Water Quality Control in Distribution Systems: Bayesian Optimization & Physics-Informed Machine Learning

Water distribution systems (WDSs) play a vital role in maintaining public health by ensuring the continuous supply of safe drinking water. One of the most important operational challenges in WDSs is maintaining adequate chlorine residuals throughout the network to prevent microbial regrowth while minimizing the formation of harmful disinfection byproducts (DBPs). Achieving this balance requires optimizing chlorine dosage strategies, both in terms of injection locations and dosing schedules, under complex, dynamic, and spatially distributed network conditions. Despite recent advancements in sensing and modeling, real-time control of the water quality (WQ) in WDSs remains difficult to achieve, primarily because of the limited WQ monitoring coverage and the intensive computational effort required for repeated WQ simulations. The latter have traditionally been performed by means of physics-based models that involve solving complex, nonlinear systems of partial differential equations (PDEs) to simulate the underlying physical processes that govern chlorine transport and decay in the WDS. This dissertation aims to address these key challenges by developing innovative WQ prediction and control frameworks to enable efficient and sustainable chlorine residual management in WDSs. To achieve this objective, this dissertation presents novel approaches for designing advanced physics-informed machine learning (PI-ML) models for WQ prediction, and implementing the Bayesian optimization (BO) technique for optimizing the WQ in WDSs. The first thrust of this dissertation introduces a novel BO-based framework for optimizing chlorine booster scheduling in WDSs. The proposed framework integrates BO with a physics-based WQ simulation model, namely EPANET-MSX (multi-species simulation engine), to determine optimal chlorine injection rates and timing. Using Gaussian Process Regression (GPR) as a surrogate data-driven model, this study systematically investigates the effects of various acquisition functions, such as Expected Improvement (EI), Probability of Improvement (PI), Upper Confidence Bound (UCB), and Entropy Search (ES), along with multiple covariance kernels, including Matérn, squared-exponential, gamma-exponential, and rational quadratic. A comprehensive sensitivity analysis further examines the influence of key BO hyperparameters, such as the kernel length scale, initial sampling size, and exploration–exploitation trade-off. The results demonstrate that BO can efficiently identify optimal disinfection schedules with substantially fewer WQ simulations than conventional methods, with the choice of acquisition function showing the most significant impact on optimization performance. The second thrust of this dissertation builds on the first thrust by rigorously benchmarking BO against traditional evolutionary algorithms (EAs) for multi-species WQ control. The latter involves modeling the complex interactions between chlorine and other chemical or microbiological species, offering a more realistic yet computationally challenging representation of water chemistry. A comprehensive comparative study was conducted to assess BO against two widely used EAs), namely the Genetic Algorithm (GA) and Particle Swarm Optimization (PSO), for optimizing multi-species chlorine dosage in a benchmark WDS. The results reveal that BO achieves superior efficiency, requiring substantially fewer simulations and less computation time per iteration while maintaining or exceeding the optimization performance of EAs. Moreover, BO exhibited lower sensitivity to variations in the constraints imposed on species concentrations, highlighting its robustness and adaptability to complex operational scenarios. The third thrust of this dissertation transitions from optimization to prediction by introducing a novel Physics-Informed Machine Learning (PI-ML) framework for simulating WQ dynamics in WDSs. While traditional physics-based models (PBMs) rely on solving partial differential equations (PDEs) representing advection, dispersion, and reaction processes, they are computationally expensive and limited in scalability. Conversely, purely data-driven ML models, although efficient, lack physical interpretability and generalizability across networks. To address this gap, this study develops an ensemble PI-ML framework that integrates fundamental physical principles into the architecture of ML models. The PI-ML ensemble effectively captures the key transport and reaction mechanisms governing the WQ dynamics while retaining computational efficiency. Overall, this dissertation advances the field of drinking water quality management by introducing a suite of computationally efficient, scalable, and physics-aware frameworks for chlorine residual prediction and optimization. The proposed BO-based and PI-ML approaches bridge the gap between physical modeling and data-driven learning, offering practical and generalizable tools for real-time water quality control. These contributions lay the groundwork for a paradigm shift towards intelligent, adaptive, and resilient water supply systems that ensure safe and efficient water delivery under varying operational conditions.

Author dc:creator

• Mohammadreza Moeini (23291662)

• Engineering, Civil
• Computer Science
• Artificial Intelligence
• Engineering, Environmental

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.31451431.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451431