检索增强的网络流量预测方法

摘要：网络流量预测是保障网络服务质量的关键技术，现有时序模型难以融合文本描述的变更事件信息。本文提出一种融合时序大模型与语言大模型的协同预测框架，实现变更事件驱动的网络流量动态预测。针对变更事件稀疏性及专业语义理解难题，设计基于检索增强生成的变更影响知识库，通过检索历史相似变更的流量影响特征，构建可解释的上下文提示。模型采用双阶段架构：首先使用时序大模型生成基础流量预测，继而由语言大模型结合检索的变更案例及当前变更描述，对预测结果进行语义推理修正。实验表明，在真实网络运维数据集上，模型在变更事件场景下的预测误差相较于仅通过时序预测的方法有明显下降。

关键词：流量预测；检索增强生成；时序预测

Abstract: Network traffic prediction is a critical technology for ensuring network service quality, yet existing time series models struggle to incorporate textual descriptions of change events. This paper proposes a collaborative prediction framework that integrates large-scale time series models and large language models to achieve change-driven dynamic network traffic forecasting. To address the sparsity of change events and challenges in professional semantic understanding, we design a Retrieval-Augmented Generation (RAG)-based change impact knowledge base. This retrieves traffic impact characteristics from historically similar changes to construct interpretable contextual prompts. The model adopts a two-stage architecture: First, a large time series model generates baseline traffic predictions; subsequently, a large language model performs semantic reasoning-based refinement of these predictions by incorporating both the retrieved change cases and the current change description. Experiments on real-world network operation datasets demonstrate that our framework significantly reduces prediction errors in change event scenarios compared to time series-only approaches.

Keywords: network traffic prediction; Retrieval-Augmented Generation; time series prediction