Learning-based moving horizon autonomous control of a chemical reactor

Abstract

This paper proposes a learning-based moving horizon autonomous control of a chemical reactor (LMHAC) approach for chemical reactor with multiple operating conditions. In the proposed LMHAC scheme, model-based control, model-free control and process modeling are integrated in a moving horizon framework. A control switching logic makes a selection between model predictive control (MPC) and adaptive dynamic programming (ADP) depending on whether the model parameters are known or unknown under the current operating condition. To be compatible with the moving horizon framework, the conventional ADP is fitted into a finite horizon composed of two different stages, namely a learning stage and a control-identification stage. In the learning stage, a constrained finite-horizon ADP (CFADP) first learns an approximated optimal controller from the collected input-state information pair generated by an initial admissible control. In the control-identification stage, the approximated optimal control is applied to the process to generate a sequence of input-state information pairs which is then utilized in turn to identify the unknown model parameters. The LMHAC framework is capable of providing the optimal or nearly optimal control for different operating conditions online and incrementally enlarge the known domain of system dynamics. The feasibility and performance of the proposed approach are illustrated via a case study.