Artificial neural network model development of blended biodiesel

The need to find a more sustainable and environment friendly biofuel has driven a lot of experimental studies in past few decades. But, the cost of physical testing and resource constraint has limited the capability to find better and low emission fuels. This study aims to bridge this gap by developing an artificial neural networks (ANNs) model that can predict the emission (i.e., nitrogen oxide (NOx), carbon oxides (COx), hydrocarbons (HC)) of biodiesel and its blends based on the composition of feedstock and its additives, physiochemical properties and engine working conditions. The ANN model used a comprehensive dataset of 424 variation of biodiesel based on 17 input variables. The final optimised ANN model is configure to have data division ratio of 60/20/20, learning rate of 0.005, batch size 64, hidden neuron of 25 and run up to 200 epochs. The Scaled Conjugate Gradient (SCG) algorithm provided better result in comparison to the Levenberg-Marquardt (LM) algorithm with remarkable overall coefficient of determination (R²) of 0.977 and mean squared error (MSE) in the range of ~ 10^-7. The predictive capability of model for feedstock alone, showed R² of 0.980 with MSE of 2.34 × 10^-7, while for blended biodiesel slightly less R² of 0.911 with MSE of 5.29 × 10^-5 is achieved due to unstandardised condition of data present in the literature. Finally, the developed ANN model can help pave the path to discover novel biodiesel and its blended versions with less emissions at low cost.