Comparative analysis of flood Forecasting using numerical Weather predictions and AI-based precipitation forecasts in the solo River

Abstract

This study evaluates the performance of the European Centre for Medium-Range Weather Forecasts (ECMWF) Numerical Weather Prediction (NWP) Integrated Forecasting System (IFS) and the Artificial Intelligence Forecasting System (AIFS) in flood forecasting. The primary objective is to compare the accuracy of precipitation forecast products in predicting flood magnitude and arrival time. Hydrological simulations are conducted using the Rainfall–Runoff–Inundation (RRI) model, with all model configurations held constant except for the rainfall input. The evaluation of rainfall forecasts indicates that, at the same temporal resolution, the AIFS generally exhibits lower root-mean-square error (RMSE) and mean error (ME) than the IFS, although both forecasts tend to underestimate rainfall. Forecasted discharge results indicate that for the initial 24-hour lead time, there is no significant difference in RMSE between the IFS and AIFS. As the forecast lead time increases, AIFS shows superior performance with a progressively larger difference in RMSE. The F1-score and Frequency Bias (FB) consistently indicate more favourable results for AIFS compared to NWP across all lead times. The F1-score for AIFS ranges from 0.216 to 0.717, while IFS ranges from 0.121 to 0.692. The confusion matrix shows that the AIFS detects peak flood events more effectively than the IFS, producing a higher number of true positives, though this improvement is accompanied by an increase in false positives during non-flood periods.