Driven by the “Dual Carbon” goals, profound changes are anticipated in urban socio-economic development, energy management, and consumption. Therefore, establishing effective models for predicting urban carbon emissions is advantageous for understanding emission variations and trends. Taking into account differences in urban economic development, industrial structure, and geographical location among different provinces, this study fully utilizes Transformer neural networks to explore carbon emission data characteristics across cities. Compared to most studies, our proposed model does not require pre-selection of carbon emission influencing factors, demonstrating strong flexibility and adaptability. Conducting numerical experiments based on data from 34 cities from 2000 to 2019, our results show that the proposed method achieves satisfactory prediction accuracy. The findings not only validate the capability of the Transformer model in handling urban carbon emission data from diverse cities but also provide a scientific basis for developing more specific and feasible emission reduction strategies that consider the heterogeneity across provinces and regions.