Carbon-Efficient Data Center Scheduling under Uncertainty: A Conformal Prediction Approach

The temporal and spatial mismatch between data center energy consumption and renewable energy supply, particularly solar energy, presents a significant challenge. During early morning and evening hours, computational load far exceeds the available renewable energy supply, while at midday, renewable energy generation often surpasses computational demand, leading to wasted renewable resources and increased carbon emissions. To address this issue, systemic task scheduling optimization strategies have been developed, which involve deferring nonurgent tasks from peak load periods to times of renewable energy surplus. However, the effectiveness of these strategies heavily relies on the accurate prediction of both renewable energy availability and computational load, both of which are subject to considerable uncertainty. This uncertainty can lead to suboptimal scheduling decisions. This study proposes a novel task scheduling optimization framework that integrates confidence prediction information derived from Conformal Prediction (CP) methods. By utilizing multi-step probabilistic prediction intervals for risk-aware optimization, the framework mitigates the impact of uncertainty on task scheduling strategies and enhances their reliability. In a simulated data center environment constructed with real-world load and solar energy data, the proposed frame-work achieved an 7.47% reduction in overall system carbon emissions, with a maximum reduction of 10.36%. This approach provides an interpretable uncertainty quantification paradigm for complex system optimization, effectively alleviating the impact of uncertainty due to time series data prediction on suboptimal policies generated by task scheduling optimization frameworks.