Research On Short-term Wind Power Forecasting Based On MKL Methods

The kernel methods such as the Support Vector Machine(SVM)are one kind of the efficient methods for solving the non-linear problem,it has been successfully applied in short-term wind power prediction.Multiple Kernel Learning(MKL)as one kind of new kernel learning method combines kernel functions with different characteristics by kernel weight coefficient.Kernel weight coefficient convert the selection problem of kernel function to the distribution problem of kernel weight coefficient and it's sparsity can enhance decision function's interpretability.The reproducing kernel Hilbert space which combined by different kernel functions can make the model have stronger generalization capability and robustness.In order to further improve the performance of short-term wind power prediction model,the dissertation follow the MKL method and applie it to short-term wind power direct prediction and indirect prediction.The main content of this dissertation is as follows:(1)The dissertation analysis the fundamentals and the steps of the Empirical Mode Decomposition(EMD)and the Empirical Wavelet Transform(EWT)which are used in data's adaptively decomposition and preprocessing.The Electrocardiograph(ECG)'s standard data set is used to comparative analysis them.The experimental results show that the number of empirical mode decomposed from EWT is obvious less than EMD's and some empirical modes decomposed by EMD have the apparently mode mixing.The dissertation go into the MKL method based on semi-infinite linear programming and the MKL-wrapper algorithm as well as the MKL-chunking algorithm on the basis theory of SVM,the dissertation then elaborated the fundamentals and implementation step of Simple MKL method.(2)The dissertation analysised the characteristics of seasonal periodicity and time continuity of a large wind farm's power in in different seasons,and four week's data in four seasons which have different character is selected as test sets.The EWT-MKL method is combined the EWT method and MKL method which is realized by MKL-wrapper,MKL-chunking and Simple MKL algorithm.The dissertation applied MKL methods in the short-term wind power direct prediction in different seasons.These experiments are also implemented in the SVM method in the same condition respectively.The experimental results show that the MKL forecasting model is better than that of the SVM's,and the prediction effect of MKL combination model by EWT's preprocess is best.The MKL model's prediction accuracy with kernel parameters determination and punishment parameters is random values in the range of set value in different seasons show that MKL has stronger generalization ability and the selection of its parameters has stronger robustness.(3)The dissertation analyzed the influence of the conversion accuracy of Wind powercharacteristic curve which is bulit by different ways.The MKL method and the EWT-MKL method are applied in a wind farm's speed prediction.The short-term wind power indirect prediction's experiments are realized by the wind power characteristic curve.These experiments are also implemented in the SVM method and the Wavelet Support Vector Machines(WSVM)method in the same condition respectively,and the results show that the precision of EWT-MKL combined forecasting model is obvious overmatch other's in the short-term wind power directly prediction,and MKL forecasting model's precision is superior to SVM's precision.