Research On Probabilistic Prediction Of Wind Power Based On Deep Learning And Nonparametric Estimation

With the aggravation of energy consumption and environmental pollution,the development and utilization of wind energy has attracted the wide attention of scholars.Due to the high volatility and intermittency of wind energy,wind power generation is random,and the grid connection of wind power will bring huge potential risks to the grid.Wind power probability prediction is a key technology to reduce the difficulty of grid dispatching and increase the connection of new energy to the grid.Most traditional wind power prediction methods are deterministic point prediction methods,that is,they can only obtain the predicted value of single point power at the next moment or several moments,and point prediction can not evaluate its reliability.However,the randomness and uncertainty of wind power generation,as well as the limitations of prediction methods themselves,make wind power prediction always have errors.The deterministic point prediction method provides limited reference information for decision makers and is not enough to meet the needs of modern power system scheduling and decision making.Therefore,it is very important to study the probabilistic prediction method of wind power generation and select the appropriate evaluation index for the safety,stability and economic operation of power grid.In order to achieve high precision capture and reasonable optimization of wind power allocation,Combining maximum information coefficient(MIC)with multi-Task learning(MTL)and long short-term memory(LSTM)networks,a method for predicting ultra-short-term points of MTL-LSTM wind power is proposed.Based on the MIC correlation analysis of wind power sequence and wind speed sequence,the MIC correlation analysis of alternative sequences with wind power sequence and wind speed sequence was conducted respectively,so as to construct the characteristic input sequence of neural network.Wind power prediction was the main task,wind speed prediction was the auxiliary task,and a multi-task learning framework was adopted to build the LSTM network.Network hyperparameters are optimized based on grid search,and wind power prediction is finally realized.Then,based on the real historical data of a wind farm in the United States,by comparing the results of LSTM modeling under the framework of multi-task learning with those of single-task LSTM modeling,BP neural network modeling and traditional ARIMA modeling,it is verified that the proposed method can achieve higher accuracy in wind power point prediction through feature reconstruction,hyperparameter optimization and model building.According to the results of wind power point prediction by MTL-LSTM,this paper modeled wind power prediction error based on non-parametric kernel density estimation,and proposed a grid search-based deletion-cross validation method to find the optimal bandwidth in kernel density estimation.It is verified that the probabilistic prediction interval of wind power obtained by this method is better than the probabilistic prediction interval obtained by non-parametric kernel density estimation of theoretical optimal bandwidth,and better than the probabilistic prediction interval obtained by parametric method.