Process structure-based fully connected neural network for the modelling of chemical processes: A comparison between global and modular configurations

Abstract

BackgroundChemical process modeling is essential for the optimization, precise control and prediction of processes performance. As the modeling based on the first principles is time-consuming and labor-intensive, utilizing neural networks for the process modeling was proposed. However, the poor interpretability of the neural network models is may a problem in the modelling of chemical processes where the accuracy and reliability of the model are highly required. MethodInspired by the functional partitioning of neural networks in human brain, in this paper, global fully connected neural network (GFNN) and modular fully connected neural network (MFNN) models are constructed for the production process of ethylbenzene and that of isopropanol with cycle process, respectively. For the chemical processes with complex cycle processes, the quality of the product can be easily affected by small feed changes due to the highly nonlinear model which is called snowball effect. To avoid snowball effect, the cycle-avoidance MFNN is constructed, and it shows better modelling performance than GFNN. Significant findingsMFNN model has more advantages than GFNN model in the modelling of large-scale chemical processes without cycle processes. In addition, it is clear that MFNN model has a partial interpretability which means the local physical meaning for different parts of the whole process.