MODEL PREDICTIVE CONTROLLER FOR PERIODIC ADSORPTION PROCESSES

V. DE LA TORRE, G. FOX. L. BIEGLER, A. BANDONI

Rio de Janeiro, Brasil

Congreso; ENPROMER 2005; 2005

Periodic Adsorption Processes (PAPs) have become a key technology for gas separation. PAPs operate at ambient conditions and require by far less capital investment than other gas separation technologies. As a result, they have become especially attractive for bulk gas separations of oxygen, nitrogen and hydrogen. Models for PAPs are multiple instances of Partial Differential Equations (PDEs) in time and space with periodic boundary conditions that link the processing steps together. The control of such systems is necessary for reducing the economic impact of low recovery, wasteful overproduction and low product specifications caused by varying production demands or incoming disturbances in feed streams. However, PAPs also represent a significant challenge to standard control techniques. Our goal is to design a model-based controller by exploiting the most important features of Model Predictive Control (MPC) to steer PAPs between desired operating levels as dictated by external demands, while maintaining maximum process efficiency. A single-bed six-step O2 Vacuum Swing Adsorption (VSA) process, which separates oxygen from air is used as a representative example; a detailed model description can be found in previous work of Jiang et al. (2003). We study the dynamics of cyclic operation of the O2 VSA process by moving the system from the starting point through intermediate optimal steady states to the target and perform identification tests using Pseudo-Random Binary Sequences (PRBS) of inputs. Despite the process nonlinearities, the entire response range could be captured with a single identified linear model. Thus, the MPC implementation was simplified and no intermediate points are needed to predict/anticipate the performance of the process. Challenges and implementation details are discussed. Closed loop simulations show a good performance of the controller for most of the tests even for the worst case scenarios.