Hyperlocal Weather Forecasting by Machine Learning: Findings, Results and Conclusions of the IEEE Region 8 Competition

Motivated by the need for climate-change resilience, this paper summarizes the results of an international hackathonstyle competition aimed at improving local weather forecasts using machine-learning post-processing of multiple forecast sources and historical data. The approaches and results of the top teams are presented and compared, including support vector regression (SVR), gradient boosting, k-nearest neighbors (KNN), long short-term memory (LSTM) networks, and transformerbased architectures. Each approach tailors model architectures, preprocessing, feature selection, and hyperparameter tuning to the characteristics of the target meteorological variables across three case studies in Africa, the Middle East, and Europe (IEEE
Region 8). Six variables are evaluated against ground-truth measurements collected over the seven days following each submission. The results show that careful architectural and featureengineering choices, informed by domain knowledge, can substantially reduce error relative to the operational ECMWF-IFS benchmark. Across the five variables with non-trivial benchmark values, the best per-variable RMSSE reductions reach 59.26%, with an average reduction of 30.46% using post-processing of historical NWP forecasts and local observations, highlighting the significant potential of this technology.