Generalizable Latent Representation Learning for Multi-Source Time Series Forecasting in Cloud Systems

Abstract

Time series forecasting in cloud computing environments faces persistent challenges related to scalability and cross-scenario adaptability, due to the continuous generation of complex time series from multi-source monitoring metrics. This paper focuses on cloud system time series forecasting and proposes a prediction framework centered on a unified latent representation. By mapping high-dimensional and heterogeneous system observations into a low-dimensional latent space, the framework enables structured modeling of system dynamics. Temporal dependency modeling and recursive forecasting are performed directly in the latent space, which effectively reduces modeling complexity and alleviates reliance on specific system configurations and metric distributions, thereby improving applicability across diverse operating environments. The framework jointly considers prediction stability, temporal structure consistency, and inference efficiency, allowing the model to maintain prediction quality while satisfying the computational constraints of large-scale cloud systems. Based on publicly available cloud time series data, comparative experiments are conducted to systematically evaluate the proposed method in terms of prediction accuracy, temporal consistency, and inference efficiency. The results demonstrate that the method achieves a more balanced performance across multiple criteria, validating its effectiveness for scalable and generalizable time series forecasting in cloud system scenarios.