Prediction of Smoke Points of Kerosene Distillates Using Simple Laboratory Tests: Artificial Neural Network versus Conventional Correlations

Abstract

In the present study, an artificial neural network (ANN) model and three well-known correlations were used to predict the smoke points of 430 kerosene distillates from their specific gravities and distillation temperatures. The ANN model was developed in MATLAB software, it is a feedforward multilayer perceptron with a single hidden layer. The optimal number of neurons in the hidden layer as well as the best training algorithm and the best values of connection weights and biases were determined by trial and error using the nftool command. The early stopping technique by cross-validation was employed to avoid overfitting of the model. The developed model composed of 17 sigmoid hidden neurons and one linear output neuron was trained with the Levenberg-Marquardt backpropagation algorithm. This model allowed the prediction of smoke points with a coefficient of determination of 0.852, an average absolute deviation of 1.4 mm and an average absolute relative deviation of 6%. Statistical analysis of the results indicated that the prediction accuracy of the ANN model is higher than that of the conventional correlations. Indeed, in addition to its effectiveness, the proposed ANN method for the estimation of smoke points has the advantages of low-cost and easy implementation, as it relies on simple laboratory tests. Thus, the developed ANN model is a reliable tool that can be used in petroleum refineries for fast quality control of kerosene distillates.