Comparison of Composition Soft Sensor Correction Methods for a Kaibel Dividing-Wall Column

Abstract

Though a Kaibel dividing-wall column (DWC) is able to save approximately 40% operation fees, the fear of controllability issues of the Kaibel DWC obstructs its large-scale industrialization. In our previous study, a composition/temperature cascade MPC-PI scheme was proposed with proportional–integral (PI) controllers stabilizing the operations and model predictive control (MPC) improving the performance. In the current paper, the MPC-PI scheme in combination with the data-driven soft sensor model replaces detecting compositions by detecting auxiliary variables, which is more applicable and practical. Although the same soft sensor model is applied, different Kalman filters used for error corrections will result in different composition estimation performances, which will consequently affect control performances for the Kaibel DWC. Soft sensors with different nonlinear Kalman filters have been studied and compared for the control scheme, containing the extended Kalman filter (EKF), unscented Kalman filter (UKF), and particle filter (PF). Research shows that the EKF can reach the setpoints of the product compositions faster than the UKF, but there are oscillations with reciprocating motion using the EKF. The speed of the PF to achieve stability is the fastest; however, the steady-state offsets remain. Therefore, the UKF is the most appropriate filter in composition soft sensors for the Kaibel DWCs, as its corresponding control performances are stable without the oscillations using the EKF and the steady-state offsets employing the PF.