Improving wind flow simulation over complex terrain for wind resource assessment through a hybrid data-driven and physics-informed generalized k– ω framework

Abstract

Accurate prediction of wind speed and turbulence over complex terrain remains challenging because terrain-induced separation and atmospheric boundary layer turbulence are not fully represented by conventional Reynolds-averaged Navier–Stokes (RANS) closure models. This study develops Mod-GEKO-ML, a hybrid framework combining adjoint optimization and machine learning (ML) calibration of spatially varying generalized k–ω (GEKO) free coefficients with physics-informed modification of selected traditional closure constants. The GEKO-ML mapping is trained on the H42 bump, tested on H38 and H31 for in-family transfer, and applied without retraining to Askervein Hill and Bolund Island. The turbulent kinetic energy (TKE) objective best balances mean flow and turbulence improvement, while reducing the input from eight features to five preserves predictive capability. Field scale evaluations demonstrate partial- and region-dependent transferability, with improved wind speed prediction in terrain affected regions and consistent TKE error reduction. Component analysis shows that ML calibration mainly improves local terrain-induced mean flow responses, whereas closure modification improves upstream TKE preservation and contributes more strongly to TKE error reduction. Feature space analysis shows that H42 training support varies across terrain regions and does not directly determine ML improvement. At Askervein Hill, Mod-GEKO-ML reduces lee side root mean square error (RMSE) by 13.7% for wind speed and 44.8% for TKE relative to standard GEKO. At Bolund Island, TKE RMSE decreases in all three terrain regions under both inflow directions. This framework provides a practical RANS-based approach for wind flow and turbulence prediction over complex terrain.