Learning-Augmented Online Control for Sustainable Data Centers and Edge Computing Systems: Fostering a Sustainable and Supportive Digital Infrastructure

Managing the energy footprint of data centers, among the fastest-growing users of electricity, has become critical as rising demand drives up both costs and emissions. Edge computing systems, complements to data centers, are expanding rapidly as billions of connected devices push workloads closer to users and further increase the overall energy footprint of digital infrastructure. The most pressing challenge arises from short, high-demand peaks that inflate capacity charges and trigger reliance on carbon-intensive backup power. Conventional strategies often fall short: forecast-based control fails when predictions drift, while worst-case hedging wastes batteries and resources. This work introduces learning-augmented online (LAO) control, a method that treats forecasts as guidance rather than certainty. By combining simple predictive signals with probabilistic timing, it balances efficiency with resilience and maintains stable performance under uncertainty. Evaluations across diverse data center and edge workloads show consistent reductions in demand peaks and higher success rates than deterministic or purely predictive policies. By aligning computing demand with electricity and carbon objectives, LAO provides a practical foundation for more sustainable and grid-supportive digital infrastructure within an increasingly integrated energy and computing ecosystem.