AN ADAPTIVE SOFT SENSOR DESIGN BASED ON A NEURAL NETWORK WITH A PREDICTIVE FILTER IN THE FEEDBACK LOOP FOR A REACTION DISTILATION TECHNOLOGICAL PROCESS

Abstract

Rectification and reactive distillation columns are the main of all units in the petrochemical and refining industry. Soft sensors consist of mathematical models that estimate of the quality of an output product in real time are used for technological processes control. In general, changes in the composition of raw materials, catalyst deactivation, etc. result in inconsistency between obtained data and the current state of the technological process. Soft sensor design obtained on such data will loss of accuracy in estimating the necessary parameters of the output product. An adaptive soft sensor design based on a neural network with a predictive filter in the feedback loop for solve the mismatching obtained data and the current state of the technological process problem is proposed. “Moving window” conception is used for size window adapting to the actual state of the technological process. Parameter estimation based on a neural network using data matching to the technological process. A predictive filter in the feedback loop for improve the estimation accuracy of the quality parameter at the cost of predicting the error of the soft sensor designed is proposed. A comparative analysis of several adaptive soft sensors based on neural networks using the "moving window" conception for estimation a by-product concentration of in the output product of the reaction-distillation column and the effectiveness of the proposed approach are shown. Application of the predictive filter in the feedback loop allows to improve the accuracy of the soft sensor based on a neural network by 12.94 % (coefficient of determination) and by 39.81 % (mean absolute error) in comparison with that of without predictive filter.