Winter cities, characterized by severe weather conditions, contribute significantly to global carbon emissions. The carbon emissions reduction of winter cities directly relates to the possibility of achieving the Paris Agreement’s temperature control target (1.5 °C). Accurate daily predictions of city-level carbon emissions can help us better understand the carbon emissions mechanism of cities, as well as the impact of urban operation on carbon emissions, and thus better guide low-carbon operation in cities. Through analysis of long-term series carbon emissions data, this study reveals that winter cities globally exhibit typical weekly cyclical fluctuations. By leveraging the analytical capabilities of large language models (LLMs), we propose a new prediction workflow that uses an STL-decomposition-based in-context learning method to guide LLMs in predicting winter city carbon emissions. The method utilizes cities’ historical daily carbon emissions data to provide accurate future predictions in a natural language conversational format without additional training. The results suggest that compared with the Multi-layer Perceptron (MLP) and Naive Bayes models (NBM), the LLMs combined with in-context learning improves the prediction accuracy of daily winter city carbon emissions.