Leveraging Deep Learning for Efficient Explicit MPC of High-Dimensional and Non-linear Chemical Processes

This paper presents an efficient Deep Learning (DL) based method for explicit model predictive control (E-MPC) of high-dimensional and/or nonlinear chemical processes for which mathematical E-MPC approaches are difficult to apply. The method uses DL models for off-line development of control laws that accurately approximate the relationship between the optimal values of the control variables to be applied in the next sampling period (SP) as a function of the values of the state variables in the current SP. The training data are generated by solving the MPC problem considering different initial values of the state variables selected using design of computer experiments (DOCE) techniques. The obtained DL-based control laws are then integrated into a closed-loop for online control of the process. A numerical validation procedure is used to evaluate the performance of the developed control laws in terms of their accuracy and computational cost. The method is applied to case studies for which a “direct” application of mathematical E-MPC techniques is difficult due to their high dimensionality and nonlinearity. The results show a high performance of the proposed method and a reduction in the complexity of the solution procedure compared to mathematical E-MPC.