Research On Short Term Wind Power Forecasting Based On Improved LSTM

As a kind of renewable energy,wind energy plays an important role in China.However,due to many natural factors,like wind speed,wind direction,temperature,humidity and pressure,wind power output inevitably has randomness and volatility,which affects the smooth operation of the power grid.The accuracy of short-term wind power forecasting affects the grid scheduling,wind farm generation planning and reserve capacity planning.So it is vital to improve the accuracy of short-term wind power forecasting.This thesis compares the forecasting results of 4 single wind power forecasting models when forecasting one day’s wind power through real case,and concludes that long-short term memory neural network(LSTM)has the best forecasting effect.In order to achieve accurate prediction,this thesis takes LSTM as the basic model,and adopts two improved methods of wind power prediction.Aiming at the problem that the wind power predicted by power data-driven model lags behind,a wind power prediction method based on CVMD-SE-MCC-LSTM is adopted.Numerical weather prediction(NWP)data is considered as the input of the model.The input features is selected by analyzing the correlation.Different meteorological factors are introduced as the input of the prediction model in spring,summer&autumn and winter.In addition,by improving the variational mode decomposition(VMD)method to decompose the historical power data,the problem of power data volatility and random component interference prediction is improved,and the adaptive decomposition is realized,and the prediction cost brought by the decomposition algorithm is reduced through the sample entropy.Finally,using the improved cost function to train LSTM prediction model can effectively improve the performance of wind power prediction model.The effectiveness of CVMD and MCC is verified by ablation experiments using the actual data of an electric field in Liaoning Province.Aiming at the optimization of power time series feature extraction,a wind power prediction method based on MFE-CNN-LSTM is proposed.This method first extracts 11 statistical features such as trend factor,sequence non correlation and Skewness from NWP data,and clusters the original data with the extracted basic features and statistical features,and establishes prediction models according to the categories,so as to improve the adaptability of the prediction model.In addition,the network architecture of LSTM is improved.Through the feature extraction ability of CNN and the nonlinear sequence prediction ability of LSTM,the historical information of wind power and NWP data are fully mined.Finally,the effectiveness of the proposed short-term wind power prediction method is verified by multi-feature extraction ablation experiment and CNN ablation experiment using wind farm data in Xinjiang province.In addition,the comparison experiment shows that the RMSE of CVMD-SE-MCC-LSTM short-term wind power prediction model is 849.68 kW less than that of EMD-LSTM.Compared with CVMD-SE-MCC-LSTM model,MFE-CNN-LSTM model can reduce 3.44% MSE value of one day power when wind power fluctuation is large.