Predictability of higher heating value of biomass feedstocks via proximate and ultimate analyses – A comprehensive study of artificial neural network applications

Higher heating value (HHV) is a key characteristic for the assessment and selection of biomass feedstocks as a fuel source. The HHV is usually measured using an adiabatic oxygen bomb calorimeter; however, this method can be time consuming and expensive. In response, researchers have attempted to use artificial neural network (ANN) systems to predict HHV using proximate and ultimate analysis data, but these efforts were hampered by varying case specific approaches and methodologies. Based on the complex ANN structures, a clear state of the art ANN understanding must be required for the prediction of biomass HHV. This study provides a comprehensive ANN application for HHV prediction in terms of how the activation functions, algorithms, hidden layers, dataset, and randomisation of the dataset affects the prediction of HHV of biomass feedstocks. In this paper we present a comparative qualitative and quantitative analysis of thirteen different algorithms, four different activation functions (logsig, tansig, poslin, purelin) with a wide range of hidden layer (3–15) for ANN models, used to predict the HHV of the biomass feedstocks. ANN models trained by the combination of ultimate-proximate analyses (UAPA) datasets provided more accurate predictions than the models trained by ultimate analysis or proximate analysis datasets. Regardless of the used datasets, sigmoidal activation functions (tansig and logsig) provide better prediction results than linear activation function (poslin and purelin). Furthermore, as training activation functions, “Levenberg-Marquardt (lm)” and “Bayesian Regularization (br)” algorithms provide the best HHV prediction. The best average correlation coefficients of 30 randomised run were observed with tansig as 0.962 and 0.876 for the ANN model developed by the UAPA dataset with a relatively high confidence levels of ∼96% for training and ∼92% for testing.