ResCast: Enhancing Global Medium-range Precipitation Forecasting with Residual Diffusion Model

Guowen Li, Yang Liu, Shilei Cao, Mengxuan Chen, Xintong Liu, Xuehe Wang, Juepeng Zheng, Yuxiang Li, Zhiyu Ye, Haoyuan Liang, Fang Wang, Meng Jin, Hong Cheng, Haohuan Fu

Keywords: Medium-range precipitation forecasting;

TL;DR: We propose ResCast, a novel method for medium-range precipitation forecasting that combines meteorological residual diffusion modeling and precipitation regression, achieving state-of-the-art performance on ERA5 data.

Abstract:

Machine learning techniques have been successfully applied to global weather forecasting, achieving significant results across various applications. However, existing data-driven machine learning methods struggle to provide accurate medium-range meteorological predictions. As a result, precipitation forecasts regressed from these predictions are less accurate, making reliable medium-range precipitation forecasting difficult. The root causes of these issues are error accumulation in meteorological variable forecasts and a lack of effective variable interaction in the precipitation regression module. In this paper, we propose ResCast, a novel approach to global medium-range precipitation forecasting by combining meteorological residual diffusion modeling and precipitation regression.
The diffusion component consists of (i) the Details Network (DetNet), which captures global features, and (ii) the Multi-Attention U-Net (AttUnet), which generates residuals for meteorological variables to reduce prediction bias. Then, a precipitation regression module quantifies the influence of residual-enhanced meteorological variables on precipitation, improving forecast accuracy. We evaluate our approach on the ERA5, an established dataset from the ECMWF, using comprehensive metrics and compare global medium-range precipitation forecasts against four state-of-the-art baselines (such as ENS, GraphCast, etc.). The results demonstrate the effectiveness and superiority of the proposed framework. The code implementation can be found in https://anonymous.4open.science/r/ResCast-78BD.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 12585