Physics-guided neural network integrating uncertainty evolution for spatiotemporal wind speed prediction

Accurate short-term wind forecasts are important for wind power operation, grid scheduling, and intelligent energy control, but wind speed changes are difficult to predict because they vary across space, time, and height. Here we show that a physics-guided neural network improves the prediction of three-dimensional wind speed fields by learning both data patterns and the physical consistency of wind evolution. The model combines information across time, space, and height, and uses physical constraints to guide the learning of wind speed, wind direction, and spatial structure. Tests show that the model predicts wind speed more accurately than current machine-learning methods within a twenty-four-hour forecast period at both ten-metre and one-hundred-metre heights. Compared with a leading numerical weather forecasting system, it gives more accurate wind speed forecasts within five hours and more accurate wind direction forecasts within twelve hours. The model also exhibits strong regional generalization and temporal robustness, providing support for practical wind power applications. Short-term forecasts of wind speed and direction across heights and regions are more accurate than those from existing data-driven models and a leading weather forecasting system, according to a physics-guided neural network for three-dimensional wind field prediction.