AI Weather Models Show Significant Promise for African Rainfall Prediction

New research demonstrates that AI-based weather prediction models are nearing the accuracy of traditional numerical weather prediction (NWP) systems, but at a significantly lower computational cost. This development holds particular importance for Africa, a continent increasingly affected by extreme rainfall events, where many forecasting centers struggle with the necessary infrastructure to run complex physical models for long-range forecasts.

The study conducted a detailed comparison of prominent AI models like GraphCast, GenCast, and the Functional Generative Network (FGN) against the established physical NWP model IFS. The evaluation focused on rainfall prediction across various African regions, considering different seasons, wet and dry conditions, elevation zones, and lead times. Both deterministic and probabilistic forecasts were post-processed and assessed using standard meteorological metrics and observational datasets such as IMERG, RFEv2, and CHIRPS.

The findings indicate that all tested models maintain predictive skill beyond basic climatology, even at extended lead times. Notably, AI models generally surpassed IFS in wet regions, while IFS showed better performance in dry, high-elevation areas, likely due to its superior resolution in capturing terrain-influenced rainfall patterns. Across diverse datasets and seasons, the AI models collectively demonstrated an average improvement of approximately 5% over IFS. GraphCast achieved calibrated skill comparable to the ensemble-based FGN, with FGN showing greater significant skill at longer lead times.

These results underscore the substantial potential of calibrated AI weather prediction to deliver accessible and computationally efficient rainfall forecasts throughout Africa. The research also emphasizes the ongoing importance of factors like spatial resolution, ensemble design, and the unique regional characteristics of the African continent in optimizing these forecasting systems.