Machine-Learning-Based Soft Sensors for Energy Efficient Operation of Crude Distillation Units

The oil refining industry is considered one of the largest energy-consuming industrial sectors worldwide and the third-largest global source of greenhouse gas emissions. In addition, increasingly restrictive environmental quality specifications for oil products worldwide require increased use of energy during the distillation process, further increasing the emissions. Therefore, energy usage reduction could help in two ways: to reduce the amount of greenhouse gas emissions and maximize the profits of the oil refining plants. The development and use of soft sensors to monitor the crude refining process in real-time enables timely insights and decision-making to ensure the products meet the required quality. Furthermore, the same approach can be used to further optimize the energy consumption over the multiple stages of the refining process. In this paper, we address the problem of predicting the energy consumption in a crude oil refinery. We do so on a real-world use case, with data obtained from a Tüpras refinery. We found the best performance was achieved with a CatBoost regressor when performing a tenfold cross-validation and assessing their significance against other models when comparing confidence intervals at a 95% level of significance.