From Traditional Simulation to Machine Learning: A Comparison of EPANET and Physics-Informed Neural Networks for Water Quality Modeling

Modeling water quality in water distribution systems (WDSs) is critical for ensuring safe drinking water and managing contamination risks. Traditional tools, such as EPANET, rely on numerical methods to simulate water quality dynamics, but face challenges in adaptability. Physics-informed neural networks (PINNs), a novel machine learning approach, embed governing equations directly into neural network architecture, enabling efficient simulations with minimal data. This paper presents a comparative analysis of EPANET and PINNs in modeling chlorine transport in a simplified WDS network. The study assumes constant chlorine concentration at the source, with no reactive losses, and evaluates the ability of PINNs to replicate EPANET results. While PINNs demonstrated consistency with EPANET simulations, the simplified nature of the case study limits broader conclusions about their performance in complex systems. Nevertheless, the study highlights PINNs as a promising framework for future applications in water quality modeling, with potential benefits such as enhanced resolution and flexibility for spatiotemporal analysis. Future research should focus on applying PINNs to more complex networks and incorporating dynamic and reactive processes to fully explore their capabilities.