Towards Energy-Efficient Cloud Datacentres: A Unified Framework for Generative and Multi-Scale Time Series Forecasting

Cloud computing, the backbone of modern-day digital infrastructure, is increasingly driven by artificial intelligence (AI). These workloads are computationally intensive, latency-sensitive, and resource-hungry, where efficient utilisation of the provisioned datacentre resources is both economically and environmentally critical. The rapid expansion of AI is significantly escalating the energy demands and operational complexity of datacentres, underscoring a pressing need for intelligent resource management strategies. Such a surge in AI workloads signifies the need for understanding the behavioural patterns of AI workloads particularly in terms of their arrival trend and energy profiles for achieving an optimised resource provisioning. Time-series forecasting driven resource optimisation is a potential strategy to workload volatility and sustaining operational efficiency under an increasingly heterogeneous workload behavioural profile. This thesis addresses these challenges by developing a progressive series of generative forecasting frameworks tailored for dynamic cloud environments. First, an empirical analysis of large-scale Google and Alibaba workload traces is conducted, revealing key temporal and statistical characteristics that influence forecasting performance. Building on these insights, the thesis introduces TR-GAN, a Transformer-Rectification-based Generative Adversarial Network that combines adversarial learning with attention-based sequence modelling to enhance long-horizon prediction accuracy. To further improve robustness and generalisability under volatile conditions, DAA-T-GAN is proposed, integrating diffusion-inspired data augmentation and transformative adversarial learning, achieving strong cross-dataset generalisation and competitive performance even without augmentation. Finally, DARCNet is developed as a lightweight, multi-scale dilated attention framework with adversarial residual correction, enabling efficient deployment in real-time scenarios and supporting additional tasks such as short-term forecasting and anomaly detection. Extensive experiments conducted on real-world Google and Alibaba datasets, as well as standard time-series benchmarks, demonstrate that the proposed frameworks consistently outperform the state-of-the-art baselines across multiple forecasting horizons and scenarios, with significant gains in accuracy, stability, and computational efficiency. For example, DARCNet achieves up to a 19.1% MSE reduction compared to the best non-generative baseline while offering a 3.6× training and 2× inference speed-up. These improvements underscore the unique advantages of generative models: by learning the underlying probability distributions of volatile workloads, they not only produce more accurate point forecasts but also offer intrinsic capabilities for uncertainty quantification and realistic scenario generation, capabilities that are essential for robust decision-making in dynamic cloud environments but are inherently lacking in deterministic non-generative approaches. These findings underscore the potential of the proposed frameworks to serve as scalable, high-performance forecasting solutions for next generation cloud datacentre, particularly in environments dominated by AI workloads.<p></p>

Author dc:creator

• Zekun Sun (21048077)

• Time Series Forecasting
• Cloud computing
• Generative forecasting frameworks

Statement dc:rights

• All rights reserved

Identifier

10779/exe.32134381.v1

OAI identifier oai:identifier

oai:figshare.com:article/32134381