A Random Forest-Based Forecasting System for temperature and precipitation in the Basque Meteorology Agency

Accurate forecasting of surface meteorological variables, such as precipitation and air temperature, is essential for operational meteorology, particularly in regions with complex topography like the Basque Country. This study presents the implementation and preliminar validation of a forecasting system developed by the Basque Meteorological Service (Euskalmet), based on Random Forest (RF) machine learning algorithms. The system is designed to support operational forecasting tasks by providing high-resolution, site-specific predictions of precipitation and temperature.

The previous system, relied on a statistical framework, was based on multiple linear and logistic regression, using adjusted R² and pseudo-R² as performance metrics. While effective, it produced forecasts at daily resolution and had limitations in capturing non-linear relationships often present in mesoscale meteorological processes. The Random Forest approach overcomes these constraints by naturally modelling complex interactions and providing hourly forecasts, leading to improved accuracy, particularly in regions with complex terrain.

The RF system is trained using a combination of predictors from a synoptic-scale model and a mesoscale model, along with historical observational data from Basque Country automatic weather stations network. The predictors include variables such as surface pressure, humidity, wind and temperature at multiple levels. Model training and validation were carried out using data from a sufficiently long time period to ensure robustness across different weather regimes. Feature selection and model tuning were performed to optimize accuracy and computational efficiency.

Performance metrics such as bias, root mean square error, and correlation score and graphics such as scatterplots and Taylor diagrams were used to compare forecasted variables with observed values

This preliminary implementation demonstrates the potential of machine learning-based systems to complement traditional numerical weather prediction outputs in an operational setting. Finally, are presented some general conclusions and new features to take into consideration in order to improve the system and possible operational implementation.