Regional wind speed prediction model based on graph attention network and meteorological coupled equations

Aiming at the challenges of spatio-temporal nonlinearity and physical consistency in regional wind speed prediction, this paper proposes graph attention network with physical constraints, a new model that fuses the graph attention network and meteorological equations. The model dynamically captures the complex relationship between meteorological stations through the graph attention mechanism and jointly optimises the loss function with the horizontal momentum equation as the physical constraint. Experiments based on high-resolution data in the Beijing-Tianjin-Hebei region from 2018-2021 show that the root mean square error and mean absolute error of graph attention network with physical constraints in 24-hour forecasts are 1.52 m/s and 1.11 m/s, respectively, which are reduced by 11.1% and 11.9% compared to the optimal baseline, and the R2 is improved to 0.948. Its excellent performance in extreme events provides a new paradigm for high-precision, interpretable weather prediction.