Leveraging AI Models for Regional Weather Prediction: A Data Pipeline for Africa

This study presents a pilot implementation of an AI-driven data pipeline for weather forecasting in Africa, leveraging GraphCast, a state-of-the-art deep learning model developed by Google DeepMind. The pipeline automates data extraction, transformation, and model inference, utilizing reanalysis data from ERA5 to generate short- to medium-range weather forecasts. The primary objective is to evaluate the feasibility and accuracy of GraphCast for regional weather prediction, particularly in climate-sensitive sectors such as agriculture and disaster preparedness. The study compares GraphCast's forecasts against ground-truth observations from 107 Trans-African Hydro-Meteorological Observatory (TAHMO) stations in Kenya over a 10-day lead time, using Root Mean Square Error (RMSE) as a performance metric. Results indicate promising forecasting potential, with an RMSE of 5.42 mm for precipitation and $2.52^{\circ} \mathrm{C}$ for temperature, but also highlight areas for model refinement. Key lessons include the need for localized fine-tuning of AI weather models and the integration of high-resolution regional climate data to improve accuracy. The findings demonstrate the potential impact of AI-based Weather Prediction (AIWP) models in improving forecasting capabilities for Africa. The study contributes to ongoing efforts to make AI-driven weather forecasting more accessible and lays the groundwork for future enhancements, including an operational web-based forecasting tool.