Hybrid surrogate of urban collection system¶
Type of project¶
M.Sc.
Contact¶
Rocco Palmitessa (rpal@dhigroup.com)
Topic¶
Traditional physics-based hydrodynamic models for urban collection systems (e.g., sewer and stormwater networks) are highly accurate but computationally expensive, limiting their use in real-time control (RTC) and early warning systems. While data-driven surrogate models can offer massive speedups, they often struggle with the complex spatial topology of pipe networks and the sudden, discrete flow changes caused by active human-controlled infrastructure (like pumps and gates), leading to unstable long-term forecasts.
Possible Research Questions GNN and Control Integration: How can Graph Neural Networks (GNNs) be effectively coupled with rule-based control logic to accurately simulate both the spatial network hydrodynamics and the discrete state changes of active assets (e.g., pumps, controllable weirs)? * Stable Multi-Step Forecasting: What training strategies or architectural modifications (e.g., autoregressive training, temporal attention mechanisms) are most effective for preventing error accumulation and ensuring stable multi-step predictions over extended forecast horizons? * Scenario Robustness:* How robust is the hybrid surrogate model when subjected to unseen operational scenarios and extreme rainfall events, and how does its computational efficiency and accuracy benchmark against a full-scale traditional hydrodynamic model?
Relevance to DHI¶
Fast and reliable simulations are critical for real-time control, digital twins, and early warning systems in urban drainage. A validated, stable hybrid surrogate model would allow DHI to offer near-instantaneous network forecasting, optimizing pump operations and reducing urban flood risks without sacrificing the accuracy of their industry-standard physics-based engines (like MIKE+).
What is the student's contribution¶
The student will improve the existing design of our GNN-based surrogate, train, and validate it using simulation data from a calibrated urban hydrodynamic model provided by DHI. They will benchmark the surrogate's multi-step stability and predictive performance against the baseline physics model across various rainfall and operational scenarios.
References¶
Garzon, A., Kapelan, Z., Langeveld, J., Taormina, R., 2025. Evaluation of graph neural
networks for urban drainage metamodeling: key components and transferability
analysis. Water Res., 125079