Soft Sensors for Kerosene Properties Estimation and Control in Crude Distillation Unit

Neural network-based soft sensors are developed for kerosene properties estimation, a refinery crude distillation unit side product. Based on temperature and flow measurements, two soft sensors serve as the estimators for the kerosene distillation end point (95 %) and freezing point. Soft sensor models are developed using linear regression techniques and neural networks. After performing multiple linear regression analysis it is determined that it is not possible to realize linear models. Within MLP neural networks the number of neurons in the hidden layer are varied and different learning algorithms are used (back propagation with variations of learning rate and momentum, conjugate gradient descent, Levenberg-Marquardt) as well as pruning and Weigend regularization techniques. Bootstrap resampling with replacement and cross-validation resampling are used for improving generalization capabilities. Statistics and sensitivity analysis is provided for both models. Two developed soft sensors will be used in crude-oil unit as on-line estimators of kerosene properties, which so far were available only as infrequent and irregular laboratory analyzers.