Application of data mining tools to extract knowledge for dry reforming of methane from published papers in literature

Abstract

The aim of this thesis is to extract knowledge for dry reforming of methane reaction using data mining techniques that are decision trees and artificial neural networks. Firstly, the experimental data were collected from 101 papers published between 2005 and 2014. The data set consisted of 5521 data points with 63 variables. The conversion of methane as a function of catalyst preparation and operational variables was modeled in MATLAB. CH4 conversion values were classified as 0-50, 50-75 and 75-100 for the decision tree analyses of total data, and Ni, Co, Pt and wet impregnation bases subsets. Training and testing errors of these trees were 21.1%-21.5%, 16.7%-19.0%, 10.2%-10.9%, 9.3%-12.6%, and 13.6%-15.1%, respectively. The neural networks were also used for modeling data; tansig was used as activation function and trainlm and trainbr were used as training algorithms for training and testing, respectively. The optimal neural network topology was found as 63-20-1 (20 neuron in hidden layer) using the prediction ability of testing result. R2 and corresponding RMSE values for training and testing were found to be 0.97-4.23 and 0.89-8.66, respectively. The relative significances of input variables were also determined using the optimal neural network topology. It was found that reaction temperature was the most significant variable, and operational variables had higher group significance than the catalyst design variables within the range of data set. Finally, the optimal neural network topology was used to predict the results of experiments and papers. 590 out of 753 experiments (78.4%) were predicted with RMSE values lower than 15. The results of 46 out of 101 papers (45.5%) were predicted with R2 values higher than 0.5 and the RMSE values of 65 out of 101 papers (64.4%) had lower than 15.