Enhancing event-based flood forecast stability and physical interpretability via Reg-TCN-LSTM coupled deep learning with discharge regularization constraints

Study region The Anyang River Basin, Henan Province. It’s a representative data-scarce, flash-flood-prone small-medium basin with mixed urban-plain flooding. Study focus To address limitations of deep learning in capturing local dynamics, long-term dependencies, error accumulation, and weak physical constraints at longer lead times, we propose a lightweight discharge-regularized coupled model, Reg-TCN-LSTM, combining TCN's local feature extraction with LSTM's long-term memory. A flow-magnitude regularization term softly constrains predicted discharge amplitudes, enhancing stability and physical consistency without added complexity. Using 20 flood events split 7:3 into training and test sets, we evaluated its 1–12 h forecasting performance against TCN, LSTM, and TCN-LSTM benchmarks with metrics including NSE , RMSE , RE , andΔt. New hydrological insights for the regions Reg-TCN-LSTM achieved the best overall performance across all lead times ( NSE : 0.574–0.969; RMSE : 7.06–26.25 m³/s), improving NSE by up to 14.34% over TCN, LSTM, and TCN-LSTM, while reducing peak and timing errors. It better preserved flood hydrographs and showed enhanced stability at 9–12 h lead times. Gradient-based attribution reveals that TCN dominates short-lead forecasts by capturing local dynamics, whereas LSTM gains importance at longer lead times via antecedent hydrological memory. The flow-magnitude regularization term was particularly effective during flood peaks. Overall, Reg-TCN-LSTM offers improved accuracy, stability, and interpretability with low complexity, providing a practical solution for real-time forecasting in data-scarce small-to medium-sized basins.