A Hybrid NWP–LSTM Framework for Seasonal Wind Speed Forecasting with Multi-Resolution Downscaling and Bias Correction

Reliable seasonal forecasting of near-surface wind speeds is essential for optimizing renewable energy production, particularly in regions with expanding wind power infrastructure. Global seasonal forecast models, despite offering valuable large-scale predictability, are limited by coarse resolution (~1°), which fails to resolve local topographic, land-surface, and boundary-layer influences critical for accurate hub-height wind predictions. This study presents a high-resolution dynamical downscaling framework using the Weather Research and Forecasting (WRF) model to enhance seasonal wind speed forecasts over a target region in India. Initial intercomparison of leading global seasonal systems (ECMWF SEAS5 and NCEP CFSv2) demonstrated superior performance by ECMWF SEAS5 in reproducing observed wind climatology over the Indian subcontinent, leading to its selection as the primary driving dataset. A three-domain WRF configuration (27 km → 9 km → 3 km) was implemented, and comprehensive sensitivity experiments identified the MYNN planetary boundary layer (PBL) scheme as the optimal configuration, yielding the lowest wind speed bias and best representation of vertical wind shear.

Downscaled hindcast simulations were rigorously validated against ERA5 reanalysis across multiple vertical levels, showing substantial improvements in hub-height wind speed skill metrics. To extend forecast skill beyond the 7-month limit of available boundary conditions, a long short-term memory (LSTM) neural network was developed and trained on 40 years of ERA5 wind time series using a sliding-window approach (7-month input → 90-day output). The model was retrained for each sliding window to adapt to evolving patterns, resulting in robust predictive performance from months 8 to 10. Finally, quantile mapping bias correction was applied to the downscaled and LSTM-extended outputs compared to ERA5, resulting in an approximately 38% reduction in root mean square error and a marked improvement in probabilistic reliability. The resulting bias-corrected, high-resolution seasonal wind speed dataset provides enhanced accuracy for wind resource assessment, power production forecasting.