WindFLOW: A Wind Foundation Model for Generative Temporal Super-Resolution

Wind power plants are increasingly expected to provide grid services such as operating reserves, ramping support, and headroom-based dispatch, all of which depend critically on accurate wind variability at minute-level time scales. However, most available wind datasets and operational forecasts are limited to coarser resolutions, creating a fundamental data gap for grid service quantification. To address this challenge, this paper introduces a novel temporal wind super-resolution framework based on time series foundation models, named WindFLOW. We leverage the pretrained TimesFM Transformer with a three-layer fully-connected neural network head to downscale 5-minute wind speed into 1-minute sequences. The WindFLOW method is evaluated using over 58 million minutes of wind observations from seven wind sites. Results demonstrate that WindFLOW consistently outperforms state-of-the-art deep learning models and classical interpolation methods across all sites, with a 0.39 m/s average mean absolute error. The WindFLOW model achieves superior pointwise accuracy while uniquely preserving empirical distributions and high-frequency spectral energy. These findings indicate that time series foundation models can reliably reconstruct minute-scale wind dynamics from coarse inputs and enable scalable high-resolution wind data generation to enhance grid reliability through grid services.