Deep learning approaches for high-resolution wind data enhancement around a nuclear power plant

When radioactive materials are released near nuclear power plants, local wind fields control the dispersion path and concentration distribution. In Korea, complex terrain, sea–land breezes, and seasonal variability amplify prediction uncertainty, underscoring the need for high-resolution wind data. This study proposes and validates a framework that combines physics-based Weather Research and Forecasting (WRF) modeling with deep learning to generate high-resolution winds while lowering computational cost. We employ a Transformer-based Uformer trained on WRF-generated datasets to perform super-resolution of wind fields, addressing the limitations of simple interpolation. To curb storage and data processing burdens from finer grids, we apply Singular Value Decomposition (SVD) for compact, low-loss compression. The super-resolved winds are coupled to a Lagrangian Patricle Dispersion Model (LPDM) to assess dispersion sensitivity and evaluate Uformer fidelity and SVD effectiveness. Experiments show horizontal resolution improved from 1.5 km to 300 m, capturing local flows dominating near-field transport. With SVD, data volume decreases by ∼32% with negligible reconstruction error ( , enabling faster storage and reuse. Overall, the pipeline delivers high-resolution winds more quickly and efficiently than physical modeling, strengthening atmospheric dispersion predictions in complex meteorology typical of Korean nuclear power plant regions and enhancing the timeliness of radiological emergency decision support.