Research On Ultra-Short-Term Wind Power Prediction Methods

The randomness and volatility of wind power is the root cause of restricting wind power integration.Accurate ultra-short-term wind power prediction can not only assist dispatching departments in real-time adjustment of dispatching plans and allocation of reserve capacity,thus reducing operating costs,but also help to know about the changing trends of wind power and reduce the adverse effects of wind power on power grids,thus improving the security and stability of power systems and promoting the large-scale integration of wind power.The measured data of the Sotavento wind farm in Spain is used to study the ultra-short-term wind power prediction methods in this paper,and the main work is as follows:(1)A simple outlier detection method based on Isolation Forest is proposed to detect wind power abnormal values.This method can discover abnormal data explicitly,and it has high operational efficiency and adaptability.The integrated learning model Extreme Gradient Boosting(XGBoost)is used for feature selection,and it can choose more effective features as the inputs of prediction models.(2)Artificial Neural Network(ANN)and Support Vector Regression(SVR)are adopted to forecast wind power.The application of the integrated learning model XGBoost and Gradient Boosting Regression Tree(GBRT)in ultra-short-term wind power prediction is studied.These four models have been tested in Sotavento wind farm,and the results show that the prediction accuracy of XGBoost and GBRT is higher than that of ANN and SVR.After a detailed comparison of XGBoost and GBRT,it is concluded that XGBoost provides more accurate predictions in the month when the wind power fluctuation is large,and GBRT provides more accurate predictions in the month when the wind power fluctuation is small,thus the prediction accuracy of the two is equivalent.However,GBRT cannot perform parallel operations,and XGBoost supports multi-threaded parallel computing and greatly reduces training time,so XGBoost's computing efficiency is much higher than GBRT.It can be seen that the XGBoost algorithm is a superior approach for ultra-short-term wind power prediction.(3)The Tree-Structured Parzen Estimator(TPE)algorithm is adopted to optimize the hyperparameters in forecasting models,and the comparison experiments with the unadjusted forecasting models have been carried out.The results show that the models optimized by the TPE algorithm have a significant improvement in prediction accuracy compared with the models which have not been optimized by the TPE algorithm.(4)The equal weighted average combination model,the error square sum minimum combination model and the combined model based on SVR are used to forecast wind power,using the above four single prediction models.In addition,a combined forecasting model based on Feature Weighted Support Vector Regression(FWSVR)is proposed in this paper.The experimental results of actual wind farms show that compared with the other three combined forecasting models,this model has higher forecasting accuracy and a better forecasting effect.