Prediction of chemical exergy of syngas from downdraft gasifier by means of machine learning


Sezer S., Kartal F., ÖZVEREN U.

THERMAL SCIENCE AND ENGINEERING PROGRESS, vol.26, 2021 (Journal Indexed in SCI) identifier

  • Publication Type: Article / Article
  • Volume: 26
  • Publication Date: 2021
  • Doi Number: 10.1016/j.tsep.2021.101031
  • Title of Journal : THERMAL SCIENCE AND ENGINEERING PROGRESS
  • Keywords: Artificial neural network, Biomass gasification, Downdraft gasifier, Aspen Plus (R), Chemical exergy, ARTIFICIAL NEURAL-NETWORK, BIOMASS GASIFICATION PROCESS, FLUIDIZED-BED GASIFIER, MUNICIPAL SOLID-WASTE, STEAM GASIFICATION, CO-GASIFICATION, HYDROGEN-PRODUCTION, AIR GASIFICATION, KINETIC-ANALYSIS, BOTTOM ASH

Abstract

The rapid consumption of fossil fuels because of the increasing energy demand caused the increase in greenhouse gas emissions. However, biomass gasification is attracting much attention as an environmentally friendly and highly efficient thermochemical conversion due to its high carbon conversion and low greenhouse gas emissions. Further, downdraft gasifiers are known as the most suitable technology for biomass gasification processes because they offer an easy-to-control working environment and low investment cost. In recent years, artificial neural network models (ANN) have been used in the literature as a machine learning approach to predict gasification parameters. In this work, the parametric study was carried out for the variation of gasifier temperature (873.15 K-1173.15 K) and steam/biomass ratio (0.1-1.5) for 22 lignocellulosic biomass samples. Thus, 32,025 different experimental conditions generated by Aspen Plus (R) were used with Bayesian regularized ANN as a machine learning approach to predict the chemical exergy of the syngas from the downdraft gasifier. The operating parameters of gasifier temperature and steam/biomass ratio were found to be highly influential on the syngas quality and chemical exergy value of the syngas. Therefore, the operating conditions and biomass properties (carbon, hydrogen and oxygen content) were selected as input parameters for the ANN model. The regression coefficients (R-2) were found to be convincingly 0.9992, 0.9991 and 0.9942 for training, test and hazelnut shell gasification data, respectively. Moreover, the results for root mean squared error (RMSE) were within satisfactory limits for the developed ANN model.