Manage Water and Sewer Infrastructure Proactively Using AI: Case Studies

This presentation compares machine learning to traditional methods of assessing the health of water and wastewater infrastructure. It will illustrate the effectiveness of machine learning to help utilities make science-driven decisions on managing water and wastewater pipes (including identifying lead water pipes). Current methods rely on the age or prior history of incidents and failures. Machine learning is shown to be at least three times more accurate in assessing pipe health and up to sixty-five times more accurate in identifying problem areas and individual assets. Machine learning is a cost-effective, resilient strategy for utilities to improve asset management and reduce their impact on the planet. Nearly 1,000 water pipes fail every day in the U.S. (more globally), causing water loss, damages, and disruptions that demand immediate repair or replacement. An estimated six billion gallons of treated water is lost each day in the U.S., enough to fill over 9,000 swimming pools. Every break creates damage and disruption, and the total cost is billions in repairs, replacements, and water loss. The EPA deemed potable water in communities like Flint and Benton Harbor, MI, unsafe. At the same time, cities across the country have been served Consent Decree orders due to their failing wastewater networks. The EPA estimates that three million Americans become ill from exposure to water contaminated by wastewater incidents per year. Research shows water main breaks have increased 27% in the past six years. Globally, utilities rank water pipe replacements as their highest priority. Utilities need tools to help them make intelligent decisions about which pipes to monitor, repair, or replace, and, as importantly, which ones to leave alone. Digging up a pipe projected to fail but finding it healthy is not a good use of resources. Many utilities try to avoid failures by proactively replacing about one percent of their pipes every year. Often, they choose the pipes to prioritize pipe replacement using educated guesses on pipes that may leak or break. They use traditional methods to prioritize pipes: - Prior failure model – if a pipe has failed before, it's likely to break again, - Pipe age model – the older the pipe, the higher the risk of failure, - Statistical simulations that use key variables and assign weights to each one, - A combination of these approaches for targeted pipe materials or size. Usually, utilities try to find the worst one percent at risk and then take proactive steps based on available resources. They would inspect, monitor, repair, or replace the top one percent and use leak detection, condition assessment, or other activities on the next 3 or 4%. Case Study 1 is a large Western water district. They have about 5,000 miles of water mains broken into 260,000 pipe segments with unique identifiers. They were interested in machine learning but were skeptical of the usefulness of this new technology. The utility was already replacing 1% of their pipes annually using traditional models to choose which pipes to replace. Their leadership wanted to know if machine learning could do a better job. They engaged a machine learning firm to conduct a pilot. That firm asked for all active pipe details along with historical failure data - but to withhold the most recent year of failures. The idea was to predict the health of all the pipe segments for that year and have the utility validate the accuracy of the predictions. The results were impressive. Machine learning ranked each segment by likelihood of failure. In the top 1% of the highest risk segments, it predicted 50% of all the failures that year. Since the utility was already replacing 1% of the pipes annually, they could have avoided half the failures if they had these predictions. Equally impressive: of the breaks identified in the top 1%, 86% never failed before. The 'Prior Failure' model alone would have been a mediocre predictor, at least for this year. Every utility is unique in the mosaic of pipe infrastructure and failure history. Each has different combinations of pipe materials, ages, sizes, soil, weather, proximity to roads, railroads and bridges, rainfall, contractor skill, and even seismic activity. Machine learning doesn't 'care' but instead simply looks for patterns that led to prior failures and assigns probabilities of future health based on those patterns. Case Study 2 is a southern utility with a history of wastewater incidents. This is a network of 4030 miles of collection mains. Annually, this utility confronts over 800 'incidents' that disrupt lives and jeopardize public health. Similar to water management, the traditional approach to protecting health and sewer assets is to make proactive maintenance decisions based on age or history of prior incidents. In a comparative study, this utility evaluated results of three planning models: age, prior incidents, and machine learning. Each model ranked every pipe segment by probability of failure in a specific year. The results were remarkable. The pipe age model accurately predicted 12 pipe failures in the top 1% of high-risk pipes. The prior failures model accurately predicted 245 pipe failures in the top 1% pure. Machine learning accurately predicted 787 incidents, 65 times more accurate than the age model and more than three times more accurate than failure history. Utilities can apply machine learning to assess the risks for every pipe segment and make decisions based on science to manage their infrastructure. They can rely on this scientific approach and continue to do what they do best – serving their communities with reliable and safe water and sewer services.