(33g) Neural Network Based Soft Sensor for Pilot-Plant Distillation Column

The development of an inferential soft sensor for pilot-plant distillation column of ethanol-water mixture using neural network (NN) method has been investigated in this work. Inferential sensors are increasingly used in the process industries to infer the value of the main quality variable utilizing much easier to measure secondary variables of the process. The lags between the input variables and the output variable vary due to changes in operating conditions. Previous studies have introduced different methods to estimate lags for input and output variables, but all of them have assumed these lags to be constant regardless of the changes in the operating conditions. In this work, an inferential sensor that can infer the composition of ethanol at the top product using time lags for the input variables and varied first-order time constant lags with the output variable is developed.

The developed inferential sensor is based on a neural network model. This NN model is built in MATLAB and Simulink. An OPC connection is established in order to communicate between MATLAB/Simulink and the Distributed Control System (Emerson DetlaV DCS) of our pilot-plant distillation column. Principal Component Analysis (PCA) and Projection to Latent Structures (PLS) methods are used in this work to remove the outliers from the input variables set and to determine the most correlated values of the input variables and their lags with the output variable (ethanol mole fraction) respectively. The model adaptively selects the correct first-order time constant lags of an output variable according to the instantaneous operating condition (the composition of ethanol is increased or decreased) and assigns a best value for each case. The prediction performance of the proposed NN models (with and without time lags for input variables) is illustrated using experimental data from a pilot-scale ethanol-water distillation column at UT’s Chemical Engineering Department. More than 400 samples were collected to create and validate the results of NN models. The proposed NN model with time lags for input variables and varied first-order time constant lags for output variable gave lower error (RSME=0.01) compared with the NN model without any time lag for input and output variables (RSME=0.06).

In summary, a high accuracy soft sensor for the ethanol composition of the top distillation product has been developed and validated. Based on this soft sensor, an inferential PI controller and Model Predictive Controller (MPC) for this pilot-plant column will be developed in future work.