A low wind output events prediction method considering power balance over a forecast horizon of up to 10 days during evening peak hours

Low wind output events (LWOEs) of wind power cluster pose a growing threat to the power supply capability as the penetration of wind power increasing, particularly during evening peak hours. Accurate prediction of LWOEs during this period is necessary. Current research predominantly focuses on the prediction of wind power continuous sequences, while lacking studies on LWOEs prediction. Besides, the existing approaches fail to consider the power deficit when identifying whether LWOE is occurred. Aiming at the above problems, a LWOEs prediction method considering power balance over a forecast horizon of up to 10 days during evening peak hours is proposed in this paper. Firstly, the daily load proportion during evening peak hours is defined based on the time-varying characteristics of load, and a LWOEs identification method considering the power balance relationship is constructed. Then, the daily load proportion and Numerical Weather Prediction results are both taken as inputs to establish the LWOEs prediction model. Besides, an adaptive weighted network that considering the contribution differences of meteorological data from each station and the dynamic relationship of source-demand is built to extract key factors from complex inputs, providing critical information for the prediction model. Focal Loss is employed to solve the problem of unbalanced sample distribution and accurately predict whether LWOE is occurred during evening peak hours in the next 10 days. Operation data of wind farms in two provinces of China is taken for the case study, and five basic models are used to verify the effectiveness and robustness of the proposed method. The results show that the proposed method has better performance under different conditions. Compared with traditional methods, the power prediction accuracy can be improved by 4.10 % and 5.50 % on average, respectively, when RMSE is used as the evaluation index; the event prediction accuracy can be improved by 12.31 % and 25.01 % on average, respectively, when F1-score is used as the evaluation index.