Answers are grounded in this paper and cite their sources.
Improving overall streamflow accuracy does not necessarily improve forecasts of extreme high flows. This study develops a hydrologically informed dual-expert framework for 1-, 3-, 5-, and 7-day-ahead streamflow forecasting at Yichang Station in the upper Yangtze River. A general-flow expert combines an LSTM with temporally ordered out-of-fold (OOF) XGBoost residual correction, while a separate LSTM trained with a Peak-Balanced Huber (PBHuber) loss serves as the peak-flow expert. A soft gate combines their predictions using information available at forecast issuance. Daily streamflow and basin-scale meteorological data for 2000–2020 were used, and each model was evaluated using the same 30 random seeds. Relative to the baseline LSTM, the framework reduced mean RMSE by 0.93%, 10.26%, 9.93%, and 4.87% and mean P95-MAPE by 5.63%, 12.78%, 17.20%, and 22.55% at the four lead times, respectively. Ablation analysis identified complementary contributions from residual correction, peak-oriented learning, and dynamic fusion, with the clearest aggregate gains at H3–H7. SHAP attribution became less concentrated on current Yichang streamflow and more distributed across upstream streamflow and meteorological predictors as lead time increased. These attributions describe model behaviour rather than causal hydrological effects. The framework thus offers a lead-time-dependent trade-off between aggregate accuracy and extreme-flow error control, with limited additional aggregate benefit at H1.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: