Accelerating hydrodynamic simulations of urban drainage systems with physics-guided machine learning

Area of Science:

• Environmental Engineering
• Hydraulic Engineering
• Machine Learning

Background:

• Urban drainage systems require accurate hydraulic modeling for design and operation.
• High-fidelity (HiFi) hydrodynamic models provide detailed simulations but are computationally expensive.
• Faster surrogate models are needed for real-time applications and interactive design.

Purpose of the Study:

• To develop and demonstrate a novel, fast, and accurate surrogate modeling approach for urban drainage system hydraulics.
• To leverage physics-guided machine learning trained on limited HiFi model data.
• To preserve the detailed simulation output of HiFi models at reduced computational cost.

Main Methods:

• Physics-guided machine learning was employed to create surrogate models.
• Surrogates were trained using a limited dataset from a high-fidelity hydrodynamic model.
• The approach focuses on simulating water levels, flows, and surcharges across the entire network.

Main Results:

• Simulation times were reduced by one to two orders of magnitude compared to HiFi models.
• R-squared values of approximately 0.9 were achieved when comparing surrogate and HiFi model time series.
• Surrogate training times were approximately one hour, with potential for further reduction.

Conclusions:

• The developed surrogate modeling approach offers a significant speed-up for urban drainage hydraulics simulations.
• It maintains a high level of detail comparable to HiFi models, making it suitable for various applications.
• The generic model formulation suggests potential for application to other water systems beyond urban drainage.