Gated Residual Fusion Framework for Mitigating Error Accumulation in Multistep NOx Prediction for Utility CFB Boilers

Abstract

Accurate multistep prediction of nitrogen oxide (NOx) emissions is essential for combustion perception and optimization in utility boilers. Under deep peak-shaving conditions, frequent operating condition transitions induce mixed operating data, strong variable coupling, and progressive error accumulation, restricting the long-horizon stability of conventional models. To address these issues, a gated residual fusion framework was constructed by using data from a 300 MW circulating fluidized bed (CFB) boiler. A preprocessing strategy integrating rolling signal denoising, correlation analysis, and Random Forest ranking selected 30 variables from 162 candidates. iTransformer and TimesNet were coupled to extract complementary dynamic representations, and sequence decomposition fusion, parallel fusion, and gated residual fusion were compared. In the proposed architecture, the dominant NOx evolution is captured by the iTransformer backbone, while the TimeNet residual branch provides correction for local deviations that are insufficiently represented by the backbone. Compared with standalone VMD-iTransformer and VMD-TimesNet, RMSE was reduced by 73.99% and 78.37% at step 1 and by 39.86% and 49.51% at step 6, respectively. Compared with eight baseline models, the proposed model achieved the best average performance, with RMSE, MAE, MAPE, and R2 values of 3.2228 mg/m3, 1.8930 mg/m3, 5.2338%, and 0.9185. Results demonstrate peak recovery, abrupt change tracking, and long-horizon stability.