A transformer-based wind power forecasting model with multi-scale power difference embedding

Abstract Effective engineering management of wind power projects requires balancing cost-efficiency with sustainable development through comprehensive resource optimization. While wind power forecasting (WPF) improves operational stability and economic performance via generation prediction, current methods face persistent challenges in responding to short-term power fluctuations and capturing long-term weather patterns. This study presents a novel multiscale prediction model combining power difference (PD) features with a hybrid Long short-term memory (LSTM)-Transformer architecture. The proposed approach employs differential operations to extract physically meaningful dynamic features, followed by a cascaded deep learning framework: bidirectional LSTM processes hourly variations, Transformer encoders learn inter-day dependencies, and positional encoding maintains temporal coherence. Experimental results show significant improvements over baseline models, with 25.81% mean squared error (MAE), 21.05% root mean squared error (RMSE), and 23.44% R-square (R 2 ) gains versus standalone LSTM, and 28.13% MAE, 26.83% RMSE, and 33.90% R 2 enhancements compared to pure Transformer. The integration of physical mechanisms with multiscale temporal modelling demonstrates superior prediction accuracy and system robustness for renewable energy applications.