Energy Digest

Next-generation AI data centers require architectural shifts to address increasing power demand, transients, and thermal stress due to rapid growth of AI workloads. Three enabling technological building blocks identified for support include high-voltage conversion-ratio DC/DC converters, facility-level low-voltage DC distribution, and medium-voltage solid-state transformers. These building blocks offer advantages but also present technical challenges that need to be addressed.

A novel explainable control framework (XCF) is proposed to provide human-understandable insights into controller behavior, enabling precise and reliable control in complex scenarios. The XCF offers model-agnostic explanations for controllers and can refine local explanations based on system response dynamics. A large language model agent-supported user interface is developed to analyze user requirements, interpret generated explanations, and provide interactive consultation.

The paper proposes a reference-free heterogeneous multi-agent reinforcement learning framework for optimizing tie-line power shaping in industrial microgrids, achieving zero production failures with minimal computational time. The framework effectively eliminates dependence on predefined reference trajectories and enables adaptive 1-min online decision-making. It reduces grid purchase costs, contract-demand exceedance times, and cumulative ramp excess by significant margins compared to original operation.

Artificial intelligence (AI) data centers can reduce electricity consumption during peak demand through software-based workload orchestration, enabling them to operate as grid-interactive assets. Modern GPU-based AI data centers can respond dynamically to power system conditions, demonstrating rapid load reduction, sustained curtailment, and carbon-aware operation while preserving service levels. These capabilities transform AI infrastructure into flexible resources supporting grid reliability, accelerating interconnection, and improving computing sustainability.

Synthetic data augmentation can improve certain score-based imbalanced classification metrics when the true minority distribution is well-specified, but its effect is limited to variance reduction and may introduce additional bias. Under misspecification, synthetic data augmentation can correct ranking errors by changing effective class balance. Simulation studies confirm these findings, with nontrivial improvements under misspecification conditions.

Gallium-nitride (GaN) power devices offer advantages in efficiency, power density, and reliability for specific stages of grid-to-load conversion in data centers, particularly high-frequency, low-to-mid-voltage stages. GaN architectures provide stage-dependent benefits, with commercial lateral GaN HEMTs excelling in these applications and vertical GaN emerging as an option for higher-voltage and higher-power conversion. The effective deployment of GaN power devices requires coordinated design of devices, topology, packaging, and thermal management to achieve optimal performance and reduce carbon emissions.