Content area
In modern process industries, precise process modeling plays a vital role in intelligent manufacturing. Nevertheless, both mechanistic and data-driven modeling methods have their own limitations. To address the shortcomings of these two modeling methods, we propose a neural network model based on process mechanism knowledge, aiming to enhance the prediction accuracy and interpretability of the model. The basic structure of this neural network consists of gated recurrent units and an attention mechanism. According to the different properties of the variables to be predicted, we propose an improved neural network with a distributed structure and residual connections, which enhances the interpretability of the neural network model. We use the proposed model to conduct dynamic modeling of a benzene–toluene distillation column. The mean squared error of the trained model is 0.0015, and the error is reduced by 77.2% compared with the pure RNN-based model. To verify the prediction ability of the proposed predictive model beyond the known dataset, we apply it to the predictive control of the distillation column. In two tests, it achieves results far superior to those of the PID control.
Details
Accuracy;
Proportional integral derivative;
Benzene;
Optimization techniques;
Artificial neural networks;
Predictive control;
Data processing;
Dynamic models;
Prediction models;
Intelligent manufacturing systems;
Catalytic cracking;
Efficiency;
Machine learning;
Simulation;
Production controls;
Product quality;
Control algorithms;
Artificial intelligence;
Digital twins;
Neural networks;
Process controls;
Variables;
Distillation;
Methods;
Chemical engineering
1 Center for Process Simulation & Optimization, Beijing University of Chemical Technology, Beijing 100029, China;