Abstract: A variational modal decomposition (VMD)-Transformer prediction model for SO₂ emission concentration in circulating fluidized bed (CFB) boiler based on whale optimization algorithm (WOA) was constructed. First, the mean impact value (MIV) method was adopted to screen input variables, outliers were eliminated through inter quartile range (IQR) method, and normalization was subsequently performed. Then, WOA was used to synchronously optimize the time step, as well as the hyperparameters of the VMD and Transformer models. Finally, the optimized VMD model was applied to extract the time series signals of SO₂ concentration, and the multi-head self-attention mechanism of the optimized Transformer model was combined to build the time-series prediction model for SO₂ emission concentration. Taking a 300 MW CFB unit as the research object, four segments of variable-load operating conditions were selected to carry out the prediction of SO₂ emission concentration. Ablation experiments further verify that the synergy between VMD for signal decomposition and Transformer for temporal modeling can improve the prediction accuracy. Results show that under variable load conditions, the minimum mean absolute error (MAE) between the predicted value and the actual value of the model is 1.336 9 mg/m³, and the determination coefficient R² is 0.960 9. The established model can provide a support for the online optimization of pollutant control in CFB units and promote the intelligent development of CFB units.