Improving streamflow prediction in the WRF-Hydro model with LSTM networks

Abstract

Researchers have attempted to use machine learning algorithms to replace physically based models for streamflow prediction. Although existing studies have contributed to improving machine learning methods, they still have weaknesses, such as large dataset requirements and overfitting. Therefore, we propose an approach that combines the Weather Research and Forecasting hydrological modeling system (WRF-Hydro) and the Long Short-Term Memory (LSTM) network, i.e., WRF-Hydro-LSTM, to improve streamflow simulations. In this approach, LSTM was employed to predict the residual errors of WRF-Hydro; in contrast, the conventional approach with LSTM predicts streamflow directly. Here, we performed numerical experiments to predict the inflow of Soyangho Lake in South Korea using WRF-Hydro-LSTM, WRF-Hydro-only, and LSTM-only. WRF-Hydro-LSTM and LSTM-only showed better results (NSE = 0.95 and R greater than 0.96) compared to WRF-Hydro-only (NSE = 0.72 and R = 0.88); however, in terms of the percent bias, WRF-Hydro-LSTM had a better value (1.75) than LSTM-only (17.36). While the LSTM-only follows objective functions and not physical principles, WRF-Hydro-LSTM simulates residual errors and efficiently decreases uncertainties that are inherent with conventional methods. Furthermore, a sensitivity test on the training dataset indicated that the correlation coefficient and NSE value were not overly sensitive, but the PBIAS value differed substantially depending on the training set. This study demonstrates that WRF-Hydro-LSTM is particularly useful for representing real-world physical constraints and thus can potentially improve streamflow prediction compared to using either of the two approaches exclusively.