Research On Wind Power Forecasting Method Based On Long Short-Term Memory Networks And Nonparametric Kernel Density Estimation

With the economic development of countries around the world,the demand for energy and electricity continues to expand.In order to alleviate the energy crisis and the shortage of electricity,the new energy power has gained the attention of all countries in the world in recent years,and has carried out various researches on the optimal control of new energy power.In China,the new energy power represented by wind power generation has received strong support from relevant government departments of China and the State Grid Corporation,and has achieved unprecedented development.Due to the randomness,volatility and intermittent nature of wind energy,not only the wind power injected into the grid on a large scale has a serious impact on the safety,stability and economic operation of the entire regional power grid,but also causes low utilization of wind energy.In order to ensure the good operation level of the power grid,improve the utilization of wind energy,improve the optimal allocation of energy and other issues,which promote the effective prediction of wind power has great research significance.According to the analysis of the wind power historical data from a wind turbine located in the wind farm,wind speed is the key factor affecting wind power.Based on the wind speedpower relationship,this paper studies the abnormal data processing method for data samples.An abnormal data processing method based on the density clustering method with noise and 3σ technology is proposed,which makes the wind speed-power in the historical data set more satisfying the reasonable fan curve.In view of the differences of wind power impact factors,this paper proposes a wind power spot prediction model based on K-Means clustering and LSTM networks.This model can make the relationship between wind power and various impact factors more detailed,thereby weakening the model’s constraints imposed by other sample data on model training.Since there is a certain error in the wind power spot prediction,in order to provide more information for the grid scheduling,energy optimization configuration and other aspects,the wind power probabilistic interval prediction has emerged.In this paper,a proportional wind power prediction error data generation method is proposed to ensure a reasonable matching between the prediction method and the error source.At the same time,a non-parametric kernel density estimation method with window width optimization is proposed,which can give the fluctuation range of wind power prediction error at different confidence levels,so that the corresponding wind power prediction interval can be obtained.The experimental data show that the model with the optimal window width optimization has better performance than the Gaussian distribution model and the random window width optimization model.