Comparative evaluation of ECMWF and GFS for operational day-ahead wind speed forecasting

Accurate wind speed forecasting is essential for renewable energy integration. However, the relative contributions of global numerical weather prediction (NWP) selection, regional downscaling, and artificial intelligence (AI) post-processing within operational forecasting chains remain poorly quantified. This study systematically evaluates ECMWF HRES and GFS within a unified framework integrating Weather Research and Forecasting (WRF) downscaling and Pyraformer-based AI correction, using three years of observations from wind farms in northwestern China's Gobi region. Three principal findings emerge. First, ECMWF HRES consistently outperforms GFS, with average advantages of 3–4 % RMSE reduction that increase systematically with forecast lead time. Second, AI post-processing contributes approximately 20 % RMSE reduction compared to 3–4 % from NWP switching. AI effectiveness diminishes at extended lead times while NWP quality differences become more pronounced. Third, forecast activity metrics expose a critical limitation invisible to traditional error measures. AI corrections systematically suppress forecast variability by 20–30 %, degrading extreme event representation essential for operational decisions. These findings demonstrate that AI post-processing delivers five-to six-fold greater improvements than premium NWP subscriptions, reshaping resource allocation priorities. The results provide practical guidance for forecasting system design and highlight the necessity of evaluation frameworks balancing error reduction with variability preservation.