Medium-Range River Stage Forecasting with a Long Short-Term Memory Network

Abstract Accurately forecasting rainfall–runoff and the subsequent rise and fall within rivers remains a challenge in hydrological modeling. While physical modeling system capabilities have improved, they can be computationally demanding and/or require extensive calibration when deploying at a new forecast point. Data-driven machine learning models can largely avoid these challenges while maintaining high accuracy. In this study, a long short-term memory (LSTM) network is used to produce river stage height forecasts based on lagged inputs of basin-averaged precipitation, modeled relative soil moisture, and gauge height. A basin-specific LSTM was developed and tested on 42 precipitation-dominated river basins. The results show that the LSTM statistically outperforms the corresponding ensemble of the medium-range National Water Model when driven by the same Global Forecast System precipitation. With a 0.14 higher average Nash–Sutcliffe efficiency and 13.1% lower peak bias over the 7-day forecast, the LSTM demonstrates high viability for operational use. Significance Statement Accurately forecasting river flooding is of high importance for protecting life and property. Traditional physics-based hydrological models that have been used by operational centers to provide forecasts to the public are computationally expensive and still have high uncertainties. In this study, we examined the use of machine learning for forecasting river flooding out to 7 days and determined that the developed machine learning framework provides better forecasts than a physics-based modeling approach and shows high practicality for operational forecasting. Ultimately, the use of more accurate and timely forecasting systems, like the machine learning framework developed here, better serves the public and reduces costs associated with flooding events.