Feed Forward Neural Network Model for Production of Isopropyl Myristate in a Semibatch Reactive Distillation: An Evaluation of Extrapolation Prediction Capability

Abstract

Since last decade, the application of neural network (ANN) has been grown in chemical industries especially for the model based control system due to its capability to solve the complex model and its feasibility for online application. In general, the development of a good ANN model is depending on the quality of the data and the model structure. However, the ANN has a limitation in predicting extrapolation data. Thus in this paper, two multiple inputs - multiple outputs (MIMO) models (MIMO1 and MIMO2) were developed for production of isopropyl myristate in a semibatch reactive distillation and the capability of predicting the extrapolation data is evaluated by z-score normalization technique. Two data sets are simulated based on two possible scenarios occurs in industry: the first scenario is when the constant reflux ratio is applied for 10 h of batch time; the second scenario is when there is a high excess of isopropanol in the reboiler at the end of the process. The result shows that by using the z-score normalization, the MIMO2 was able to predict top composition (x_d) better than bottom composition (x_b) Since the MIMO2 model shows the ability to generalize extrapolation data for both x_b and x_d with low mean square error (MSE) and high coefficient of determination (R²) value, thus it proves that using the z-score normalization method can facilitate the model to extrapolate the data satisfactorily.