A novel hybrid approach for leakage area detection in water distribution networks

The demand for drinking water has noticeably increased due to ever-increasing population growth and climate change.  In addition, a significant amount of water is lost from water distribution networks (WDNs) through leakages because of pipe aging, corrosion, structural flaws, etc. Consequently, detecting leakage in WDNs can play an essential role in maintaining sustainable water-supply infrastructures.Recently, hybrid approaches combining different methods, such as statistical, hydraulic, and machine learning (ML) methods, have attracted significant attention among the scientific community. In light of the above, this study presents a hybrid method for leakage detection that relies on two steps: i) a robust ML-based procedure to identify the leakage area, and ii) an optimization algorithm to further narrow down the suspect leakage area detected in the first step.In the first step, the WDN is divided into leakage areas using the k-means clustering algorithm. Then, a Support Vector Machine (SVM) algorithm is used to identify the suspect leakage area. Pressure differences (i.e., pressure differences between 24h leakage scenarios and the daily average pressure of the leakage-free scenario as baseline condition) are considered as input, and leakage areas are assumed as targets for the SVM algorithm. Moreover, noise related to demand pattern uncertainty and pressure sensor inaccuracy is added to the model. In the second step, a two-step optimization algorithm is applied. It relies on: a minimization process based on a derivative-free optimizer that reduces the difference between simulated and measured data at the pressure/flow sensors placed in the WDN, and a filtering-clustering-ranking algorithm that eliminates nodes where the leaked volume is assumed to be negligible by giving a priority list of nodes for further inspection.  The proposed method was tested on L-Town, a large-scale WDN used as benchmark for the Battle of the Leakage Detection and Isolation Methods (BattLeDIM). The preliminary results indicate that the proposed approach can effectively identify leakage areas, especially in large-scale WDNs, potentially offering a practical tool to water utilities managing complex distribution networks.