High flow prediction model integrating physically and deep learning based approaches with quasi real-time watershed data assimilation

Abstract

When operating a flood forecasting system, it is crucial to prioritize both promptness and accuracy, along with reflecting the latest watershed conditions. This study proposes a hybrid model for rainfall-induced high flow prediction that combines an instantaneous physically-based model and a deep learning-based rainfall loss model assimilating with quasi real-time watershed conditions. Its primary objective is to present immediate and reliable high flow predictions, ensuring stability in prediction procedures. The surface runoff component of the model is based on nonlinear instantaneous unit hydrograph (IUH) derived using a dynamic wave model, which considers the nonlinearity between the rainfall excess intensity and the time to peak and peak flow of the IUH. To account for rainfall loss, which can significantly influence rainfall-runoff quantity, Long Short-Term Memory (LSTM) model is employed to establish link between key rainfall loss parameters and the latest watershed conditions obtained from the Global Land Data Assimilation System (GLDAS). The application of the proposed model to 28 high flow events reveals several noteworthy outcomes. Firstly, unlike conventional models that depend on fixed parameter values specific to each event, the proposed model offers immediate, stable, and highly accurate high flow predictions without the need for manual calibration. Secondly, the prioritization of parameters reflecting watershed conditions varies between different watersheds and depends on the rainfall events. These findings indicate the necessity to estimate rainfall loss parameters individually for each rainfall-runoff event in each watershed, as no universal value exists for these parameters. Lastly, despite the limited number of test cases conducted in this study, the LSTM model achieved sufficiently accurate rainfall loss estimation for high flow prediction using only 50 datasets.