Abstract. Since the start of the 21st century, the European Alpine region has faced unprecedented low-flow conditions and drought events, severely impacting sectors dependent on reliable water availability, such as hydropower production, agriculture, and transportation. The growing frequency and severity of these low-flow conditions have led to a need for early warning systems. In this study, we present a novel machine learning (ML) aided hybrid forecasting framework designed to enhance sub-seasonal low-flow predictions in the European Alps. By harnessing the statistical power of ML and integrating diverse data sources, we trained 11 models using the Temporal Fusion Transformer (TFT) algorithm. These models incorporate features such as European Atlantic Weather Regimes (WR) for capturing large-scale atmospheric circulation patterns, in-situ streamflow observations for initial conditions, and process-based predictions from the European Flood Awareness System (EFAS). Our results show that the hybrid framework, even when using only WR data, outperforms climatology. The best results are achieved by combining observational data with process-based model data (raw EFAS output), underscoring the value of integrating diverse data sources. The models effectively capture initial condition persistence and correct biases in the raw EFAS output. Based on the Continuous Ranked Probability Skill Score (CRPSS), the best model effectively extends the skilful forecast horizon by 5 days on average across all stations during low flow periods. Furthermore, the interpretability of the TFT model provides valuable insights, identifying glacier coverage as a key catchment feature influencing model performance. Future research should further explore the connections between hydrological features and prediction skill, as well as the framework's applicability in ungauged areas and other regions.
Chang et al. (Mon,) studied this question.