Modeling of pyrolysis product yields by artificial neural networks

Abstract

Artificial neural network (ANN) needs to be applied to the complex, multivariate, and highly variable biomass and pyrolysis data to define optimum input variables and develop effective models.  In this study, two different ANN methods, the feed-forward network (FFN) and the cascade-forward network (CFN), were applied to model pyrolysis product yields (biochar-BC, bio-oil-BO, and gas mixture-G) from 11 biomass and pyrolysis variables through hierarchical modeling approach.  Both methods were supplied with two subsets of data, with two-thirds being used for training and one-third for testing the performances of the methods, after normalizing all data (72 samples).  The performances of both ANN methods were evaluated by using three statistical parameters.  In general, FFN and CFN methods had very similar performances in training and testing.  Both methods had mean R² of 0.91, 0.96, and 0.95 for training BC, BO, and G, respectively.  For testing of all FFN and CFN models, the R² values of BC and G were less than 0.50, but the R² values of BO were over 0.50 (up to 0.81) for only the last 5 models of FFN and CFN.  Both types of ANNs are promising tools in predicting pyrolysis product yields.