Long short-term memory networks for enhancing real-time flood forecasts: a case study for an underperforming hydrologic model

Flood forecasting systems play a key role in mitigating socioeconomic damage caused by flood events. The majority of these systems rely on process-based hydrologic models (PBHMs), which are used to predict future runoff. Many operational flood forecasting systems additionally implement models aimed at enhancing the predictions of the PBHM, either by updating the PBHM's state variables in real time or by enhancing its forecasts in a post-processing step. For the latter, autoregressive integrated moving average (ARIMA) models are frequently employed. Despite their high popularity in flood forecasting, studies have pointed out potential shortcomings of ARIMA-type models, such as a decline in forecast accuracy with increasing lead time. In this study, we investigate the potential of long short-term memory (LSTM) networks for enhancing the forecast accuracy of an underperforming PBHM and evaluate whether they are able to overcome some of the challenges presented by ARIMA models. To achieve this, we developed two hindcast-forecast LSTM models and compared their forecast accuracies to that of a more conventional ARIMA model. To ensure comparability, one LSTM was restricted to use the same data as ARIMA (eLSTM), namely observed and simulated discharge, while the other additionally incorporated meteorologic forcings (PBHM-HLSTM). Considering the PBHM's poor performance, we further evaluated if the PBHM-HLSTM was able to extract valuable information from the PBHM's results by analyzing the relative importance of each input feature. Contrary to ARIMA, the LSTM networks were able to mostly sustain a high forecast accuracy for longer lead times. Furthermore, the PBHM-HLSTM also achieved a high prediction accuracy for flood events, which was not the case for ARIMA or the eLSTM. Our results also revealed that the PBHM-HLSTM relied, to some degree, on the PBHM's results, despite its mostly poor performance. Our results suggest that LSTM models, especially when provided with meteorologic forcings, offer a promising alternative to frequently employed ARIMA models in operational flood forecasting systems.