Comparison of Dynamical and AI Model Forecasting Skills for Three Extreme Rainfall Events in 2025 over Northern China

Accurate forecasting of heavy rainfall remains a formidable challenge in weather prediction. This study systematically assesses the forecasting performance of five dynamical and five data-driven artificial intelligence (AI) models for three heavy rainfall events over northern China in 2025. The results indicate severely limited forecasting skill across all evaluated models, with the dynamical models performing slightly better than the AI models. For the Yuzhong extreme heavy rainfall on 8 August 2025, the average forecast intensity of the five dynamical models reaches only 19% of the observed precipitation intensity, compared with only 13% for the five AI models. Forecast skill for locating extreme heavy rainfall is generally deficient across all evaluated models. However, regional high-resolution dynamical models outperform global models. Although the evaluated models capture rainband evolution to varying degrees, they exhibit pronounced systematic biases in spatial placement and intensity distribution relative to observations. At the rainstorm threshold, the dynamical models show a slight advantage over the AI models. In contrast, at the moderate- and heavy-rain thresholds, the AI models show better forecast consistency and longer predictability windows than the dynamical models. These findings provide quantitative benchmarks for improving heavy rainfall forecasting techniques.