| Abstrakt: |
In this study, six feedforward neural network models were developed to estimate the net heat of combustion of light kerosene distillates. These networks use different sets of physicochemical properties of the distillates as input variables and are all composed of 8 sigmoid hidden neurons and one linear output neuron. The networks were designed in MATLAB software with 205 data points using the nftool command. Determining the relative importance of input variables in the networks revealed the significant effect of density on the estimates. The developed models as well as two correlative methods taken from the literature were used to predict the net heat of combustion of 40 other samples. The statistical analysis of the results was carried out by calculating for each estimation method the absolute errors, the mean absolute error, the standard deviation of the absolute errors and the coefficient of determination. It was found that the most accurate method is the neural network model based on the density, viscosity, aromatics content and sulfur content of the distillates. The least efficient method is the neural network that does not include density in its inputs, which once again indicates the importance of this property. Consequently, density should be taken into account to ensure high prediction ability of estimation methods. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink