Energy Digest

PowerAgentBench-Dyn is a benchmark designed to evaluate Agentic AI systems on power system dynamic-analysis tasks, targeting problems that require reasoning, tool usage, and iterative experimentation. The benchmark includes two initial tasks: Dynamic Model Quality Review Benchmark and Dynamic Security Risk Screening Benchmark, which assess agents' ability to validate models, diagnose faults, identify security risks, and propose mitigation measures. The framework provides a reproducible evaluation metric for assessing Agentic AI systems in power system operation and planning.

Large-scale power outages are caused by dynamic interactions between network dynamics and component failures in power grids, allowing for the study of cascading failures through a new model that integrates node and line failure dynamics with oscillator models. The study reveals two novel mechanisms driving system fragility: high inertia can amplify cascade sizes when not balanced with other properties, while increasing node robustness paradoxically leads to larger cascades. This understanding is crucial for achieving resilient future power grids.

Ramping procurement co-optimization under net-demand uncertainty can lead to generator under-compensation, requiring discriminatory bid cost recovery. Locational marginal pricing (LMP) yields discriminatory energy prices but may not be ideal for price-taking generators, whereas maximum dispatch cost pricing (MDCP) and maximum temporal locational marginal pricing (MTLMP) can provide truthful bidding incentives with some trade-offs in producer profits and consumer payments. Single-interval co-optimization is advantageous under high forecast uncertainty, while multi-interval co-optimization excels when net-demand forecasts are accurate and ramp needs are challenging.

Gradient Difference-based Graph Unlearning (GDGU) is a method for graph-level multi-label classification tasks in electric vehicle charging networks that removes the influence of deleted training data through first-order parameter correction. GDGU achieves state-of-the-art localization utility and forgetfulness fidelity, outperforming second-order GU baselines, while reducing computational cost and memory usage by 10-12 times compared to full retraining. This method enables efficient and private data sharing in EVCS cyberattack localization tasks.

A new open-source Python package called ev-flow generates synthetic plug-in electric vehicle charging behavior for eight U.S. regions, grounded in 2017 National Household Travel Survey microdata and regional sales-mix models, with a focus on replicability and reproducibility. The package produces behaviorally realistic populations of individual charging profiles, addressing the lack of real-world charging telemetry, and is licensed under MIT. It fills a niche for U.S.-focused, NHTS-grounded charging behavior in contrast to European generators and simulators.

Model-free reinforcement learning controllers demonstrate improved resilience to cyberattacks, including false data injection and denial-of-service attacks, with Lyapunov reward offering best results for accuracy and low tracking error, while RL-MPCs require longer training times. Exponential mode provides a good trade-off between resilience and moderate training conditions. Proximal Policy Optimization outperforms Deep Deterministic Policy Gradient with a significant reduction in KPI variance.

Deep neural networks can predict failure time distributions for systems with competing risks by associating neural network structure with data structures, allowing different covariate groups to impact prediction through separate sub-networks. The Structured Segmented Hazard Deep Neural Network (SSH-Net) outputs cause-specific hazard functions and utilizes a penalized log-likelihood loss function. SSH-Net is validated using simulation studies and Titan GPU failure time data, demonstrating improved accuracy in predicting cause-specific cumulative incident functions.

A new defense framework against False Data Injection Attacks (FDIA) in Power Grids uses pseudo-feature padding to strengthen Deep Neural Networks (DNNs) by adding input layer padding, making adversarial attacks computationally infeasible due to non-transferable and unpredictable perturbations. This method is lightweight, model-agnostic, and requires no core architecture modifications, making it highly deployable in real-world CPS settings. The framework significantly improves DNN robustness with negligible impact on performance, effectively mitigating attacks that bypass conventional defenses.

Shared mobile storage for electric vehicle fleets offers techno-economic viability for demand charge reduction by minimizing costs and maximizing energy efficiency, with real-world data from San Francisco showing significant demand charge savings achievable through modest fleet sizes. A proposed mixed-integer linear program framework jointly minimizes demand charges and total cost of ownership, while a marginal-value-based heuristic algorithm achieves near-optimal performance at high computational efficiency. The analysis reveals how tariff structures, fleet size, and cost components influence overall profitability.