Physics-Guided Deep Learning for Short-Term Offshore Wind Power Forecasting
Abstract
Given wind energy’s growing significance in the world’s energy structure, the demand for high-precision forecasting is more urgent than ever. However, wind power’s inherent non-stationarity is linked to complex and variable meteorological conditions, which pose significant challenges for accurate forecasting. The accuracy of short-term wind power forecasts hinges on estimated future wind speed. Systematic biases often degrade the accuracy of Weather Research and Forecasting (WRF) forecasts. A hybrid LSTM–LightGBM correction model is proposed to correct the WRF wind speed bias. The wind correction model significantly reduces the systematic wind speed bias in the WRF model and achieves a notable reduction in RMSE across the entire wind speed range, with the highest RMSE decreasing by 7%. A new physics-guided and deep learning-integrated model is designed for 24 h short-term wind power forecasting, which integrates physical laws with data-driven features, effectively alleviating the “black-box” problem of purely data-driven models and making prediction results more consistent with engineering practice. The model’s predicted wind power has an MAE of 103.94 kW and an R2 of 0.9853, demonstrating superior prediction capability. The results provide reliable technical support for the real-time scheduling of wind farms and support the reliable, stable operation of the power system.