Research On Short-term Wind Power Prediction Based On Multiple Combined Model

Aiming at carbon neutralization to reduce carbon emission,it has become an inevitable trend to rebuild the power system by new energy sources.In recent years,the installed capacity of wind power in China continues to increase.However,the wind power is evidently random and fluctuating by the instability of wind energy,which introduces the impact to the power system by large-scale wind power grid connection.Wind power prediction could significantly improve the utilization rate of wind energy and ensure the safety,economic and stable operation of the power system to promote the reform of the energy structure.First of all,the research scheme of “Data preprocessing-Predictive model building-Example verifying” is proposed based on the overall review of the research background and current status of wind power prediction in detail.By analyzing and summarizing the method,process and error sources of wind power prediction,and the performance evaluation index of the forecasting model is studied.Secondly,the Laida criterion,wavelet transform and CEEMDAN are selected for data preprocessing to reduce the adverse effects of abnormal data and random noise on the accuracy of wind power.By eliminating abnormal data,the concentration of data distribution and the reliability of data are improved.This study takes the wind farm actual data in Hebei Province as an example to verify the preprocess effects of the three methods.Combining the BP neural network with the data preprocessing algorithms,CEEMDAN is selected for its high prediction accuracy to provide reliable sample data for the combined prediction model.Finally,the CEEMDAN-PSO-BP neural network,CEEMDAN-PSO-SVM and CEEMDAN-PSO-LSTM are constructed to solve problems of poor convergence,easy to fall into local optimum and low prediction accuracy of BP neural network.The SVM and LSTM network prediction models are individually verified,which are based on the preprocessing of CEEMDAN output data and optimizing model parameters by the PSO.In view of the characteristics of wind farm data volatility under different wind speed changes,three sets of sample data with different volatility levels are calculated to confirm the combined prediction model.By comparing the prediction accuracy of CEEMDAN-PSO-BP neural network,CEEMDAN-PSO-SVM and CEEMDAN-PSO-LSTM prediction models in the short-term wind power prediction with the actual wind farm data in Hebei Province,the results illustrate that the CEEMDAN-PSO-LSTM neural network model has more advantages in short-term wind power prediction.