State prediction of distributed parameter systems based on multi-source spatiotemporal information. (November 2022)

• Record Type: Journal Article
• Authors: Mu, Guoqing; Chen, Junghui; Liu, Jingxiang; Shao, Weiming; Zhao, Dongya
• Is Part Of: Journal of process control. Volume 119(2022)
• Page Start: 55
• Page End: 67
• Publication Date: 2022-11
• Subjects: Distributed parameter system -- Ethylene oxychlorination reaction -- Multi-source -- Process state forecasting -- Spatiotemporal; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2022.09.009
• Languages: English
• ISSNs: 0959-1524

Abstract

Abstract: To predict the process states of nonlinear distributed parameter systems (DPS) using various process variables, a novel multi-source spatiotemporal modeling method is proposed in this paper to improve conventional temporal modeling methods. To tackle the challenge of modeling industrial process data of DPS with multi-source spatiotemporal characteristics, a multi-source spatiotemporal network (MS-STN) model, which is the integration of a long short-term memory (LSTM) network to extract information containing temporal characteristics and a convolutional long short-term memory (ConvLSTM) network to extract information containing spatiotemporal characteristics, is constructed. The comprehensive use of various process information is beneficial to improving the fitting accuracy of the model, and the generalization ability of the model can be enhanced at the same time. To prevent the gradient vanishing or gradient explosion in presenting complex spatiotemporal data due to the increase of network layers, a model structure of the residual network based on ConvLSTM is proposed in performing model training. Finally, the industrial ethylene oxychlorination reaction process is taken as an example. The experimental results of predicting the temperature of the reaction tube well demonstrate the effectiveness of the proposed method. Highlights: Temporal models cannot meet requirements of distributed parameter system (DPS). The multi-source spatiotemporal modeling method enhances state prediction of DPS. LSTM and ConvLSTM extract temporal and spatiotemporal information, respectively. ConvLSTM-based residual networks prevent gradient vanishing or gradient explosion.