavelet frame can build a group of basis insquare integral space only by translation.A short-term load forecasting method uses the hybrid model of wavelettransform and support machine is given. At first, based on waveletmulti-resolution analysis the load series is decomposed into the series withdifferent frequency characteristics, then according to the features of thedecomposed components different SVMs are constructed to forecast thecomponents, finally the forecasted signals of the components are reconstructedto obtain the ultimate forecasting result. Simulation results show that the givenforecasting method is effective.Based on having successfully developed an electric forecasting software, amethod about how to develop a short-term load forecasting software is given.