Enhancing biodiesel cetane number prediction: Comparative study of neural network architectures with integer and conformable transfer functions

Abstract

The precise prediction of biodiesel properties, particularly cetane number, presents an ongoing challenge due to the complexity of its composition and combustion behavior. Consequently, this study is dedicated to enhancing the prediction of biodiesel properties by employing integer and conformable transfer functions in the hidden layer of artificial neural networks. The ELU, SWISH, SOFTPLUS, LOGSIG, and TANSIG functions and their conformability were evaluated to predict the cetane number. A dataset of experimental observations was obtained, encompassing 16 input variables. Based on molar ratios, these variables represent the Fatty Acid Methyl Esters (FAME), ranging from C8 to C24. The experimentation incorporated a slope-intercept test with a confidence level of 99% as an initial filter for model selection, demonstrating a commitment to improving biodiesel property prediction. The findings uncovered that a conformable artificial neural network (CANN) model featuring a 16:4:1 architecture and employing the CTANSIG transfer function in the hidden layer and linear function in the output layer achieved outstanding performance. It exhibited a strong correlation coefficient of 0.9384 and RMSE 0.0408, effectively balancing precision and computational efficiency. Additionally, this study streamlined ANN models using conformable transfer functions, thereby reducing complexity and enhancing predictive accuracy.