Machine Learning-Based Non-Revenue Water Management in Smart Cities: A Systematic Review of Leakage Detection, Prediction and Decision Support

Water entering a distribution system may be lost physically through leaks, lost economically through metering and billing issues, or used without being invoiced, making Non-Revenue Water (NRW) a recurring concern for metropolitan water utilities. In order to identify anomalous network behavior, locate leaks, and assist with operational decisions, machine learning (ML), the Internet of Things (IoT), advanced metering infrastructure (AMI), hydraulic modeling, and cloud/edge computing are rapidly being integrated. With a focus on leakage detection, localization, prediction, and decision support, this paper summarizes the development of data-driven NRW management. Data sources, pre-processing and feature engineering, algorithm families, validation procedures, and operational deployment are the five dimensions around which the evidence is arranged in the review, which follows an organized systematic-review approach based on PRISMA 2020. In addition to hydraulic-model residuals and network architecture, pressure, flow, acoustic, and AMI data are analyzed. While CNN, LSTM, and graph-based models are increasingly employed for spatial, temporal, and topological patterns, traditional machine learning models like support vector machines, random forests, and gradient-boosting techniques are still helpful when datasets are moderate and manufactured features are instructive. Moving from high-performing laboratory or simulated trials to resilient utility-scale deployment with sensor sparsity, noise, demand unpredictability, network changes, and limited labeled leak occurrences is another important research problem, according to recent studies. Thus, a closed-loop NRW management system that connects sensing, data quality control, machine learning inference, hydraulic validation, field verification, repair priority, and model retraining is proposed in the review. Standardized datasets, leakage-specific assessment methods, cross-network validation, explainable and physics-informed machine learning, cybersecurity, model-drift management, and economic decision support are important research needs. The study comes to the conclusion that machine learning (ML) should not be used as a stand-alone substitute for hydraulic engineering and field practice, but rather as a decision-support layer inside a larger water-loss-control program.