Soft-sensor models to estimate the efficiency of H₂S removal from an oil refinery stream of nonphenolic sour water

ABSTRACT

A large set of results of concentration of hydrogen sulfide in the feed and bottom product streams of a sour water stripper found in a typical oil refinery was experimentally obtained. The readings of H₂S concentrations were at several different operating conditions in terms of the main process variables that classically have significant effects on the efficiency of H₂S removal (E). In particular, the considered factors were the mass flow rate, and temperature of sour water fed into the stripper, the mass flow rate of external steam injected into the reboiler, the difference between the temperature of the product stream leaving and entering the reboiler, and the difference of pressure at the two ends of the tower. Three different soft-sensor models were suggested to describe the observed variation in E from 63 to 97%, namely, an equilibrium, statistical and an artificial neural network model. The best of them was the neural network one with three input variables, four neurons in the one-hidden layer, and a hyperbolic tangent function for both the output and one-hidden layers. The mean absolute relative deviation between measured and calculated E by involving this model was only approximately 2.5% with negligible tendency for the residuals. It confirms the reliability of this approach as a tool to inferential estimation of the efficiency of removal of H₂S from the sour water by stripping.

KEYWORDS:

• Efficiency of H₂S removal
• equilibrium model
• hydrogen sulfide
• neural network
• oil refinery
• sour water
• statistical model
• stripper