The paper deals with the design of neural based soft sensors to improve product qualitymonitoring and control in a refinery by estimating the stabilized gasoline concentration (C5) in the top flow and the butane (C4) concentration in the bottom flow of a debutanizer column, on the basis of a set of available measurements. Three-step predictive dynamic neural models were implemented in order to evaluate in real time the top and bottom product concentrations in the column. The soft sensors designed overcome the great time delayintroduced bythe corresponding gas chromatograph, giving on-line estimations that are suitable for monitoring and control purposes.