Physics-constrained spatiotemporal transformer network: A hybrid paradigm for meteorologically consistent wind speed prediction

Accurate wind speed prediction plays a pivotal role in optimizing wind energy utilization and ensuring stable power grid operations. Despite advancements in data-driven techniques, many existing models generate predictions that deviate from established physical laws, limiting their real-world applicability. To bridge this gap, our study introduces a dual-methodological framework grounded in physics-aware machine learning. To address this, our work introduces two innovations: (a) A physics-informed hybrid architecture combining ConvNeXt-based spatiotemporal attention with a pyramidal transformer, designed to extract features while adhering to fundamental principles such as energy conservation, thermodynamic equations, and buoyancy effects. (b) A constrained optimization framework that enhances predictive accuracy while enforcing meteorological consistency. Evaluations across four seasonal datasets show superior performance over existing methods, with notable gains in both single and multi-step forecasts. For example, in dataset 1, our model attains a mean absolute percentage error of 0.204 (14.3% lower than the next-best approach) and a coefficient of determination (R2) of 0.913 (3.6% higher). Physical constraints further bolster reliability, reducing errors by up to 19.6% (dataset 3) and ensuring compliance with atmospheric physics. Interpretability studies reveal that the model prioritizes physically meaningful variables: 80 m/10 m wind speed ratios dominate predictions, while temperature and humidity contributions vary seasonally.