Forecasting the research octane number in a Continuous Catalyst Regeneration (CCR) reformer

Gasoline is still one of the most consumed crude oil derivatives in the global market. Its grade is established by the Research Octane Number (RON), which is measured by laboratorial analysis at a lower rate in comparison to process variables. This delay represents a bottleneck to process operation (monitoring, control, and optimization) and quality (less consistency due to poor process control), with impact in the company's bottom line results. In order to mitigate the effects of such delay, we address the problem of predicting RON using real industrial data from a refinery located in Portugal. The dataset was collected at Matosinhos Refinery of Petrogal, SA, during an extended period of operation (2 years). We report the performance of a wide range of state-of-the-art linear and non-linear methods able to cope with high-dimensional data, which are assessed through a framework based on Monte Carlo Double Cross-Validation. Overall, 25 methods and their variants were tested which, to the best of authors’ knowledge, is the largest number of methods ever considered in studies dedicated to the prediction of RON. Furthermore, a ranking system for efficiently comparing the methods is also proposed, facilitating the identification of the top performing ones. According to this methodology, kernel partial least squares and tree-based ensemble methods stand out among the methods tested, signalling the presence of non-linear relationships in the datasets.