Comparing different Physics-Informed Neural Network models for chlorine modeling in EPANET under varying initial and boundary conditions

Accurate water quality modeling in water distribution systems (WDS) is essential for ensuring safe and reliable drinking water. While numerical solvers such as EPANET provide robust simulations, their computational cost increases substantially for real-time or large-scale applications, particularly when boundary and initial conditions vary over time. Existing Physics-Informed Neural Network (PINN) approaches face limitations in handling such changing conditions, despite their prevalence in real WDS operations. This study focuses on enhancing the adaptability of PINNs for chlorine modeling under diverse and dynamic scenarios. The proposed framework embeds the governing Advection–Reaction (AR) equation into a deep learning architecture and introduces targeted modifications to the formulation of boundary and initial condition losses. Training data are generated using EPANET simulations, and the framework is evaluated under multiple scenarios, including constant and time-varying velocities as well as fixed and dynamic boundary and initial conditions. Results demonstrate that a PINN model explicitly designed for boundary-condition adaptability can accurately reproduce EPANET water quality simulations while reducing computational demands. Key factors influencing performance, such as proper PDE specification, loss balancing, and data preprocessing, are identified. Although the analysis is conducted on a single-pipe testbed to isolate these effects, the findings establish an essential foundation for extending adaptive PINNs to full WDS networks. The primary contribution of this work is the development and demonstration of a PINN architecture capable of reliably adapting to varying boundary and initial conditions, addressing a critical gap in current PINN-based water quality modeling research.