Energy Digest

A new study proposes a temperature-aware scheduling approach to reduce carbon emissions and water consumption of Large Language Models (LLMs) in large-scale geo-distributed edge data centers, leveraging ambient temperature diversity to optimize cooling energy costs. The method co-optimizes LLM energy costs, carbon emissions, time-to-first token, and water consumption using a distributed optimization algorithm. It demonstrates reductions in cooling energy consumption and improves overall cost efficiency for geo-distributed cloud environments.

Stability of economic model predictive control can be proven under the assumption of two-storage strict dissipativity, which requires two storage functions to satisfy dissipativity and be separated by a positive definite function. This condition is more easily verifiable than traditional strict dissipativity and is related to optimal control via value functions. Two-storage strict dissipativity is both sufficient and necessary for asymptotic stability in finite-horizon economic model predictive control systems.

Eigenvalue patterns and participation analysis of symmetric renewable energy power systems investigate the dynamics of ideally-, quasi-, and group-symmetric systems, which can facilitate stability analysis due to their homogeneous nature. Two types of modes are defined: inner-group modes describing interactions among subsystems within a group, and group-grid modes describing interactions between groups and the external grid. A new concept, group participation factor, is proposed to extend conventional participation factors for repeated and close modes.

The coordination between Transmission System Operator (TSO) and Distribution System Operator (DSO) is crucial for energy transition, enabling the utilization of flexibility from Distributed Energy Resources (DERs). Effective coordination schemes are necessary to balance the system while avoiding network congestion. A broad range of schemes have been analyzed, including their pros and cons, with a focus on optimizing TSO/DSO use of flexibility resources.

A novel model-free control framework leveraging policy distillation is presented to address the trade-off between control performance and computational burden in power inverters, using an error energy-guided hybrid reward mechanism and adaptive importance weighting. The approach compresses the heavy DRL policy into a lightweight neural network while retaining desired control performance, overcoming computational bottlenecks during deployment. Experimental validation on a kilowatt-level experimental platform shows significant reduction in inference time and improved transient response speed and parameter robustness.

A unified, real-time building-level carbon-aware energy management system (CAEMS) is proposed to tackle climate change by optimizing grid imports, energy storage, and flexible demand simultaneously, directly integrating time-varying marginal carbon intensity signals into the EMS objective. The CAEMS model uses a mixed-integer linear program with a Transformer-based forecaster to predict electricity prices and carbon intensity, achieving significant emissions reductions while minimizing cost increases. Simulation results show a 22.5% reduction in emissions with only a 1.7% increase in cost when modest carbon prices are used.

Generative agents powered by large language models (LLMs) can relax the rigidity of traditional mathematical models for human decision-making in power dispatch and auction settings. By incorporating an in-context learning module, LLMs can learn to prioritize post-blackout energy reserves over short-term profit in home battery management tests. In simultaneous ascending auctions, LLM agents with structured prompting can both reproduce economically rational strategies and exhibit systematic behavioral deviations when given strategic objectives.

European Member States are introducing national capacity mechanisms but isolated CMs are inefficient and prone to under- or over-investment. A novel conceptual design proposes a coupled European capacity market using flow-based market coupling, which reduces system costs by harnessing available capacity in neighboring market zones. This approach ensures deliverability with respect to network constraints in all scarcity situations.

VB-NET is a physics-constrained gray-box deep learning framework that transforms air conditioning systems into virtual batteries, allowing for demand-side flexibility in renewable energy grids. The framework overcomes challenges such as parameter acquisition, interpretability, and data scarcity through strict enforcement of physical laws and multi-task learning. VB-NET outperforms conventional models in state of charge tracking and recovers underlying thermodynamic laws to yield physically consistent parameters.

A novel quantum annealing model predictive control-based power allocation framework is proposed to accelerate optimization problems in large-scale household energy scheduling with hydrogen storage. The framework addresses complex challenges posed by multiple fuel cells and electrolyzers, determining startup/shutdown times, output power, and hydrogen generation rates. This approach effectively solves large-scale optimization problems, particularly for scenarios with many connected households.