Abstract: The circulating fluidized bed (CFB) semi-dry flue gas desulfurization (FGD) system is widely used for sintering flue gas treatment in the iron and steel industry. However, due to its complex structure and the coexistence of gas-solid two-phase flow and multi-step reactions, traditional mechanistic models are inadequate for accurately predicting dynamic variation patterns in the outlet SO₂ concentration. To guide the stable operation of the desulfurization system, a hybrid prediction model named DCNN-BiLSTM-Attention was constructed, integrating Deep Convolutional Neural Networks (DCNN), Bidirectional Long Short-Term Memory (BiLSTM) networks, and an Attention mechanism. Sliding-window IQR and Savitzky-Golay filtering were applied for data processing, and seven key variables were selected as inputs using the mutual information method. The model was validated based on actual operating data from a steel plant. The results demonstrated that the model achieved high accuracy on the test set, with mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) of the proposed DCNN-BiLSTM-Attention model being 0.4047 mg/m³, 0.5257 mg/m³, and 6.62%, respectively, which were 33.84%, 32.89%, and 36.69% lower than those of the CNN-BiLSTM-Attention model. The coefficient of determination (R²) reached 0.9799, representing a 2.56% improvement. Furthermore, the model maintained high prediction accuracy and stability even under significant SO₂ concentration fluctuations. This study provides a reliable technical reference for guiding the stable operation and optimized control of CFB-FGD systems.