融合分解算法和注意力机制的云负载预测模型

Cloud Resource Prediction Model Integrating Decomposition Algorithm and Attention Mechanism

摘要:

为了提高云平台资源分配的均衡性，针对云资源负载数据的非线性、高噪声以及动态性特点，提出一种融合CEEMDAN分解算法及注意力机制的Transformer-BiLSTM负载预测模型。该模型使用CEEMDAN分解算法对负载序列数据进行分解，得到不同频率的分量，降低数据复杂度；通过Transformer编码层构成的编码器对各分量进行编码，获取数据的全局信息，并把得到的编码输出通过注意力模块进行权重的自适应分配；采用BiLSTM构成的解码器解码得到预测结果。实验结果表明，相较于主流模型，所提出的模型在不同预测步长的误差均有降低，验证了预测方法的有效性。

Abstract:

In order to improve the balance of cloud platform resource allocation, a Transformer-BiLSTM load forecasting model is proposed, which integrates CEEMDAN decomposition algorithm and attention mechanism, according to the nonlinear, high noise and dynamic characteristics of cloud resource load data. In this model, CEEMDAN decomposition algorithm is used to decompose the load sequence data to obtain components with different frequencies, thus reducing the data complexity. Each component is encoded by an encoder composed of the Transformer coding layer, and the global information of the data is obtained, and the obtained encoded output is adaptively distributed by the attention module. The prediction results are obtained by decoding with a decoder composed of BiLSTM. The experimental results show that, compared with the mainstream model, the error of the proposed model in different prediction steps is reduced, which verifies the effectiveness of the prediction method.