A dynamic prediction model of sulfur dioxide concentration is proposed to address the challenge of accurate measurement of sulfur dioxide emission concentration at the exit of a limestone-gypsum wet flue gas desulfurization system under deep peaking.The model integrates a convolutional neural network (CNN), a bidirectional gated recurrent unit (BiGRU), and an attention mechanism to predict the sulfur dioxide concentration. The model developes utilizes kernel principal component analysis to determine seven characteristic variables, which are then used as inputs to the model.The attentional mechanism is combined with CNN and BiGRU to construct a model for predicting SO2concentration at the outlet of the desulfurization system. A simulation experiment is conducted with the FDG system of an operating 600 MW supercritical unit as the research object. The simulation results demonstrate that the average absolute error MSE of the dynamic model established in this paper is 0.706 4 mg/m3, the root mean square error RMSE is 0.912 5 mg/m3, and the average relative error is 6.27%, which is 25.07%, 23.45%, and 17.28% lower compared with CNN-BiGRU, and even lower than CNN and BiGRU; The coefficient of determination of the dynamic model was 96.74%, which was 3.91%, 5.26%, and 9.66% higher than CNN-BiGRU, BiGRU, and CNN models respectively. This outcome indicates that the dynamic model based on CNN-BiGRU-Attention exhibits high prediction accuracy and learning ability, and can accurately predict the trend of SO2 concentration at the outlet of the desulfurization system.