Probabilistic modelling for methane leak detection in gas distribution networks (Papers Track)
Katherine Green (Guidehouse); Rubab Atwal (Guidehouse)
Abstract
Methane leaks from gas distribution pipelines in the UK contribute significantly to the country's total greenhouse gas emissions. Machine learning methodologies can be employed to improve timely detection of leaks, allowing them to be fixed sooner, therefore reducing emissions. Here we present a probabilistic machine learning framework, based on a Wasserstein autoencoder and Bayesian inference, which has been developed to detect, localise, and quantify leaks within a UK-based gas distribution system with limited data availability.