(745e) A Convolutional Neural Network Model for Biomass Gasification in Fluidized Bed

Biomass is a renewable resource and its utilization has received great attention due to its life cycle carbon-neutrality and the potential to substitute fossil fuel to produce a variety of energy-related products. Thermochemical gasification is an important route for conversion of biomass that results in a product gas consisting H2, CO, CO2, CH4 and other light hydrocarbons that can be used as fuel gas to generate power or as well as raw material to produce variety of chemicals. Among the existing gasifiers, fluidized beds offer many advantages such as high conversion efficiency and great flexibility regarding feedstock type and size.

Modeling and simulation of FB gasifier is a problem of considerable theoretical and computational difficulty mainly due to presence of several chemical and physical processes in the reactor and the lack of proper characterization of solid fuel. The largest challenge of the existing modeling works is lack of generality and predictive capability. Here, we present a neural based (black-box) model that can predict the composition of product gas, amount of tar and solid conversion under a wide range of operating conditions and different types of feedstock.

Initially in this study, elemental composition (C, H, N, O, S), proximate analysis (Volatile matter, Fixed carbon, Ash), structural component (Cellulose, Hemicellulose, Lignin) and ash composition of 35 types of biomass were analyzed using principal component analysis (PCA), Factor analysis (FA), and a variety of machine learning approaches. A new method of biomass characterization is proposed based on composite statistical ensembles to improve understanding of nature of biomass.

Subsequently, over 200 data sets of biomass gasification in fluidized bed were collected, structured and analyzed. Multiple measures of associations between product gas characteristics, biomass constituents, and operating condition were analyzed to detect underlying features of the system that were most useful in developing the predictive model. The architecture of the neural network was founded based on the subjective knowledge of the process of biomass gasification in fluidized bed and statistical inferences of the data. The model was trained using the Bayesian regularization technique that is known for yielding more general models. It should be noted that the outputs of BBM are designed to only be used as constitutive equations and in the last step the equilibrium formulation is used as the physical framework of the model to explicitly close the mass and energy balance.