Temporal PFD-guided graph convolutional networks: a novel approach to process modeling

The present study proposes a novel methodology to construct a regression model of process systems, namely, temporal PFD-guided graph convolutional networks (GCN). The approach integrates domain knowledge derived from process flow diagrams (PFDs) and controller configurations into a GCN framework, enabling enhanced state estimation in chemical processes. We introduce a process topology with temporal propagation derived from PFDs to construct robust graph structures for GCNs. The proposed method integrates causal relationships among process variables and their time-series dependencies, enhancing prediction accuracy and adaptability. A case study of a concentration estimation on the Tennessee Eastman Process (TEP) demonstrates the effectiveness of the PFD-guided GCN. The results indicate significant improvements in prediction accuracy compared to 1D-CNN, especially under abnormal operating conditions and when limited training data is available. This approach provides a practical and generalizable solution for process state estimation and soft sensor applications in dynamic and data-sparse industrial environments.