Physics-Informed Machine Learning for Wind Forecasting using WRF

Physics-informed machine learning (PIML) integrates physical knowledge into data-driven models to prevent them from producing unrealistic results. Various techniques exist to perform such integration of physics laws. We applied these methods to the domain of wind forecasting, incorporating physical constraints, tuning wind components, and deriving important features. The results improved when these elements were tuned together carefully. Our baseline model was enriched with physical information, with the goal of achieving better mean absolute wind direction metrics compared to the baseline. At a specific location and for a 6-hour forecast horizon, the average wind direction error was reduced by up to 65° in particular cases of high wind speeds, and by 11.9° on average compared to the weather model inputs alone. Compared to the baseline data-driven solution, the average improvement was 9.7°.