Physics-motivated low-rank transformer for wind power forecasting

Wind power forecasting is essential for the safe, stable, and economic operation of power systems. Although deep learning models have shown strong capability in capturing spatiotemporal dependencies, they still face high computational cost in long-sequence forecasting and insufficient use of physically meaningful variable relationships. To address these issues, this paper proposes a physics-motivated low-rank Transformer framework for wind power forecasting. The method treats wind speed as the core endogenous variable and organizes other meteorological and temporal features as exogenous variables, so that physical knowledge is used to guide the input organization of the data-driven model. Meanwhile, a Nyström-based low-rank attention implementation is adopted in the Transformer sequence modeling process to reduce the computational burden of long-horizon forecasting. Experimental results show that, on the 24 h forecasting task of the Hunan wind farm dataset, using wind speed as the core endogenous variable reduces MAE and RMSE by up to 21.3% and 20.2%, respectively. On the public SDWPF dataset, compared with a representative Transformer-based baseline, the proposed method reduces MAE and RMSE by 52.7% and 43.3%, while reducing epoch time and peak GPU memory consumption by approximately 34.3% and 41.6%, respectively.