Ultra-short-term Wind power prediction algorithm based on bidirectional neural controlled differential equations

Against the backdrop of the continuous growth of the new energy electricity trading market, improving the accuracy of ultra-short-term electricity forecasting and reducing lag are crucial for new energy enterprises . This article proposes a ultra-short-term wind power prediction model (Bi-NDCE-UPF) based on bidirectional neural control differential equations to explore ways to improve prediction accuracy and lag. It has two innovative points: 1. A bidirectional neural controllable ordinary differential model for ultra-short-term power forecasting has been proposed. Compared with Bi-GRU and Bi-LSTM, this model has significantly improved the accuracy and delay of the third point in ultra-short-term forecasting. 2. A prediction delay mitigation structure has been designed to effectively alleviate the lag and distortion of prediction data. This algorithm has been validated in four wind farms in central China and has unique advantages. We use four metrics to evaluate all models: MSE, MAE, Dynamic Time Warping (DTW), and Time Distortion Index (TDI). Compared with Bi-GRU, Bi-LSTM, and CNN-LSTM, the text model has significantly improved in terms of MSE and DTW. Compared with DLlinear and PatchTST models, the DTW and TDI models in this paper have better advantages. • Proposed a wind power ultra-short-term prediction algorithm based on bidirectional controllable neural differential equations, which was first applied to wind power prediction. • Analyze the third point lag issue in ultra short term wind power forecasting and design a time delay structure to enhance model attention. • The Hubei Four Wind Field Experiment shows that Bi-NDCE-UPF outperforms models such as Bi-GRU in terms of MSE and MAE, and significantly improves time delay. • When predicting item 16th point, this algorithm has advantages over cutting-edge algorithms such as DLinear in TDI and DTW, and is expected to become one of the industry’s basic algorithms.