A Comprehensive Review of Machine Learning Techniques for Optimized Urban Water Systems (UWSs)

Urban Water Systems (UWSs) across the world are under mounting pressure as cities continue to expand, infrastructures age, and climate change introduces new levels of uncertainty into water availability and distribution. The traditional tools and hydraulic models that once guided urban water management are increasingly unable to cope with the highly dynamic, non-linear behavior of today's networks. These limitations have prompted a shift toward more adaptive and intelligent approaches capable of handling complex data environments. This review explores how Machine Learning (ML) is emerging as a powerful instrument for addressing these modern challenges. By examining research published largely in the past five years, the paper provides a structured overview of how ML has been applied to major UWS functions— such as forecasting future water demand, identifying leaks and pipe failures, monitoring water quality, and optimizing day-to-day system operations. A diverse range of techniques is discussed, from established learning models like Random Forests (RF) and Support Vector Machines (SVM), to more sophisticated deep learning methods including Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs). Across multiple studies, ML consistently surpasses traditional modelling approaches by learning intricate spatial and temporal relationships that conventional tools fail to capture. Certain algorithms have demonstrated notable advantages—for example, LSTM networks excel in predicting time-dependent water usage, while CNNs show strong performance in analyzing acoustic signals for leak detection. Beyond summarizing existing applications, the review highlights emerging themes and persistent gaps. Key concerns include the uneven availability and reliability of operational datasets, growing demands for data privacy and cybersecurity, and the ongoing challenge of interpreting decisions made by complex deep learning models. Additionally, hybrid frameworks, which combine the strengths of data-driven ML models with physically-based hydraulic models, are gaining interest as a promising direction for future research. Ultimately, the paper emphasizes the need for UWSs to evolve towards intelligent, autonomous, and sustainable systems. By integrating ML into standard practice—supported by robust data collection, careful preprocessing, and informed model selection—urban areas can significantly improve resource allocation, reduce water loss, and enhance overall resilience. This study demonstrates how data-driven strategies, when aligned with the realities of urban infrastructure, can play a pivotal role in shaping the next generation of efficient and sustainable urban water management.