Optimizing Fully-Linear Deep Learning for Pipe Burst Localization in Water Distribution Networks: A Comparative Analysis of Particle Swarm Optimization and Population-Based Training

This paper presents a comparative study of hyperparameter optimization techniques - Particle Swarm Optimization (PSO) and Population-Based Training (PBT) - for fully-linear deep learning architectures, specifically FL-ResNet and FLDenseNet, applied to pipe burst localization in Water Distribution Networks (WDNs). Accurate and efficient burst detection in WDNs is critical due to the potential for substantial water loss and service disruptions, yet traditional detection methods are often costly, labour-intensive, or limited in scalability. Recent advancements in deep learning offer enhanced diagnostic and predictive capabilities for WDN management. This study leverages PSO and PBT to optimize the performance of FL-ResNet and FL-DenseNet models. Results indicate that PSO consistently achieves higher accuracy with lower variance and faster convergence compared to PBT, with PSO-FL-ResNet reaching a mean accuracy of 98.92 % in significantly fewer training epochs. Statistical analysis validates PSO's superior stability and efficiency across both models. These findings suggest that PSOoptimized models could provide a more reliable and resource-efficient solution for pipe burst localization in WDNs.