Explainable AI for Seasonal Low-Flow Forecasting and Water Management in Austria

Droughts and low-flows are significant hydrological and environmental hazards that threaten a wide range of water-related sectors, such as navigation, hydropower production, and water management in general. Under climate change scenarios, the increasing risk of severe and persistent low-flow events will lead to rising costs for economy and society. Improved forecasting of low-flows for lead times of 1 to 6 months would be vital for many sectors, as it would allow for a more proactive water management. Streamflow forecasting is common in areas as North-America, Africa or Australia, but less frequent in Europe. Here the links to atmospheric modes are weaker, which poses a particular challenge. Seasonal forecasting of low-flow is even rarer, although hydrological drought has the virtue of being a slowly evolving process that is not substantially influenced by short-term precipitation events and is therefore likely easier to predict. For Austria, a seasonal low flow forecasting is completely missing. Our proposed project aims to fill this gap, by developing a probabilistic seasonal low-flow forecasting framework for five main river basins in Austria (Danube, Inn, Salzach, Drau, Mur). The study area was selected in initial communication with stakeholders to increase the societal impact of the study. The approach is innovative in many aspects: • It evaluates different spatial and temporal aggregates of the predictor variables (e.g. soil moisture, groundwater) and their value for forecast accuracy. • It develops bias-adjusted and downscaled climate forecasts and assess their relative performances with respect to meteorological variables and the added value for low-flow forecasting over various lead times. • It assesses the value of simpler single-site data-driven models compared with complex multi-site data-driven space-time models. • It combines process-based with data-driven models to explore relative merits that help to improve the models. • It evaluates the user value for stakeholders with particular emphasis on navigation and hydropower production. The results of the proposed project will directly feed into the emerging needs of the water sector due to climate change for critical infrastructure such as hydropower production, transport, navigation or water quality-related issues. A probabilistic forecasting framework for low-flow would enable water managers to act with foresight, increase society’s resilience to droughts and reduce the economic costs of this hydrological hazard. The outcome of this study will not only be beneficial for society and economy, but will have a significant impact to advance scientific knowledge, by: i. Improving seasonal climate forecasts for Austria. ii. Quantifying the user value of seasonal low-flow forecasting for Austria. iii. Advancing methods and knowledge about complex space-time statistical models and their relative merits compared to single-site forecasting schemes.