Short-term Prediction Study Of Wind Power Based On Hybrid Neural Network And Multiple Signal Decomposition

Wind energy is green energy and renewable energy.The development of wind energy not only meets the increasing demand for energy but also solves the problems of climate pollution,ecological destruction and energy exhaustion caused by the use of traditional energy.Since wind power generation is unstable,random and volatile,large-scale grid-connected operation will destroy the safety,reliability,stability and economy of power grid.In order to reduce the adverse impact of wind power non-stationarity on the operation of power system,the wind power forecasting technology is adopted to predict the wind power output in a period of time in the future and help the power grid dispatch.Accurate prediction can significantly enhance the security,stability,economy and controllability of power system.Due to the non-stationarity of wind power signal and the instability of prediction model,the prediction accuracy of ultra-short-term wind power is still difficult to meet the needs of wind farm and power grid real-time scheduling.In order to improve the prediction accuracy of ultra-short-term wind power,the prediction of ultra-short-term wind power is improved from two aspects of wind power signal decomposition processing and deep prediction model.The main research contents and innovations are as follows:In order to reduce the influence of complex nonlinear signals of wind power on the prediction results,the wind speed and wind power signals are processed by signal decomposition algorithm.In order to avoid the influence of mode mixing caused by EMD and the influence of white noise added by EMD,the complex unsteady wind speed and wind power signals were decomposed into a series of sub-sequences with different frequencies by the improved total EMD.Sample entropy is used to quantify subsequence complexity.In order to reduce the influence of the high frequency subsequence generated by the improved EMD on the prediction results,the wavelet packet decomposition is used to decompose the high frequency subsequence into several subsequences.The experimental results show that under the same prediction model,compared with the ensemble empirical mode decomposition,using the improved complete overall empirical mode decomposition reduces the MAE by about 15.85%,the MSE by about 8.96%,and the RMSE by about 8.94%,the training time increased the 1199.491 s.The improved complete overall empirical mode decomposition can improve the model prediction accuracy.Using the wavelet packet to decompose the high-frequency subsequences twice,under the same prediction model,the MAE is reduced by about 50.80%,the MSE is reduced by about45.74%,and the RMSE is reduced by about 45.68%,the training time increased the1633.599 s.In order to fully extract implicit characteristics of wind power data,using convolutional neural network and gated cyclic unit were used to build the prediction model,and the prediction was based on wind speed sub-series,wind direction sub-series and wind power sub-series.The convolutional neural network was used to mine the coupling characteristics among wind speed sub-series,wind direction sub-series and wind power sub-series to improve the prediction accuracy of the model.The experimental results show that under the same data,the CNN-GRU model is compared with the prediction results of the CNN model and the BP model,the CNN-GRU model are better than other models under the same data,MAE,MSE,and RMSE are reduced by 35.82%,34.02% and 34.06%,and the training time increased the 510.703 s and 902.401 s,respectively,compared with traditional neural networks.By comparing the prediction error and training time of model with different convolution layers,the longest time is 1365.624 s when five-layer convolution is adopted.Compared with other network structures unchanged,the model prediction error and training time of different convolution layers are used.When five-layer convolution is used,the longest time is 1876.327 s.Five-layer convolution and two-layer convolution,three-layer convolution,Compared with the prediction results of the four-layer convolution,the MAE is decreased by about 64.97%,35.36% and 50.54%,the MSE is decreased by about 61.71%,26.13% and44.21%,and the RMSE is decreased by about 61.69%,26.08% and 44.23%,the training time increased the 1198.202s、922.745 s and 861.973 s.This study shows that the ICEEMDAN-SE-WPD-CNN-GRU model can predict ultra-short-term wind power with high accuracy,and the related research results can provide a reference for the wind power signal preprocessing method and the deep learning model architecture for wind power prediction.