Data-driven accident consequence assessment on urban gas pipeline network based on machine learning

Abstract

The accidents in Urban Pipeline Network (UPN) may cause enormous economic loss and serious threats to society and environment. The daily operation and maintenance of UPN is usually associated with many aspects of data. How to take full advantage of these multi-source data in combination with advanced data mining techniques to assess the post-event risk of pipeline accidents is of great significance to management of resilient urban systems. This work first summarizes the factors affecting accident consequence of gas UPN and establishes the risk evaluation indicators. A traditional risk assessment model based on the Kent index method and the analytic hierarchy process is then employed to determine the relative risk value of each pipeline. To reduce the dependency on experts’ subjective judgements or calculation of probability events in a Bayes decision procedure, a data-driven model based on graph embedding and clustering algorithm is proposed. The Graph Convolutional Network (GCN) technique is used to extract the topological features of pipeline network as a complement to the common attribute features considering the top pipelines usually bear comparable level of risks. A case study on a real gas pipeline network consisting of more than 6500 pipelines verifies the effectiveness of the proposed model.