The Optimization Of Simulated Moving Bed Separation Process Based On Machine Learning

Simulated moving bed(SMB)is a continuous preparative chromatography technology,which is widely used in petrochemical,pharmaceutical and fine chemical industries to achieve high-purity separation of products.However,solving the mechanism model of SMB is too time-consuming so as to unable to be used for on-line optimization and process control.Therefore,several machine learning models for the substitution of the mechanism model to accelerate the optimization of SMB operation parameters were constructed in this research,and the online optimization of the SMB process by using the deep learning model was also investigated.Taking stevia separation with linear adsorption isotherm and naphthol enantiomeric separation with bi-Langmuir adsorption isotherm as the model systems,the data sets were generated by using the mechanism model.The inputs were the flow rates of all the zones and the switching time,and the outputs were the product purities of the extract and the raffinate lines.Then,support vector machine algorithm,k-nearest neighbor algorithm,decision tree algorithm,random forest algorithm and neural network algorithm were used to build machine learning models.The results showed that the prediction accuracies of the random forest model and the neural network model were satisfactory,and the mean absolute errors(MAE)in the testing set were lower than 0.19%and 0.08%respectively.Finally,the machine learning models were applied to optimize the SMB operation parameters.It was found that the generalization ability of the neural network model was better than that of random forest model,and the optimization result was consistent with that obtained based on the mechanism model.Considering the decay of the stationary phase in use,an online optimization strategy for the SMB process based on the deep learning model was proposed in this research.Taking stevia separation with linear adsorption isotherm as the model system,the effectiveness of the online optimization strategy was verified by simulation experiments.The control unit of the strategy contained two parts,the model parameter estimator and the operation parameter optimizer.The model parameter estimator was realized by a 1D convolution neural network-long short term memory network model.Its inputs were the operation conditions(the flow rates of all zones and the switching time)and the product concentration profiles of extract and raffinate in a switching period,and the outputs were the model parameters,including Henry coefficients and mass transfer coefficients.The operation parameter optimizer was mainly composed of a neural network model.The inputs were the model parameters and operation conditions,and its outputs were the product purities of extract and raffinate lines.Based on this model,the optimal operation conditions under specific model parameters could be quickly obtained by the optimization algorithm.During the operation,when the purity of the product is lower than the required purity due to the decay of the stationary phase,the estimator predicts the model parameters at that time and passes them to the optimizer.The optimizer gives the optimal operating conditions,and so the operating conditions can be adjusted accordingly to achieve a high purity separation of the products.Both models gave high prediction accuracies on the testing sets,the mean absolute percentage error(MAPE)of parameters predicted by the 1D convolutional neural network-long short term memory network model on the testing set was lower than 0.2%,and the MAE of product purities predicted by the neural network model was lower than 0.04%.The results of subsequent simulation experiments showed that the on-line optimization strategy based on the deep learning model can work well under different decay rates of the stationary phase and maintain the high purity separation of the products.