A novel NOx emission prediction model for multimodal operational utility boilers considering local features and prior knowledge

Abstract

This study proposed a data-driven NOx modelling framework that could capture the multimodal operational characteristics of a utility boiler, improve the training sample quality and strengthen the model interpretability. With this framework, to obtain a high-quality steady-state dataset from operational data, a multivariate F-test algorithm, which enhanced robustness by introducing the minimal number of stable units (MNSU) and the variational mode decomposition (VMD) outlier detection method, was first proposed. Based on the obtained sample, a Gaussian mixture model (GMM) was established to classify the data by their operational modes. In addition, the least absolute shrinkage and selection operator (LASSO) algorithm was employed to select specific features for each mode. Finally, the NOx prediction model was constructed using the light gradient boosting machine (LightGBM) model integrated with prior knowledge. The mechanistic relationships between the selected features (i.e., the oxygen content and separated overfire air damper opening) and target variable (i.e., NOx) were considered as the constraint conditions. By taking a 660 MW utility boiler as the research object, the proposed modelling framework obtained R2, RMSE and MAE values of 0.979, 3.573 mg/m3, and 2.586 mg/m3, respectively, performing better than five comparison models.