Physics-Constrained Neural Networks Enhance Short-Term Weather Forecasting Accuracy

This research focuses on improving short-term weather forecasting through advancements in physics-constrained neural networks (PCNNs). The study introduces several innovations to the WeatherGFT architecture, aiming for greater accuracy and stability in hybrid forecasting models. Key enhancements include an upgraded numerical solver that incorporates a fifth-order weighted essentially non-oscillatory scheme (WENO-5), a beta-plane approximation, and subgrid-scale viscosity. This allows for a fourfold increase in the integration time step to 1200 seconds and a reduction in daily mean squared error by up to 26%. Additionally, a unified autoregressive hybrid block replaces multiple specialized modules, which helps prevent overfitting to specific lead times. The physical core of the model is also integrated with two state-of-the-art neural backbones, resulting in new models named PI-PredFormer and PI-IAM4VP. Evaluations on the WeatherBench South Pacific dataset show that these hybrid models reduce root mean squared error by 8-22% for 1-12 hour lead times compared to purely neural models, while also maintaining better physical consistency. These findings suggest that refining hybrid components is a practical path to more accurate and efficient short-range weather forecasting.