Physics-informed spatio-temporal network with trainable adaptive feature selection for short-term wind speed prediction

Wind speed prediction (WSP) is essential for optimizing wind power generation, enhancing turbine performance and ensuring grid stability. Wind speed prediction models face challenges in effectively integrating complex spatio-temporal data with physical principles. This limitation reduces their accuracy and reliability for optimizing wind power generation and ensuring grid stability. To solve the issue, this study proposes a physics-informed spatio-temporal network (PISTNet) for short-term WSP that effectively integrates spatio-temporal data with physical principles. The proposed model designs a dynamic feature adapter (DFA) module that dynamically emphasizes relevant temporal and spatial information through adaptive masking mechanisms. It also incorporates an extended advection equation-based physical modeling (EPM) module, which provides physics-based WSP fused with neural network-extracted features using a feature fusion module (FFM). An adaptive physics penalty loss (APPL) function is proposed to enhance model’s accuracy to selectively enforce physical constraints based on significant prediction deviations. Comprehensive experiments conducted on four diverse datasets from Hamburg, Herning, Palmerston North, and Silkeborg demonstrate that the proposed model consistently outperforms six state-of-the-art prediction methods across multiple evaluation metrics, achieving up to a 7.1% improvement in RMSE, 8.0% in MAE, 2.3% in $1/R^2$ score, and 9. 5% in SMAPE compared to the closest competitors. The findings highlight the potential of combining data-driven and physics-informed approaches to achieve accurate and reliable WSP, thereby contributing to the advancement of wind energy systems and grid management.