TF-LLM: Enhanced time series analysis with time-frequency large language models

In recent years, pre-trained large language models (LLMs) are gradually introduced into time series analysis, and researchers have observed their significant potential for multi-task reasoning, particularly in handling complex symbolic sequences. However, how to effectively and deeply leverage the contextual reasoning capabilities of LLMs in time series data remains a key challenge. To address this challenge, this study proposes the TF-LLM framework to handle various tasks in time series, such as forecasting, classification, imputation, and anomaly detection. We innovatively integrate the strengths of both time and frequency domains: frequency representations simplify data complexity and enhance the capture of global and local periodic patterns, while time modeling addresses fine-grained dependencies, mitigating the effects of non-stationarity. Additionally, to enhance the model’s reasoning capabilities, we introduce prompt learning to enrich the contextual information of inputs and help the LLMs better understand time series data. We conduct extensive multi-task experiments on seven benchmark datasets, covering tasks like forecasting, classification, imputation, and anomaly detection. The results indicate the superior performance of the proposed TF-LLM in handling complex time series tasks, outperforming several existing methods.