Research On Short-term Wind Power Prediction Based On Deep Learning Methodology

Wind power occupies a very important position in renewable energy power generation.However,due to the intermittent,random and fluctuating nature of wind power,the connection to grid of large-scale wind power poses a great challenge to the safe and stable operation of the power system.Wind power prediction(WPP)is extremely important.With the development of artificial intelligence and big data of electric power in recent years,it has become a new research hotspot to combine wide-area spatiotemporal big data with deep learning technology to predict wind power.On the basis of summarizing the current status of wind power prediction research,the problems and challenges faced by wind power prediction are clarified in the thesis.In view of the deficiencies of the existing research,the combined model based on convolutional neural network(CNN),recurrent neural network(RNN)and stacked denoising autoencoder(SDAE),and the ensemble learning technologies are utilized to predict wind power.And the prediction results are analyzed.The research work is shown as follows:(1)A short-term WPP model based on wavelet decomposition and convolutional neural network is proposed.Firstly,the principles of CNN and wavelet decomposition(WD)are introduced,and the WPP modeling process based on wavelet decomposition and convolutional neural network is analyzed.Then,the influence of different network parameters of CNN on the prediction effect is explored through an example,and the influence of different number of input sequence lags and different WD decomposition levels on the prediction performance of WDCNN is analyzed.Finally,the validity of WD and CNN is verified through actual examples,and the proposed model is compared with traditional prediction models.(2)A short-term WPP model based on variational mode decomposition and recurrent neural network is proposed.Firstly,the principles of RNN and variational mode decomposition(VMD)are introduced,and a short-term WPP model based on VMD-RNN is constructed.Then,the influence of different input models on the prediction performance of RNN is analyzed through an example,and the influence of the number of different decomposition modes of VMD on the prediction performance of VMD-RNN is explored.Finally,the validity of VMD and RNN is verified through actual examples,and the proposed model is compared with traditional prediction models.(3)A short-term WPP model based on SDAE,support vector machine and bat optimization algorithm is proposed.Firstly,the principle of SDAE is introduced.And the short-term WPP modeling process based on SDAE,support vector machine(SVM)and bat optimization algorithm(BOA)is analyzed.A short-term WPP model based on SDAE-SVM-BOA is constructed.Then,the prediction performance of SDAE-SVM-BOA is analyzed through an example and compared with that of the traditional prediction model.Finally,the effectiveness of the BOA is demonstrated through a practical example.(4)Based on modern signal processing technology and stacking technology,two novel ensemble deep learning prediction models are proposed.On the one hand,a deep learning ensemble model based on WD and VMD is proposed.Firstly,the multi-model ensemble strategy based on WD and VMD is introduced,and the architecture of the core prediction model based on convolutional neural network and gated recurrent unit is explained.Then,the constructed deep learning ensemble prediction model based on modern signal processing technology is explained.Finally,the effectiveness of the proposed model is illustrated by numerical examples.On the other hand,a deep learning ensemble model based on stacking technology is proposed,which integrates the combined deep learning models proposed in the previous chapters and the constructed deep residual network.Firstly,the principles of stacking technology and deep residual network are introduced.Then,the effectiveness of the stacking ensemble model and the constructed deep residual network are demonstrated through the examples.