Fluctuation pattern recognition based ultra-short-term wind power probabilistic forecasting method

Abstract

-Probabilistic wind power forecasting includes more detailed information than deterministic forecasting, which can provide reliable guidance for the optimal decisions of power system scheduling operation. However, there are certain laws in the magnitude and direction of the forecasting errors corresponding to different power series fluctuations, which leads to different predictability and forecasting accuracy of different power fluctuation patterns. As most studies still focused on the model algorithm improvement and pay less attention to the law of power data itself, this paper proposes a novel probabilistic forecasting method based on the swinging door algorithm (SDA), fuzzy c means (FCM) clustering method, long short-term memory (LSTM) neural network, and nonparametric kernel density estimation (KDE), considering the correlation between wind power fluctuation patterns and forecasting errors. SDA and FCM are used to assign appropriate pattern labels to the power fluctuations, and then LSTM and KDE are used to introduce pattern recognition results in probabilistic forecasting models, excavating the inherent law of the data for classification modeling. Simulation shows that the proposed model can adapt to different error distribution patterns, and the models introduced fluctuation pattern recognition can improve the skill score of probabilistic forecasting by 36.50% on average than those without pattern recognition.