Energy Digest
Summary
Summary
Summary
Summary
Summary
Summary
Summary
Summary
Summary
Technical Papers & Research
AI-curated academic research for power system engineers
Grid Operations & Resilience 2 papers
The MADWMPG framework proposes a dual-policy hybrid imitation and safe reinforcement learning method to safely mitigate grid imbalance in active distribution networks. The framework uses a wide-margin constraint barrier recourse module to learn proactively suppress exploratory violations, addressing operational volatility with a novel adaptation of noncrossing quantile regression distributional critics. The framework was trained and tested on a modified multiphase IEEE 123-bus test feeder, demonstrating superiority over conventional reinforcement learning and SRL methods.
AI agents executed a predefined ride-through workflow on a synthetic Texas grid with open-source tools, assessing a 133 MW battery's performance. The agents failed to meet ride-through criteria due to the battery absorbing reactive power during sags and the data-center load returning late, missing ERCOT's load-return requirement. The agents logged errors, and their performance was compared to results from the same machines, showing a significant difference in RMS and EMT values.
Energy Storage & Markets 3 papers
Data centres can make different contributions to supply adequacy despite reducing grid demand by the same amount, depending on workload scheduling and usable battery energy. Longer task deadlines increase capacity credit by allowing operators to defer and recover work, while fixed evening commitments and protected backup reserves have limited capacity credit. Assessing recovery deadlines and dispatchable energy is crucial to accurately determining capacity credit.
Long-term power system planning in Great Britain must account for uncertainty in electricity demand and generation, and coordinated transmission and energy storage planning can significantly reduce costs and improve outcomes. The study's framework integrates multiple energy transition pathways and captures the characteristics of transmission and storage investment over time, finding substantial benefits from joint planning. Coordinated investment can reduce transmission capacity needs by 2%, cumulative load shedding by 97%, and total system cost by 36%.
Researchers developed a coupled model to optimize dynamic wireless charging tolls and prices, accounting for displaced residential charging and network losses, to reduce grid loading. The proposed model resulted in lower grid loading violations and lower revenue reductions compared to a conventional model. Peak-aware pricing further reduced daily peak loading by 3.7%.
Other 5 papers
Deep Reinforcement Learning (DRL) offers a data-driven control approach for efficiently managing underfloor heating (UFH) systems and heat pumps (HPs), optimizing energy consumption and thermal comfort. A DRL-based controller was developed and deployed in a real residential building, successfully maintaining indoor temperatures within comfort limits while dynamically adjusting HP supply temperature and UFH valve positions. The findings demonstrate the feasibility of deploying DRL controllers in occupied buildings.
MARS proposes a feature-guided physics-informed deep learning framework for real-time battery health and lifetime prediction, achieving improved state-of-health (SOH) accuracy with mean absolute error/root mean square error values of 0.72%/0.92% and 0.63%/0.80% on NASA and CALCE benchmarks. The framework couples a dual-branch temporal encoder with an adaptive Kalman filter, providing calibrated prediction intervals for SOH estimation and conditional RUL prediction. The system's conditional RUL prediction intervals achieve high coverage probabilities, with 100.00% and 99.24% coverage on NASA and CALCE, respectively.
A data-driven model order reduction approach is proposed using Generalized Bases of Orthonormal Transfer Functions to identify a revised equation error formulation. The model is parameterized using GBOFs, and the frequency-domain fit is adjusted by selecting the basis poles, which are guided by a fit indicator. This approach aims to reduce the order of a system while maintaining an accurate representation.
Predictions from the GridSFM-Open model were found to be feasible AC-OPF operating points in some cases, but not necessarily feasible, with large gaps in others. Even with fine-tuning, the model's predictions were only marginally useful as initial values for a nonlinear solver, with some grids requiring significant speedups to converge. The model's performance varied between grids, with one grid showing faster convergence when seeded with fine-tuned predictions.
The proposed online optimization rule assigns tasks a start period to optimize energy cost, shifting flexible tasks to low-price periods while keeping high-priority tasks on schedule. Simulations show that the rule keeps GPU and CPU use within capacity while scheduling tasks at low cost, achieving a lower cost than a multi-timescale optimization benchmark. The rule also reduces computation time by eliminating the need for a solver.
Was this digest helpful?