The accurate prediction of SO_2 and NO_x emission mass concentrations can effectively guide the control of pollutants emissions

which is of great significance for the environmental protection operation of circulating fluidized bed(CFB) units. A 330 MW CFB unit is taken as the research object

and the Pearson coefficient is used to realize the screening of input variables

and the interquartile range(IQR) method is applied to screen the outliers and replace them with the normalization at the same time

to complete the data preprocessing. Subsequently

the features of input variables are extracted by convolutional neural network(CNN)

and by entering into the gate-recurrent unit(GRU) the time-series features are processed. The multi-head self-attention(MHA) mechanism is introduced to capture the important relationships between features

and the model output is obtained after training. Finally

the results of the test set are evaluated using the mean absolute error(MAE)

mean absolute percentage error(MAPE)

and the coefficient of determination(R~2). The results show that the model is able to predict the pollutants mass concentration in CFBs more accurately and achieve good prediction results

and the superior performance of the model is proved by the comparison of ablation experiments with the model. The proposed CNN-GRU-MHA model can realize the monitoring and optimization guidance of pollutants emissions CFB units

so that the power plant can adjust the operation parameters in time to ensure that the pollutants emissions meet the standards.