Temporal Graph Convolutional Network Soft Sensor for Chemical Process with Distributed Outputs

To explore data-driven soft sensors and enhance the model explainability, a temporal graph convolutional network is proposed for chemical processes with distributed outputs. It can model process variable relationships in the form of a topology graph. Firstly, the subgraph from prior knowledge is constructed and supplemented using mutual information. In this way, the advantages of both prior knowledge and process data are explored to construct the topology graph. Then, the spatial and temporal dependencies are captured by a graph convolutional layer and a gated recurrent unit, respectively. Finally, a soft sensor based on the spatiotemporal graph convolutional network is established for quality prediction of the chemical process with distributed outputs. The effectiveness of the proposed model is verified by molecular weight distribution prediction of a styrene radical polymerization process.