Carbon dioxide (CO2) is a major contributor to global warming. Accurate and high-resolution prediction of CO2 emissions is critical to achieving carbon neutrality around the world. Previous methods used traditional statistical models or machine learning models, which can only predict the annual CO2 emissions. With the development of deep learning models, such as GRU, they have been used for prediction. However, there is a lack of predictions for fine time granularity for CO2 emissions. Thus, in this paper, we propose a carbon emission prediction model at fine time granularity. Autoformer is used to improve the prediction accuracy. A carbon emission dataset that contains 3 years of China’s over 300 cities on a daily basis is used for prediction. The results show that compared with other methods, our model can achieve the highest prediction accuracy. Our model provides high-quality, fine-grained CO2 emissions data to support global emission monitoring across various urban content.