"Application of Neural Network in developing Virtual Analyzer of Reformate Research Octane Number"

The interest of this Final Year Research Project covers the topic of Application of
Artificial Neural Networks for developing virtual analyzer for petroleum quality,
Research Octane Number. In general, the work deals with the potential application of
neural network technology to Research Octane Number of Reformate estimation. This is
done by presenting the system with a representative set of examples describing the
problem, namely pairs of input and output samples; the ANN will then extrapolate the
mapping between input and output data. The trained network was able to accurately and
efficiently estimate the Research Octane Number at a given time. Statistical analysis was
also conducted to verify if the key variables tor estimating the Research Octane Number
are suitable for network training. The selected key variables in predicting Research
Octane Number are, teed flow rate, recycle flow rate, coil outlet temperature of furnace
and equivalent temperature bed of reactors.
1068 sample data points are used tor modeling the Research Octane Number which then
are divided selectively intro three sections; training, validation and testing data. For this
case study, Backpropagation Network and Levenberg Algorithm are used. To evaluate
the performance of the neural network model, the trained network was simulated using
data that the network has not been trained before. The optimum configuration tor the
network is 2 hidden layers which 16 and 4 neurons respectively with R-squared is equal
to 0.75. The design of the model is described in depth and further improvement is done
for increasing the R -squared, and the MA TLAB source codes are included in appendices.