Learning-Based Analogous Flood Recommendation Model

Identifying hydrologically similar historical floods is critical for advancing flood prediction and risk management. Current approaches often lack robustness in feature representation and similarity quantification for high-dimensional flood series. This study introduces a novel hierarchical model that synergizes clustering with a deep Siamese network to address these gaps. The proposed framework operates in a coarse-to-fine manner: first clustering 27 flood characteristics via PCA‑driven K‑means to define six representative flood types, followed by a Siamese network with Bi‑LSTM and contrastive learning to perform refined similarity matching within each cluster. Validated over 20 flood events across four stations in the lower Jinsha River, the model consistently achieves high-performance metrics—mean correlation coefficient of 0.95, peak flow error below 1.8%, NSE above 0.85, and R² exceeding 0.90—with perfect matching in flood duration. The results confirm the model’s capability to reliably capture dynamic flood signatures and its transferability across diverse basin settings, thereby offering an effective and generalizable tool for operational flood forecasting and risk-informed decision-making.