Water Main Likelihood of Failure Machine Learning Model

Water Main Likelihood of Failure Machine Learning ModelAbstractAging water infrastructure can result in unexpected water main breaks, adversely affecting utility service reliability and leading to substantial costs. To address this challenge, we developed a predictive machine learning model to assess the Likelihood of Failure (LoF) for water mains proactively. Using a comprehensive dataset from the Virginia Beach Department of Public Utilities-comprising historical main break data, hydraulic metrics, and environmental factors-this model identifies patterns and predictors of water main breaks, assigning a probability of failure to each uninspected main with a foundational accuracy exceeding 85%. Importantly, our approach embraces a 'human-in-the-loop' philosophy, not only in the machine learning design but also in a detailed sensitivity analysis of key parameters. Additionally, the city conducted a Consequence of Failure (CoF) assessment to enhance understanding and recognize data limitations. Such collaboration enhances the efficacy of the ML model derived from utility data. Figures 1 to 3 show some of the explanatory analysis results for the data. Beyond its predictive capabilities, the model fosters a deeper understanding of factors influencing failure, including pipe length, material, meteorological data, and soil metrics. A notable outcome was the utilization of Shapley Additive Explanation (SHAP) plots (Figure 4). These plots revealed, for instance, that pipe diameter plays a significant role in predicting failure: smaller diameters correlate with a higher failure probability in Virginia Beach's dataset. The final variable weightings will be refined based on dataset limitations and engineering expertise. The insights garnered from understanding significant variables and predictions of water main failures offer multiple advantages for utilities: 1.Enhanced service reliability through targeted attention to high-risk zones 2.Decreased unplanned maintenance and labor expenses 3.Informed, strategic infrastructure investment decisions Machine learning empowers utilities to process, analyze, and cross-reference intricate datasets efficiently. Compared to traditional mechanistic models, ML models elucidate variable interrelations more quickly and thoroughly. With the potential for delivering precise outcomes swiftly, utilities can transition from a reactive stance to a more proactive maintenance and operational approach. As standardized practices evolve for leveraging utility data, machine learning is poised to become an indispensable tool, driving service quality, cost-effectiveness, and operational efficacy.This paper was presented at the WEF/AWWA Utility Management Conference, February 13-16, 2024.SpeakerRoostaei, JavadPresentation time08:30:0010:00:00Session time08:30:0010:00:00SessionManagement and Oversight of New Frontiers in Artificial IntelligenceSession number16Session locationOregon Convention Center, Portland, OregonTopicDigital Transformation including AI and ChatGPTTopicDigital Transformation including AI and ChatGPTAuthor(s)Roostaei, JavadAuthor(s)J. Roostaei1, P. Wilson1, M. Turner1, S. Cook2, D. Piron2, K. Bilyk1, R. Cagir1Author affiliation(s)Hazen and Sawyer 1; Virginia Beach Public Utilities 2;SourceProceedings of the Water Environment FederationDocument typeConference PaperPublisherWater Environment FederationPrint publication date Feb 2024DOI10.2175/193864718825159317Volume / Issue Content sourceUtility Management ConferenceWord count9