Abstract

Abstract
Failure Predicting Model for Oil Pipelines
Bassem Abdrabou

Oil and gas pipelines are considered the safest means to transport petroleum products comparing to railway and highway transportations. They transport millions of dollars’ worth of goods every day. However, accidents happen every year and some of these accidents inflict catastrophic impact on the environment and result in great economic loss. In order to maintain safety of the pipelines, several inspection techniques have been developed in the last decades. Despite the accuracy of these techniques, they are very costly and time consuming. Similarly, several failure predicting and condition assessment models have been developed in the last decade; however, most of these models are limited to one type of failure, such as corrosion failure, or mainly depend on expert opinion which makes their output seemingly subjective.
The present research develops an objective model of failure prediction for oil pipelines depending on the available historical data on pipelines' accidents. Two approaches were used to fulfill this objective: the artificial neural network (ANN) and the Multi Nomial Logit (MNL). The ANN is used to develop a model to predict failure due to mechanical, corrosion or third party, which collectively account for 88% of oil pipeline accidents. This model had a prediction accuracy of 68.5%. Another ANN model is developed to predict only corrosion or third party failure with a prediction accuracy of 72.2%. The Average Validity Percentage (AVP) for the two models is 73.7 and 72.8, respectively.
The MNL approach is used to develop a model that predicts failures caused by mechanical, corrosion or third party elements with a prediction accuracy of 68.4% and Pseudo R Squared of 0.42. The Average Validity Percentage (AVP) for this MNL approach is 73.7%. This model also generates a probability equation for each type of failure.
The three developed models show convincing results, since they are based on solid historical failure data for the last 38 years, with no subjectivity or ambiguity. These models could easily be used by oil pipeline operators to identify the type of failure threatening each pipeline so that appropriate preventive and corrective measures can be planned. The models also help to prioritize in-line inspection of different pipeline segments according to the predicted type of failure.