Soft-sensor of Carbon Content in Fly Ash based on LightGBM

Abstract

The soft-sensor method of carbon content in fly ash is to predict and calculate the carbon content of boiler fly ash by modeling the distributed control system (DCS) data of thermal power stations. A novel data-driven soft-sensor model that combines data pre-processing, feature engineering and hyperparameter optimization for application in the carbon content of fly ash is presented. First, extract steady-state data by data mining technology. Second, twenty characteristics that may affect the carbon content in fly ash are identified as variables by feature engineering. Third, a LightGBM prediction model that captures the relation between the carbon content in fly ash and various DCS parameters is established and improves the prediction accuracy by the Bayesian optimization (BO) algorithm. Finally, to verify the prediction accuracy of the proposed model, a case study is carried out using the data of a coal-fired boiler in China. Results show that the proposed method yielded the best prediction accuracy and closely approximates the non-linear relationships between variables.