Load Prediction in Double-Channel Residual Self-Attention Temporal Convolutional Network with Weight Adaptive...

Area of Science:

• Cloud Computing
• Machine Learning
• Resource Management

Background:

• Rapid fluctuations in cloud resource demand necessitate timely container cluster scaling for service quality.
• Accurate resource load prediction remains a significant challenge in widespread cloud adoption.

Purpose of the Study:

• To propose a novel method, DSTNW, for rapid and accurate cloud computing load prediction.
• To enhance the responsiveness and precision of container cluster scaling in cloud environments.

Main Methods:

• Developed a Double-channel Temporal Convolutional Network (DTN) for long-term dependency capture and feature extraction.
• Integrated a residual temporal self-attention mechanism (SM) to focus on significant features.
• Combined DTN and SM into a dual-channel residual self-attention temporal convolutional network (DSTN).
• Implemented an adaptive weight strategy for optimizing single and stacked DSTN accuracy.

Main Results:

• The DSTNW method demonstrated superior prediction performance compared to existing state-of-the-art approaches.
• Achieved an average improvement of 24.16% on the Container dataset and 30.48% on the Google dataset.

Conclusions:

• The proposed DSTNW method offers a significant advancement in cloud computing load prediction.
• This approach enhances the ability of container clusters to adapt to dynamic resource demands, ensuring service quality.