A Bayesian inferential sensor for predicting the reactant concentration in an exothermic chemical process

In many chemical reactors, concentration measurements are conducted off-line in a laboratory, which involve manual work and can therefore be conducted only infrequently. We propose a Bayesian inferential sensor to predict the reactant concentration in the inlet stream of an exothermic chemical process. The inferential sensor is based on the Bayesian inverse approach and the autoregressive integrated moving average (ARIMA) model. It enables the prediction of the reactant concentration at the frequency of automated on-line measurements, which is typically much higher than that of laboratory measurements. We demonstrate the method on real industrial process data from catalytic hydrogenation of aromatic compounds. The predicted aromatics concentration in the inlet stream, generated based on the latest on-line measurements and two-week-old laboratory data, has a coefficient of determination of 0.936 and a root mean square error of 0.654 mass-%.