Tower-Constrained Post-Processing of WRF Hub-Height Wind Speed for Regional Wind-Resource Screening in Hami, Xinjiang, China

We evaluate WRF 90 m wind speeds over Hami, China, using hourly observations from 12 towers during June 2016–May 2017. The study combines a tower-point seasonal hindcast, a four-model comparison at six WRF anchors, a bidirectional cross-farm evaluation, and monthly residual interpolation with 2016–2018 wind-power-density (WPD) scenarios. The March–May 2017 hindcast uses causal feature construction and training-period-only fitting on 26,367 matched tower-hour samples. Its two-stage ExtraTrees correction reduces central error relative to its power-law baseline but detects none of the 128 observed hours at or above 25 m−1, revealing substantial loss of the high-wind tail. In a separate leave-one-WRF-anchor-out comparison, ET, DLinear, TCN, and CVAE use identical 48 h power-law wind histories at the six anchors; their unweighted anchor mean RMSE values span 3.521–3.540 m−1 and all strongly underdetect 25 m−1 h. In the cross-farm evaluation, shared-anchor observations were excluded from fitting: all models predict zero of 223 YW exceedance hours when trained with BLK observations, while the reverse direction gives low and variable POD values. For the monthly product, all towers at a held WRF coordinate were excluded from IDW donors. Across six anchors and 72 anchor-month comparisons, the unweighted mean RMSE decreases from 1.04 to 0.77 m−1. The published power-2, fixed-density central WPD remains 726.7 W m−2; across 162 sensitivity cases varying IDW power, residual, cubic factor, density, and penalty weight, fixed-density and ISA means span 457.3–1102.5 and 411.0–988.5 W m−2, respectively. Only 8.36% of product cells lie inside the six-anchor convex hull. Together, these analyses distinguish tower-point prediction, monthly interpolation, and regional WPD comparison. The maps support regional comparison and identify locations for follow-up measurement within the one-year tower record and six-anchor network.