Neural Network Based Inferential Model For Ethane Steam Cracking Furnace
The product yield distribution of ethane steam cracking is typically obtained using analysers and lab sampling. Since both methods take time to produce results, primarily depending on them to determine main product yield will hinder immediate control action on the process. In order to resolve this i...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2019
|
| Subjects: | |
| Online Access: | http://eprints.usm.my/46726/1/Neural%20Network%20Based%20Inferential%20Model%20For%20Ethane%20Steam%20Cracking%20Furnace.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The product yield distribution of ethane steam cracking is typically obtained using analysers and lab sampling. Since both methods take time to produce results, primarily depending on them to determine main product yield will hinder immediate control action on the process. In order to resolve this issue, an inferential sensor is required. In this study, a neural network based inferential model is developed. The ethane steam cracking process has been modelled using ASPEN Plus and validated with industrial data taken from literature. The relative error (RE) of the model outputs obtained are less than 10%. The ASPEN Plus model is used for input variable selection, nonlinearity assessment, and data generation for neural network modelling. The input variable selection study found that five variables are significantly influential to the ethane and ethylene yields, namely reactor pressure, coil outlet temperature, steam-hydrocarbon ratio, feed composition, and fuel composition. Nonlinearity assessment of the process shows that the process exhibit asymmetrical response and input multiplicities characteristics, and thus, can be classified as a nonlinear process. Data generated from the ASPEN Plus model is used for training, validation, and testing. Two methods have been used to generate the data which are sequential excitation and simultaneous excitation. Four variables are individually excited and combined to make a sequential excitation profile. Data from sequential excitation is divided into training and validation while data from simultaneous excitation is used solely for testing. Three neural network model, namely the Feedforward Neural Network (FFNN), the Generalized Regression Neural Network (GRNN), and the Extreme Learning Machine Neural Network (ELM-NN) are developed and they are evaluated in terms of prediction accuracy and computational time. The evaluation results show that ELM-NN prediction accuracy is higher than FFNN and GRNN. To train, the best model for ELM-NN, GRNN, and FFNN models require 0.0068 seconds, 0.35 seconds, and 12 seconds respectively. In terms of computation time of new set of input data sample, all three models require less than 0.05 seconds to compute one sample of data. However, computation time of the trained GRNN model increases exponentially with the increasing amount of data samples in a batch while for trained FFNN and trained ELM-NN model, the increment is not significant.
Out of the three models, the ELM-NN gives the best performance in terms of prediction accuracy and computational time. The R2 of the ELM-NN model is 91.3% and 82.6% for ethane and ethylene yield respectively. The model requires 0.0068 seconds to train and 0.0001 seconds to compute ethane yield and ethylene yields from a new set of input data. This makes the model suitable for applications in real time inferential control system. |
|---|