Limestone-gypsum wet flue gas desulfurization (WFGD) is the main method of flue gas desulfurization in thermal power plants and plays an important role in atmospheric environmental protection, but it can also suffer from corrosion and fouling problems that affect the operational efficiency. In order to optimize the operation of the wet WFGD system, a data-driven approach is used to model the SO2 flue gas emissions dynamically. Firstly, the SO2 emission data are decomposed using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to obtain several intrinsic mode functions (IMFs). The intrinsic mode functions containing noise are denoised using wavelet threshold denoising to obtain the pure components. Then a deep learning model combining convolutional neural network (CNN) and bi-directional gated recurrent unit (BiGRU) is designed for SO2 emission prediction. After comparing the two schemes of predicting the components separately and then reconstructing them and reconstructing the components and then predicting them, it is found that the root mean square error and the mean absolute error of the former are reduced by 0.135 7 and 0.284 3, respectively, compared with the latter. Experiments are conducted based on the first scheme in comparison with other benchmark models, the root mean square error and the mean absolute error of the proposed model are 0.699 6 and 0.355 3, which are the lowest. The results indicate that the proposed model has significant advantages in predicting SO2 emission concentration.