Abstract
This study aims to address the critical need for refined, highresolution seasonal climate forecasts in the Alpine region in support to risk management and decision-making processes, especially in the challenging context of climate change. By leveraging regression-based Machine Learning (ML) algorithms, a Perfect Prognosis approach is applied to statistically downscale the daily fields of 2-metre temperature and total precipitation of ECMWF SEAS5 seasonal forecasts over the Alps. In particular, four different ML methods are considered: Random Forest, Light Gradient Boosting Machine (LGBM), Adaptive Boosting (AdaBoost) and Extreme Gradient Boosting (XGBoost). The daily fields of the European CERRA reanalysis (5.5 km) are used as reference target, while a set of meteorological predictors from the coarser grid are considered. In a preparatory phase, all ML methods and configurations are implemented and validated starting from the predictor fields of the ERA5 reanalysis. LGBM displayed the best results during training and validation for both temperature and precipitation, with superior computational speed and efficiency with respect to the other methods. It demonstrates prowess in capturing daily variations, with R2 scores of 0.95 mean temperature and 0.67 for precipitation, with generally low biases (-0.05 °C and 5.34% for daily mean temperature and precipitation, respectively, as yearly averages). Further optimization to increase the prediction accuracy of extreme values and annual precipitation averages are discussed. The best performing LGBM method is finally applied to downscale the SEAS5 seasonal forecast data and will represent a crucial component of a drought predicting model for the Alps in the framework of the EU-funded interTwin project (101058386).