Identifying Leaks in Water Distribution Networks Using Deep Learning Neural Network and Frequency Ratio Models

In recent years, researchers and policymakers have focused on leaks in water systems as a critical issue because of their negative impact on human society. Most classical methods can only provide approximate leakage locations, typically identifying the general area of a node or pipe section in the network. In this paper, frequency ratio (FR) and deep learning neural network (DLNN) models were used to identify water leaks in water distribution networks (WDNs). Eight predictor variables were used to assess leakage susceptibility in the WDN of Khenchela City, in northeastern Algeria. According to both models, the two most significant predictor variables in the studied WDN are pipe material and age. The predictive ability of these models was evaluated using the receiver operating characteristic (ROC) curve, based on 339 recorded water leakage locations. The results were very satisfactory, with the DLNN model showing slightly higher accuracy than the FR model, achieving area under the curve (AUC) values of 86.7% and 82.3%, respectively. Although the FR model was applied for the first time in this field, it demonstrated strong potential as a decision-support tool for water leak detection.