Machine Learning Algorithms for Failure Prediction Model and Operational Reliability of Onshore Gas Transmission Pipelines

A transmission pipeline is the safest and most effective way of transporting large volumes of natural gas over long distances. However, if not maintained efficiently, failures of gas transmission pipelines can occur and cause catastrophic events. Therefore, an accurate prediction of pipe failures and operational reliability is required to determine the optimal pipe replacement timing such that the incidence of pipe failures can be prevented. Nowadays, computer-assisted technology helps businesses make better decisions, and machine learning is among the excellent techniques that can be utilized in predicting failures. In this study, two machine learning algorithms, i.e., random forest and binary logistic regression, are developed, and their prediction abilities are compared. The model is developed based on a decade of unstructured and complex historical failure data of the onshore gas transmission pipelines released by the United States Department of Transportation. The modeling process begins with data pre-processing followed by model training, model testing, performance measuring, and failure predicting. Both algorithms have demonstrated excellent results. The random forest model achieved an AUC of 0.89 and a predictive accuracy of 0.913, while the binary logistic regression model outperformed with an AUC of 0.94 and a prediction accuracy of 0.949. The trained model is further employed to predict future failures on a 11900-mile natural gas pipeline spanning from Louisiana to the northeast section of the United States. We show the location of the pipes that will be broken in the interval of five years and estimate that 29%/63%/83% of the pipes will break by 2025/2030/2035.