Time-Frequency and Multi-Scale Learning Network for Long-Term Cloud Workload Prediction

Abstract

Accurate long-term workload prediction (LTWP) is crucial for effective resource management in cloud environments, where workload dynamics are characterized by high variability, noise, and long-range temporal dependencies. Existing methods, such as linear models and PatchTST-based models, often struggle with capturing long-range dependencies, handling noise, and adapting to diverse workload patterns. To address these issues, this paper proposes TFMSNet, a novel time-frequency and multi-scale learning network for LTWP. TFMSNet integrates low-frequency wavelet features to capture global workload trends and high-frequency components to model local variations and change points, enabling more comprehensive temporal representation. Its multi-branch architecture allows for efficient capture of multi-scale dependencies, enhancing adaptability to diverse and dynamic workloads. Extensive experiments on five real-world datasets—Azure-2017, Alibaba-2018, Google-2019, Alibaba-2020 and Alibaba-2021—demonstrate that TFMSNet consistently outperforms state-of-the-art models in performance metrics of MAE, MSE, RMSE and MAPE across multiple prediction horizons. Ablation studies confirm the critical roles of frequency-enhanced attention and multi-branch architecture in improving accuracy. Moreover, the model’s robustness under noisy conditions and its efficiency with shorter input sequences further highlight its practical applicability for real-time, resource-constrained cloud environments.