Dual-constraint physics-informed LSTM-XAJ hybrid model for 10-minute flood forecasting in small-to-medium basins: A case of the Xiang River Basin, China

Study region Luolingqiao (326 km²) and Wenjiashi (164 km²), two small-to-medium sub-basins of the Xiang River Basin in Hunan Province, China. Study focus Reliable flash flood forecasting in small-to-medium basins (SMBs) requires both high temporal resolution and physically constrained long-lead prediction. This study develops a dual-constraint physics-informed framework that couples the Xin'anjiang (XAJ) model with an encoder-decoder LSTM to produce 10-minute runoff forecasts at lead times of up to 6 h, forced exclusively by quality-controlled X-band radar quantitative precipitation estimation (QPE). XAJ-derived physical trajectories are supplied to the LSTM decoder, and a dual-constraint loss penalizes deviations from both observed discharge and XAJ-simulated runoff. The framework is evaluated on 28 flood events under blocked chronological cross-validation, in hindcast mode with observed QPE over the forecast horizon. New hydrological insights for the region At the 6-h lead time, the hybrid model significantly outperforms the standalone XAJ model: relative to XAJ, NSE increases by 20%, RMSE decreases by 25%, and peak error decreases by 46%. Continuous evaluation from 1 to 6 h reveals a crossover near 3 h: the LSTM dominates at shorter horizons, whereas the hybrid model becomes optimal beyond 3–4 h, where physical regularization suppresses recursive error accumulation. Coupling process-based constraints with deep learning at 10-minute resolution is therefore feasible and valuable for flash flood early warning in SMBs with concentration times shorter than 3 h.