24AI Deep Learning-Based Methods for Precipitation Bias Correction

Accurate precipitation forecasting is vital for both economic stability and societal well-being. However, the complexity of precipitation processes and the inherent limitations of numerical weather prediction (NWP) technology lead to increasing errors in short- to medium-term forecasts (1 to 10 days) as the forecast period lengthens. Post-processing techniques are therefore essential to correct biases in model-generated precipitation forecasts. Traditional deep learning models have improved overall forecast accuracy, but they often fall short in correcting heavy precipitation events, resulting in generalized and less precise corrections. This study investigates the use of U-Net and diffusion models for bias correction in short- to medium-term daily precipitation forecasts produced by the National Centers for Environmental Prediction’s Global Forecast System (NCEP-GFS). By designing specialized loss functions and incorporating key forecast variables (such as humidity) along with geographical data (such as topography), the study evaluates the effectiveness of these corrections. The findings demonstrate that the customized loss function, combined with relevant precipitation-related factors (e.g., topography and humidity), significantly enhances the correction of heavy precipitation events. Furthermore, using U-Net’s correction results to guide the diffusion model further refines the distribution of forecast outcomes, bringing them closer to actual precipitation patterns. This approach effectively addresses the generalization issues observed in U-Net corrections, leading to more precise and accurate bias correction.