PRADA: Prompt-guided Representation Alignment and Dynamic Adaption for time series forecasting

Time series forecasting endeavors to construct models capable of predicting future values and trends grounded in historical observations. However, current LLM-based approaches migrate the inference power of LLM to the time series forecasting through prompt guidance, but ignore the modality gap between time series and natural language. This gap arises from the fact that time series have periodic and non-periodic patterns that are not present in natural language, hindering the capabilities of LLM-based models. In addition, the potential statistical property drift in time series makes the model rely on spurious correlation features, limiting the capture of spatio-temporal dependencies. To tackle the unique problems, we introduce the Prompt-guided Representation Alignment and Dynamic Adaption (PRADA) method, which harnesses multi-view Text-Series Adaptive Alignment (TSAA) guided by learnable prompts to capture the representations of different patterns. Specifically, we first decompose the input time series into different components and align orthogonal prompts consisting of learnable context vectors with time series embeddings independently for LLM’s input adaption. Furthermore, the time-frequency dual constraint is introduced to empower the model to capture the overlooked label autocorrelation from both the time and frequency domains. Through multi-view adaptive alignment guided by learnable prompts, PRADA is able to dynamically model spatio-temporal dependencies and adapt to the semantic gap between time series and natural language, which enables LLM-based models to obtain more robust times series representations in real scenarios. Experiments on multiple public datasets demonstrate the state-of-the-art (SOTA) performance of PRADA in time series forecasting. The code will be available at https://github.com/HowardLiu28/PRADA.