Multiple leaks location in water distribution networks combining physically-based and machine learning approaches using high-performance computing

Water loss has increasingly become a major concern for water utilities. Loss reduction is an opportunity to reduce intermittency and improve coverage in water-scarce locations, enhance the utility’s sustainability, and reduce operating costs and risks. However, localising leaks is a complex task. On the one hand, although the reactive strategy of repairing leaks when they become visible is needed, most of the leaked water remains invisible. As a consequence, Non-Revenue-Water values could be persistently rising, and these are difficult to reduce. On the other hand, utilities that have implemented a proactive strategy still face the challenge of selecting an appropriate network of a limited number of conveniently-placed sensors. Unfortunately, enough and well-located instrumentation is only one part of an effective leak location result; the other part involves the proper use of data to obtain information that eases the pinpointing activities. In this regard, extensive research has been done using different approaches, such as inverse models, pressure residuals and inverse transient. This study proposes an approach to the leak location problem by combining analytic and data-based methods, resulting in a physically-based machine-learning procedure. It involves a representative hydraulic model of a water distribution network as a basis for pressure sensitivity analysis and classification algorithms to identify the likely leaking area. Three stages can be recognised in the procedure. The first stage corresponds to single-leak location, which is essentially based on the current research practice. For this matter, the hydraulic model is used in the first place to obtain pressure data from the non-leak scenario. Then, single leaks are placed in all nodes, one at a time, altering the leak flow rate. Later, the difference between the leak and non-leak situations is used to construct pressure residual vectors at the sensor nodes. Finally, a modified classification algorithm based on KNN, selected after several experiments, identifies the most likely leaking area. The second stage pushes forward the current practice by exploring the location of multiple leaks. In principle, two leaks are placed simultaneously using High-Performance Computing. Although the pressure residuals are calculated following the same procedure used for a single leak, an additional step is included, consisting of grouping the nodes based on the impact of their leaks on the sensor nodes. For this matter, a clustering procedure involving Hierarchical Agglomerative clustering is proposed. This methodology is applied to the case study of Brăila, Romania. A District Metered Area (DMA) containing approximately 24.6 km pipes, multiple inlets, and eight pressure sensors has been selected within the NAIADES Project for this purpose. When the 95th percentile was utilized as a selection criterion, prediction accuracies of 30% were achieved for single leaks between 0.05 and 0.5 L/s, while leaks exceeding 10 L/s were accurately found in all cases. However, when two simultaneous leaks were modelled, accuracy levels dropped. In the best-case scenario, 76% of the leaks were discovered when the total flow rate was 16–18 L/s. Then, it was concluded that accuracy values prediction is affected by three factors. First, the leak magnitude, as there is a direct relationship between the leak flow rate and the prediction accuracy. Second, the place where the leak occurs, as some nodes may be undetectable. Third, the allocation and coverage of the pressure sensors. Overall, it can be concluded that the presented approaches have a good potential for locating leaks in WDS of various types. However, putting them into practice is challenging as they require complete and calibrated models, which are not always readily available. Furthermore, networks with multiple inlets are redundant, and leak detection may result in less accurate predictions than single-inlet systems since the pressure drop caused by a particular leak is often smaller. It is acknowledged that the proposed approaches should be tested in further case studies or benchmarking networks for performance assessment. In addition, further research is required to push forward these approaches to a higher number of simultaneous leaks, considering the increase in the volume of data and computational requirements