Spatial Predictor Selection for Next-Day Minimum Temperature Forecasting: An Automated Machine Learning Framework Applied Across European Climate Regimes

Eric Duhamel

Accurate prediction of daily minimum temperature (Tmin) is critical for frost protection, energy management, and public health preparedness. While numerical weather prediction models have improved substantially, their performance for Tmin forecasting remains limited by difficulties in representing fine-scale nocturnal processes. This study presents an automated framework for identifying optimal spatially-distributed predictors for next-day Tmin forecasting, applied to eight climatically diverse sites across Western Europe.
Using 26 meteorological variables from ERA5 reanalysis data spanning 2004–2024, we systematically explored a search space of approximately 45,000 candidate predictors within a 540 km radius around each target station. An iterative optimization algorithm guided by mean absolute error (MAE) identified 90-predictor configurations for each site. Three regression models—linear regression, LightGBM, and XGBoost—were evaluated, with XGBoost consistently achieving optimal performance.
Results demonstrate substantial skill across all sites, with MAE ranging from 0.81°C (Nice, Mediterranean) to 1.34°C (Brest, oceanic), representing 35–54% improvement over persistence and 51–64% over climatological baselines. The analysis revealed both universal patterns—near-surface air temperature dominated predictive gain at all sites (37–66%)—and distinctive climate-specific signatures: Mediterranean stations exhibited strong persistence signals (30% contribution from previous-day Tmin), oceanic climates showed enhanced dewpoint importance (16%), and continental sites featured significant soil temperature contributions (14%).
Predictor selections proved highly stable at the variable level (23–24 of 26 variables consistently selected across independent runs), while spatial autocorrelation caused greater variability in specific grid-point selections. Importantly, 80% of predictive gain originated from just 4 predictors and 90% from 12 predictors, suggesting that substantially reduced configurations could achieve comparable performance for operational applications.
While the absolute MAE values reflect the idealized nature of reanalysis data and are not directly transferable to operational contexts, the methodology for predictor identification remains valid and applicable to numerical weather prediction outputs. This framework provides a systematic, reproducible approach to spatial predictor selection that can be adapted to other forecasting variables and domains.