Abstract: To address peak–valley misalignment and increasing errors in multi-step prediction of nitrogen oxides (NO_X) from circulating fluidized bed (CFB) boilers, a prediction model based on TE time-lag reconstruction and Ridge–Mamba residual fusion is proposed. Based on the combustion and denitrification mechanisms of the boiler, intermediate variables, including instantaneously combustible carbon mass, dynamic O₂ concentration, and CO concentration, are constructed and combined with historical NO_X values and unit operating parameters to form the candidate inputs. The change in future NO_X concentration relative to the current state is used as the prediction target, and transfer entropy (TE) is employed to identify effective historical positions and determine the optimal time lag. Subsequently, Ridge regression estimates the baseline variation component of the difference target, while Mamba further learns the complex dynamic residuals not captured by Ridge. The two outputs are finally fused to predict the future NO_X concentration. Ablation results indicate that Ridge-based baseline variation estimation effectively reduces error accumulation in multi-step prediction, while TE-based time-lag reconstruction and the difference target further improve the model’s ability to track dynamic variation trends.