Prediction of flue gas properties using artificial intelligence : Application of supervised machine learning by utilization of Near-Infrared Spectroscopy on solid biofuels

This degree project studies implementation and comparison of different AI models to predict (1) solid biofuel elements including carbon, hydrogen, nitrogen, and oxygen as well as moisture content, ash content, and higher heating value (HHV) of the fuel and (2) flue gas compositions such as concentration of carbon dioxide, carbon monoxide, nitrogen, nitrogen oxides, and water content using near-infrared spectroscopy and chemometric approaches. The study executes these predictions by simulating the operation of a combined heat and power plant (CHP) that is equipped with carbon capture and storage (CCS). The focus of this study is to investigate the possibility of using Near-Infrared spectroscopy (NIRs) technology to predict the emissions from a CHP plant, which can further improve the performance of the CCS system by providing the necessary fuel data in real time. The acquired NIR data is used to develop the Artificial Intelligence (AI) models using several regression algorithms including Linear regression, Gaussian process, Support Vector Machine, Artificial Neural Network, Ensemble Trees, and Tree regression. NIR data has been pre-processed using Savitzky-Golay (SG) and Multi scatter correction (MSC) techniques. Highest accuracy has been achieved for moisture content of the fuel using Exponential Gaussian Process, where an RMSE of 2.5687 and an R2-value of 0.9 has been obtained. Indeed, only a handful of regression algorithms have shown reasonable accuracy when predicting the fuel elements, where the HHV of the fuel has been predicted poorly as none of the algorithms have been able to execute the prediction successfully which leads to negative values of R2. Prediction of flue gas composition has been done resulting in reasonable accuracies for CO2 fraction with values of 0.1051 and 0.6 for RMSE and R2 respectively. Furthermore, the computational time of the algorithms has been investigated, which indicates that some of the pre-processing techniques could improve the computational time for a certain regression model, but not for all of them. It is conclusively possible to predict fuel elements and flue gas compositions based on data acquired from NIR spectroscopy. However, great effort must be put into model development including data treatment and AI model calibration to achieve desirable precision and reliable results