Abstract: Under the "dual carbon" policy, the accurate prediction of carbon emissions of thermal power plants is conducive to the formulation of reasonable carbon emission reduction schemes. A short-term carbon emission prediction model for thermal power plants based on CEEMDAN-ISSA-BILSTM algorithm is presented. First, the carbon emission sequence is decomposed into sub-sequences of different frequencies using full set empirical Mode decomposition (CEEMDAN). Secondly, the high-frequency components are denoised by wavelet threshold denoising technology, and the denoised sub-sequences are reconstructed. Finally, in order to obtain the optimal hyperparameters of the prediction model, the multi-strategy modified Sparrow search algorithm (ISSA) is used to optimize the hyperparameters of bidirectional long short-term memory network (BiLSTM). Based on the historical data of a 200MW unit of a thermal power plant, three evaluation indexes are compared with the single prediction model and the combined model. The results show that the proposed prediction model can accurately predict 100days of carbon emissions and change trend. Combining all the evaluation indexes, the combined prediction model is better than other comparison.