(470a) Inferential Control of Distillation Column By Successive Update of Softsensor and Control Algorithm

In the industrial processes, quality variables should be monitored in real-time, and directly controlled to suppress the variation. The smaller variation minimizes the risk of violation of the quality criterion and more economical process operation is achieved. However, it is often difficult to measure quality variables, such as concentration, in real-time. Softsensors are powerful means to predict the values of the difficult-to-measure variable in real time, from the measurement of easy-to-measure variables. The control system with such a softsensor is called inferential control system and useful for quality control.

The operation data used for building a softsensor is usually obtained from a plant which is under the influence of an existing control system. However, the controlled variables of the existing control system are different from those of the inferential control system. The process dynamics under an inferential control can be different from those under the existing control system. When the process dynamics is largely different, the prediction accuracy of the softsensor will deteriorate in the inferential control system. As a result, the quality control performance will not improve. On the other hand, when the difference in operating condition is small, the quality variation will not reduce by introducing the inferential control system.

Most of the researches on inferential control did not take the above-mentioned problem into account, and did not show the condition where the model construction data was obtained [1–7]. In references [8–15], the model construction data was prepared in impractical condition. Some authors obtained data for model building from a plant without a control system [8–12]. Qian et al. obtained data for the model construction under sustained excitation condition [13]. Bidar et al. changed the setpoint frequently when they obtained the model construction data [14]. Clementi et al. assumed the variable which cannot be measured in real-time in reality to be measurable in real-time and controlled on simulation to obtain the model construction data [15]. Kano et al. obtained model construction data under a practical condition where the existing control system was used and no intentional oscillation to generate the data with a broader range was added [16]. However, even in reference [16], the effective method to improve the quality control performance was not shown.

As mentioned above, the quality control performance does not improve, whether the operating condition is changed largely or not. In the present work, we proposed a novel method to introduce the inferential control system into the existing plant. In the proposed method, the limit for the change in operation condition is set according to the applicable domain size of the softsensor. In addition, the softsensor is updated to enlarge the range of the applicable domain, by using the new data obtained when the inferential control system is used. The new operating data covers a larger domain because the former controlled variable in the existing control system fluctuates intensely in the inferential control system. Consequently, the softsensor updated with this newly obtained data is expected to have a larger applicable domain.

We use a cascade control system where the quality variable and the former controlled variable are controlled in the primary and secondary control loops, respectively. This is because the cascade structure is known to be superior to the ordinary single-loop feedback structure in the prediction accuracy of the softsensor [16]. In addition to using the cascade control system, the upper and lower limit constraint is imposed on the setpoint of the secondary control loop in order to restrict the variation of the secondary controlled variable. When the softsensor is updated, the constraint is relaxed to change the operating condition and to enhance the quality control performance. As the number of new data increases, we can obtain the inferential control system with higher control performance and prediction accuracy.

A case study was conducted through the control simulation of the vinyl acetate monomer (VAM) plant [17, 18]. The main objective in the control simulation was to keep the water mole fraction at the bottom of the distillation column (x_B) constant. The inferential control system which was designed and updated by the proposed method (proposed control system), was compared with that of the existing plant (conventional control system 1) and with the inferential cascade control system where the softsensor or the control algorithm are not updated (conventional control system 2). x_B was directly controlled in the proposed control system and conventional control system 2, and the 18th tray temperature was controlled to keep x_B constant in conventional control system 1. The proposed control system was best of the three in the x_B control performance. Mean absolute errors of the proposed control system and control system 2 were 3.95×10^-2 mol% and 7.05×10^-2mol% respectively, and the proposed method improved the x_B control performance by 44 %, compared to control system 2. The proposed method was found to be useful in introducing the inferential control system.

References

[1] H. Guan, L. Ye, F. Shen, and Z. Song. Economic operation of a fluid catalytic cracking process using self-optimizing control and reconfiguration. Journal of the Taiwan Institute of Chemical Engineers, 96:104–113, 2019.

[2] S. Bosca, A. Barresi, and D. Fissore. On the use of model-based tools to optimize in-line a pharmaceuticals freeze-drying process. Drying Technology, 34(15):1831–1842, 2016.

