跳转至

TS-RAG：用于时间序列预测的检索增强生成

文章背景与核心概要

TS-RAG 是一项将检索增强生成（RAG）技术引入时间序列预测领域的创新深度学习框架。尽管 RAG 在大语言模型（LLM）中通过整合外部参考信息显著提升了模型表现，但将其应用于时间序列预测时，长期面临模型参数规模较小、训练数据受限以及缺乏原生生成能力等挑战。

针对简单拼接检索序列效果不佳的问题，TS-RAG 提出了一种专门设计的“参考标记”（reference tokens）机制。该机制能够将输入序列与检索到的相似历史模式进行无缝融合，从而更有效地捕捉复杂的时间动态。实验结果表明，TS-RAG 在多个主流基准测试中均达到了行业领先水平（SOTA）。

TS-RAG：用于时间序列预测的检索增强生成

执行摘要

TS-RAG 是一种新颖的深度学习框架，它将检索增强生成（RAG）引入了时间序列预测领域。虽然 RAG 通过整合外部参考资料成功增强了大语言模型（LLM），但将其应用于时间序列时，传统上一直受到模型参数较小、训练数据有限以及缺乏原生生成能力的阻碍。

TS-RAG is a novel deep learning framework that introduces Retrieval-Augmented Generation (RAG) to time series forecasting. While RAG has successfully enhanced large language models (LLMs) by integrating external references, applying it to time series has traditionally been hindered by smaller model parameters, limited training data, and a lack of native generative capabilities.

简单地拼接检索到的时间序列片段并不能达到预期的效果。为了克服这一问题，TS-RAG 利用专门设计的“参考标记”（reference tokens）将输入序列与检索到的相似历史模式无缝融合，从而更有效地捕捉复杂的时间动态，并在各大基准测试中取得了最先进的结果。

Simply concatenating retrieved time series sequences fails to achieve the desired outcomes. To overcome this, TS-RAG utilizes specially designed reference tokens to seamlessly fuse input sequences with similar retrieved historical patterns, capturing complex temporal dynamics more effectively and achieving state-of-the-art results across major benchmarks.

论文元数据

  • arXiv 标识符： arXiv:2608.06223 [cs.AI]
  • 主要学科： 人工智能 (cs.AI)
  • 次要学科： 机器学习 (cs.LG)
  • 提交日期： 2026年8月6日
  • 作者： Yixiong Xiao, Congxi Xiao, Jingbo Zhou
  • arXiv Identifier: arXiv:2608.06223 [cs.AI]
  • Primary Subject: Artificial Intelligence (cs.AI)
  • Secondary Subject: Machine Learning (cs.LG)
  • Submission Date: August 6, 2026
  • Authors: Yixiong Xiao, Congxi Xiao, Jingbo Zhou

摘要

尽管深度学习模型（特别是基于 Transformer 的架构）在时间序列预测中表现出色，但检索增强生成（RAG）在该领域的应用仍然有限。由于 RAG 已被证明可以通过结合相关的外部信息来增强大语言模型的能力，因此检索相似的时间序列序列作为参考，也可能提高时间序列预测任务的准确性。

While deep learning models, particularly transformer-based architectures, have shown impressive performance in time series forecasting, the application of retrieval-augmented generation (RAG) in this domain remains limited. Since RAG has proven effective in enhancing the capabilities of large language models by incorporating relevant external information, retrieving similar time series sequences as references might also improve accuracy in time series forecasting tasks.

然而，大多数时间序列模型受到训练数据有限、参数规模较小以及缺乏大语言模型所具备的广泛生成能力的限制。像语言模型那样简单地将参考序列拼接到提示词中，可能无法产生预期的结果。

However, most time series models are constrained by limited training data, smaller parameter scales, and a lack of the extensive generative capabilities found in large language models. Simply concatenating reference sequences into the prompt, as done in language models, may not yield the expected results.

为了解决这些挑战，我们提出了一种新颖的方法——TS-RAG，它利用 RAG 来增强预测性能。该框架引入了专门设计的参考标记，以有效地融合来自输入序列和检索到的相似序列的信息，从而更稳健地捕捉复杂的时间动态。实验结果表明，TS-RAG 在多个现实世界的预测基准测试中均实现了持续的最先进性能。

To address these challenges, we propose a novel approach, TS-RAG, which leverages RAG to enhance forecasting performance. The framework introduces specially designed reference tokens to effectively fuse information from the input sequence with that from retrieved similar sequences, enabling a more robust capture of complex temporal dynamics. Experimental results demonstrate that TS-RAG achieves consistent state-of-the-art performance across several real-world forecasting benchmarks.

获取资源