Towards transferable metamodels for water distribution systems with edge-based graph neural networks

Area of Science:

• Hydraulic Engineering
• Artificial Intelligence
• Network Science

Background:

• Data-driven metamodels accelerate simulations of water distribution systems.
• Graph Neural Networks (GNNs) show promise for modeling these systems due to their graph structure.
• Current GNN metamodels struggle with transferability due to limited edge-level process representation.

Purpose of the Study:

• Introduce Edge-Based Graph Neural Networks (EB-GNNs) to enhance metamodel transferability.
• Improve the representation of physical processes occurring at the pipe (edge) level.
• Evaluate EB-GNN performance against traditional GNNs for water distribution system emulation.

Main Methods:

• Developed an EB-GNN architecture incorporating edge-level physical process representation.
• Emulated steady-state EPANET simulations for water distribution networks.
• Compared EB-GNNs and traditional GNNs on benchmark systems for accuracy and speed-up.
• Assessed transferability by testing models on unseen network topologies.

Main Results:

• EB-GNNs more accurately capture pipe-level physical processes compared to node-based GNNs.
• Achieved high generalization performance on unseen networks (R² up to 0.98 for flowrates, 0.95 for heads).
• Demonstrated significant speed-up compared to traditional physics-based simulations.

Conclusions:

• EB-GNNs offer improved accuracy and transferability for water distribution system metamodeling.
• The proposed architecture effectively represents link-level dynamics, addressing limitations of traditional GNNs.
• EB-GNNs show potential for efficient design, control, and optimization of water networks, especially with scarce data or numerous layout evaluations.