A terrain-stratified comparison of regional MPAS-A (v8.3.1) and nested WRF (v4.7.1) for short-term near-surface wind forecasting

Abstract. Accurate deterministic forecasts of 10 m wind over complex terrain remain challenging, and the relative performance of nested regional and variable-resolution modelling strategies is still insufficiently documented. This study presents a controlled comparison of five deterministic forecast products over Türkiye: four nested WRF configurations at 12 km parent and 4 km child domain resolution with one-way and two-way nesting, and regional MPAS-Atmosphere with an approximately 4 km refined inner region. All experiments use identical 0.25° GFS forcing, are initialised daily at 00:00 UTC, and integrated for 48 h. After excluding the first 6 h as spin-up, verification is performed over lead times 6–47 h using 10 m wind speed observations from 509 TSMS stations during January and June 2023. Terrain effects are explicitly addressed through a Terrain Ruggedness Index computed from SRTM GL3 elevation data and a three-class terrain stratification. Model performance is assessed using mean error, mean absolute error, root mean square error, and Pearson correlation coefficient.

MPAS achieves the lowest error across all aggregate metrics, with ME of 0.76 m s−1 and RMSE of 2.16 m s−1, compared to 0.99 m s−1 and 2.27 m s−1 for the best WRF configuration. All models exhibit systematic positive bias. The performance gap between MPAS and WRF is strongly terrain-dependent: negligible in low-complexity terrain but reaching 41 % ME reduction at high-complexity sites. Increasing WRF resolution from 12 km to 4 km does not consistently reduce bias, with the coarser 12 km two-way configuration outperforming the finer 4 km two-way configuration across both months and all terrain classes, pointing to grey-zone boundary layer limitations and lateral boundary constraints. Two-way nesting provides meaningful improvement at 12 km but offers diminishing returns at 4 km. MPAS’s aggregate ME advantage partly reflects a compensating bias structure: reduced overprediction at calm and moderate wind speeds is offset by stronger underprediction at high wind speeds, reaching ME of −3.28 m s−1 at high-complexity sites. This wind speed dependence also explains the spatial reversal along the Aegean coast, where WRF 4 km configuration outperforms MPAS at 70 % of stations. These results indicate that neither modelling approach offers a universal advantage; relative performance depends on terrain complexity, local wind climate, and the verification metric of interest.