Abstract

Water loss in distribution networks represents a critical challenge in urban water management, with significant environmental, economic, and public health implications. Around 126 million cubic meters of treated water are globally lost annually through distribution network leakages. This thesis presents a hybrid deep-learning approach for leakage localization that minimizes the distance between predicted and actual leak locations in water distribution networks. The developed algorithm integrates hydraulic modelling with deep learning techniques through three main phases: (1) Hydraulic Model Modification, (2) Water Demand Adjustment and Simulations, and (3) Leak Localization. A hybrid CNN-LSTM model was developed, combining CNN for spatial feature extraction and LSTM for temporal pattern recognition. Additionally, a hybrid loss function was designed to optimize both classification accuracy and distance in leak localization. The algorithm was validated using the Battle of L-Town Water Distribution Network benchmark. Results demonstrated significant improvements in localization accuracy, achieving average errors of 120.15 meters for background leakages and 98.76 meters for bursts when using the hybrid loss function - representing 29% and 39% reductions, respectively, compared to standard loss functions. The main contributions of this thesis are two-fold: the accuracy of leakage localization has been significantly improved through the development of a hybrid CNN-LSTM model with a customized hybrid loss function, and the search area for leak identification has been reduced, with the distance-based loss function helping to minimize the physical distance between predicted and actual leak locations. This reduction in search area translates directly to decreased inspection time and costs for water utilities.