Soft sensor for better control of carbon dioxide removal process in ethylene glycol plant

Abstract

Low Carbon-dioxide in cycle gas loop of Ethylene Glycol (EG) plant improves catalyst selectivity and overall economics of the plant. Carbon-dioxide produced as a bi-product in ethylene oxide reactor is removed by Benfield process. In this process the carbonate and bicarbonate ratio in lean carbonate solution is considered as an important quality control (QC) variable as the efficiency of carbon-dioxide removal is largely depends on it. In the event of a process malfunction or operating under suboptimal condition, the CO2 content in cycle gas loop will continue to rise until corrective action is taken after obtaining lab results for carbonate and bicarbonate ratio. This time consuming sampling process can be overcome by implementing a technological solution in form of an accurate and robust mathematical model capable of real time QC variable prediction. For well understood processes, the structure of the correlation for QC variables as well as the choice of the inputs may be well known in advance. However, Benfield process is too complex and the appropriate form of the correlation and choice of input variables are not obvious. Here, the processes knowledge, operating experience and statistical methods were applied in developing the soft sensor. This paper describes a systematic approach to the development of inferential measurements of carbonate and bicarbonate ratio using Support Vector Regression (SVR) analysis. Given historical process data, a simple SVR-based soft sensor model is found capable of identifying and capturing the cause and effect relationship between operating variables (model inputs) and QC variables (model outputs). Special care was taken to choose input variables, so that the final correlation and regression coefficient make senses from process engineering point of view. The developed soft sensor is implemented in commercial ethylene glycol plant in an Exaquantum interface and found satisfactorily predicting the carbonate and bicarbonate ratio in real time.