Temporospatial graph attention networks with applications for industrial dynamic soft sensor modeling

Deep learning has found considerable application in soft sensor modeling due to its capacity for feature extraction and nonlinear fitting. However, deep learning models are black-box models, and explaining their predictive behavior and incorporating prior knowledge into the model pose considerable challenges. Such limitations hinder the applicability of deep learning models in actual industrial processes. In addition, industrial process data contain substantial process information and are highly nonlinear. They involve dynamic time series and reflect the trends of random variables over time. Therefore, the incorporation of dynamic process information into soft sensors is paramount. In this article, we propose temporospatial graph attention networks (TSGAT) for industrial dynamic soft sensor modeling. This model utilizes the graph attention network to introduce prior knowledge into the model and construct the observable nonlinear relationship among variables. The graph attention network can aggregate information to extract spatial features at each time step. The long short-term memory network is used to extract temporal features. The proposed method is verified through the high–low transformer process within a real-life ammonia synthesis process. The results demonstrate that the TSGAT can achieve high predictive accuracy and a capacity to incorporate prior knowledge into the model. © 2023 Chinese Academy of Sciences. All rights reserved.