Establishment of a neural network model for ethylene production from naphtha feedstock
Advancing the chemical engineering fundamentals
Chemical Reaction Engineering (T2-2P)
Keywords: Artificial Neural Network, Steam Cracking, Naphtha, Ethylene
Thermal cracking or steam pyrolysis of hydrocarbons converts them into unsaturated raw materials such as ethylene and propylene, which can be used in the petrochemical industry for polymer production. The reaction mechanisms of naphtha cracking are generally accepted as free-radical chain reactions. Unfortunately, the absence of a simple predictive applied model of pyrolysis is an obstacle to the development of practical methods of conversion.
Neural networks have been used as a promising opportunity, when complex reaction systems can not be well identified, or in the case of lack of basic knowledge of reaction mechanisms. It has been claimed that Artificial Neural Networks (ANN) are 120-5000 times faster than phenomenological models, and can therefore lead to significant reductions in computation times. Various aspects of kinetic modeling of chemical reactors with multilayer feed forward networks have been studied. Most published works on ethylene synthesis and kinetics are based on mechanistic models. No attempts have been made to incorporate the use of neural networks in modeling such process.
In this paper, the steam cracking process with naphtha feedstock is modeled by a multilayer, feed forward, fully connected artificial neural network. Flow residence time, steam ratio and coil outlet temperature were the input variables to the network. These input variables will help to generalize the model, because the model does depend on the reactor size or feed flow rate. The output variables of the network were hydrogen, methane, acetylene, ethylene and ethane yields. Levenberg-Marquardt algorithm was used for training of network. All ANN calculations were carried out using MATLAB7 mathematical software with ANN toolbox for windows.
Since so many data are needed for the neural network training, the experimental data are used only for model validation. In order to generate the training data for the network, a running program is written in the Visual Basic Editor of EXCEL. In this program, the output variables are evaluated by changing the input variables in proper ranges and executing the mathematical model of the reactor. The program is linked to HYSYS in order to make use of the property package. It is also linked to MATLAB software to transfer the generated training data.
The most suitable neural network topology turned out to be a multi-layer perceptron with three nodes in the input layer (reaction conditions), fifteen nodes in the hidden layer (with sigmoid activation functions) and five nodes in the output layer (reaction results). The input variables are chosen in a way to help generalize the neural network model, so that it can be used for a similar reactor of different geometry.
See the full pdf manuscript of the abstract.
Presented Tuesday 18, 13:30 to 15:00, in session Chemical Reaction Engineering (T2-2P).
