Energy Digest

Quantization-enabled demand response is being explored to manage increasing energy loads from large language models (LLMs) in data centers. A proposed framework maps model quantization configurations to dispatchable parameters and accounts for instance switching, request routing, and precision selection in a two-stage optimization process. This approach reduces total data-center operating cost by 34.3% without curtailing served token volume.

Small modular reactors' (SMRs) load following capabilities rely on a thermodynamic coupling between primary and secondary loops. A hybrid dynamic framework successfully coupled an equation-based model with a physics-based secondary steam cycle to accurately assess load following, showing that partial control strategies are insufficient for efficient operation. Full action of three actuators stabilizes steam pressure, limits thermal excursions, and maintains safe operating margins during load changes.

A new framework called BR-FedMAPPO is proposed to protect interconnected Microgrids from Stealthy False Data Injection Attacks. The framework uses a triple-surface Moving Target Defense and an adaptive isolation strategy to learn a policy that mitigates attacks, contains cascading disruptions, and maintains cost-aware dispatch performance. Simulation results show effective mitigation of attacks and containment of disruptions in four interconnected MGs based on the IEEE 30- and 118-bus test systems.

A learn-to-optimize (LTO) architecture is proposed for distributed optimal power flow, combining data-driven and model-based methods, achieving near-instantaneous interpretable decision-making. The LTO architecture surpasses state-of-the-art solvers in optimality and excels over existing data-driven approaches in feasibility. It outperforms existing solutions through comparative case studies underpinning its effectiveness.

PowerAgentBench-SS is a steady-state benchmark framework that evaluates the performance of tool-using agents in power system operation and planning studies. The benchmark assesses an agent's ability to execute a workflow involving multiple tasks, including inspecting grid cases, proposing mitigations, and producing auditable evidence trails. The results highlight the limitations of solver-only or answer-only evaluations, showing that tool-use efficiency, validation-budget use, and mitigation behavior are also critical metrics for evaluating agent performance in this domain.

A comparative simulation study evaluated two model predictive controller strategies with different control objectives, finding that one minimising quadratic heating power consumed less energy than the other prioritising indoor temperature tracking for thermal comfort. The comfort-oriented controller achieved lower total heat consumption while maintaining high comfort levels without structural modifications to the building envelope. This study demonstrates the potential of intelligent heating control strategies to reduce heat demand in buildings with lower investment and faster implementation.

A novel LEO SBSP system model was developed and simulated, showing that peak DC power delivery reaches 1.986 MW with mean per-site delivery ranging from 40 to 75 kW. The simulation considered orbital propagation, eclipse cycles, satellite power chain, and atmospheric attenuation. The incident peak power density at the rectenna remained within safe limits, suggesting potential for realistic per-site delivery of 50-100 kW.

Electricity markets in Europe are complex systems driven by nonlinearities and high-dimensional interactions, with increasing interdependence across regions. A combination of DNN models and explainable artificial intelligence (XAI) techniques was used to analyze drivers of electricity prices across 39 European bidding zones, identifying renewable energy sources as a disproportionate influence on price formation. Gas prices remain a dominant driver across electricity markets, while interconnections significantly shape price dynamics in the integrated EU-wide market scenario.

A decision-focused reinforcement learning framework is proposed to address the challenges of smart control of electric vehicle (EV) charging with unknown departure times. The framework trains a forecaster end-to-end with feedback from the charging policy actions, resulting in higher-quality actions and improved overall performance. This approach yields superior charging decisions, achieving up to 14% improvement in total reward and 55% reduction in unsupplied energy.

INDEQS, a graph-based Neural Controlled Differential Equation (NCDE) forecasting method, incorporates prior knowledge of a directed graph at distinct architectural positions to improve performance on time series forecasting tasks. The method offers two variants: a lightweight graph-constrained variant and an expressive variant that learns additional graph connections from data. Informedness consistently improves mean absolute error over uninformed NCDE methods, particularly on larger graphs.