Abstract: To address the challenge of predicting SO₂ concentration under quick load change conditions for circulating fluidized bed (CFB) units, from the principles of combustion, the mechanisms of SO₂ generation and reduction within the furnace were analyzed, and models were constructed for the combustion of char, dynamic oxygen balance, and reactive limestone. Based on these models, an SO₂ concentration prediction model was established. The model took coal feed rate, bed temperature, total air flow, and limestone feed rate as input variables. By integrating multidimensional correlation analysis, key intermediate parameters such as char combustion rate, dynamic O₂ concentration, and reactive limestone were introduced as extended inputs to reconstruct the dataset. The reconstructed dataset was trained using the Kolmogorov-Arnold network (KAN) to build a hybrid input Kolmogorov-Arnold network (HIKAN). The model hyperparameters were optimized using the meta-learning optimizer (MLO), resulting in the proposed MLO-HIKAN prediction model. Case study results demonstrate that this model achieves high accuracy in predicting SO₂ emission concentration under quick load change conditions, providing a new method for the environmentally friendly operation of CFB units.