Li et al. (2025) Intelligent and interpretable forecasting method for ice-jam flood disaster levels based on fusion model

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-06
• Authors: Yu Li, Hongwei Han, Fuchang Tian, Ximin Yuan, Dongxu Yang
• DOI: 10.1016/j.jhydrol.2025.134730

Research Groups

• State Key Laboratory of Hydraulic Engineering Intelligent Construction and Operation, Tianjin University, Tianjin, China
• School of Civil Engineering, Tianjin University, Tianjin, China
• School of Water Conservancy and Civil Engineering, Northeast Agricultural University, Harbin, China
• Heilongjiang Provincial Key Laboratory of Water Resources and Water Conservancy Engineering in Cold Region, Northeast Agricultural University, Harbin, China
• Ningxia Hui Autonomous Region Center for Flood and Drought Disaster Prevention and Control, Department of Water Resources of Ningxia Hui Autonomous Region, Yinchuan, China

Short Summary

This study proposes an intelligent and interpretable forecasting framework for ice-jam flood (IJF) disaster levels, integrating generative modeling, feature selection, and ensemble learning to address data scarcity and model interpretability challenges. The developed fusion model significantly improves forecasting performance and provides localized interpretations of risk scenarios.

Objective

• To develop an intelligent and interpretable forecasting framework for ice-jam flood (IJF) disaster levels, overcoming challenges of limited data, complex formation mechanisms, and low model interpretability.

Study Configuration

• Spatial Scale: Heilongjiang River (implied context of cold regions).
• Temporal Scale: Forecasting of disaster levels; specific historical data range or forecast horizon not detailed in the provided text, but focuses on "early warning systems."

Methodology and Data

• Models used: Dynamic Conditional Adversarial Variational Autoencoder with Gradient Penalty (DynaCondAVAE-GP), Mantel test, Recursive Feature Elimination with Cross-Validation (RFECV), SHapley Additive exPlanations (SHAP), Anti-Overfitting Tree-structured Parzen Estimator (AOTPE), Stacking ensemble with Validation-Weighted ElasticNet (VW-ENet), XGBoostR-FECV, k-means clustering.
• Data sources: Scarce historical observational datasets (enriched by DynaCondAVAE-GP).

Main Results

• The proposed DynaCondAVAE-GP–XGBoostR-FECV–VW-ENet fusion model achieved substantial performance gains.
• Root Mean Square Error (RMSE) was reduced by 86.74 % on the test set and 55.71 % on the validation set compared to the Random Forest (RF) model.
• The model surpassed the Stacking model with ElasticNet as the meta-learner.
• Dominant hazard factors identified were maximum ice thickness and downstream cumulative negative temperature.
• SHAP with k-means clustering enabled localized interpretation, revealing distinct feature contribution patterns for different risk scenarios.

Contributions

• Introduction of DynaCondAVAE-GP, a novel generative model for enriching scarce hydrological datasets with diverse, physically consistent samples.
• Development of an intelligent and interpretable IJF disaster level forecasting framework integrating generative modeling, robust feature selection, and ensemble learning.
• Identification and validation of key hazard factors (maximum ice thickness, downstream cumulative negative temperature) using Mantel test, RFECV, and SHAP.
• Enhancement of model generalization through a newly designed Anti-Overfitting Tree-structured Parzen Estimator (AOTPE) and adaptive ensemble learning (VW-ENet).
• Provision of localized interpretation of forecasted disaster levels using SHAP and k-means clustering, aiding in understanding risk scenarios.

Funding

• Not specified in the provided text.

Citation

@article{Li2025Intelligent,
  author = {Li, Yu and Han, Hongwei and Tian, Fuchang and Yuan, Ximin and Yang, Dongxu},
  title = {Intelligent and interpretable forecasting method for ice-jam flood disaster levels based on fusion model},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134730},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134730}
}