Machine learning modeling and experimental study to forecast the pressure of Very High-Pressure (VHP) steam in an industrial steam cracking process

Abstract

The steam cracking process converts hydrocarbons such as ethane and propane into ethylene, propylene and etc. In this study, a number of crucial features, which can affect the furnace output of Very High-Pressure steam (VHP), are studied. Thus, a set of experiments is conducted at a petrochemical plant to determine the effect of the coil’ outlet temperature of cracked hydrocarbon gas, the feed flow rate, dilution steam (DS), negative pressure in the furnace, and flow rate of injected water into steam on the pressure of VHP steam. In addition, the credibility of industrial data is confirmed using the Leverage method. It is found that a maximum VHP steam pressure for each feed flow rate is obtained when the ratio of dilution steam to the feed rate roughly equals 0.37. The mentioned ratio also controls the coke formation in the reactor coils. In the following, three effective machine learning methods are developed, based on 1077 industrial data samples covering a wide operating conditions' range, for estimating process' VHP steam pressure. The model built by radial basis function (RBF) shows sprier agreement with plant data with an average absolute relative deviation (AARD), standard deviation (SD), and R2 values of 0.11%, 0.79% and 99.33%, respectively, for total data points. It is revealed that the developed models are reliable for controlling the amount of VHP steam pressure by changing feed flow rates, the flow rate of injected water into steam, dilution steam, and the temperature of outlet cracked hydrocarbon gas. In addition, a sensitivity analysis is accomplished using Spearman's correlation coefficient based on the most accurate RBF model.