A recurrent neural network model for biomass gasification chemistry

Abstract

Detailed chemical kinetic models involving hundreds of species and thousands of reactions have recently been developed for biomass thermochemical conversion. The high computational cost of these kinetic models makes them impractical even for simple reactor geometries. In this work, we develop a recurrent neural network (RNN) model for the secondary gas-phase reactions of biomass gasification in an inert environment in the temperature range of 800–1000 °C. A gated recurrent unit (GRU) based RNN architecture is used to ensure accurate predictions over the entire range of time in the reactor. A compact kinetic model reduced from a detailed kinetic scheme using an automated reduction algorithm is employed as the reference kinetic scheme for the gas-phase reactions. A comprehensive range of biomass compositions and reactor conditions are used to generate the training data ensuring a wide range of the model applicability. The developed GRU-based RNN model can predict the temporal evolution of important reactants and products during biomass gasification in the freeboard region of a fluidized bed reactor. The model reduces the computational cost associated with the reference kinetic scheme by four orders of magnitude.