Short-term wind power prediction based on multiscale numerical simulation coupled with deep learning

Accurate wind power prediction is crucial for optimizing wind resource use and maintaining grid stability due to the uncertainty of wind energy. Currently, multiscale numerical simulation combining Weather Research and Forecasting (WRF) with Computational Fluid Dynamics (CFD) can achieve satisfactory prediction accuracy, but high computational costs and long processing times limit its real-time application. To address these challenges, this study presents an improved wind power prediction method that integrates multiscale numerical simulation coupled with deep learning to enhance both prediction accuracy and efficiency. This approach employs WRF for meteorological prediction, followed by a deep learning-based error correction strategy to refine wind speed predictions. Subsequently, the corrected data is input in to a CFD power database which is constructed by simulating the full-direction wind field. The proposed deep learning strategy preprocesses the data using Random Forest, Variational Mode Decomposition, and Principal Component Analysis, and corrects the data by combining Temporal Convolutional Networks, Bidirectional Long Short-Term Memory networks, Squeeze-and-Excitation Networks and Global Attention Mechanisms. The proposed model demonstrates excellent performance in short-term wind power prediction for an actual wind farm located in complex terrain. Compared to the reference model with the lowest mean absolute percentage error, the wind speed prediction error is reduced by 2.28 %, and the wind power prediction error is reduced by 2.62 %.