Estimation of the main air pollutants from different biomasses under combustion atmospheres by artificial neural networks

Abstract

The production of biofuels to be used as bioenergy under combustion processes generates some gaseous emissions (CO, CO2, NOx, SOx, and other pollutants), affecting living organisms and requiring careful assessments. However, obtaining such information experimentally for data evaluation is costly and time-consuming and its in situ obtaining for regional biomasses (e.g., those from Northeast Brazil (NEB) is still a major challenge. This paper reports on the application of artificial neural networks (ANNs) for the prediction of the main air pollutants (CO, CO2, NO, and SO2) produced during the direct biomass combustion (N2/O2:80/20%) with the use of ultimate analysis (carbon, hydrogen, nitrogen, sulfur, and oxygen). 116 worldwide biomasses were used as input data, which is a relevant alternative to overcome the lack of experimental resources in NEB and obtain such information. Cross-validation was conducted with k-fold to optimize the ANNs and performance was analyzed with the use of statistical errors for accuracy assessments. The results showed an acceptable statistical performance for all architectures of ANNs, with 0.001–12.41% MAPE, 0.001–5.82 mg Nm−3 MAE, and 0.03–52.30 mg Nm−3 RMSE, highlighting the high precision of the emissions studied. On average, the differences between predicted and real values for CO, CO2, NO, and SO2 emissions from NEB biomasses were approximately 0.01%, 10−6%, 0.14%, and 0.05%, respectively. Pearson coefficient provided consistent results of concentration of the ultimate analysis in relation to the emissions studied and effectiveness of the test set in the developed models.