Abstract: To enhance the prediction accuracy of SO₂ emissions during sludge incineration process and optimize incineration and flue gas treatment conditions, an efficient and stable hybrid prediction model for SO₂ emissions was prorosed. First, using a bubbling fluidized bed sludge incineration system as the research subject, static and dynamic flame features were extracted from flame images and the input features were set combined with distributed control system (DCS) parameters, while SO₂ emission concentration was set as the model output. Subsequently, mutual information (MI) was employed to determine the optimal lag time between SO₂ and each input feature, guiding data reorganization. Finally, the extremely randomized trees (ERT) model based on Bayesian optimization-TPE (BO-TPE) was constructed and compared with multiple mainstream prediction models. Results show that the BO-TPE-optimized ERT model achieves correlation coefficient R² of 0.93 with mean absolute percentage error(MAPE) below 3%, making it suitable for online prediction and process optimization control of SO₂ emissions in sludge incineration systems.