CycleLLH:一种基于周期性整合的新型网络流量预测模型  

CycleLLH:A New Network Traffic Prediction Model based on Cycle Integration

作　　者：唐文杰 肖一磊 孔祥宇 齐恒[1] 刘秀龙 李克秋[2] TANG Wen-Jie;XIAO Yi-Lei;KONG Xiang-Yu;QI Heng;LIU Xiu-Long;LI Ke-Qiu(School of Computer Science and Technology,Dalian University of Technology,Dalian,Liaoning 116024;College of Intelligence and Computing,Tianjin University,Tianjin 300350)

机构地区：[1]大连理工大学计算机科学与技术学院,辽宁大连116024 [2]天津大学智能与计算学部,天津300350

出　　处：《计算机学报》2024年第12期2867-2888,共22页Chinese Journal of Computers

基　　金：国家自然科学基金面上项目(No.62072069)资助.

摘　　要：精准的网络流量预测是实现网络精细化和智能化管理的关键,也是网络运营商、云服务提供商等实现网络智能运维及应用服务保障的重要支撑,属于当前业界研究的热点.网络流量预测问题一般可被视为一种时间序列预测问题,现有时间序列预测模型虽然能起到一定作用,但这些通用模型很少考虑流量数据集本身特点,从而无法在网络流量预测性能上取得突破.为此,本文重点研究了网络流量数据集中的自然周期特征,提出了一种能有效利时间序列周期性特点的网络流量预测通用模型——Cycle Little Linear Head(CycleLLH).该模型主干为Transformer的编码器,其中两个关键设计在于:(1)周期整合.将流量序列按照一个特定周期划分步长划分为不同的周期块,然后将这些周期块对应相位的时间节点分别嵌入到不同输入令牌;(2)小线性层.由多个多层感知机组成,并且多层感知机单独作用于每个相位的时间特征.周期整合使得模型具有两个优点:更有利于模型提取数据集在一个周期内的特征;注意力矩阵的计算和内存复杂度可以看作是和周期划分步长二次方有关的常数,使得模型可以使用更大的回溯窗口而仅增加少量计算资源.通过在公共流量数据集上进行大量实验,本研究表明:与当前最先进的模型相比,CycleLLH在流量预测精度方面表现出显著优势,在六个数据集上的预测精度分别提升了12.3%、8.4%、29.9%、5.8%、8.3%和2.0%.代码可从https://github.com/wenjietang218/CycleLLH.git中获取.Accurate network traffic prediction is not only a key link to realize network refinement and intelligent management,but also an important support for network operators and cloud service providers in intelligent operation and maintenance and application service guarantee.This field has become a research hotspot in the current industry,and has a wide range of application prospects.Network traffic prediction technology plays an important role in many fields,including base station power management,base station overload prevention,unmanned aerial vehicle temporary base stations,new network deployment and base station construction,5G network slicing,Software Defined Network resource scheduling,and resource management in multi-access edge computing.Network traffic prediction problem is usually regarded as a time series prediction problem.With its self-attention mechanism,Transformer can efficiently compute the dependencies between each time node in parallel,and performs well in the field of time series prediction.However,although the existing time series prediction models,including the Transformer-based models,can play a certain role,they often ignore the unique periodic characteristics of traffic data sets,so it is difficult to make a breakthrough in network traffic prediction performance.Therefore,this paper focuses on the natural cycle of network traffic data sets,and proposes a general network traffic prediction model,Cycle Little Linear Head(CycleLLH),which can effectively exploit the periodicity of time series.The backbone of the model is the encoder of the Transformer,and the two key designs are as follows:1.Cycle Integration:we divide the traffic sequence into different cycle blocks according to a specific period,and then embed the time nodes corresponding to the phase of these cycle blocks into different input tokens;2.Little Linear Head:it is composed of multiple multi-layer perceptrons,and each multi-layer perceptron operates separately on each feature node.Cycle Integration makes the model have two advantage

关 键 词：网络流量预测 时间序列预测 周期性 周期整合 小线性层 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]