Energy Digest

The rapid growth of artificial intelligence (AI) computing is transforming data centers into large electrical loads that require stable and flexible power delivery architectures to operate reliably during fast grid disturbances. To address this, researchers are exploring technological trends such as workload orchestration, cooling systems, on-site resources, and energy storage, as well as higher-voltage DC distribution, solid-state transformers, and liquid cooling. A three-level stability framework is being developed to connect dominant instability mechanisms with suitable modeling, assessment, and mitigation approaches for grid-to-chip co-design in scalable AI infrastructure.

The paper develops a critic-free policy iteration method for continuous-time linear zero-sum games by characterizing saddle-point policies directly in the joint policy space and introducing a new policy game Riccati equation. The solution to this equation is in one-to-one correspondence with the symmetric solutions of the game algebraic Riccati equation, allowing for direct policy iteration without the need for critic identification. This approach has been shown to have several benefits, including reduced computational and memory requirements.

A novel pricing methodology is proposed to incentivize generators to provide voltage stability services in power systems with high penetration of Inverter-Based Resources. The method uses a primal-dual formulation and has been shown to consistently produce revenue-adequate shadow prices, enabling units to recover their costs without additional uplift payments. This addresses the limitations of previous methods that failed to guarantee cost recovery for voltage-stability service providers.

Active Distribution Networks (ADNs) are increasingly vulnerable to grid instability due to growing Inverter-Based Resources (IBRs), which can lead to unintended electrical islands. When DERs operate at fixed power setpoints, the risk of sustained islanding is limited, but enabling frequency and voltage regulation on Synchronous Generators (SGs) can sustain an island indefinitely without triggering conventional protection. The persistence of islanding depends on SG regulation mode and system inertia.

Two commonly used unit commitment aggregations, p-clustered unit commitment (PCUC) and tight unit aggregation (TUA), cannot guarantee intertemporal consistency of unit-level output allocations due to a limitation where they do not admit feasible disaggregation. A counterexample is provided that satisfies all aggregate constraints but admits no feasible disaggregation. This highlights the challenge of developing exact aggregation models for slow-ramping units that can be solved efficiently.

Temporal networks with fixed switching sequences impose stricter herdability conditions than conventional switched systems; herdability depends on network topology, switching durations, and edge weight magnitudes; a specific graph-theoretic condition involving π-graphs is sufficient for structural sign herdability in temporal networks.

A frequency-domain, impedance-based sensitivity methodology is presented to assess and enhance the stability of line-commutated converter HVDC links in weak-grid conditions using grid-forming converters. The approach integrates various techniques, including black-box identification, multivariable stability assessment, and modal impedance decomposition. Integration of a grid-forming voltage source converter substantially increases the stability margin of standalone LCC-HVDC links.

The Explainability Assistant is an open-source conversational AI system that provides flexible natural language interaction to facilitate interpretation of energy consumption models, achieving 94% intent-parsing accuracy. It adapts to different ML problem types without task-specific fine-tuning and surpasses traditional XAI dashboards in usability and consistent task accuracy. The system was preferred unanimously by energy domain specialists for practical use over traditional XAI interfaces.

LoaDiff is a diffusion-based generative model that creates synthetic smart-meter load curves for energy analytics. The model can be conditioned on static household attributes and dynamic contextual variables such as calendar information and outdoor temperature. LoaDiff generates realistic and diverse load profiles while balancing generation quality, memorization risk, and utility for downstream applications like load forecasting and appliance detection.