Influence of fuel and operation mode on air pollutants emission from pulverized coal-fired power plant: Field experiments and ML predictions

• Experimental investigations of air pollutant emissions at TPP “Kolubara” conducted. • Coal proximate/ultimate analysis, steam production and boiler power served as input. • Random Forest, XGBoost and ANN algorithm applied on formed measurements database. • ANN showed the best RMSE and MAPE for prediction of CO, NO x and SO 2 emissions. • Feature importance explored and new application for emissions prediction deployed. Experimental investigations of pollutant emissions are of great importance for the control of power plant operation to comply with emission limits, while development of machine learning models in the field of energy has received increasing attention. In this study, methodologies for experimental determination of proximate and ultimate coal analysis and air pollutant emissions at thermal power plant “Kolubara” are presented. Based on the measurement results, a database is generated consisting of input parameters (proximate and ultimate analysis of lignite, power output of thermal power plant Unit A1 and steam production) and output features (CO, NO x and SO 2 emissions). Three machine learning models (Random Forest, Extreme Gradient Boosting and Artificial Neural Network) are applied on the database for the emissions prediction. Artificial Neural Network results show the best possibility of a reliable assessment of the air pollutant emissions. The feature importance analysis procedure is performed to better understand which features are most/least relevant to the emissions of CO, NO x and SO 2 . In addition, a new software/application for the prediction of air pollutant emissions is deployed. This application could serve as a control tool for expert engineer groups to improve emissions monitoring and achieve more environmentally friendly operation of the power plant.