This survey explores the application of large language models (LLMs) in time series analysis. It provides a systematic overview of existing methods that leverage LLMs for time series tasks, including forecasting, classification, imputation, and anomaly detection. The paper discusses the challenges of training general-purpose time series models, such as the non-stationarity of time series data and the difficulty of adapting models to changing data distributions. It highlights the potential of pre-trained LLMs to capture complex dependencies in time series data and facilitates various applications. The survey categorizes existing methods into five groups: direct query, tokenization, prompt design, fine-tuning, and model integration. It also discusses the applications of LLMs in both general and spatial-temporal time series data, tailored to specific domains. The paper concludes with a discussion of future research opportunities to further advance time series analysis with LLMs, including improvements in tokenization and prompt design, interpretability, multi-modality, domain generalization, scaling laws, and the use of LLMs as agents. The survey emphasizes the importance of developing novel tokenization methods and prompt designs to enhance model performance, as well as the need for interpretable and safe LLMs in critical applications. The paper also highlights the potential of LLMs in various domains, including finance, healthcare, and computer vision, and discusses the challenges and opportunities in integrating LLMs into time series analysis.This survey explores the application of large language models (LLMs) in time series analysis. It provides a systematic overview of existing methods that leverage LLMs for time series tasks, including forecasting, classification, imputation, and anomaly detection. The paper discusses the challenges of training general-purpose time series models, such as the non-stationarity of time series data and the difficulty of adapting models to changing data distributions. It highlights the potential of pre-trained LLMs to capture complex dependencies in time series data and facilitates various applications. The survey categorizes existing methods into five groups: direct query, tokenization, prompt design, fine-tuning, and model integration. It also discusses the applications of LLMs in both general and spatial-temporal time series data, tailored to specific domains. The paper concludes with a discussion of future research opportunities to further advance time series analysis with LLMs, including improvements in tokenization and prompt design, interpretability, multi-modality, domain generalization, scaling laws, and the use of LLMs as agents. The survey emphasizes the importance of developing novel tokenization methods and prompt designs to enhance model performance, as well as the need for interpretable and safe LLMs in critical applications. The paper also highlights the potential of LLMs in various domains, including finance, healthcare, and computer vision, and discusses the challenges and opportunities in integrating LLMs into time series analysis.