elopment of Neural Network Model for Predicting Crucial Product Properties or Yield for Optimisation of Refinery Operation

Refinery optimisation requires accurate prediction of crucial product properties and yield
of desired products. Neural network modeling is an alternative approach to prediction
using mathematical correlations. The project is an extension of a previous research
conducted by the university on product yield and properties prediction using non-linear
regression method. The objectives of this project are to develop a framework for the
application of neural network modeling in predicting refinery product yield and properties,
to develop neural network model for three case studies (predicting crude distillation yield,
diesel pour point and hydrocracker total gasoline yield) and to evaluate the suitability of
using neural networkmodelingfor predicting refinery product yield and properties.
The project methodologies used are literature research and computer modeling using
MATLAB neural network toolbox. The framework development for neural network
modeling include aspects such as process understanding, data collection and division, input
elements selection, data preprocessing, network type selection, design of network
architecture, learning algorithm selection, network training, and network simulation using
new data set. Various configurations of neural network model were tested to choose the
best model to represent each case study. The model selected has the smallestmean squared
error when simulated using test data.
The results are presented in the form of the network configuration that gives the smallest
MSE, plots comparing the actual output with the output predictedby the neural network, as
well as residual analysis results to determine the range of deviationbetween the actual and
predicted output. Although the accuracy of the output predicted by the neural network
model requires further improvement, in general, the study has shown the tremendous
potential for the use of neural networkfor predicting refinery product yield and properties.
Suggestions for future study in the area include improvement of the model accuracy using
advanced methods such as cross-training and stacked network, integration of neural
networkwith plant's Advanced Process Control as inferential property predictor, and study
on inverted network for use in a neural network-based controller.