A Physical-Machine Learning Hybrid Model for Predicting Peak Streamflow During Extreme Flooding

Hydrological models have limited skill for predicting peak river flows during large floods. Here we use a hybrid of a physically-based hydrological model (LISFLOOD) and machine learning model (Long Short Term Memory) for predicting hourly streamflow values during an extreme flooding event caused by the 2019 Typhoon Hagibis in Japan. The LISFLOOD model (benchmark) was calibrated using hourly streamflow data (2014-2018) from several hundred gauges across Japan and its skill for simulating peak flows was evaluated. The LSTM network was trained to estimate the errors of the physical model using the difference between the observed and model-simulated streamflow as the target variable. The predicted error from the LSTM model is used to improve the accuracy of the hydrological model prediction. Then the accuracy of the hybrid model during the test period (2019) was evaluated by comparing against the skill of the benchmark model. In this presentation, we discuss the utility of the LSTM for estimating the hydrological model error in predicting the historically large flood peaks. We also highlight the existing challenges of simulating extreme floods.