Energy Digest

The increasing penetration of Distributed Energy Resources (DERs) in Low-Voltage distribution networks requires accurate knowledge of DER capacity to improve operational and planning processes. However, limitations in data availability hinder reliable estimation of DER installed capacities, prompting the development of aggregated DER installed capacity as a practical solution. Aggregated DER capacity can enhance forecasting, congestion management, flexibility quantification, hosting capacity assessment, and monitoring of DER adoption without requiring customer-level monitoring.

The proposed matrix zonotope perturbation framework uses matrix perturbation theory to analyze how noise-induced distortions affect model dynamics. It provides interpretable Cai-Zhang bounds for matrix zonotopes and extends them to constrained matrix zonotopes, offering a more scalable reachable-set update method. The experimental results show that the new approach is faster and produces less conservative reachable sets than existing methods.

A state-based scheduling framework allows for the exploitation of scheduling flexibility in slot-timed servers with long-run guarantees, constraining it to a polytope of feasible schedules. The framework fully characterizes this polytope, enabling flexible exploitation but requiring computational complexity when fully utilized. A specialized class of services, dual-curve services, can be efficiently specified and updated, providing near-practical viability while maintaining essential features.

The proposed solution, Asyn-DYNA, resolves synchronization bottlenecks in peer-to-peer energy management tasks, enabling asynchronous coordination while ensuring data privacy through distributed coordination. Asyn-DYNA converges to optimal solutions with a non-asymptotic linear convergence rate, making it suitable for dynamic and decentralized energy management tasks. The algorithm's effectiveness is validated through numerical experiments over P2P transactive networks.

A Digital Twin is used to assess real-time security in distribution networks with high penetration of distributed energy resources (DERs), identifying challenges such as voltage regulation and thermal loading limits. The Digital Twin integrates network topology and smart meter measurements to perform security assessments and determine corrective actions, effectively mitigating operational limit violations through coordinated power control. This approach enhances system security and operational efficiency, highlighting the potential of DT-based solutions for future distribution networks.

A novel hierarchical reinforcement learning framework with runtime safety shielding is proposed for power grid operation, decoupling long-horizon decision-making from real-time feasibility enforcement to ensure safe and generalizable control actions. The framework achieves better performance than flat reinforcement learning policies and safety-only methods under stress tests and unseen grid conditions. It provides a practical approach toward deployable learning-based controllers for real-world energy systems.

Stochastic electricity auctions address the challenge of communicating uncertainty about renewable energy production by conditioning contracts on both time/location and world state (e.g., weather conditions). This approach is based on equilibrium under uncertainty from microeconomic theory, requiring precise definitions of the world state. The concept is illustrated using a case study of offshore wind farms in the European North Sea.

The paper compares Fitted Dynamic Programming (DP) and Reinforcement Learning (RL) methods in finite-horizon dynamic pricing problems across increasing structural complexity. It evaluates their performance in environments with heterogeneous demand, inter-temporal revenue constraints, and multiple product types, highlighting trade-offs between explicit optimization and trajectory-based learning. The analysis focuses on revenue performance, stability, constraint satisfaction, and computational scaling.

A new framework called Physics-Informed Hybrid CNN-BiLSTM outperforms complex Transformer-based architectures in solar irradiance forecasting by integrating domain knowledge and explicitly incorporating physical constraints, achieving a Root Mean Square Error (RMSE) of 19.53 W/m^2 compared to 30.64 W/m^2 for complex attention-based baselines. This approach challenges the traditional "complexity-first" paradigm and offers a more efficient and accurate alternative in high-noise meteorological tasks. The model achieves this using a lightweight, Bayesian-optimized architecture that incorporates engineered features such as Clear-Sky indices and Solar Zenith Angle.