TS-PTLM: Enhancing General Time Series Analysis with Pre-Trained Language Models

Pre-trained language models (PTLMs) have demonstrated adaptability to a wide range of natural language processing (NLP) tasks through methods such as prompting and fine-tuning. However, significant challenges persist in applying these models to time series analysis, largely due to the limited availability of time series data. As a result, the prevailing methodology often relies on the development of specialized models for specific tasks (e.g., forecasting, anomaly detection), which constrains the generalization capabilities of these models. Additionally, there are inherent differences between textual sequences and time series data. Directly inputting time series data into PTLMs frequently leads to suboptimal performance. To address these challenges, we propose a novel model, TS-PTLM, designed for time series analysis, capable of handling long-term forecasting, short-term forecasting, and anomaly detection. TS-PTLM employs a cross-modal knowledge transfer strategy to align textual sequences with time series data, leveraging both static and dynamic knowledge embedded in PTLMs to improve performance and generalization. Furthermore, the model incorporates frequency and temporal information from the time series data, with specialized adapters for fine-tuning, which further enhances its predictive accuracy. The effectiveness of TS-PTLM is demonstrated through extensive experiments conducted across three distinct time series tasks.