Online optimization of simulated moving bed processes based on deep learning models

Abstract Due to the lengthy computation times required by the mechanistic model of the simulated moving bed (SMB) separation process, applying it directly to online optimization and control is challenging. To address this issue, this paper proposes replacing the traditional mechanistic model with a deep learning–based surrogate, enabling real‐time optimization of the SMB process. The optimization strategy's control unit comprises two components: a model parameter estimator and an operational parameter optimizer. The model parameter estimator employs a dung beetle optimization (DBO) algorithm to tune a convolutional neural network combined with a bidirectional long short‐term memory network and a multi‐head attention mechanism (DBO‐CNN‐BiLSTM‐MHA). The operational parameter optimizer consists of a deep neural network (DNN) and a multi‐objective dung beetle optimization (MODBO) algorithm. During operation, when product purity falls below the required level due to stationary‐phase degradation, the model parameter estimator predicts the current model parameters and passes them to the operational parameter optimizer. The optimizer then determines the optimal operating conditions to simultaneously maximize purity and productivity. Simulation results demonstrate that both models achieve high prediction accuracy on the test set and that the proposed online optimization strategy can continuously adapt operating parameters in response to constant‐rate stationary‐phase degradation, thereby maintaining high product purity and productivity.