A Prompt-Enhanced Pre-Training Transformer Model for Time Series Forecasting

Time series forecasting is vital in fields such as meteorology and urban computing. Inspired by the success of Generative Pre-trained Transforms (GPT) models in natural language processing and their applications in computer vision, recent research has explored extending GPT-type models to time series analysis. However, because pre-training models are trained on text data-significantly different from time series data-they cannot be directly used for forecasting. This paper proposes a new methodology, leveraging GPT-type models by first applying STL decomposition to characterize time series data. Then, specialized prompts for trend, seasonal, and residual components are designed to enhance GPT-type models' processing ability for time series tasks. Additionally, an innovative Instruction Learning Module (ILM) based on the Transformer structure is introduced to better capture and integrate domain-specific information, improving the GPT-type model's adaptability. The model is evaluated on a range of datasets, with results showing significant improvements in accuracy and robustness, indicating the potential of GPT-type models in advancing time series forecasting.