Comparative Analysis of Precipitation Forecasts between the ECMWF Artificial Intelligence System (AIFS) and Its Integrated Forecast System (IFS)
Pan Liujie, Hongfang Zhang, Mian Liang, Peirong Li, Congqi Cao, Lili Du, Jing Liu · Weather and Forecasting · 2025
Abstract This study utilizes precipitation observations from 2415 stations in China and ECMWF’s ERA5 reanalysis data to conduct a detailed comparative analysis of the precipitation forecasting performance of ECMWF’s Artificial Intelligence Forecasting System (AIFS) and Integrated Forecasting System (IFS). The main conclusions are as follows: 1) The RMSE of AIFS’s 1–5-day precipitation forecast is significantly lower than that of IFS. The 120-h RMSE of IFS is roughly equivalent to the 24-h RMSE of AIFS. The mean error (ME) of AIFS precipitation forecasts is generally better than that of IFS. IFS’s cumulative rainfall aligns better with observations in southern China and the Tibetan Plateau. 2) In southern China and the Tibetan Plateau, neither AIFS nor IFS can accurately represent the observed variation characteristics of the standard deviation (STD) of precipitation. AIFS shows less fluctuation in STD across the three regions, underestimates precipitation amounts, and has a weaker ability to forecast extreme values. 3) AIFS consistently overpredicts the frequency of light precipitation, leading to significantly lower equitable threat scores (ETSs). However, it shows better performance in terms of threat score (TS) and ETS for heavy precipitation events exceeding 25.0 mm. In terms of the stable equitable error in probability space (SEEPS), the IFS demonstrates superior performance. 4) The integrated water vapor transport (IWVT) and its STD forecasted by AIFS are relatively weak, and its vertical velocity is smaller. This may be a key direct reason for the underprediction of precipitation extremes and STD by AIFS. Significance Statement With the rapid advancement of artificial intelligence, data-driven large-scale weather forecasting models are continually emerging. This study provides a detailed analysis of the precipitation forecasting performance of the ECMWF Artificial Intelligence Forecasting System (AIFS) and IFS. Through error analysis and score verification, we found that the AIFS precipitation forecasts exhibit smaller RMSE and STD values. The IFS exhibits superior performance in TS, ETS, and stable equitable error in probability space (SEEPS) for precipitation events exceeding light rain, whereas the AIFS demonstrates enhanced TS and ETSs for moderate and above precipitation. The underestimation of STD is similarly observed in element forecasts for pressure levels. Insufficient forecasts of vertical velocity in the upper-air pressure levels and integrated water vapor transport (IWVT) may be direct reasons affecting precipitation forecasts. The aim of this study is to enhance the application capabilities of AI forecasting systems and provide new perspectives for constructing and improving precipitation algorithms within these systems.