Energy-aware job scheduling for green data centers under uncertainty: A structure-compressed spectral risk control approach

As electricity expenditures become a dominant component of operating costs for green data centers (GDCs), energy-aware job scheduling provides an effective means to reduce expenses. Nevertheless, practical scheduling remains hindered by dependent workloads with structural redundancy and environmental uncertainty. To address these challenges, this study proposes a structure-compressed spectral risk control approach for scheduling in GDCs, termed SCS-RC. First, the scheduling problem is formulated as a Markov decision process that models heterogeneous jobs and their internal task dependencies, while capturing energy consumption via power usage effectiveness under time-varying electricity prices and temperature conditions. To mitigate redundancy before decision making, an inter-task similarity detection and merging algorithm compresses task dependency graphs, by which similar substructures are identified and merged to form a compact task graph. Then, the fully parameterized quantile function is adopted to estimate the return distribution for each state–action pair via adaptive quantile fractions, accommodating skewed and long-tailed returns under uncertainty. To limit adverse tail exposure, the learned quantiles are further aggregated by spectral risk measures to guide action selection. Case studies show that SCS-RC improves profit stability and reduces energy consumption compared with representative baselines, while achieving more reliable training convergence and maintaining efficient online inference under increasing job-arrival scales.