Explanatory and predictive analysis of naphtha splitter products

Abstract

Refinery operations are always prone to optimization, and due to the increasingly adverse effects of COVID-19 on energy sectors, its importance increased significantly. This work aims to predict the naphtha column KPI parameters with high accuracy so that operators make corrective actions efficiently. Although linear regression provides acceptable results for prediction, this is not the case for top and bottom product C7 and C6 prediction in the central Naphtha Splitter column. First, we did gather all the available data to overcome this problem, which can affect the top and bottom products. Including upstream units that feed the column. Instead of one common technique (linear regression), we used five additional machine learning methods: Adaboost, support vectors, kNN, random forest, XGboosting. Since there are many measurements, however, very few samples need to reduce dimensions before modeling. We used BorutaSharp to select the essential features. We also use classification machine learning methods to categorize bottom products since there is no need to predict the value instead of whether the value is higher or lower than a constant. Overall, we achieved 30% higher accuracy than the traditional ways for the top product, and we reached to predict C6 content in the bottom with higher accuracy than 80%. Xgboost provides the best regression model, and stochastic gradient boosting yields the best classification model. After our implementation, the energy consumption is decreased significantly, and 100k$/month is saved since we can monitor top and bottom products simultaneously.