Yin et al. (2026) Exploring a process-aware spatiotemporal graph-based surrogate for integrated urban drainage simulation

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-20
• Authors: Boyan Yin, Ruidong Li, Baoxiang Pan, Guangheng Ni
• DOI: 10.1016/j.jhydrol.2026.135350

Research Groups

• State Key Laboratory of Hydroscience and Engineering & Department of Hydraulic Engineering, Tsinghua University, Beijing, China
• National Key Laboratory of Earth System Numerical Modeling and Application, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China

Short Summary

This study proposes PAST, a Process-Aware Spatio-Temporal graph-neural-networks-based surrogate model, to efficiently simulate integrated urban drainage processes by holistically representing rainfall–runoff-routing and incorporating regulation effects. PAST achieves high performance and physical explainability, significantly outperforming baseline models, especially under regulated and extreme rainfall conditions.

Objective

• To develop an efficient, integrated, and interpretable surrogate model for urban drainage systems that addresses limitations in representing rainfall–runoff-routing processes, encoding drainage network regulation, and providing physical explainability.

Study Configuration

• Spatial Scale: A residential area with moderate imperviousness (specific dimensions not provided).
• Temporal Scale: Event-based, real-time simulation of rainfall-runoff and hydrodynamic routing processes within urban drainage systems.

Methodology and Data

• Models used:
  • Process-Aware Spatio-Temporal (PAST) graph-neural-networks-based surrogate model.
  • Long Short-Term Memory (LSTM) networks for rainfall–runoff simulation.
  • Graph Attention Networks with LSTM (GAT-LSTM) for hydrodynamic routing.
  • Adaptive connectivity layer for embedding element-level control rules.
• Data sources: Simulated urban drainage data (rainfall, runoff, water depth, total inflow) from an underlying Urban Drainage Model (UDM).

Main Results

• PAST achieved high performance with Nash-Sutcliffe Efficiency (NSE) values of 0.96 for water depth and 0.88 for total inflow simulations.
• The model substantially outperformed baseline models, particularly under regulated conditions and during extreme rainfall events.
• Attention-based explainability analysis revealed that PAST learns physically plausible dynamics by allocating more attention weights to nodes contributing higher inflows.

Contributions

• Advances integrated, interpretable, and efficient surrogate modeling for urban drainage systems.
• Introduces a process-aware graph neural network that holistically simulates hydrological and hydrodynamic processes.
• Incorporates an adaptive connectivity layer to plausibly embed element-level control rules, enhancing simulation of regulation effects.
• Provides physical explainability through attention mechanisms, revealing relationships between neural attention and inflow contribution.

Funding

• Not specified in the provided text.

Citation

@article{Yin2026Exploring,
  author = {Yin, Boyan and Li, Ruidong and Pan, Baoxiang and Ni, Guangheng},
  title = {Exploring a process-aware spatiotemporal graph-based surrogate for integrated urban drainage simulation},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135350},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135350}
}