Learning MPC for Repetitive Operation Processes

ALEJANDRO GONZÁLEZ; EDUARDO J. ADAM; DARCI ODLOAK; JACINTO L. MARCHETTI

Foz de Iguazú

Congreso; XVIII Congreso Brasilero de Ingeniería Química; 2010

Several industrial systems work by executing a sequence of similar or almost identical finite-time operations, which allows learning from previous runs to improve control performance. The operation of batch and semi-batch reactors are typical examples of these systems in the chemical industry. In this work, a model predictive controller (MPC) based on the concept of shrinking horizon is proposed to deal with finite-time operations. In addition, the proposed controller includes a learning mechanism leading to improve the performance as the sequence of operations progresses. When a control strategy is developed for these systems, two convergence analyses are necessary: the convergence of the closed-loop responses during the time-interval the single operation lasts and the convergence from operation to operation, i.e., the convergence of the learning mechanism that provides high performance when the number of trials increases. All these features found with the proposed MPC are illustrated by simulating a case study.