A Data-Driven Framework for Probabilistic Estimates in Oil and Gas Project Cost Management: A Benchmark Experiment on Natural Gas Pipeline Projects

Abstract

Nowadays, the Oil and Gas (O&G) industry faces significant challenges due to the relentless pressure for rationalization of project expenditure and cost reduction, the demand for greener and renewable energy solutions and the recent outbreak of the pandemic and geopolitical crises. Despite these barriers, the O&G industry still remains a key sector in the growth of world economy, requiring huge capital investments on critical megaprojects. On the other hand, the O&G projects, traditionally, experience cost overruns and delays with damaging consequences to both industry stakeholders and policy-makers. Regarding this, there is an urgent necessity for the adoption of innovative project management methods and tools facilitating the timely delivery of projects with high quality standards complying with budgetary restrictions. Certainly, the success of a project is intrinsically associated with the ability of the decision-makers to estimate, in a compelling way, the monetary resources required throughout the project’s life cycle, an activity that involves various sources of uncertainty. In this study, we focus on the critical management task of evaluating project cost performance through the development of a framework aiming at handling the inherent uncertainty of the estimation process based on well-established data-driven concepts, tools and performance metrics. The proposed framework is demonstrated through a benchmark experiment on a publicly available dataset containing information related to the construction cost of natural gas pipeline projects. The findings derived from the benchmark study showed that the applied algorithm and the adoption of a different feature scaling mechanism presented an interaction effect on the distribution of loss functions, when used as point and interval estimators of the actual cost. Regarding the evaluation of point estimators, Support Vector Regression with different feature scaling mechanisms achieved superior performances in terms of both accuracy and bias, whereas both K-Nearest Neighbors and Classification and Regression Trees variants indicated noteworthy prediction capabilities for producing narrow interval estimates that contain the actual cost value. Finally, the evaluation of the agreement between the performance rankings for the set of candidate models, when used as point and interval estimators revealed a moderate agreement (a=0.425).