Numerical weather prediction (NWP) models are essential for precipitation forecasting but are constrained by coarse spatial resolutions (10–50 km), which fail to capture fine-scale variations required for regional disaster prevention, particularly in complex terrain. While statistical and machine learning downscaling methods have been developed to bridge this resolution gap, they predominantly operate as “black boxes” without explicit physical guidance, leading to predictions that violate meteorological principles and systematic underestimation of extreme precipitation events. To address these limitations, this study aims to develop a Physics-Informed Machine Learning framework that explicitly integrates multi-scale topographic modulation and physical consistency constraints into precipitation downscaling. Specifically, a Random Forest model enhanced with Multi-Scale Structural Similarity (MS-SSIM) loss and Physical Constraint Enhancement (MSSSIM-PCE-RF) was constructed. The model introduces elevation gradient weights at low-resolution layers and micro-topographic parameters (slope, surface roughness) at high-resolution layers, while enforcing physical consistency between precipitation intensity, radar reflectivity, and ground observations via the Z-R relationship. Based on hourly data from 2252 meteorological stations in Jiangxi Province (2021–2022), coupled with topographic factors (DEM, slope, aspect) and Normalized Difference Vegetation Index (NDVI), a technical framework of “data fusion–feature synergy–machine learning–spatial reconstruction” was established. Results demonstrate that the MSSSIM-PCE-RF model achieves a validation R² of 0.9465 and RMSE of 0.1865 mm, significantly outperforming the conventional RF model (R² = 0.9272). Notably, errors in high-altitude, steep-slope, and high-vegetation areas are reduced by 45.3%, 42.0%, and 43.1%, respectively, with peak precipitation period errors decreasing by 37.2%. Multi-scale topographic analysis reveals significant orographic lifting effects at 250–1000 m elevations, peak precipitation at 12–15° slopes, and abundant precipitation on south/southeast aspects. By explicitly embedding topographic modulation and physical consistency constraints, the model effectively alleviates systematic underestimation of extreme precipitation in complex terrain, providing high-resolution data support for transmission line disaster prevention and micro-meteorological risk assessment.

High-resolution wind-field prediction over complex terrain is important for wind-energy assessment, grid safety, and hazard mitigation. This study uses a coupled Weather Research and Forecasting–Computational Fluid Dynamics (WRF–CFD) workflow for the Askervein Hill benchmark. Driven by the fifth-generation European Centre for Medium-Range Weather Forecasts Reanalysis dataset, WRF provides mesoscale atmospheric conditions, and CFD resolves the terrain-induced flow features at high spatial resolution. Because repeated WRF and CFD simulations are computationally expensive, we develop a time-aware neural-network surrogate model, termed the multilayer perceptron-prior gated temporal correction (MLP-GTC) model, to predict the resulting three-component wind field. The model first learns the local wind response associated with terrain and inlet conditions, and it then applies a gated temporal correction based on recent inlet conditions and their corresponding pointwise predictions. The data are divided chronologically into 683 training, 146 validation, and 147 test time steps, with 5187 fixed spatial points evaluated at each time step. The coupled WRF–CFD simulations reproduce the observed wind structure with a maximum Pearson correlation coefficient of 0.96 and a 10 m wind-speed root mean square error of 0.84 m/s. For the three-component velocity field, MLP-GTC achieves validation/test root mean square errors of 0.282/0.310 m/s and mean absolute errors of 0.179/0.199 m/s. It improves on the multilayer perceptron model (0.486/0.516 m/s) and the matched five-step single-layer long short-term memory network (0.334/0.438 m/s).