Improving multi-model ensemble streamflow forecasts by combining lumped, distributed and deep learning hydrological models

Deep learning has recently shown promise for hydrological applications. This study investigates the accuracy of a hybrid multi-model neural network approach for streamflow forecasting, using nine conceptual hydrological models (eight lumped, one semi-distributed) and one deep learning (DL) model. It aims to evaluate if the Long Short-Term Memory (LSTM) DL model within the multi-model framework improves short-term streamflow forecasts over a Canadian catchment. By integrating traditional hydrological models with the LSTM, the study addresses the operational challenges and enhances the forecast skill, especially for the early lead-times up to 9 days. Results indicate that the combination of LSTM with other models leads to better performance, suggesting that DL achieves optimal results when paired with different modeling approaches. LSTM models appear to be promising predictive tools for hydrological forecasting when integrated within a hybrid multi-model framework, facilitating gradual adoption in operational settings.