Development of soft sensor with neural network and nonlinear variable selection for crude distillation unit process

Abstract

Abstract This study focuses on the development of a neural network-based soft sensor for the estimation of the product properties for real-time monitoring and control in the crude distillation unit (CDU) process. There are a large number of predictor variables displaying a high level of cross-correlation in the CDU process, which increase complexity of the model and lower the model accuracy. Therefore, a novel variable selection method for neural network that can be applied to describe nonlinear industrial processes is developed to solve the problem. The proposed method is an iterative two-step approach. Firstly, a multi-layer perceptron (MLP) is constructed. Secondly, the least absolute shrinkage and selection operator (LASSO) is introduced to select the input variables that are truly essential to the model with the shrinkage parameter is determined using a cross-validation method. Then, variables whose input weights are zero are eliminated from the dataset. The algorithm is repeated until there is no improvement in the model accuracy. The results show that the model constructed by the proposed soft sensor could successfully follow the dynamics of the CDU process. In addition, the superiority of the proposed approach is illustrated by the comparison with other state-of-art methods. It turns out that the proposed approach can build a more compact model and present higher level of prediction accuracy than other existing methods.