Abstract
The combined use of an equilibrium model and artificial neural network (NN) regressions has been investigated for modelling biomass gasification. The benefits of this approach are to improve the accuracy of equilibrium calculations, and to prevent the NN model from learning mass and energy balances, thereby minimising experimental data requirements. A complete stoichiometry is formulated, and corresponding reaction temperature difference parameters computed under the constraint of the non-equilibrium distribution of gasification products determined by mass balance data reconciliation. The NN regressions relate temperature differences to fuel composition and gasifier operating conditions. The application of Bootstrap and Bayesian regularisation validation algorithms has been investigated to prevent the NN from overfitting the data, and for estimating prediction intervals (PI). Given the prior knowledge available from experimental data, PI become of particular interest for determining whether a regression is indeed required, or whether it is reasonable to consider a given reaction temperature difference independent of composition and operation variables. The results of a preliminary investigation, illustrated with atmospheric air gasification fluidised bed reactor data, indicate that for the reactions relating to the equilibrium of major gas phase species (the water gas shift reaction and ammonia formation from nitrogen and hydrogen) the temperature difference could be constant. Furthermore, the shift reaction might be at equilibrium. Char, light hydrocarbon, and tar formation reaction temperature differences appear to be more strongly correlated to changes in operating conditions.
