Research On Deep-learning Based Short-term Wind Power Forecasting Method

Wind power forecasting is predicting the wind power output based on its patterns getting from the analysis of historical data.High-accuracy short-term wind power forecasting is not only the data basis for grid scheduling decisions,but also a technical supporting mean for complementary coordinated control on hybrid energy systems containing wind energy.It is also one of the key technologies for large-scale wind power integrating into electrical grid.Wind power generation is of high randomness and low predictability,coupled with the characteristics of multiple features and unstable.Therefore,improving the accuracy of wind power forecasting methods has been a hot and difficult research topic for long.Three short-term wind power forecasting models based on deep-learning method with different improvement schemes are presented in the article,then their performance is tested on wind power dataset with data from a certain wind turbine in Bengaluru from January 2007 to July 2021.The main content is as follows:Aiming to deal the nonlinear and unstable characteristics of wind power data,a short-term wind power forecasting model based on decomposition is proposed,which combines variational model decomposition and generative adversarial networks.Variational mode decomposition is introduced to disperse the nonlinearity in the data and convert the forecasting task on complex sequences into several prediction tasks on simple sequences;Generative adversarial network is constructed to make the best use of the ability of fitting nonlinearity;A activation function and a loss function are designed to improve the instability of the traditional generative adversarial network model.According to the experimental results of verifying its predictive performance,the model has achieved better prediction results,the mean square error decreased by 65.1%,79.65%,and 46.54% compared to VMD-ARIMA,LSTM,and VMDLSTM,respectively.At the same time,the parameters of the activation function are analyzed.In response to the problem of multiple features and difficulty in feature selection,a shortterm wind power prediction model is constructed by combining attention mechanism and generative adversarial network.Attention mechanism is introduced to simultaneously select features.An Encoder-Decoder model is used to complete the features fusion,and WGAN is constructed to replace the original generative adversarial network to fit the nonlinear relationships.It is showed in the experiment of the predictive performance that the mean square error decreased by 90.46%,83.71%,and 66.37% compared to ARIMA,LSTM,and WGANgp,respectively.It indicates that the model can adaptively complete feature fusion,selection,and fitting the nonlinearity in wind power data.A short-term wind power forecasting model is generated by integrating the two models mentioned before and applying improved sparrow search algorithm.The model makes use of the improvement methods of the two models mentioned above,variational model decomposition is introduced to handle the instability and nonlinearity,attention mechanism is utilizing to achieve adaptive feature selection,Encoder-Decoder model is constructed to complete feature fusion,and WGAN is applied to solve the problem of model collapse.The result of experiment shows that the model can integrate the advantages of the two models.Its predicted mean square error decreased by 78.76% and 73.46% compared to two models designed before respectively,and the prediction accuracy is improved.