Accurate short-lead river discharge forecasting is critical for effective flood risk mitigation in small mountainous basins, where rapid hydrological responses pose significant challenges. In this study, we focus on the Fuhu Stream in Emeishan City, China, and utilize high-resolution 5-min time series of upstream precipitation, stage, and discharge to predict downstream flow. We benchmark three data-driven models—SARIMAX, XGBoost, and LSTM—using a dataset spanning from 7 June 2024 to 25 October 2024. The data were split chronologically, with observations from October 2024 reserved exclusively for testing to ensure rigorous out-of-sample evaluation. Lagged correlation analysis was employed to estimate travel times from upstream to the basin outlet and to inform the selection of time-lagged input features for model training. Results during the test period demonstrate that the LSTM model significantly outperformed both XGBoost and SARIMAX across all regression metrics: it achieved the highest coefficient of determination (R2 = 0.994) and the lowest prediction errors (RMSE = 0.016, MAE = 0.011). XGBoost exhibited moderate performance, while SARIMAX showed a tendency toward mean reversion and failed to capture low-flow variability. Accuracy evaluation reveals that LSTM accurately reproduced both baseflow conditions and multiple flood peaks, whereas XGBoost and SARIMAX failed. These results highlight the advantage of sequence-learning architectures in modeling nonlinear hydrological propagation and memory effects in short-term discharge dynamics. Feature importance analysis indicates that the LSTM model was highly effective for real-time forecasting and that the WSQ/LY sites served as critical monitoring nodes for achieving reliable predictions. This research contributes to the operationalization of early warning systems and provides support for decision-making regarding downstream flood disaster prevention.
Guo et al. (Tue,) studied this question.