Abstract:

The simulated moving bed model serves as the foundation for process optimization. To overcome the respective limitations of mechanistic and data-driven models, a hybrid modeling approach is proposed by embedding neural networks into the mechanistic framework. First, data were collected from twin-column experiments, and a dual-branch neural network was employed to learn the nonlinear mapping between system states and adsorption rates. Then, differential evolution combined with grid search was used to identify the model structure and parameters, upon which a sequential simulated moving bed system was constructed. Finally, the proposed method was validated through the purification of xylo-oligosaccharides. Results demonstrate that the model can accurately predict the outlet concentrations in both twin-column experiments and sequential simulated moving beds. Furthermore, process optimization based on this model yielded a continuous and uniformly distributed Pareto front, with relative prediction errors of product purity and recovery below 2%.

Key words: simulated moving bed, chromatography, separation, hybrid modeling, process optimization, neural networks

摘要：

模型是模拟移动床工艺优化的基础。针对机理模型和数据驱动模型各自的不足，提出一种在机理模型中嵌入神经网络的混合建模方法。首先，通过双柱实验采集数据，利用双分支神经网络学习系统状态与吸附速率之间的非线性映射；然后，结合差分进化与网格搜索辨识模型结构与参数，构建顺序式模拟移动床系统；最后，以低聚木糖的纯化过程进行验证。结果表明，该模型能准确预测双柱实验与顺序式模拟移动床的出口浓度，并在工艺优化中获得连续均匀的帕累托前沿，目标产物纯度和收率的预测误差均低于2%。

关键词: 模拟移动床, 色谱, 分离, 混合建模, 工艺优化, 神经网络