Energy Digest

A new executable framework for bridging AI and power systems education has been developed to address the gap in existing works, providing a reusable material for newcomers and interdisciplinary learners. The framework consists of open modules that follow established engineering and power-system domain rules, covering core concepts such as deep neural networks, convolutional neural networks, and physics-informed neural networks. The modules are released as Jupyter notebooks and have been successfully delivered through an IEEE online course and webinar series, demonstrating a need for this type of education in the field.

This paper studies distributed online control for linear dynamical systems with adversarial disturbances and time-varying convex costs over a network of agents, competing against centralized policies in hindsight. Each agent uses spectral controllers that convolve past disturbances with leading eigenvectors of a Hankel matrix, updating parameters through distributed online gradient descent. A sublinear regret bound of $O(\frac{\sqrt{T}\text{poly}(\log T)}{γ^3})$ is established under standard assumptions.

A standard set of inputs was created to facilitate reproducible benchmarking in local flexibility markets, including a modified CIGRE MV network with base-load and stress-load conditions, synthetic flexibility offers, wholesale prices, and transformer parameters. This input set allows for the comparison of different clearing methodologies on the same data. The associated data are publicly available on Zenodo.

A new technology uses an "inter-router circuit" featuring a parallel inductor as a temporary energy buffer to enable fully controllable bidirectional power transmission, overcoming limitations of conventional power packet routers. This allows for more flexibility in power supplies for battery-powered autonomous systems. The proposed system uses time-division multiplexing and has been verified through experiments using prototype hardware.

LLM inference serving's power dynamics can be controlled by chunked prefill scheduling, which regulates power ramp rate without affecting peak power. Splitting long prompts into smaller steps actually reduces the mean ramp rate substantially, especially at heavy loads or with large requests. This technique could reduce a grid operator's need for fast-ramping reserve capacity by 20-23%.

Decentralized control synthesis for integrating multiple inverter-based resources (IBRs) into power systems is proposed using block-diagonal dominance (BDD) theory, ensuring small-signal stability and decentralized multi-input multi-output (MIMO) control design. A new numerical metric is introduced to quantify the conservatism of the decentralized stability certificate, with a guaranteed minimum decay rate. The approach is validated through a case study on the IEEE 9-bus test system.

An AI-based decision-support pipeline for day-ahead photovoltaic forecasting was developed to improve reliability of low-carbon energy systems. The pipeline corrects timestamp conventions, constructs new features, adds atmospheric context, and combines predictors through validation-learned stacking, resulting in significant reductions in forecast errors. It achieves performance improvements by 6.6% and 9% compared to individual machine-learning models under different evaluation protocols.