Hierarchical Multi-Scale Patch Alignment for cross-site probabilistic wind speed forecasting

Under the carbon neutrality target, wind resource assessment requires forecasting models with cross-site transferability and uncertainty quantification. However, terrain heterogeneity and mesoscale dynamical processes induce distribution shifts that limit existing models’ generalization. To address this, we propose a Multi-Scale Patch Alignment (MSPA) framework for cross-site probabilistic wind speed forecasting. Specifically, historical sequences are segmented into multi-temporal patches to capture both short-term fluctuations and long-term rhythms. During domain adaptation, scale-specific local feature alignment and temporal consistency constraints are performed to mitigate cross-domain discrepancies. A sparse attention mechanism is further introduced to adaptively select similar features for alignment, effectively suppressing negative transfer. In the inference stage, a probabilistic head using quantile regression loss precisely calibrates the coverage and sharpness of predictive intervals. Extensive experiments on onshore and offshore datasets demonstrate that MSPA consistently outperforms strong baselines in key metrics, including CRPS, Pinball Loss, and Predictive Entropy. Furthermore, the framework exhibits superior robustness during wind speed peaks, troughs, and abrupt variation periods, providing a practically deployable solution for reliable cross-domain wind speed forecasting in secure energy systems.