Machine Learning Driven Source Identification, State Estimation and Sensor Optimization in Water Systems

Water distribution and sewer systems are vital infrastructure systems that play a key role in safeguarding public health and environmental quality. In water distribution systems, operators face difficulty maintaining disinfectant residuals within a narrow range to ensure microbiological safety while avoiding toxic disinfection byproducts. Meanwhile, sewer system operators are challenged by identifying and eliminating illicit discharges that could compromise wastewater treatment processes or contaminate receiving water bodies. In general, both systems share a fundamental constraint: limited coverage of water quality sensors, resulting from their high deployment costs, technical challenges, and logistical constraints. This constraint hinders comprehensive knowledge of systems’ parameters and obstructs operational decision-making. This dissertation addresses these challenges by introducing innovative machine learning (ML) driven frameworks to tackle three problems: i) Source Identification (SI) of contaminants in sewer systems, ii) State Estimation (SE) of water quality parameters in water distribution systems, and iii) Sensor Placement Optimization (SPO) in both systems. Traditional methods proposed to address these problems have been constrained by a number of simplifying assumptions that limit their practical applicability, such as considering only single, non-reactive contamination sources or static sensor configurations. In addition, these methods typically rely on computationally intensive physics-based models like EPA-SWMM and EPA-NET. As a result, these conventional methods lack the scalability required for real-world applications. This dissertation advances sewer system monitoring by introducing a novel Multi-Layer Perceptron Neural Network (MLP-NN) surrogate model that effectively emulates the physics-based EPASWMM, enabling computationally efficient water quality simulations. This MLP-NN model is then integrated within a Genetic Algorithm (GA) optimization framework to enable real-time SI with high identification accuracy. The SI model is subsequently incorporated into an SPO framework that introduces two key performance metrics: observability and reliability, providing valuable insights into the inherent trade-offs between the system's ability to detect contamination events (observability) and source characterization accuracy (reliability). The SPO framework demonstrates how these metrics vary with sensor location and network configuration. In addition, the developed SPO framework offers practical solutions for optimal sensor deployment. In the field of water distribution systems, this dissertation presents one of the first attempts to apply Graph Neural Networks (GNN) to estimate water quality parameters at unmonitored junctions. This was achieved by developing two GNN models. The first is a Static Prediction GNN (SP-GNN) model, which provides accurate state estimation for fixed sensor configurations. The second is a Dynamic Prediction GNN (DP-GNN) model, which achieves generalized state estimation across any sensor configurations without the need for retraining. The DP-GNN model enables the implementation of one universal model to diverse sensor designs and lays the foundation for GNN application in SPO. This dissertation also proposes an advanced Temporal Graph Neural Networks (TGNN) model that integrates Long Short-Term Memory (LSTM) within GNNs to assimilate data from mobile sensors. This TGNN model enables comprehensive spatiotemporal coverage, which is required to perform unsteady state estimation of chlorine concentrations resulting from dynamic water demand patterns. This dissertation advances water and wastewater infrastructure research by demonstrating the capabilities of utilizing ML techniques in water quality monitoring in drinking water and sewer networks. Additionally, this thesis contributes to the practical application of ML techniques in real-life water systems by bridging the gap between academic research and field deployment, reducing computational costs, and considering real-life challenges. Generally, the frameworks developed in this dissertation provide water operators with practical tools for enhanced system monitoring while establishing a foundation for future research. Ultimately, this work represents a step toward safer, more efficient, and resilient water systems.

Author dc:creator

• Aly Khaled Aly Salem (23291299)

• Machine Learning Driven Source Identification
• State Estimation
• and Sensor Optimization in Water Systems

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.31451038.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451038