An artificial neural network estimator design for the inferential model predictive control of an industrial multi-component distillation column

---

Thesis Type: Postgraduate

Institution Of The Thesis: Orta Doğu Teknik Üniversitesi, Graduate School of Natural and Applied Sciences, Graduate School of Natural and Applied Sciences, Turkey

Approval Date: 2003

Thesis Language: English

Student: ALMİLA BAHAR

Supervisor: CANAN ÖZGEN

Abstract:

An inferential control methodology, that utilizes an artificial neural network (ANN) estimator for a model predictive controller, is developed for an industrial multi-component distillation column. In the column, propane and butane is separated from a mixture of propane, n-butane, i-butane, and i-pentane with a top product purity of 96% propane and a bottom product purity of 63% n- butane. Dual composition control of the column must be used in a multivariable model predictive controller for an efficient operation. Feedback control systems necessitate the knowledge of the control variables. However; in distillation columns, the on-line measurements of top and bottom product compositions are difficult. Therefore, temperature measurements must be used to predict the product compositions in an inferential way. Temperature measurement points for the inferential control of the column are selected by the help of SVD analysis and also by considering the column dynamics, using a dynamic simulation program, called modified-DAL. A moving window neural network estimator, which incorporates the system dynamics into account, is designed to find out the product compositions from temperature measurements on trays. In this study, the data for the ANN is collected using the modified-DAL instead of the real plant. Since in a normal plant operation, compositions are measured discretely by taking samples for analysis, the estimator results are further corrected in 30 minutes intervals with the composition data taken from the simulation of the column. A Multi Input Multi Output (MIMO) MPC is used with the developed ANN estimator to achieve the dual composition control of the column. The performance of the developed control system utilizing ANN estimator is tested considering set-point tracking and disturbance rejection performances for the unconstrained and constrained cases. The comparison of the responses resulting from the controller utilizing ANN estimator, with the case where direct composition values are used in the control system shows that there is not much difference between the two; IAE scores are similar and the shapes of response curves are similar. Keywords: Distillation Column, Inferential Control, SVD, Artificial Neural Networks, Model Predictive Controller.