Application of cascade forward neural network and group method of data handling to modeling crude oil pyrolysis during thermal enhanced oil recovery

Abstract Oil recovery during in situ combustion is majorly controlled by hydrocarbon oxidation and pyrolysis reactions, which govern fuel formation and heat evolution. Fuel deposition, in turn, can be accurately predicted in part through crude oil pyrolysis using thermogravimetry analysis (TGA). The theoretical models based on TGA runs, however, may be limited to crude oil samples at hand. In this study, we develop more general models to predict the residual mass during crude oil pyrolysis based on multi-layer perceptron (MLP), cascade forward (CFNN), generalized regression (GRNN), and radial basis function (RBF) neural networks. More than 2000 experimental data spanning wide range of weight percentages of asphaltenes and resins as well as oAPI gravities, heating rates, and temperatures are used. Moreover, six optimization algorithms; including Bayesian regularization (BR), scaled conjugate gradient (SCG), Levenberg-Marquardt (LM), conjugate gradient backpropagation with Fletcher-Reeves updates (CGF), resilient backpropagation (RB), and conjugate gradient backpropagation with Polak-Ribiere updates (CGP) are used to improve the performance and prediction ability of the MLP and CFNN neural networks. The CFNN model optimized with the LM algorithm best fits all the experimental data with a mean absolute percent relative error of 1.04%. Lastly, a mathematical correlation is developed utilizing the group method of data handling (GMDH) to estimate the residual mass of crude oil pyrolysis in TGA. Despite its simplicity, the correlation also provides very good estimates. Sensitivity analysis showed that temperature followed by asphaltenes and resin content showed the highest effect on mass loss during crude oil pyrolysis. Outliers estimation applying the Leverage approach suggested only 1% of the data points could be doubtful.