Combined Modeling of Static and Sequential Data for Petrochemical Process

The traditional modeling of petrochemical processes only uses static data, but fails to fully consider the impact of sequential information on modeling indicators. In this study, a combination network of static and sequential data (CNSS) for petrochemical process was proposed. The feedforward neural network was used to extract the information of the static data. The bidirectional long short-term memory (Bi-LSTM) and the self-attention mechanism were used to extract the chronological logic and cross information of the operating variables, respectively. Through the combination of the information from static and sequential data, a better prediction performance of CNSS can be obtained. The model was verified through the research octane number (RON) prediction of refined gasoline from a S Zorb unit and the concentration prediction of outlet NO_x from flue gas denitration system in fluid catalytic cracking (FCC) unit, respectively. The results show that CNSS has a much higher prediction accuracy than the machine learning model with static data. Its average absolute error and average absolute percentage error for RON prediction of refined gasoline are 0.1091 units and 0.12%, and those for NO_x outlet mass concentration prediction are 2.4430 mg/m³and 5.60%, respectively. The CNSS model can be an important reference to establish a machine learning model for petrochemical process, in which timing information should be considered due to the large fluctuations in the values of process parameters.