A novel deep learning approach for regional high-resolution spatio-temporal wind speed forecasting for energy applications

Abstract

Accurate spatio-temporal wind speed forecasting is crucial for optimizing wind energy production. Traditional forecasting relies on numerical weather prediction (NWP) models, which are computationally intensive, especially when implemented on large high-resolution grids. Recently, Deep Learning (DL) has emerged as an efficient alternative, utilizing historical data to learn patterns and predict future conditions. This work develops a regional DL-based forecasting system that reduces the computational burden of physical models, by using a long-term reanalysis dataset for the Arabian Peninsula (AP). The system forecasts hourly wind speed at 5 km spatial resolution up to 48 h ahead. We focus on vertical levels, corresponding to the hub heights of wind turbines for energy production. We explore two approaches: recursive forecasting, which advances the system’s state at a fine scale over time, and downscaling, which refines coarse-resolution forecasts into high-resolution counterparts. Furthermore, we propose merging both approaches by combining the propagation of spatio-temporal dynamics at fine-scale with coarse-scale predictions. The performance of the frameworks was evaluated qualitatively and quantitatively. Results show that the recursive approach accumulates errors over time steps, whereas the downscaling approach effectively generates high-resolution forecasts. Combining both approaches resulted in a more robust framework, demonstrating notably improved performance and stabilized error evolution.