Dynamic Resource Load Forecasting in Cloud Computing: A Representation Learning Perspective

Abstract

In cloud computing environments, resource loads exhibit strong temporal volatility, non-stationarity, and multi-source coupling. Accurate prediction of their evolution is a fundamental requirement for efficient resource management and service quality assurance. To address the limitations of traditional methods in capturing complex temporal dependencies and load generation mechanisms, this paper proposes a dynamic cloud resource load prediction approach based on deep time series representation learning. The method models multivariate historical resource monitoring data as unified time series inputs. Through temporal encoding and representation aggregation, it automatically learns key patterns in load variation. This enables a robust characterization of load evolution in latent space. Under a unified data setting, comparative analysis with several representative time series prediction models shows clear advantages in error control and overall stability. The results indicate that a representation-centered modeling paradigm effectively mitigates the impact of noise and distribution fluctuations. It is therefore more suitable for complex and dynamic load behaviors in real cloud environments. This study provides a systematic solution for cloud resource load prediction and offers valuable reference for intelligent resource allocation, capacity planning, and cloud platform optimization.