Development of inferential sensor and real-time optimizer for a vacuum distillation unit by recurrent neural network modeling of time series data

Abstract

In lube base oil production, a vacuum distillation unit (VDU) is a crucial unit operation which must be controlled tightly to meet final product specifications. To address this requirement, model-based optimal control can be employed, which requires an accurate model relating to the operation and quality variables involved. However, the mathematical modeling of VDU is hampered by the lack of process knowledge and the limited online data acquisition capability for key quality variables. As a recourse, data-based modeling can be attempted to develop inferential sensors and controllers but one is faced with the problem of temporal credit assignment due to the complex characteristics of the VDU process dynamics, e.g., varying time-lags among the variables. To overcome this difficulty, the use of a stacked RNN structure with sequence length is proposed for the prediction of kinematic viscosity. In addition, regression models for the product's distillation temperatures are constructed by testing a large number of regression model structures. Then, these quality predictors are applied in developing a real-time optimizer which adjusts key operating variables in order to satisfy the product specifications despite various disturbances. Accuracies of the soft sensor and the optimization model are evaluated, first offline and then, online through a commercial plant test performed at SK Ulsan Refinery Complex in Korea. The overall approach and framework are general and can be used for the development of inferential sensors and controllers in other similar situations.