Sun et al. (2026) Traceable Risk Evolution Forecasting for Irrigation Districts Driven by Enhanced Spatiotemporal Attention (ESTAM)

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-01-01
• Authors: Xinjuan Sun, Yongchao Zhu, Hairui Li
• DOI: 10.1007/s11269-025-04489-9

Research Groups

• School of Electronic Engineering, North China University of Water Resources and Electric Power, Zhengzhou, China
• Advanced Research Institute of Digital Twin Water Conservancy, North China University of Water Resources and Electric Power, Zhengzhou, China

Short Summary

This study proposes an enhanced spatiotemporal attention model (ESTAM) for traceable risk evolution forecasting in large-scale irrigation districts. The ESTAM achieves 91.88% prediction accuracy and provides causal diagnosis of primary risk drivers, significantly outperforming baseline models and enabling proactive risk management.

Objective

• To develop an enhanced spatiotemporal attention model (ESTAM) that can accurately forecast dynamic risks in large-scale irrigation districts, capture complex spatiotemporal dependencies, and critically, diagnose the causality of primary risk drivers to enable proactive and fine-grained risk management.

Study Configuration

• Spatial Scale: A large-scale Yellow River water diversion irrigation district covering 3600 km², with a primary canal network spanning 78 km, regulated by 13 main control gates and 39 branch gates, and typical operational flows of approximately 25 m³/s. The system is represented as a graph with N key monitoring units (nodes).
• Temporal Scale: Operational states simulated from January 1, 2023, to December 31, 2023, sampled at a 15-minute frequency. The model forecasts risk levels for future time steps based on historical observation sequences.

Methodology and Data

• Models used:
  • Proposed: Enhanced Spatiotemporal Attention Model (ESTAM)
  • Baseline models for comparison: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Graph Convolutional Network (GCN), Graph Attention Network (GAT), Spatiotemporal Graph Convolutional Network (ST-GCN).
• Data sources:
  • A high-fidelity digital twin of a large-scale Yellow River water diversion irrigation district, rigorously grounded in real-world physical parameters and calibrated operational data.
  • Synthesized dataset generated by simulating the system's operational states from January 1, 2023, to December 31, 2023, at a 15-minute frequency, integrating data from over a thousand sensors.
  • Dataset features include static, temporal, and real-time sensor data (e.g., water level, flow rate, rainfall), alongside historical, derived, and augmented spatiotemporal features.
  • Data augmentation techniques: Mixup for class imbalance and low-magnitude random noise addition to key sensor features.

Main Results

• The ESTAM achieved a superior risk level prediction accuracy of 91.88%, significantly outperforming baseline models (e.g., GCN at 90.0% and ST-GCN at 85.6%).
• ESTAM demonstrated the lowest prediction errors with a Root Mean Square Error (RMSE) of 0.367 and a Mean Absolute Error (MAE) of 0.103.
• For auxiliary diagnostic tasks, the model achieved high performance: 94.8% accuracy and 0.945 weighted F1-score for risk cause diagnosis, and 97.5% accuracy and 0.975 F1-score for upstream heavy rainfall assessment.
• Case studies confirmed ESTAM's interpretability and robustness:
  • Accurately reproduced spatiotemporal risk propagation from upstream rainfall events, with time lags consistent with expected hydraulic travel times.
  • Captured the nuanced relationship between long-term sediment accumulation and risk levels, showing sensitivity to short-term remediation efforts.
  • Successfully traced the causal chain of 'gradual-to-abrupt' gate failure, diagnosing both the latent root cause and its manifest secondary hazard (e.g., excessive water level).

Contributions

• Proposes ESTAM, a novel deep learning model featuring a parallel spatiotemporal attention architecture that effectively captures both topological propagation and dynamic temporal dependencies of risks in irrigation districts.
• Introduces a multitask learning paradigm that enables not only accurate risk prediction but also the crucial causal diagnosis of primary risk drivers, addressing the "black-box" limitation of conventional models.
• Achieves state-of-the-art prediction accuracy and robust diagnostic capabilities, significantly outperforming established baseline models in complex spatiotemporal forecasting tasks.
• Provides a traceable framework that translates complex data into actionable causal insights, enabling proactive, fine-grained interventions and enhancing the operational efficiency and resilience of water supply systems.
• Validates the model's practical utility and interpretability through realistic case studies, demonstrating its ability to handle diverse risk scenarios like rainfall propagation, siltation, and equipment failures.

Funding

• National Key R&D Program of China (2024YFC3210800).

Citation

@article{Sun2026Traceable,
  author = {Sun, Xinjuan and Zhu, Yongchao and Li, Hairui},
  title = {Traceable Risk Evolution Forecasting for Irrigation Districts Driven by Enhanced Spatiotemporal Attention (ESTAM)},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04489-9},
  url = {https://doi.org/10.1007/s11269-025-04489-9}
}