Integrating machine learning ensembles and flood classification for enhanced flood forecasting with dynamic parameter weighting

It is a recognised problem that the structure and parameters of hydrological models are often not transferable in time. This may reduce model performance, with negative implications for applications in water resources modelling or flood forecasting. We hypothesise that integration of time-varying model parameters has the potential to better account for different active meteorological or hydrological processes that lead to different flood event types. To enhance flood forecasting the parameter sets should therefore be weighted dynamically according to the event type. In addition, ensemble approaches can be employed to reduce parameter uncertainty and to increase simulations accuracy for both continuous and event-based flood forecast models. Here, we introduce a machine learning based flood-type specific dynamic parameter weighting for ensemble flood forecasting. The potential of this method is demonstrated for an example case study application, employing five different machine learning approaches. Four types of flood events were considered, heavy-rainfall floods, long-duration-rainfall floods, sequence-of-rain floods and minor events with no clear generating processes. The results show that applying flood-type-specific parameter sets outperforms the other methods. Moreover, the performance consistently matched or exceeded the benchmark models. This improved performance is particularly relevant for practical applications in operational flood forecasting.