Search for: All records

The rapid rise of AI applications has driven datacenters to unprecedented energy demands, which has prompted major tech companies to adopt on-site nuclear power plants (NPPs) alongside grid electricity. While existing research focuses on off-site NPPs in multi-energy systems optimized for investment returns, recent advances in small modular reactors (SMRs), particularly load-following SMRs (LF-SMRs), offer flexible, reliable power tailored for datacenter co-location. However, LF-SMRs are governed by a set of physical constraints, such as ramp rate and stability limits, making them unsuitable as fully dispatchable sources. This paper proposes a novel day-ahead workload scheduling approach that jointly coordinates datacenter operations and LF-SMR output, explicitly modeling these constraints. We develop a two-stage formulation that forecasts carbon-free grid energy from the grid using conformal prediction in the first stage and then optimizes LF-SMR output and workload scheduling via mixed-integer programming in the second stage. Evaluation on real workload traces shows that our method reduces carbon-based energy consumption by up to 43.44% compared to baselines that omit nuclear integration or ignore SMR limitations. 

Buildings produce a significant share of greenhouse gas (GHG) emissions, making homes and businesses a major factor in climate change. To address this critical challenge, this paper explores achieving net-zero emission through the carbon-aware optimal scheduling of the multi-energy building integrated energy systems (BIES). We integrate advanced technologies and strategies, such as the carbon capture system (CCS), power-to-gas (P2G), carbon tracking, and emission allowance trading, into the traditional BIES scheduling problem. The proposed model enables accurate accounting of carbon emissions associated with building energy systems and facilitates the implementation of low-carbon operations. Furthermore, to address the challenge of accurately assessing uncertainty sets related to forecasting errors of loads, generation, and carbon intensity, we develop a learning-based robust optimization approach for BIES that is robust in the presence of uncertainty and guarantees statistical feasibility. The proposed approach comprises a shape learning stage and a shape calibration stage to generate an optimal uncertainty set that ensures favorable results from a statistical perspective. Numerical studies conducted based on both synthetic and real-world datasets have demonstrated that the approach yields up to 8.2% cost reduction, compared with conventional methods, in assisting buildings to robustly reach net-zero emissions.