Application of robust statistics on a crude distillation unit

Abstract

Refineries are highly complex and integrated systems, separating and transforming crude oil into valuable products. One of the most important processes in refineries is the Crude Distillation Unit (CDU) process, in which raw crude oil is separated into various fractions to be further processed in other parts of the refinery. In the refinery, Heavy Diesel (HD) T95 value is very important quality indicator. In the current study, conventional and robust statistical methods were employed on the historical data of a CDU process in TUPRAS İzmit Refinery for monitoring and HD T95 prediction purposes. Process data consisted of online measurements of process variables and laboratory measurements of HD T95 values for a one-year period. In the first part of the study, trajectories of process variables were analyzed to identify relations between process variables and to distinguish normal from abnormal operating conditions in the distillation history. For this purpose, skipped- Principal Components Analysis (PCA) and Minimum Covariance Determinant (MCD)+PCA methods were applied to process data and MCD+PCA method was found as more efficient method in detecting disturbances in the operation conditions. In the second part of the study, Monte Carlo (MC) simulations were applied by creating clean and contaminated datasets to evaluate predictive performances of LS and various robust regression methods, and to assess the metrics (RMSE, MAE) for evaluating the quality of predictions under contamination. LTS10%+LS and LTS20%+LS were found as best predictive models, and RMSE was found to be reliable in assessing models when 70%- 90% of the highest absolute prediction errors were taken into account. In the last section, LS and robust regression methods were applied and compared to select the most convenient prediction method for HD T95 values. The best predictive performance was obtained by LTS30% model with 97.5% CL. By applying this method to historical dataset, 15% of training dataset was detected as outliers and when these outliers were excluded from dataset, the model can predict HD T95 value with a maximum 7 0C error.