Forecast of NOx Emissions for a 660MW Coal-Fired Boiler with Multilayered Gradient Boosting Decision Tree Considering Multiple Operating Modes
Ziwei Wang, Yongzan Zhou, Yukun Zhu, Haiquan Yu, Wei Fan
Abstract
High Resolution Image Download MS PowerPoint Slide Nitrogen oxides (NOx) are among the primary pollutants emitted by coal-fired power plants. Accurate prediction of NOx concentrations at the boiler outlet is crucial for optimizing unit control and reducing emissions. This study introduces a data-driven NOx emissions prediction methodology based on a multilayered Gradient Boosting Decision Tree (mGBDT) framework. Initially, Kernel Independent Component Analysis (KICA) is employed to eliminate nonlinear correlation among collected auxiliary variables. Subsequently, high-quality variables, grounded in physical mechanisms, are integrated with the extracted independent features. The robust Gaussian Mixture Model (RGMM) is then applied to capture the intrinsic multimode operational characteristics from these integrated features. Finally, local mGBDT-based NOx emissions prediction models are developed for each identified mode. The optimal hyperparameters for each local model are determined using Particle Swarm Optimization (PSO) and 10-fold cross-validation techniques. Utilizing historical measurements from the studied boiler, the proposed framework achieved a square of correlation coefficient ( R 2 ) of 0.947, a root-mean-square error (RMSE) of 6.09 mg/m 3, and a mean absolute error (MAE) of 4.009 mg/m 3, outperforming five comparison models.