A novel <scp>VMD</scp> ‐ <scp>LassoNet</scp> ‐ <scp>iTransformer</scp> framework with enhanced feature fusion for dynamic <scp>NOx</scp> forecasting in flexible utility boilers

Abstract Accurate prediction of NOx emissions in flexible utility boilers is essential for both environmental protection and operational optimization. This study proposes a novel variational mode decomposition (VMD)‐LassoNet‐inverted Transformer (iTransformer) framework for dynamic NOx forecasting. First, an improved VMD with adaptive thresholding is used to reduce the noise in the original data. Next, a LassoNet network with L1‐regularized skip‐layer connections is utilized for feature selection in the first stage, reducing input dimensions from 33 to 28 critical parameters while maintaining interpretability. In the second stage, the iTransformer model performs deep feature extraction using a temporal attention mechanism, enabling accurate dynamic modelling of NOx emissions by capturing multivariate correlations and long‐term dependencies. Experiments on a 660 MW utility boiler dataset demonstrate that the proposed model achieves state‐of‐the‐art performance, achieving an R 2 value of 0.95, a MAE of 2.63 mg/m 3 , and a RMSE of 3.71 mg/m 3 in single‐step forecasting on the test data, surpassing LSTM, Transformer, and TimesNet. In multi‐step forecasting, the model reduces prediction errors by 7.46% and 3.4% compared to Transformer and TimesNet, respectively. The framework's hierarchical feature fusion and temporal inversion mechanisms effectively handle load‐varying scenarios, providing a solid foundation for real‐time NOx emission control.