PERFORMANCE COMPARISON OF DIFFERENT ACTIVATION FUNCTIONS IN NEURAL NETWORKS FOR BIOMASS ENERGY CONTENT PREDICTION
Keywords:
Artificial neural networks, biomass, higher heating value, proximate analysis, transfer functionAbstract
The application of artificial neural networks to solve complex linear and nonlinear problems such as predicting biomass higher heating value (HHV) requires meticulous choice of the number of layers and neurons per layer, the training algorithms, and the activation functions among other hyper-parameters. Although some studies have examined these hyper-parameters, the effects of different activation functions on biomass HHV prediction have not attracted credible research attention. This study, therefore, employs three distinct activation functions (logsig, tansig, and purelin) in artificial neural networks for biomass HHV prediction based on proximate analysis. A 3-10-1 network architecture was used and the variation of the hidden layer and output layer activation functions yielded nine models (M1-M9) whose performances were assessed and compared using statistical indices. The results showed that the best performance was observed by model M2 which utlized the logsig function in the hidden layer and the tansig function in the output layer. This model had the highest determination coefficient of 0.8814 and the lowest mean square error and mean absolute error of 0.0017 and 0.0281 respectively. Understanding how these hyper-parameters influence biomass HHV prediction would guide the energy community to identify an optimal pathway to bioenergy production.
Published
How to Cite
Issue
Section
FUDMA Journal of Sciences