Objective Due to the irregularity and nonlinearity of carbon emissions this study proposes a carbon emission prediction method based on multi-head attention CNN-LSTM to enhance the accuracy of carbon emission predictions. Methods Leveraging the spatial feature extraction capabilities of CNN networks and the effective utilization of long-term and short-term dependencies by the LSTM algorithm a multi-head attention mechanism was introduced to allocate different weights. A CNN-LSTM model guided by the multi-head attention mechanism was employed to predict carbon emissions in China. Results The research selected a dataset of China?? s carbon emissions over the past 1 724 days and used the model proposed in the paper to conduct training and prediction on this dataset. After testing the values of the RMSE MASE and MAPE of the model on the test set were 0. 080 2 0. 302 0 and 0. 105 7 respectively. Comparing with the BP neural network algorithm ARIMA 2 1 4 and GM 1 1 the model designed in the paper outperformed the other prediction algorithms in terms of these three evaluation indicators. Conclusion The carbon emission prediction method based on multi-head attention CNN-LSTM can effectively improve the prediction accuracy and is suitable for time series prediction.