Leak identification and quantification in gas network using operational data and deep learning framework

Abstract

In this study, we introduce an innovative deep learning framework designed to achieve precise detection, localization, and rate estimation of gas distribution pipeline system leakages. Our method surpasses conventional statistical approaches, particularly those based on Bayesian inference, by accommodating the system’s intricate behaviors, including variable usage and production from both sources and sinks. Notably, our approach demonstrates remarkable accuracy in localizing leakages even amidst multiple occurrences within the system. Specifically, achieving over 98% accuracy in single-leakage scenarios underscores its effectiveness. Furthermore, through data augmentation involving the introduction of noise into the training dataset, we significantly enhance the model’s performance, particularly when tested against real-world-like noisy data. This study not only showcases the efficacy of our proposed deep learning framework but also underscores its adaptability and robustness in addressing complex challenges in gas pipeline systems.