Multi-Scale Transformer for CPU Utilization Time Series Forecasting in Cloud Computing

Abstract

This paper addresses the complexity and challenges of CPU utilization prediction in cloud computing environments and proposes a time series modeling method based on a multi-scale Transformer. The method first preprocesses the input utilization sequence and performs window segmentation. It then constructs embeddings at multiple scales to capture both short-term fluctuations and long-term trends. Next, the model applies a multi-head self-attention mechanism to effectively capture global dependencies in the sequence. Cross-scale interaction and weighted fusion strategies are further introduced to enhance the comprehensive representation of multi-level dynamic features. In the prediction stage, the model uses the fused feature representations to estimate future CPU utilization. Training is performed with mean squared error as the optimization objective to ensure stability and accuracy. The experimental study includes multiple comparative analyses and sensitivity tests, covering mainstream methods such as LSTM, BILSTM, 1DCNN, and Transformer. The results show that the proposed model achieves better performance in terms of mean squared error, root mean squared error, mean absolute error, and the coefficient of determination. Further sensitivity experiments on hyperparameters and the multi-scale mechanism verify the effectiveness and adaptability of the method in complex cloud computing environments. The overall results demonstrate the advantages and value of the multi-scale Transformer in CPU utilization prediction tasks.