Sub-seasonal Hydrological Drought Forecast – the use of a Hybrid Forecasting System for Streamflow and Lake Level Predictions in Switzerland

Switzerland and the nearby alpine countries are not commonly associated with the occurrence of droughts, but in recent years, Switzerland has experienced several unprecedented drought events. As many sectors in the European Alps depend heavily on the water resources, e.g. for hydropower production, navigation and transportation, agriculture, and tourism, it is important for decision makers to have early warnings of drought. Machine learning (ML) approaches have shown potential to compete with traditional hydro-meteorological models. By combining a traditional hydrological model and a ML model in a hybrid setup, the forecasting system is able to benefit from the statistical power of ML while maintaining the understanding of physical processes from the traditional model.

The objective of this study is to investigate the predictability of a hybrid forecasting system composed of a traditional hydrological model PREVAH and an ensemble of ML algorithms to provide sub-seasonal streamflow and lake level predictions for major rivers and lakes in Switzerland. Uncertainty of the hydro-meteorological prediction chain is accounted for by using 51 hydrological ensemble members, and the ML uncertainty is accounted for by performing different rounds of initial randomization. We also investigate different drivers of the drought predictability by considering input features such as initial conditions, European weather regime forecasts and a hydropower proxy.

We are able to demonstrate that the proposed hybrid forecasting system is able to perform runoff routing scheme and provide sub-seasonal forecasts of streamflow and lake level for Swiss basins. Informed ML models with additional input features achieve better performance than those obtained using hydrological model outputs only. In the first half of the forecast period (weeks 1 and 2), model performance is improved by the initial conditions and in the second half (weeks 3 and 4) by the hydropower proxy. Lake level prediction shows promising skill for different basin sizes, whereas the streamflow prediction skill is linked to basin size. This study shines light on the use of hybrid forecasting for sub-seasonal drought prediction to provide useful information for medium- to long-term planning from an integrated risk management perspective.