Neural network prediction of biodiesel kinematic viscosity at 313 K

• An ANN method was developed to predict the kinematic viscosity of biodiesel at 313 K.
• The ANN method avoids the need of the kinematic viscosities of the individual FAMEs.
• The ANN method has a better performance than two previous prediction methods.
• The ANN method can be trained to account for the effect of double bond position on viscosity.

Viscosity is an important fluid property because of its direct relation with the fuel injection process for engines. The kinematic viscosity of biodiesel at 313 K should satisfy the range specified by the international biodiesel standards. In this study, an artificial neural network (ANN) method was developed to predict the biodiesel kinematic viscosity at 313 K with the experimental data of 105 biodiesel samples collected from the literature. The ANN method only uses the mass fractions of 19 fatty acid methyl esters (FAMEs) as inputs, which avoids the need of the kinematic viscosities of the individual FAMEs required by the prediction methods using the mixing equations. Two previously reported methods based on empirical equations, the Knothe–Steidley method with the experimental or predicted (when experimental data are unavailable) viscosities of FAMEs as inputs and the Ramírez-Verduzco method with the predicted viscosities of FAMEs as inputs were also extensively evaluated to compare with the proposed ANN method. Results indicate that the proposed ANN method is able to predict overall more accurate biodiesel kinematic viscosities at 313 K with the mean squared error (MSE) of 0.0099 compared with the Knothe–Steidley method and the Ramírez-Verduzco method. In most cases, both the Knothe–Steidley method and the Ramírez-Verduzco method tend to under-predict the kinematic viscosities. The underestimation might be attributed to the viscosity contributions of minor mono-, di- and triacylglycerols resulting from incomplete transesterification and minor or trace amount of other FAMEs to the experimental viscosities, as well as the neglect of the interactions between the individual components in calculation of the mixture kinematic viscosities. However, the ANN method could learn to account for the viscosity contributions from the minor components and the interactions between the individual components to some extent during the training process. Overall, the ANN method realizes the best accuracy for the prediction of biodiesel kinematic viscosity with the highest correlation coefficient of 0.9774.