Hybrid Forecasting of Recent Flood and Drought Events in Switzerland

Flood and drought events can lead to severe socio-economic impact and damages. Thus, there is a need for early warnings of such extreme events, especially for decision-makers in sectors like hydropower production, navigation and transportation, agriculture, and hazard management. To improve the predictability of sub-seasonal streamflow, we propose the approach of a hybrid forecasting system, where a conceptual hydrological model PREVAH is combined with a machine learning (ML) model. The PREVAH model provides catchment level hydrological forecasts and the role of the ML model is to emulate a runoff routing scheme. Such a hybrid setup allows the forecasting system to benefit from the statistical power of ML while maintaining the understanding of physical processes from the hydrological model.

The objective of this study is to investigate the predictability of a hybrid forecasting system to provide monthly streamflow predictions for three recent extreme events. These include the drought event in summer 2018, the drought event in spring 2020, and the flood event in summer 2021 in selected large Swiss rivers. We also investigate different predictability drivers by considering additional input features to the ML model, such as initial streamflow, European weather regime indices, and a hydropower proxy.

We demonstrate that the proposed hybrid forecasting system has the potential to provide skillful monthly forecasts of the interested events. Informed ML models with additional input features achieve better performance than results obtained using hydrological model outputs only. This study sheds light on using hybrid forecasting for sub-seasonal hydrological predictions to provide useful information for medium-term planning at a monthly time horizon and reduce the impact of flood and drought events.