Study On Forecasting Methods For Short-term Wind Speed In Wind Farms Based On Support Vector Machine

Timely and accurate prediction of wind speed can effectively reduce the negative effects onwind power on the entire grid. In this thesis, on the basis of the research to the wind speed onthe wind farm in recent years, the dates which record on the wind farm of north east is used toanalysis the characteristics of wind from January2006to December2007.The main contents areas follows:Firstly, considering the difficulty to determine the parameters of Least Squares support vectormachine(LS-SVM), three-tier inferred using the Bayesian evidence is proposed to optimize theLS-SVM parameters: the regularization parameter and kernel function, and the short term windspeed on wind farm is predicted based on this method. Comparing to the method of Grid searchto optimize the parameters of LS-SVM, the results show that LS-SVM which based on Bayesianframework can predict more accuracy than that of Grid search to optimize the parameters.Secondly, according to the characteristics of wind speed sequence, the method of usingwavelet transform which can smooth the processing of wind speed dates is proposed, the variouscomponents obtained are predicted by Bayesian framework based on the LS-SVM modeling,then each component of the prediction results can be wavelet reconstructed so as to getpredictive value of the true wind speed. Experimental results show that LS-SVM predictionmodel which under the framework of wavelet and Bayesian compare to other forecasting modelshas better prediction.Finally, In order to reduce the impact of single larger prediction error, a confidence intervaljudgment method in a certain confidence level can be proposed in this thesis. The wind speedforecasting range under certain confidence level is obtained by calculating, the predictive valueof wind speed which within an acceptable range can be directly used for the wind farm nextdecision-making and analysis, unacceptable wind speed forecast value wouldn’t be not used, inorder to reduce the too large a single error which can bring the decision-making risk.