Machine Learning Approach for the Investigation of Metal Ion Concentration on Distillate Marine Diesel Fuels through Feed Forward Neural Networks

Abstract

In this work, a set of Feed Forward Neural Networks (FNN) for the estimation of the metal ion concentration of diesel fuels is presented. The dataset vector is obtained through in situ measurements from distillate marine diesel fuel storage tanks all over Greece, in order to reduce the selection bias. It has been demonstrated that the most correlated ions among them are Aluminum (Al), Barium (Ba) and Calcium (Ca). Moreover, the FNN models are the most reliable models to be used for the model construction under discussion. The initial L2 error is relatively small, in the vicinity of 0.3. However, after removing a small dataset that includes 1–2 data points significantly deviating from the model trend, the error is substantially reduced to 0.05, ensuring the reliability and accuracy of the model. If this dataset is cleared, the estimated error is substantially reduced to 0.05, enhancing the reliability and accuracy of the model. The correlation between the sum of the concentrations of the model in relation with the Density and Viscosity are, respectively, 0.15 and 0.29 which are characterized as weak.