DeepWind: A foundation model for zero-shot wind power forecasting

Wind power forecasting suffers from poor cross-regional generalization, as existing models require site-specific retraining and fail to capture the non-stationary, multi-regime nature of wind dynamics. To address these challenges, we propose DeepWind, a domain foundation model pre-trained on a newly curated corpus of approximately 560 billion time-series data points from diverse wind sites across different climates and terrains. DeepWind employs a decoder-only Transformer with a Regime-Aware Mixture-of-Experts (MoE) framework, enabling automatic specialization across distinct physical operating states (idle, ramping, and saturation) without explicit supervision. Decoupled Time and Variate Attention mechanisms further enable effective integration of heterogeneous meteorological covariates for improved zero-shot generalization. To benchmark generalization rigorously, we introduce WindBench, an eight-dataset evaluation suite covering diverse climates, resolutions, and operational scales, ensuring no overlap between training and evaluation data to guarantee unbiased zero-shot assessment. In zero-shot evaluation across WindBench, DeepWind-Base reduces nMAE by 20.86% on average over the best competing foundation model and outperforms fully supervised baselines across all forecasting horizons (1–12 h), while few-shot adaptation of DeepWind-Small with less than 3% of parameters updated achieves an average of 5.0% improvement in accuracy over full-shot baselines. The code and model weights are publicly available at https://github.com/Hexian-2001/DeepWind.