[3] A. Escalante-Sanchez, J. Barrera-Cortes, H.M. Poggi-Varaldo, T. Ponce-Noyola, and I.S. Baruch. A soft sensor based on online biomass measurements for the glucose estimation and control of fed-batch cultures of Bacillus thuringiensis. Bioprocess and Biosystems Engineering, 41(10):1471–1484, 2018.

[4] P. Cao, X. Luo, and X. Song. Feasibility analysis and online adjustment of constraints in model predictive control integrated with soft sensor. Chinese Journal of Chemical Engineering, 25(9):1230–1237, 2017.

[5] A. Abusnina, D. Kudenko, and R. Roth. Improving robustness of gaussian process-based inferential control system using kernel principle component analysis. Proceedings - 2014 13th International Conference on Machine Learning and Applications, ICMLA 2014, pages 99–104, 2014.

[6] Q. Jiang, Y. Cai, S. Jia, Z. Cao, B. Chen, and H. Zhou. Optimal product quality control in a hydrocracking fractionator with process simulation approaches. Industrial and Engineering Chemistry Research, 54(17):4805–4814, 2015.

[7] A.K. Jana and S. Banerjee. Neuro estimator-based inferential extended generic model control of a reactive distillation column. Chemical Engineering Research and Design, 130:284–294, 2018.

[8] G. Pannocchia and A. Brambilla. Consistency of property estimators in multicomponent distillation control. Industrial and Engineering Chemistry Research, 42(20):4452–4460, 2003.

[9] A. Bahar and C. Zgen. State estimation and inferential control for a reactive batch distillation column. Engineering Applications of Artificial Intelligence, 23(2):262–270, 2010.

[10] S.R. Vijaya Raghavan, T.K. Radhakrishnan, and K. Srinivasan. Soft sensor based composition estimation and controller design for an ideal reactive distillation column. ISA Transactions, 50(1):61–70, 2011.

[11] V. Sakhre, S. Jain, V.S. Sapkal, and D.P. Agarwal. Modified neural network based cascaded control for product composition of reactive distillation. Polish Journal of Chemical Technology, 18(2):111–121, 2016.

[12] K. Chiewchanchairat, P. Bumroongsri, V. Lersbamrungsuk, A. Apornwichanop, and S. Kheawhom. Implementation of Model Predictive Control in Industrial Gasoline Desulfurization Process. Computer Aided Chemical Engineering, 37:1619–1624, 2015.

[13] X. Qian, K. Huang, S. Jia, H. Chen, Y. Yuan, L. Zhang, and S. Wang. Composition/temperature cascade control for a Kaibel dividing-wall distillation column by combining PI controllers and model predictive control integrated with soft sensor. Computers and Chemical Engineering, 126:292–303, 2019.

[14] B. Bidar, M.M. Khalilipour, F. Shahraki, and J. Sadeghi. A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instrumental variable (LIV) technique. Journal of the Taiwan Institute of Chemical Engineers, 84:49–59, 2018.

[15] L.A. Clementi, R.B. Suvire, F.G. Rossomando, and J.R. Vega. A Closed-Loop Control Strategy for Producing Nitrile Rubber of Uniform Chemical Composition in a Semibatch Reactor: A Simulation Study. Macromolecular Reaction Engineering, 12(2), 2018.

[16] M. Kano, K. Miyazaki, S. Hasebe, and I. Hashimoto. Inferential control system of distillation compositions using dynamic Partial Least Squares regression. Journal of Process Control, 10(2):157–166, 2000.

[17] Y. Machida, S. Ootakara, H. Seki, Y. Hashimoto, M. Kano, Y. Miyake, N. Anzai, M. Sawai, T. Katsuno, and T. Omata. Vinyl Acetate Monomer (VAM) Plant Model: A New Benchmark Problem for Control and Operation Study. IFAC-Papers OnLine, 49(7):533–538, 2016.

[18] Omega Simulation Co., Ltd. Visual Modeler Trial Version, 27 Feb 2020. www.omegasim.co.jp.