Monetary Consequence Prediction for Hazardous Liquid Pipelines in US

Abstract

There are more than 230,000 miles of hazardous liquid pipelines in the US, and more than 4,000 significant incidents have been reported since 2010 with 3.1-billion-dollar damage. The cost consequence estimation of pipeline failures can be utilized in the risk-based integrity management of aging pipelines.

The objective of this research is to develop monetary consequence prediction models of hazardous liquid pipelines using multivariate Lasso regression and artificial neural network (ANN). Hereby, the incident data reported by PHMSA (Pipeline and Hazardous Materials Safety Administration) are used for the analyses. In particular, the target response refers to the cost sum of public and non-operator private property damage as well as operatorﾒs property damage and repairs. The independent variables include general information (e.g., age, pipe geometry) and incidents features (e.g., failure mode, explosion status).

For the regression analysis, a hybrid step-wise methodology is first implemented where a forward selection and backward elimination by significance testing are used. Each potential variable, its higher orders, and its interaction variables are added to the model one at time; when any of the variables in the model becomes statistically insignificant, the corresponding variable will be removed from the model. Then the selected variables are used in a Lasso regression approach to derive the final model, where the Lasso regression is able to avoid overfitting or discourages fitting a complex model when there is a large database. In addition, the data is split into 80% and 20% for model development and testing, respectively.

For ANN, the network architecture contains input, hidden, and output layers of neurons connected through links. These links have learning rates with self-modification ability to reduce prediction error each time a new input is introduced. A back-propagation algorithm is adopted to adjust the learning rate of links using feedback signals for the model optimization. For model development, 60% and 20% of data are used for model training and fitting evaluation, respectively; then the rest of 20% data is for testing.

The performance of the regression and ANN models are evaluated and compared by measuring the residual standard error (RSE) of the testing data. While two models can satisfactorily predict the cost consequence, ANN model outperforms in terms of RSE of the training data and predictability of the testing data. The results of this research will be also applied in a case study of maintenance planning of pipelines for optimal inspection and repair scheduling.