Prediction of the Unburned Carbon Content of Fly Ash in Coal-Fired Power Plants

Prediction of carbon content in fly ash is a valuable index in both designing and optimizing coal-fired boilers. In this study, methods of multiple regression (MR), decision trees (DT), and support vector regression (SVR) were used for prediction of unburned carbon content of fly ash in a 300-MW power plant with a tangentially fired furnace using a low-NOx concentric firing system. Twenty-one boiler operational and coal petrological variables from 21 experiments in a previous benchmark study were used. A variables correlation matrix was used for feature (independent variables) selection in order to reduce the dimensions (number) of input variables from 21 (e.g., O2 volume, damper position, coal properties, burner tilt) to five (oxygen concentration in the flue gas, over-fire air, as-received volatile content, as-received ash content, and as-received net heat value) and to three (oxygen concentration in the flue gas, as-received volatile content, and as-received net heat value) for MR and DT, respectively. Both the MR and DT models were able to predict the carbon content with a reasonable accuracy. Furthermore, an SVR model was built considering all 21 input variables, which resulted in almost exact predictions of the unburned carbon content of fly ash. The advantage of MR and DT in this study was using fewer input variables (oxygen concentration in the flue gas, over-fire air, as-received volatile content, as-received ash content, and as-received net heat value). A lesser number of explanatory variables resulted in more computationally time-efficient models. The advantage of the SVR was its very high accuracy in the predictions of the unburned carbon content of fly ash.