基于MVMD-MHAT-BiLSTM的云资源负载预测方法
Cloud Resource Load Prediction Method Based on MVMD-MHAT-BiLSTM
史爱武 1罗干 1李林逸 1黄河1
作者信息
- 1. 武汉纺织大学 计算机与人工智能学院,湖北 武汉 430200
- 折叠
摘要
目前,云计算服务提供商在预测大规模工作负载和资源使用方面面临着巨大挑战,由于难以捕捉非线性特征,传统的预测方法通常无法对资源负载数据实现较高的预测性能.此外,原始的时间序列中存在大量噪声,如果不采用平滑算法对这些时间序列进行去噪,预测结果可能无法满足提供者的要求.为此,提出一种MVMD-MHAT-BiL-STM组合预测模型,该模型首先使用改进的灰狼优化算法优化VMD参数,之后通过变分模态分解的信号分解方法,将复杂、非线性的原始时间序列分解为低频本征模态函数;接着在BiLSTM中引入多头注意力机制捕获多层次、双向特征;最后使用注意力机制探索不同输出维度的重要性.在阿里的公开数据集中验证该模型性能,与BiLSTM、Pa-BiLSTM、CNN-BiLSTM、MHAT-BiLSTM和VMD-MHAT-BiLSTM模型相比,该模型的均方根误差下降了8.6%～19.3%,实现了更高的预测精度.
Abstract
Cloud computing service providers currently face huge challenges in predicting large-scale workloads and resource usage.Due to the difficulty in capturing nonlinear characteristics,traditional prediction methods usually cannot achieve high prediction performance for re-source load data.In addition,there is a lot of noise in the original time series.If smoothing algorithms are not used to denoise these time se-ries,the forecast results may not meet the provider's requirements.To this end,this paper proposes a MVMD-MHAT-BiLSTM combined pre-diction model.This model first uses the improved gray wolf optimization algorithm to optimize the VMD parameters,and then uses the varia-tional mode decomposition signal decomposition method to decompose the complex,nonlinear original The time series is decomposed into low-frequency intrinsic mode functions;then a multi-head attention mechanism is introduced in BiLSTM to capture multi-level,bidirectional fea-tures;and finally the attention mechanism is used to explore the importance of different output dimensions.Taking the CPU usage of machines in Alibaba Cloud Cluster Data as an example,compared with theBiLSTM,Pa-BiLSTM,CNN-BiLSTM,MHAT-BiLSTM and VMD-MHAT-BiLSTM,the RMSE of of the model proposed in this article decreased by 8.6%to 19.3%,achieving higher prediction accuracy.
关键词
云资源负载预测/灰狼优化算法/变分模态分解/多头注意力机制/双向长短记忆网络/注意力机制Key words
cloud resource load prediction/GWO/VMD/multi-head attention mechanism/BiLSTM/attention mechanism引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据