Towards operational sub-seasonal and seasonal low-flow forecasting in the Adige river basin using deep learning

The Alpine region is a hotspot of climate change, where projected increases in the frequency and intensity of droughts are expected to exacerbate competition over water resources. Sub-seasonal and seasonal (S2S) hydrological forecasts are a key component of drought early warning systems; however, their operational uptake in water-related sectors remains limited due to high uncertainty and low forecast skill. In mountain catchments, hydrological memory can represent an important source of predictability in addition to boundary conditions provided by seasonal climate forecasts. While previous studies have highlighted the role of hydrological memory in controlling predictability in the Alps (Staudinger et al. 2014; Stergiadi et al. 2020) using either Ensemble Streamflow Prediction (ESP), reverse-ESP or End Point Blending (EPB) approaches, the analyses have relied on conceptual hydrological models and evaluated forecast skill against model-generated reference. As a result, the reported predictability depends on model fidelity and reflects model-dependent potential predictability, rather than practical predictability verified against observations. More recently, relevant works on Alpine catchments forced process-based models with extended range probabilistic forecasts (32 days) and found that streamflow forecast skills lasted up to four weeks but decreased quickly after just few days in high elevation Alpine basins (Bogner et al. 2022). In recent years, deep learning models have attracted increasing interest due to their flexible handling of diverse input data and strong predictive performance (Chang et al 2025). These models can be trained on observation targets such as streamflow and can jointly exploit past hydrological states and seasonal climate forecasts. In this study, we evaluate S2S hydrological forecasts in several sub-basins of the Adige River basin at two forecast horizons (45 and 210 days). We generate forecasts by using a Temporal Fusion Transformer (TFT) driven by operational datasets representing some of the key components of the hydrological system (snow water equivalent, streamflow, and surface soil moisture), together with downscaled seasonal climate forecasts (SEAS5), enabling an integrated assessment of their relative contributions to forecast skill. We assess performance with a focus on low flows and discharge warning thresholds defined by the basin’s authority, using discharge climatology and the European Flood Awareness System (EFAS) as benchmarks. Finally, we assess the interpretability of the TFT layers for characterising the drivers of low-flow conditions. In this presentation, we show preliminary results on when and under which conditions the proposed approach improves forecast performance.

Bogner, K., Chang, A. Y. Y., Bernhard, L., Zappa, M., Monhart, S., & Spirig, C. (2022). Tercile Forecasts for Extending the Horizon of Skillful Hydrological Predictions. Journal of Hydrometeorology, 23(4), 521–539.

Chang, A. Y.-Y., Harrigan, S., Ramos, M.-H., Zappa, M., Grams, C. M., Domeisen, D. I. V., and Bogner, K.: Exploring Hybrid Forecasting Frameworks for Subseasonal Low Flow Predictions in the European Alps, EGUsphere [preprint]

Staudinger, M., & Seibert, J. (2014). Predictability of low flow - An assessment with simulation experiments. Journal of Hydrology, 519(PB), 1383–1393.

Stergiadi, M., di Marco, N., Avesani, D., Righetti, M., & Borga, M. (2020). Impact of geology on seasonal hydrological predictability in alpine regions by a sensitivity analysis framework. Water (Switzerland), 12(8).