Artificial neural network prediction of biodiesel properties

• Main Author: Lebes, Frederick B.
• Format: text
• Language: English
• Published: Animo Repository 2009
• Subjects: Biodiesel fuels—Properties; Biodiesel fuels; Energy crops; Feedstock; Neural networks (Computer science); Chemical Engineering
• Online Access: https://animorepository.dlsu.edu.ph/etd_masteral/6062; https://animorepository.dlsu.edu.ph/context/etd_masteral/article/13169/viewcontent/CDTG004817_P.pdf
• Institution: De La Salle University

Biodiesel has emerged as an increasingly important renewable biofuel due to its domestic availability and compatibility with modern day diesel engines. However, testing some of its fuel properties can be laborious, expensive and infeasible in some cases. Neural network (NN) models were developed in this study to predict the cetane number and the kinematic viscosity of biodiesel. The models can be a screening tool in classifying potential feedstocks for biodiesel production. By its pattern recognition and learning ability, NNs are known to fit nonlinear data and perform well in prediction tasks. Using the chemical properties of the feedstock, a pool of NN architectures with 8 input nodes and 2 output nodes were trained and were tested to determine the architecture with the optimum number of hidden nodes by trial and error. The work was done using a neural network prediction software (NNpred) running in Visual Basic. Two types of NN architectures were developed in this study: one hidden layer neural network and two hidden layer neural network. The cetane number prediction results of the NN models were compared to a simple linear model proposed by Krisnangkura (1986). The NN models were robust, but the linear model underestimates cetane number at increasing unsaturation of the feedstock. Hence, the NN models were found more favorable under these conditions. Similarly, the kinematic viscosity prediction results of the NN models were compared to a nonlinear model proposed by Allen et al (1999). The results showed that the NN models were less accurate than the nonlinear model, but can give a practical rough estimate when the complete fatty acid profile of the feedstock is not available