A Super-resolution Downscaling Approach for Ensemble Severe Precipitation Forecasts Using Transfer Learning and a Lightweight Neural Network

The Ensemble Prediction System (EPS) provides reliable precipitation forecasts. However, its relatively coarse spatial resolution, which is constrained by computational resources, directly limits its capability in predicting high-impact severe rainfall. Given that downscaling based on super-resolution offers a computationally efficient and highly practical solution to enhance forecast resolution, this study develops a Self-Attention-Enhanced Convolutional Neural Network (SAECNN) for downscaling coarse ensemble precipitation forecasts on North China where severe rainfall occurred quite frequently in recent years. By integrating a self-attention mechanism and inception-style module, the SAECNN is trained by a two-step process using high-resolution (HR) and low-resolution (LR) precipitation pairs under a composite loss function. The model is trained using 3-hour cumulative summer precipitation data from 2010 to 2019 obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5-Land reanalysis dataset. Subsequently, taking global ensemble prediction system of ECMWF (ECMWF-GEPS) as example, the SAECNN is applied to the LR GEPS and produce HR precipitation forecasts. Ablation experiments demonstrate that the combination of Huber loss and mean absolute error with minimized missed rate, together with two-step training strategy, effectively reduce forecast errors. Independent validation, against bilinear-interpolated forecasts of the ECMWF-GEPS during 2020–2021, demonstrates that the SAECNN yields realistic and detailed precipitation forecasts, reducing the probabilistic forecast bias (Ranked Probability Score) by 8–10%. A case study of Zhengzhou severe rainfall in 2021 further reveals that the model can improve the accuracy of percentile forecasts and neighborhood probability forecasts. Overall, the SAECNN provides a computationally efficient approach for downscaling and improving coarse precipitation forecasts of EPS, while enhancing probabilistic forecasting skills.