Dynamic NOx emission prediction in coal‐fired power plants based on joint multi‐head attention CNN‐GRU hybrid model

The flexible operation of coal‐fired power plants under deep peak‐shaving conditions imposes significant challenges on accurate NOx prediction for SCR systems. Given the inherent complexities of the SCR denitrification process, characterized by dynamic nonlinearity, temporal variability, and multivariable coupling in nitrogen oxide emissions, this study proposes a convolutional neural network‐gated recurrent unit (CNN‐GRU) hybrid model integrated with multi‐head attention (MA) mechanisms to address these system‐specific characteristics for precise NOx prediction. The model combines the local feature extraction capability of CNNs, the long‐term temporal dependency modelling strength of GRUs, and the adaptive feature weighting functionality of MA mechanisms, achieving dynamic weight allocation across feature channels and temporal scales to enhance robustness and feature representation. Furthermore, a sparrow search algorithm (SSA) is introduced to optimize model parameters adaptively, improving prediction accuracy and generalization performance. Experimental validation using real operational data demonstrates the model's superior performance, with mean absolute error (MAE) below 0.5 mg/m 3 and mean absolute percentage error (MAPE) below 2%. Ablation experiments confirm the effectiveness of the proposed architecture, showing over 28% prediction accuracy improvement compared to Transformer‐based models while maintaining enhanced generalization capability.