Estimating near-surface wind over complex terrain using deep learning with short-term observations

Accurately predicting near-surface winds over complex terrain remains challenging due to terrain-induced flow patterns and limited observational data. Numerical weather prediction models such as the Weather Research and Forecasting (WRF) model are commonly used to simulate wind fields, but they often fail to resolve fine-scale terrain effects, resulting in significant errors in wind speed and direction. A deep learning framework is introduced to correct WRF-derived wind fields using a convolutional neural network (CNN). Horizontal wind components at 10 m above ground level were extracted from WRF outputs as CNN inputs. The CNN was trained using short-term observational datasets (one month), while data from the remaining eleven months were used for testing. To ensure robustness and reduce overfitting, a 5-fold cross-validation scheme was employed during training. The CNN-predicted wind speeds showed lower errors than those from WRF, and the model improved wind direction estimates, which WRF often poorly captures. Our findings demonstrate that the CNN model, trained with only short-term observations, can effectively correct near-surface wind speeds across extended periods. This study highlights the importance of proper WRF domain configuration and the potential of combining data-driven models with mesoscale simulations for complex terrain applications.