Abstract: To address the problems of limited application scope and low prediction accuracy of data-driven prediction models caused by narrow operating condition coverage and high noise of plant measured data in the real-time optimization (RTO) system for crude oil atmospheric-vacuum distillation, a study was conducted with a 8.0 Mt/a crude oil atmospheric-vacuum distillation unit of a petrochemical enterprise as the research object. A dynamic model of the unit was established by the HYSYS software to generate high-quality simulation data covering a wide range of operating conditions, which was then fused with plant measured data to construct a multi-source dataset. On this basis, four machine learning algorithms including KNN, MLP, RNN and PI-GCN+LSTM were adopted to build product prediction models. The results showed that the fused data significantly expanded the value range of key process parameters and supplemented samples of extreme and transient operating conditions. All the four models trained by the fused data exhibited better prediction performance than those trained by pure plant measured data, among which the KNN model achieved a coefficient of determination of 0.96 with second-level response speed, and the PI-GCN+LSTM model provided a new direction for model optimization under complex operating conditions. This method effectively makes up for the defects of plant measured data and improves the adaptability of models under variable operating conditions, which can provide reliable prediction support for the stable operation of the RTO system of crude oil atmospheric-vacuum distillation units and an engineering pathway to solve the data bottleneck in the industrial application of data-driven prediction models.

Key words: crude oil atmospheric and vacuum distillation, machine learning, dynamic modeling, model predictive control, neural network