Enhancing predictive monitoring of ethylene oxychlorination reactor states through spatiotemporal coupling analysis

The production of polyvinyl chloride (PVC) encounters challenges stemming from the temporal and spatial coupling characteristics inherent in the fixed bed ethylene oxychlorination process. Consequently, the implementation of enhanced safety measures and risk reduction strategies becomes imperative. This study introduces a pioneering methodology leveraging a spectral temporal graph neural network. By leveraging reactor temperature data, spatial variable decoupling facilitated by the Fourier transform, and a self-attentive mechanism within graph neural networks, the proposed approach adeptly forecasts future reactor states. The model's seamless alignment with the physical knowledge of reaction processes, validated through the adjacency matrix and hotspot region identification, underscores its efficacy in ensuring process safety and mitigating operational risks in PVC production. Empirical findings further validate the effectiveness of the approach, with predictions demonstrating an error margin of less than 0.5°C in forecasting future reactor temperatures.