Short-term Wind Speed Prediction Based On Combinatorial Models Of Machine Learning

The wind power industry is booming with the support of the government,and the development and utilization of wind power is getting more and more attention from the international community.The conversion and management of wind energy is closely related to wind speed.Unstable and uncontrollable wind speed will seriously affect wind power generation,and then affect power quality and supply and demand balance,which brings huge challenges to the dispatching of power system.For the safe and stable operation of power grid,an effective way is to carry out short-term wind speed prediction.In order to improve the accuracy of wind speed prediction,this paper starts from two aspects: mining the predictability of wind speed data and optimizing the performance of prediction model.The main work content of this paper is as follows.The data preprocessing methods and some reference models are introduced,and the evaluation indexes of wind speed prediction model performance are listed.For the residual in wind speed prediction,a short-term wind speed forecasting model is proposed based on residual,variational mode decomposition(VMD),extreme learning machine(ELM)and long short-term memory(LSTM).The VMD algorithm is used to reduce the complexity of original data.The ELM is employed as the initial prediction engine to extract the characteristics of each wind speed subsequence,and the preliminary prediction results are obtained.The LSTM network can remember the long-term periodic component and irregular trend factor,which is selected to model the residual sequences of the preliminary prediction results.Thus,the prediction error is reduced and the prediction performance is improved.For the high fluctuation characteristics of wind speed data,the Hodrick-Prescott(HP)filtering method is proposed to extract the trend sequences of the original wind speed sequences.As for the wave sequences after trends sequence extraction,the singular spectrum analysis(SSA)algorithm is utilized for decomposing its main periodic components to obtain the periodic term and noise term of the wind speed sequence.The ELM model is constructed for training and predicting.The established HP-SSA-ELM short-term wind speed prediction model effectively excavates the important features of wind speed series,and takes the advantages of good generalization performance of ELM network,so as to improve the overall wind speed prediction accuracy.For the possibility of some non-stationary subsequences in the decomposed subsequences,a short term wind speed prediction model of CEEMDAN-RLMD two-stage decomposition and bidirectional long short-term memory(Bi LSTM)is established.Firstly,complete ensemble empirical mode decomposition adaptive noise(CEEMDAN)is used to decompose the wind speed sequence into multiple intrinsic mode functions,and the complexity of the sub-sequences is detected according to the permutation entropy.Secondly,robust local mean decomposition(RLMD)is utilized for decomposing the intrinsic mode function with maximum entropy into a series of product functions.Then,the Bi LSTM prediction model is constructed for each sub-sequence.Finally,the sub-sequence prediction results are summed to obtain the final prediction values.The proposed model can improve the accuracy of wind speed prediction to some extent.At the end of the thesis,the work done in this thesis is summarized,and a reasonable and feasible scheme is put forward.