(539b) Distributed Model Predictive Control of a Benchmark Chemical Plant

Mercangöz, M., University of California, Santa Barbara

Efficient plantwide control of chemical processing plants enables closer operation to optimization constraints, reduces off-specification production, and decreases the number of shut-downs. In addition to these economic benefits, efficient plantwide control can also provide environmental and operational safety. However, the complex, multivariable nature of chemical plants make the design and operation of plantwide control systems a non-trivial task. Decentralized strategies, based on heuristics, process knowledge or mathematical analysis, provide flexible and scalable solutions with simple single-input single-output (SISO) controllers [1]. However, for many chemical plants with strongly interacting sub-units, centralized, multivariable controllers provide significantly better performance [2], which in turn suffer from potential problems associated with computations, commissioning and maintenance due to their size, and a higher risk of failure due to their centralized nature.

In practice, modern chemical plants typically utilize control and optimization systems in a hierarchical structure, where SISO controllers enable stable operation of most unit operations. These controllers are then connected to multivariate systems spanning several unit operations to control important quality variables, which are in turn, interfaced with economic optimization and scheduling modules as shown in Figure 1 [3].


Figure 1. Hierarchical organization of optimization and control tasks in a chemical plant.

The information flow in these hierarchical structures is in a vertical direction and the systems at the same hierarchical level are typically not aware of the existence of their counterparts, even though they may be interacting. Developing a framework to integrate decentralized control and economic optimization modules has attracted research attention and a number of formulations based on coordination layers [4] or communication-negotiation rules have been developed [5]. In the present work, we present a methodology to provide a horizontal connection among decentralized control systems in a chemical plant at the multivariate control level. The industry standard for multivariate control is model predictive control (MPC) and the proposed methodology provides a network structure for autonomous, decentralized MPC applications to improve the efficiency of both state estimation, and control calculations. Such a distributed control network is depicted in Figure 2.


Figure 2. A chemical plant with a distributed model predictive control network.

The proposed methodology has its roots in decentralized estimation and control, and distributed MPC (DMPC) [6, 7, 8] and consists of two stages, namely a design and an execution stage. In the design stage, the chemical plant is partitioned into several subsystems. In this DMPC methodology, we require the autonomous DMPC units to preserve their functionality by retreating to a completely decentralized mode in the case of a disruption in the network or in the case of a failure in a neighboring unit. This self-sufficiency requirement is tested at the design stage. Initially, a physical partitioning of process manipulated and controlled variables is carried out, followed by an analogous portioning of the discretized mathematical model of the process. If the sub-models in the given portioning do not satisfy the controllability and observability requirements for self-sufficiency, the physical partitioning is revised until a self-sufficient set of sub-models is found. The methodology proceeds with the identification of explicit manipulated variable interactions and overlapping states between sub-systems, both due to physical relationships and also due to the overlap introduced by discrete sampling. Next, local state estimators are designed for every-sub system and the accuracy of the estimators for the estimation of overlapping states is determined. According to this analysis a reliability factor is assigned to every estimator for the estimation of overlapping states.

The execution stage of the methodology starts when the DMPC units get their allocated measurements and estimate the states in their sub-systems. The estimates for the overlapping states are broadcast to neighboring DMPC units, which share the same state, and corresponding overlapping state estimates are collected from the neighbors. The estimates from neighboring DMPC units are weighted with corresponding reliability factors and assimilated with local state estimates. MPC calculations are carried out with the assimilated states at every DMPC unit. At this point, the DMPC units broadcast the calculated, ?candidate? control action for the next time step to their neighbors. Using this information, the MPC calculations are repeated by considering explicit manipulated variable interactions (identified at the design stage). Finally, control action from the improved MPC results are applied to the physical system and the execution stage restarts with the arrival of new measurements. A block diagram representation of the execution stage is provided in Figure 3.


Figure 3. On-line estimation and control calculations in a DMPC node.

The proposed DMPC methodology was successfully demonstrated on an experimental four-tank system [9]. In the present work, we extend the application to an industrial benchmark problem with multiple species, several reaction and separation steps, and recycle streams. We consider possible disturbance scenarios and compare the control performance of the DMPC methodology with the performance of an equivalent, but completely decentralized MPC strategy, as well as with the performance of a centralized MPC. The results indicate that the proposed distributed methodology outperforms the completely decentralized strategy and provides an efficient and scalable alternative to the centralized strategy, without the risk and burdens associated with large scale algorithms.


[1] W.L. Luyben. Simple Regulatory Control of the Eastman Process. Ind. Eng. Chem. Res., 35 (10): 3280-3289, 1996.

[2] J.J. Castro, F.J. Doyle III. A Pulp Mill Benchmark Problem for Control: Application of Plantwide Control Design. Journal of Process Control, 14: 329-347, 2004.

[3] S. Skogestad. Control Structure Design for Complete Chemical Plants. Comp. Chem. Engr., 28, 219-234, 2004.

[4] R. Cheng, J.F. Forbes and W.S. Yip. Price-driven Coordination for Solving Plantwide MPC Problems. In Proceedings of the 16th IFAC World Congress, Prague, Czech Rep., 2005.

[5] A. N. Venkat, J.B. Rawlings and S.J. Wright. Stability and Optimality of Distributed Model Predictive Control. In Proceedings of the European Control Conference, Seville, Spain, 2005.

[6] A.G.O Mutambara. Decentralized Estimation and Control for Multisensor Systems. CRC Press. Boca Raton, Florida, 1998.

[7] R. Vadigepalli, F.J. Doyle III. A Distributed State Estimation and Control Algorithm for Plantwide Processes. IEEE Transactions Contr. Syst. Technol., 11, 119-127. 2003.

[8] E. Camponogara, D. Jia, B. H. Krogh, and S. Talukdar. Distributed Model Predictive Control. IEEE Ctl. Sys. Mag., pages 44?52, 2002.

[9] M. Mercangoz, F.J. Doyle III. Distributed Model Predictive Control of a Four-Tank System. In Proceedings of the 8th IFAC Advanced Control of Chemical Processes Symposium, Gramado, Brazil, 2006.