(124d) Autonomous Maintenance of Advanced Process Control: Application to an Industrial Depropanizer

Autonomous maintenance of advanced process control:

Application to an industrial depropanizer

H. Guidi¹, C. A. Larsson², Q. N. Tran³, L. Özkan³, A.C.P.M. Backx³

¹SASOL Synfuels, Secunda, 2302, South Africa

hernan.guidi@gmail.com

²Department of Automatic Control and ACCESS Linnaeus Center,

School of Electrical Engineering, KTH, 100 44 Stockholm, Sweden

chrisla@kth.se

³Eindhoven University of Technology

Den Dolech 2, 5612 AZ Eindhoven, The Netherlands

n.q.tran@tue.nl; l.ozkan@tue.nl; a.c.p.m.backx@tue.nl

Although Model Predictive Control (MPC) has been widely accepted as a main technology for Advanced Process Control (APC) due to its ability of operating the system closely to the constraints, proper maintenance of MPC systems is still a challenge. Like any other model-based operating support systems, the performance of MPC depends heavily on the accuracy and calibration of the model. Base-layer problems, temporary changes in disturbances or gradual deviation in plant characteristics can all cause performance degradation and appropriate corrective action is often difficult to determine. This may cause the MPC to be shut down completely or temporarily and significantly reduce the potential benefits of MPC.

Based on this observation, this research aims to develop a novel approach to the autonomous maintenance of MPC. A schematic overview of the automated strategy is depicted in Figure 1. The system is commissioned at the optimal closed-loop bandwidth while the performance is monitored online. Once performance deterioration is detected, the system is diagnosed so as to find out the reason for the degradation. Based on the cause of the performance drop, suitable corrective action is followed. In this work, retuning and re-identification are considered as corrective action to retain the performance of the system after a change in the plant dynamics causes performance degradation.

FT-depropanizer

The FT-depropanizer forms part of the Sasol Synthetic fuels refinery in Secunda, South Africa. It is a 56-tray tower with total reflux and a side draw section above tray 38. The purpose of the FT- depropanizer is to separate C₃ and lighter components from C₄ and heavier components in the feed stream. The variables used to control the fractionation are primarily the side draw flow rate and the pressure difference between the top and the bottom of the column.

The control objectives of the FT-depropanizer are to maximize the side-draw product (C₃s) while maintaining the product quality (no impurities such as C₄s). Additional constraints present themselves through limits on the column pressure and the pressure difference which if violated result in flaring and flooding respectively. Optimal operation of the column is achieved by operating closely to these constraints.

An MPC is commissioned to satisfy the control objectives within these constraints. Maximum side draw production without violating product specification is achieved through allowing the MPC to manipulate the side-draw to feed ratio. C₄ impurities are prevented by maintaining the bottom temperature below 88^oC. The MPC additionally allows for optimal fractionation by maintaining the pressure difference as close to 40 kPa without going beyond this limit which defines the flooding constraint.

Scenario.

The maintenance loop is first applied to a simulator of the FT-depropanizer. Built on Honeywell's UniSim Operations Suite, the simulator was constructed to realistically and accurately reproduce the process dynamics that characterize the FT-depropanizer.

The linear model used in the MPC to control the depropanizer is obtained through prediction-error identification. A model is first identified in an open-loop experiment using a random binary signal with significant low-pass characteristics. The excitation is added to all three channels simultaneously and separate multiple input, single output (MISO) models are identified for each output channel. This model serves as an initial model used in the MPC as well as a reference (?ground truth?) for the model identified in closed loop. This step corresponds to the first modeling efforts in any MPC commissioning. The MPC uses the identified model in its prediction to fulfill the control objectives.

The automated tuning method used in this study aims at achieving  minimum variance of the C₄ content in the side draw, while keeping the side draw ratio as high as possible. The minimum variance is obtained when the system operates at the optimal balance between robustness and nominal performance. It was shown in [1] that increasing the closed-loop bandwidth reduces the output variance up to certain bandwidth thanks to better disturbance rejection. However, increasing the bandwidth from the optimal balance between robustness and nominal performance will raise the output variance due to the effect of modeling errors.

After the commissioning phase, a performance drop is constructed by changing the level of C₂s in the feed to the column. Re-tuning and closed-loop re-identification are performed to retain the performance. The re-tuning aims to find the new optimal bandwidth, which is lower than the original one due to the change in the plant dynamics. The re-identification is performed using Model Predictive Control with eXcitation (MPC-X) first presented in [2]. This reformulation of MPC includes the possibility of creating sufficiently informative closed-loop data by adding the excitation signal requirements as a constraint in the optimization. It combines MPC with ideas from least-costly identification and applications-oriented input design enabling identification from closed-loop data while trying to disturb the process as little as possible. The resulting identified model is guaranteed with high probability to fulfill the performance requirements when used in the MPC.

The next step is the implementation of autonomous maintenance loop on the actual distillation column. In this work, the test results from the simulator and the actual column are presented and discussed.

Figure 1. Autonomous maintenance loop (Steps considered in the paper are in bold)

References

[1] Quang N. Tran, Leyla Özkan, A.C.P.M. Backx. MPC tuning based on impact of modeling uncertainty on closed-loop performance. In Proceedings of the 2012 AICHE annual meeting. Pittsburg, PA, USA.

[2] Larsson, C.A., Annergren, M., Hjalmarsson, H., Rojas, C.R., Bombois, X., Mesbah, A., and Modén, P.E. (2013). Model predictive control with integrated experiment design for output error systems. In Proceedings of the 12th European Control Conference. Zürich, Switzerland.