Just Getting Started—Beyond AI Main Break Prediction

The success of the machine learning (ML) approach for main break predictions offers several benefits to water utilities. Successful predictions supported by proper utility responses like replacement and monitoring offer the best chance to reduce the number of main breaks and the magnitude of leakage, a form of nonrevenue water (NRW). This presentation focuses on improvements on the consequence portion of the risk assessment, emphasizing water loss reduction. Individual pipe segments (usually defined by GIS) are given both probability and consequence of failure values. Machine learning has proven effective in prediction, especially identifying pipes with no prior break history with a high chance of failing. It is possible to focus on pipes most likely to break and subsequently look at the attributes of those pipes (material, size, vintage) and project the type of failure that is likely to occur based on the nature of the pipe, including the utility break history with that type of pipe. Some pipes with a high likelihood of failure will fail with a sudden potentially damaging break (like bell fractures of cast iron or longitudinal cracks of PVC). Other pipes will break with a lower flow with the potential to be long-running hidden breaks (like pinhole leaks or circumferential breaks). The analysis can then be advanced to identify geographical clusters of such breaks that might suggest acoustic monitoring in areas of hidden breaks or a renewal program in regions of potentially damaging breaks. Alternatively, anticipated clusters of breaks may be the sudden high flow, relatively low total volume type of break that causes significant damage to merit a high priority. The decisions to address either of these types of breaks remain at the discretion of the utility, but with a better prediction of not only which pipes will break but how they might break, the utility can strike an appropriate balance between NRW and main break reduction in both the short-term and the long-term. The applications of machine learning to forecast and address water system issues have two basic requirements: (1) a problem that involves assessing multiple elements that contribute to its occurrences and (2) the availability of pertinent and accurate data that can be applied to an analysis of past events to predict and as input to forecasting. This methodology has been successfully applied to predicting main failures, but this paper will look at other potential non-plant applications. This presentation will highlight how machine learning main break prediction is enhanced by predicting how pipe might fail. This step will advance a utility’s ability to address anticipated breaks by being aware of the probability the leak has the potential to be a highly damaging leak. Some types of leaks can produce service interruptions. Other leaks can be low flow, long-running leaks that generate significant water loss. Other opportunities to enhance asset management through machine learning processes are highlighted. These opportunities include the practical organization of main replacement projects, pressure management, and leak detection strategies designed to mitigate water losses more effectively.