Applications of machine learning in drinking water quality management: A critical review on water distribution system

As the final and crucial link in delivering clean water to consumers, the water distribution system faces the risk of water quality deterioration. Conventional water quality parameter monitoring and simple analysis may not adequately reflect complex changes in distribution. Machine learning (ML) excels at uncovering the intricate relationships among these. Although some reviews exist on ML in water resources, a systematic assessment of water quality in water distribution systems is lacking. The current review offers the first critical and comprehensive review of the application of ML in water quality management within water distribution systems, including water quality prediction, anomaly detection, and contamination source identification, and addresses the associated challenges and future directions. To be specific, for water quality prediction, the focus is on chlorine, disinfection by-products, microbial indicators, heavy metals, and sensory properties. The implementation of ML has the potential to reduce the cost of water quality monitoring and offer knowledge discovery. For anomaly detection, semi-supervised, supervised, and unsupervised models are reviewed. Changes in one or more surrogate water quality parameters measured by low-cost sensors can effectively indicate anomalous events. For contamination source identification, ML demonstrates its superiority in rapidly and accurately locating contamination sources. Additionally, dataset availability, interpretability, generalization capability, integrated models, real-time response and proactive decision have been identified as key areas for implementing ML. This review helps bridge the knowledge gap and provides a reference for the intelligent development of water quality management in distribution systems. [Display omitted] •Comprehensive review of the application of machine learning in WDS water quality management for the first time.•Machine learning uses conventional water quality parameters to analyze resource-intensive processes or estimate the system's state.•Graph Neural Networks show great potential in combining topology and water quality data for water quality management.•Key challenges include dataset availability, interpretability, generalization capability, integrated models, real-time response and proactive decision.