This is the authors' abstract. We don't add key points for this paper.
Accurate soft sensing of NOx emission concentration in coal-fired boilers is important for clean combustion control. However, real CFB operating data contain heterogeneous noise, multiscale dynamic responses, and stage-dependent dominant variables, which hinder extraction of NOx-formation dynamics and degrade soft-sensing accuracy. To address these issues, a NOx emission soft sensing method is proposed. First, training samples consistent with practical disturbances are constructed from differences in the statistical distributions of process variables. Representation consistency between noisy and original data is then constrained in the latent space to reduce abnormal-disturbance effects on basic features. After obtaining basic representations, a variable-scale prior weight matrix is designed using data characteristics and CFB combustion mechanisms. This matrix is combined with data-driven scale weights to fuse multiscale features and extract dynamic information better matched to the actual combustion process. To capture dynamic changes in variable effects, learnable proxy vectors aggregate NOx-related key segments with reduced sequence interaction cost. Variable gating then generates variable weights from proxy temporal representations, dynamically identifying dominant process variables across operating stages. Experiments on practical operating data from a 350 MW CFB unit show that the proposed method achieves an R2 of 0.953 on the test set. Under compound disturbances comprising random, abnormal, and drift noise, the R2 remains 0.896, indicating strong noise robustness. SHAP analysis and internal-weight visualization show that the identified key variables and their time-varying effects are generally consistent with CFB combustion mechanisms. Overall, the proposed method improves the accuracy, robustness, and interpretability of NOx emission soft sensing, and provides a practical auxiliary tool for NOx emission monitoring in CFB boilers.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Tian et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: