Explicit machine learning-based MPC for distributed control of nonlinear processes

This work develops an explicit machine learning-based model predictive control method (explicit ML-MPC) for real-time dynamic control of chemical processes. Traditional ML-MPC for dynamic control of chemical processes in Aspen Plus Dynamics often struggles with computational efficiency. While explicit ML-MPC reduces computation time by converting ML-based optimization problems of implicit ML-MPC to mixed-integer quadratic programming (MIQP), the computational complexity grows rapidly with the dimensionality of the decision space for complex chemical process networks. To overcome this limitation, a distributed control strategy within the explicit ML-MPC framework is introduced, using multiprocessing for parallel computation to further improve the real-time performance of explicit ML-MPC. A k-d tree-based search algorithm is also developed to construct MIQP problems efficiently in real time. The entire control framework is developed as an open-source, generalizable code repository that allows users to easily design, implement, and customize real-time ML-MPC controllers within Aspen Plus Dynamics, which could significantly accelerate the adoption of advanced control strategies in the chemical industry. Finally, the effectiveness of the proposed approach is demonstrated through the closed-loop control of a chemical process network in Aspen Plus Dynamics, showing notable improvements in computational efficiency.