AI algorithms to enhance weather forecasts over a topographically complex Mediterranean region

The rapid advancement of artificial intelligence (AI) techniques has significantly increased interest within the scientific community, particularly regarding their application in weather forecasting. In recent years, the volume of research focused on integrating AI into meteorological prediction has grown substantially. This trend has been further amplified by the operational deployment of systems such as GraphCast, which have brought increased visibility to this research domain.

This study investigates the potential of several Machine Learning (ML) and Deep Learning (DL) models, including Artificial Neural Networks (ANN), Random Forests (RF), Convolutional Neural Networks (CNN), and Graph Neural Networks (GNN), to enhance short-term (1-day lead time) precipitation forecasts generated by a physically-based Numerical Weather Prediction (NWP) system. Since January 2020, the CeSMMA laboratory (Study Center for Environmental Monitoring and Modelling – University of Calabria) has issued daily forecasts for southern Italy, available at https://cesmma.unical.it/cwfv2/. The forecasting framework employs the WRF (Weather Research and Forecasting) model, with boundary and initial conditions derived from the GFS (Global Forecast System). AI models are applied as post-processing, using correction factors derived from a two-year training period based on observations from a dense regional monitoring network comprising approximately 150 rain gauges.

The results demonstrate that AI-based post-processing substantially improves daily precipitation forecasts relative to ground-based measurements. Considering the full study area (approximately 15,000 km²), the ANN reduces the Mean Squared Error (MSE) by about 29%, while the RF achieves a 21% reduction, both relative to the original WRF output. Moreover, the GNN applied to a smaller subregion with data from 22 rain gauges achieves an even greater reduction in MSE, up to 35%, during periods of intense rainfall.

Beyond improving forecast accuracy, the AI-enhanced outputs produce spatial precipitation patterns that are physically consistent and able to capture key processes, such as orographic enhancement. Such improvements stem from using AI models as post-processing tools that enhance, rather than replace, the physically-based forecasts, thus retaining the underlying dynamical consistency and capturing essential physical processes.