Abstract To address the issues of untimely and unstable control during rapid load fluctuations in the selective non‐catalytic reduction (SNCR) denitrification systems at thermal power plants, which lead to potential exceedances of pollutant emission limits, this study proposes a comprehensive predictive model to forecast NO x emission concentrations at the system outlet. Firstly, raw data collected from industrial sites undergoes filtering and noise reduction. The multi‐dimensional time series is then dynamically calibrated and reduced in dimension utilizing the maximum information coefficient (MIC). Subsequently, combining the advantages of feature extraction and fusion from cross‐attention, differential convolutions, and long short‐term memory (LSTM) gating, this study proposes a deep integration architecture and introduces a multi‐angle improved crested porcupine optimizer (ICPO) algorithm to optimize network hyperparameters adaptively. This approach enhances the model's expressive capability while minimizing computational complexity. Experimental results based on operational data from a 350 MW supercritical circulating fluidized bed (CFB) boiler in a thermal power plant demonstrate that the proposed model outperforms comparison models under various operating conditions, accurately predicting NO x emission concentrations at the SNCR outlet.
Wang et al. (Mon,) studied this question.