Energy Digest

A new probabilistic safety certification framework has been developed to ensure the safe deployment of AI-based grid-edge coordination systems. The framework reduces the complexity of system operation to a binary unsafe outcome under an operator-defined safety specification, providing a tight upper bound on the probability of false safety certification. The method is evaluated through case studies with 1,000-agent AI models and demonstrates the value of integrating adversarial attacks into the deployment process.

EvEMTBench is an open benchmark for machine learning in power system protection that addresses the challenges of comparing different studies by defining 12 protection functions as 24 scored tasks, supporting structured evaluation across various observability conditions, distribution shifts, and transfer between grids. The benchmark provides a common basis for comparing future methods with committed partitions, leakage controls, and reproducible reporting. EvEMTBench reveals that wider observability is not uniformly beneficial and that cross-grid transfer is stronger for fault detection than for fault localization.

Rapid growth in data centers has led to high levels of computational load that can cause grid events, such as transmission disturbances, resulting in large amounts of power being lost or transferred. A new framework is proposed to analyze and predict these events, combining multiple lines of work into a cycle-space certificate model for grid synchronization and transient stability. The model provides stronger static tests that reveal more loading capacity in some cases and demonstrate the importance of balancing power supply location on transient margin.

Power swings in large Data Centers running AI workloads can induce forced oscillations in power systems, leading to flicker, equipment disconnection, or blackouts. These fluctuations can be amplified by poorly damped modes, causing unbounded oscillations depending on frequency and magnitude. The impact of DTC load fluctuations on grid stability is analyzed, providing insights for system operators to define new regulations on maximum allowed load fluctuations.

Converter terminal characteristics are reshaped using virtual two-port control instead of virtual impedance, which uses four coordinated transfer channels to reconstruct terminal impedance for more accurate dynamic interactions with the network and steady-state power sharing in microgrids. Virtual two-port control transforms the original impedance through a matrix linear-fractional map, enabling stable and proper reconstruction. The proposed method offers advantages over conventional virtual impedance in applications such as passivation with reduced control effort and seriesshunt power-flow regulation.

A new algorithm called DT4X+ has been developed that enhances a previous diagnosis algorithm by modifying its construction and loss function to create fully efficient fault indicators. This results in improved interpretability, robustness, and performance compared to the original algorithm. The enhancements lead to more informative splits and better performance on dynamic-system datasets.

A fully solderless GaN-based variable speed drive was successfully designed and tested, demonstrating reliable performance with no degradation in effective on-state resistances after over 120 thermal cycles. The solderless assembly showed resistance to vibration and power cycling without electrical failure, allowing for non-destructive component replacement and re-use. Initial life-cycle assessments indicated a higher embodied carbon footprint but prospectively evaluated injection-molding scenarios could reduce it to near that of the soldered reference.

GA-Agent, a system combining genetic algorithms with large language models, achieves 100% success on eight diverse control case studies, outperforming traditional GAs in solution quality and sample efficiency while reducing function evaluations by one to two orders of magnitude. The use of a compact memory buffer and cost-effective LLM backbones results in superior performance without significant added computational overhead. GA-Agent requires no manual hyperparameter tuning or bilevel optimization, decoupling high-level semantic reasoning from low-level numerical search.