On-Line Monitoring Device for Gas Phase Composition Based on Machine Learning Models and Its Application in the Gas Phase Copolymerization of Olefins

This study addresses the challenges of time-delay and low accuracy in online gas-phase composition monitoring during olefin copolymerization processes. Three flowmeters based on different mechanisms are installed in series to measure the real-time exhaust gas flow rate from the reactor. For the same gas flow, the three flowmeters display different readings, which vary with the properties and composition of the gas mixture. Consequently, the composition of the mixed gas can be determined by analyzing the reading of the three flowmeters. Fitting equations and three machine learning models, namely decision trees, random forests, and extreme gradient boosting, are employed to calculate the gas composition. The results from cold-model experimental data demonstrate that the XGBoost model outperforms others in terms of accuracy and generalization capabilities. For the concentration of ethylene, propylene, and hydrogen, the determination coefficients (R2) were 0.9852, 0.9882, and 0.9518, respectively, with corresponding normalized root mean square error (NRMSE) values of 0.0352, 0.0312, and 0.0706. The effectiveness of the online monitoring device is further validated through gas phase copolymerization experiments involving ethylene and propylene. The yield and composition of the ethylene and propylene copolymers are successfully predicted using the online measurement data. To overcome the time-delay and low precision in on-line gas phase composition monitoring during olefin copolymerization, three flowmeters based on distinct principles are employed to detect exhaust gas in real-time. Machine learning models are used to calculate the gas composition. The efficacy of the online monitoring device is validated through gas phase copolymerization experiments of ethylene and propylene.image