Artificial Neural Network Model for the Prediction of the Products of a Crude Distillation Column

Abstract

An artificial neural network (ANN) model for predicting the flowrates and temperatures of the products of a crude distillation column in a refinery in West Africa was developed. The ANN model was based on the Bayesian Regularization Back Propagation (BRBP) algorithm, with 16 input and 10 output variables. Ninety sets of plant data were used for training and validating the model, while ten sets of data, different from those used for training, were used for testing the model. Another five sets of data with ±2 or more standard deviations from the mean of a variable were then used to evaluate the extrapolation capacity of the model. Model predictions for the testing dataset showed average percentage deviations between 0.08 and 0.49 for product flowrates, and between 0.04 and 0.09 for temperatures, respectively. Regression analysis on these data also indicated a good model fit for the predicted variables. For the sets of extrapolation data, average percentage deviations were between 2.67 and 19.95 for predicted flowrates, and between 0.67 and 1.98 for temperatures, respectively. Regression analysis showed a lack of model fit for both flowrates and temperatures. While the ANN model accurately predicted product flowrates and temperature for interpolating data, its limited capacity for extrapolation suggests potential challenges beyond the training dataset. These findings emphasise the importance of cautious interpretation and application of the model in scenarios beyond the scope of the training data.