Machine learning-aided optimal design and distributed model predictive control of reactive dividing wall column

Abstract

Reactive dividing wall column (RDWC) integrates the advantages of high conversion efficiency, low energy consumption, and reduced investment. However, this further intensification of reaction and separation increases system coupling and significantly complicates process optimization. To address this challenge, this work proposed a machine learning-aided multi-objective optimization (ML-MOO) framework for determining the optimal RDWC design. Taking the dichlorosilane anti-disproportionation RDWC process as a case study, random forest (RF), backpropagation neural network (BPNN), and support vector machine (SVM) were integrated with multi-objective particle swarm optimization (MOPSO) algorithm to optimize the steady-state operating parameters of RDWC. The hyperparameters of three ML models were tuned using Bayesian optimization algorithm (BOA) with 5-fold cross-validation. The results showed that, compared with the rigorous Aspen simulation-based optimization, the SVM surrogate model reduces total annual cost, flow rate of silicon tetrachloride, and environmental impact potential of energy by 5.3 %, 23.5 %, and 7.6 %, respectively, while reducing computation time by 19.3 %. Additionally, due to the existence of internal reactions, the dynamic behavior of RDWC is constrained by both product quality and safety redundancy. To address this, a distributed MPC (DMPC) strategy was proposed, using two sub-MPC controllers to separately control inventory and quality loops, thereby enhancing system fault tolerance. Dynamic response results indicated that benefiting from communication between sub-controllers, the integral absolute error (IAE) value of linear DMPC structure based on linear time-invariant state space (LTI-SS) model differs from that of centralized MPC (CMPC) structure by only 3 % to 10 %, demonstrating similar dynamic response performance while achieving enhanced safety.