Development of a deep learning-based schedule-aware controller: Toward the integration of scheduling and control

Abstract

Most of the existing approaches for integrating the scheduling and control layers in the process industry suffer from the solution complexity of the resulting Mixed Integer Nonlinear Programming (MINLP) program. This complexity stems from two main obstacles: i) the consideration of nonlinear closed-loop dynamics involved in the control layer and ii) the way of exchanging the information between the control and the scheduling layers. To tackle these two obstacles, this work proposes the use of a schedule-aware controller (SAC) based on deep learning (DL) models, which receives, as input, the scheduling information (i.e., production sequence, targets, and set points) along with the process conditions (values of the state, output, and control variables) to predict the optimal control actions to be applied at the next sampling periods. The method is applied to a benchmark, showing very good performance in terms of the ability to adapt the control actions in accordance with the scheduling decisions, high prediction accuracy, and reduction in the computation cost.