Research On Short-term Wind Speed Forecasting Based On CNN And Improved BP Neural Network

Short-term wind speed prediction is an effective foundation for power grid system dispatching and maintenance.High prediction short-term wind speed forecasting directly affects various indicators of wind power station.So,it is necessary to study wind speed prediction.Back-Propagation Neural Network(BPNN),as a feed-forward network with error correction,is good at dealing with nonlinear problems with small sample datasets.The BP algorithm used by Convolutional Neural Network(CNN)for weight update is the most profound expansion and application of Gradient Decrease(GD).In this study,a short-term wind speed prediction model,which combined with data preprocessing,is built based on CNN and improved BPNN.The historical wind speed data are taken as the basis to capture its internal variation rules,so as to improve the prediction accuracy of short-term wind speed.Firstly,a hybrid data preprocessing method combining adaptive singular spectrum analysis and empirical wavelet transform(EWT)is proposed in this study.The details are adaptive singular spectrum analysis is used to reduce the noise of wind speed data to improve the signal-to-noise ratio.Then,the wind speed series after de-noising is decomposed into multiple time series components by EWT,and Pearson coefficients among the components are calculated.Based on this,relevant components are combined to obtain the wind speed sub-series,which are adopted as the data basis of the subsequent modules.A hybrid network based on CNN and improved BPNN is built as the prediction module.CNN is introduced as the bottom feature extractor,and its convolution operation is utilized to extract the feature of wind speed subseries,so that the feature of the original sequence is enhanced.And the idea of pre-training is introduced to improve BPNN.The details are the initial weight and bias of BPNN are optimized by modified whale optimization algorithm(MWOA),and a group of initial parameter values close to the optimal values will be obtained.At this point,the process of BPNN pre-training is completed.Then,the parameters of BPNN are fine-tuned through supervised learning.Finally,the wind speed is predicted by the trained network.To address the problems of slow convergence and easy to embed local optimum of the standard whale optimization algorithm(WOA),apply logistic mapping to generate the initial population instead of random generation and introduce differential evolution operator to increase the particle diversity in global search stage,which effectively improves the problem of falling into local optimum and accelerates the convergence rate.Finally,the performance of the Modified WOA(MWOA)is analyzed by compared with different algorithm.And through realizing a variety of different prediction models,the superiority of the ensemble prediction model is analyzed and verified according to the accuracy,stability,difference,average running time etc.performance indicators.