How to Train an Artificial Neural Network to Predict Higher Heating Values of Biofuel

Abstract

Plant biomass is one of the most promising and easy-to-use sources of renewable energy. Direct determination of higher heating values of fuel in an adiabatic calorimeter is too expensive and time-consuming to be used as a routine analysis. Indirect calculation of higher heating values using the data from the ultimate and proximate analyses is a more rapid and less equipment-intensive method. This study assessed the fitting performance of a multilayer perceptron as an artificial neural network for estimating higher heating values of biomass. The analysis was conducted using a specially gathered large and heterogeneous dataset (720 biomass samples) that included the experimental data of ultimate and proximate analysis on grass plants, peat, husks and shells, organic residues, municipal solid wastes, sludge, straw, and untreated wood. The quantity and preprocessing of data (namely, rejection of dependent and noisy variables; dataset centralization) were shown to make a major contribution to prediction accuracy improvement. In particular, it was demonstrated that 550 samples are sufficient to ensure convergence of the algorithm; carbon and hydrogen contents are sufficient ultimate analysis data; and volatile matters can be excluded from proximate analysis. The minimal required complexity of neural network is ~50 neurons.