(371n) Estimator Based Inferential Control for C4 Splitting Process Using t-Sne Algorithm and Artificial Neural Network

Distillation is one of the most complex and energy intensive process in chemical process industries. So, there have been studies to reduce energy consumption and improve product quality by optimizing the distillation control. However, control of the distillation column is a highly challenging task due to the complexity of nonlinear systems with multiple variables and in the presence of many disturbances. To predict product composition, on-line measurements are used by direct composition analyzers, but these are expensive and can have significant time delays. Tray temperature control is frequently used instead because composition is a function of temperature. In a multi-component distillation column, however, it is not effective method since the tray temperature does not represent the product compositions exactly. Therefore, secondary measurements are used to adjust the values of the manipulated variables and estimate the product quality. In this study, a t-distributed stochastic neighbor embedding (t-SNE) algorithm has been applied to select the optimal net input vector based on the Distributed Control System (DCS) data from a C4 splitting process which was obtained every 2 seconds for 5 years about 40 variables. And then, an Artificial Neural Network (ANN) estimator is proposed to infer product composition. In the designed estimator network, Batch Normalization (BN), which minimizes internal covariate shift, was employed to overcome gradient vanishing problems with back-propagation. Overall, the t-SNE algorithm successfully selected the most suitable secondary variables compared to typical method like Principal Component Analysis (PCA). In addition, the developed ANN estimator that was developed demonstrated the ability to predict the product composition of the C4 splitting process, notably including the changes in reflux rate required various feedstock changes. Therefore, an improvement of the estimator performance will greatly contribute to making decisions appropriate for operators. Based on the developed ANN estimator, optimization work will be done to minimize operation cost and maximize production yields in future work.