Uncertainty-Aware Decarbonization for Datacenters

TL;DR

This work tackles the problem of uncertainty in carbon-intensity forecasting for datacenter decarbonization by identifying temporal and spatial sources of uncertainty and introducing a conformal-prediction–based framework (SPCI) to produce reliable prediction intervals. The approach converts point carbon-intensity forecasts into confidence sets with coverage $1-\alpha$, accommodating temporal dependencies via a feedback mechanism and autoregressive quantile modeling. Evaluation on real production traces demonstrates that the method achieves target coverages across regions and significance levels, and case studies show that incorporating uncertainty into load-shifting decisions mitigates emission increases in both temporal and spatial contexts. Overall, the work provides a practical path toward uncertainty-aware scheduling in datacenters, yielding measurable emissions reductions and guiding future integration of uncertainty quantification in carbon-aware infrastructure management.

Abstract

This paper represents the first effort to quantify uncertainty in carbon intensity forecasting for datacenter decarbonization. We identify and analyze two types of uncertainty -- temporal and spatial -- and discuss their system implications. To address the temporal dynamics in quantifying uncertainty for carbon intensity forecasting, we introduce a conformal prediction-based framework. Evaluation results show that our technique robustly achieves target coverages in uncertainty quantification across various significance levels. We conduct two case studies using production power traces, focusing on temporal and spatial load shifting respectively. The results show that incorporating uncertainty into scheduling decisions can prevent a 5% and 14% increase in carbon emissions, respectively. These percentages translate to an absolute reduction of 2.1 and 10.4 tons of carbon emissions in a 20 MW datacenter cluster.

Figures (8)

• Figure 1: Average 24-hour prediction accuracy from July to December in 2022 across three regions. The whiskers indicate standard deviations.
• Figure 2: Average 24-hour prediction accuracy for a representative week across three seasons in CISO.
• Figure 3: Average prediction accuracy in 4 temporal groups from July to December in 2022 across three regions. The whiskers indicate standard deviations.
• Figure 4: Average 24-hour prediction accuracy for November 2022 in three regions: CISO, ERCO, and ISNE.
• Figure 5: The feedback mechanism (the red arrow), where $\hat{C}_{t-1}(x_t)$ is updated based on the residuals at each step.