SIMULATION-BASED DYNAMIC OPTIMIZATION OF DISCRETELY CONTROLLED CONTINUOUS PROCESSES

MARIANO DE PAULA; ERNESTO C. MARTINEZ

Mar del Plata

Congreso; VI Congreso Argentino de Ingeniería Química; 2010

AAIQ

Discretely controlled continuous processes (DCCPs) constitute a special subclass of hybrid process plants, which is of great practical relevance. Such process systems are found in many dynamic optimization problems such as throughput maximization when integrating batch operations with continuous processing, optimal sequencing of changeover operations in polymer production with varying qualities and optimal control of safety-critical systems that implement drug (e.g., insulin) infusion protocols. In this work, a novel simulation-based approach to dynamic optimization of DCCPs is proposed using Gaussian Process Dynamic Programming (GPDP). GPDP is an approximate value function method that integrates reinforcement learning with Gaussian Processes for seeking an optimal control policy over a continuum of observables states and control actions. The GPDP method resorts to active Bayesian learning using simulation runs of the DCCPs in such a way that only the relevant part of the observable state/decision spaces is explored. As a representative DCCP,  the interface between the batch and continuous sectors in a Solvay PVC production line is considered. To maximize the average plant productivity a buffer tank is used to guarentee profitable, yet stable operation by throughput maximization using modes for accumulating and draining-off.