Energy Digest

A new data-driven modeling framework proposes a privacy-preserving approach for managing electric vehicles (EVs) in frequency regulation tasks, leveraging aggregated EV data without individual user information. The framework uses a bilinear hidden Markov model to accurately estimate power outputs and flexibility of aggregated EVs, while ensuring scalability and practicality. Simulation results demonstrate the method's accuracy and effectiveness under SOC data inaccuracies, outperforming existing models.

Graph Neural Ordinary Differential Equations for Power System Identification employ message-passing graph Neural Ordinary Differential Equations (NODEs) to identify coupled systems, offering improved generalization through structural inductive bias. The proposed method, MPG-NODEs, incorporates local node and edge embeddings, autoregressive schemes, and transfer learning options to infer latent representations of unmeasured states from past measurements. This framework outperforms state-of-the-art machine learning architectures when applied to identify voltage and frequency dynamics of power systems.

A data center's uninterruptible power supply (UPS) disconnects during voltage/frequency disturbances and then reconnects while the bulk grid is settling. Poorly timed reconnection can amplify oscillations, deepen frequency deviations, and lead to repeated "flapping." An analytical framework characterizes a safe reconnection time for large DC loads after disconnection to prevent flapping.

Energy forecasting research faces a comparability gap due to varying model evaluations and benchmarks, with reported accuracy gains often not directly comparable. A new dynamic benchmarking platform called the Energy-Arena has been introduced to address this issue, providing a continuously updated reference point for operational energy time series forecasting. The platform operates as an open submission system with standardized challenge definitions and leaderboards, improving transparency by preventing information leakage.

The current state of scientific research, as represented by traditional paper publications, imposes two structural costs: the Storytelling Tax and the Engineering Tax, which result in a loss of critical implementation details when implemented. A new protocol called Agent-Native Research Artifact (Ara) aims to address this issue by structuring research packages around four layers, preserving failures and evidence grounding every claim in raw outputs. Ara has shown promising results in improving question-answering accuracy and reproduction success in AI agent performance on certain benchmarks.

GradMAP proposes a decentralized learning method for large populations of grid-edge devices that respects three-phase AC distribution-network physics without parameter sharing or communication between agents. It achieves this by training independent neural-network policies with embedded power-flow models and using proximal surrogate methods to speed up training. GradMAP results in significant training speed-ups and lower operating costs compared to other methods, learning decentralized policies within 15 minutes on a single GPU.

Robust forecast aggregation combines predictions from multiple sources to perform well in the worst case across all possible information structures, allowing for unknown state spaces and prior knowledge. A simple log-odds aggregator is proposed, achieving nearly tight minimax-regret guarantees across three knowledge regimes with worst-case regret of 0.0255 or lower. The aggregator also outperforms existing methods, with a regret upper bound strictly less than 0.0226 in the classical setting with known state space {0,1}.

Machine learning models, including GBRT with an R squared value of 0.88, outperform traditional LSTM and SVR models for short-term electricity price forecasting in Australia's National Electricity Market, but all models struggle with high accuracy above 90% mean absolute percentage error. However, these same tree-based models excel at demand prediction tasks, achieving higher R2 values and lower errors than LSTM and SVR. Hybrid models such as tree plus transformers and data augmentation for extreme events are needed to improve price forecasting accuracy.

A novel consensus analysis method is proposed for multi-agent systems with heterogeneous time-varying delays, which can lead to conservatism when using homogeneous delay bounds. The new approach, known as individual-delay-reflected generalized consensus, uses a Lyapunov-Krasovskii functional that separates the integral term into intervals containing different delay values. This results in reduced conservatism and improved accuracy of the analysis.