TFGformer: Multivariate Time Series Forecasting via Time-Frequency Graph Learning and Covariate Fusion

2026-07-31

AI Summary

This work addresses the challenge of inaccurate long-term forecasting in heterogeneous IoT sensor multivariate time series, where static models fail to capture dynamic historical patterns. To overcome this limitation, we propose CrossRAG, a novel framework that effectively integrates retrieval-augmented generation into multivariate time series prediction. CrossRAG introduces shape-level retrieval robust to amplitude variations through Shape-Aware Memory and RevIN normalization, selects future-consistent historical references via Future-Consistent Contrastive learning, and fuses historical–future reference pairs at the representation level using Cross-Attention Temporal Fusion. Extensive experiments demonstrate that CrossRAG significantly outperforms existing parametric and retrieval-augmented methods across seven public benchmarks, achieving more accurate long-term forecasts.

Abstract

Large-scale multivariate time series from heterogeneous IoT sensors demand accurate long-term forecasting for resource scheduling and predictive maintenance. While recent time series foundation models exhibit strong generalization, they rely on static parametric knowledge and lack dynamic access to external historical patterns during inference. Retrieval-Augmented Generation (RAG) offers a potential remedy, yet its application to time series forecasting is challenged by magnitude variations across heterogeneous sources and the mismatch between historical similarity and future consistency. We propose CrossRAG, a retrieval-augmented forecasting framework that integrates Shape-Aware Memory (SAM) with RevIN normalization for magnitude-robust shape-level retrieval, Future-Consistent Contrastive (FCC) learning to distinguish informative references from hard negatives with similar history but divergent futures, and Cross-Attention Temporal Fusion (CATF) to fuse retrieved historical--future reference pairs into the backbone's representations at the representation level. Experiments on seven public benchmarks show that CrossRAG consistently outperforms both parametric-only baselines and existing retrieval-augmented forecasting methods.

Problem

• multivariate time series forecasting
• retrieval-augmented generation
• heterogeneous IoT sensors
• future consistency
• historical similarity

Innovation

• Retrieval-Augmented Forecasting
• Shape-Aware Memory
• Future-Consistent Contrastive Learning
• Cross-Attention Temporal Fusion
• Multivariate Time Series