Global Crude Oil Refinery Models: Parametric and Non-Parametric Methods

There is a dearth of public literature on reliable, evidence-based modeling representation of global crude oil refineries, this has led to the lack of transparency and consistency on how business and government policy processes are informed in many energy systems and energy transitions modeling scenarios. This paper uses actual field data to identify that 40 unique refining configurations account for the global refinery landscape. A machine learning algorithm training extremely randomized trees predictors is applied to develop a non-parametric production model of the refineries. To facilitate computational convenience and suitable deployments in diverse use cases, a multivariate linear regression model of the refineries is also presented. For all refined products captured, the machine learning model demonstrate superior predictive performance with coefficients of determination of over 90%, even though both models are useful. Other performance metrics are also assessed for both models.