Long Short-Term Memory Networks for Real-time Flood Forecast Correction: A Case Study for an Underperforming Hydrologic Model

Flood forecasting systems play a key role in mitigating socio-economic damages caused by flooding events. The majority of these systems rely on process-based hydrologic models (PBHM), which are used to predict future river runoff. To enhance the forecast accuracy of these models, many operational flood forecasting systems implement error correction techniques, which is particularly important if the underlying hydrologic model is underperforming. Especially, AutoRegressive Integrated Moving Average (ARIMA) type models are frequently employed for this purpose. Despite their high popularity, numerous studies have pointed out potential shortcomings of these models, such as a decline in forecast accuracy with increasing lead time. To overcome the limitations presented by conventional ARIMA models, we propose a novel forecast correction technique based on a hindcast-forecast Long Short-Term Memory (LSTM) network. We showcase the effectiveness of the proposed approach by rigorously comparing its capabilities to those of an ARIMA model, utilizing one underperforming PBHM as a case study. Additionally, we test whether the LSTM benefits from the PBHM's results or if a similar accuracy can be reached by employing a standalone LSTM. Our investigations show that the proposed LSTM model significantly improves the PBHM's forecasts. Compared to ARIMA, the LSTM achieves a higher forecast accuracy for longer lead times. In terms of flood event runoff, the LSTM performs mostly on par with ARIMA in predicting the magnitude of the events. However, the LSTM majorly outperforms ARIMA in accurately predicting the timing of the peak runoff. Furthermore, our results provide no reliable evidence of whether the LSTM is able to extract information from the PBHM's results, given the widely equal performance of the proposed and standalone LSTM models.