Energy Digest

A proposed Physics-Informed Sparse Machine Learning (PISML) framework is introduced to discover the unknown governing equations of grid-connected inverters from external measurements. The framework achieves a tractable mapping from black-box data to explicit control equations, reducing error by over 340 times compared to baselines and restoring analytical tractability for rigorous stability analysis. PISML also enables the compression of heavy neural networks into compact explicit forms, reducing computational complexity.

This paper proposes a projection-free power-limiting droop control for grid-connected power electronics, which results in semi-globally exponentially stable dynamics that coincide with projection-free primal-dual dynamics under certain conditions. The proposed control method is associated with a constrained flow problem and provides a bound on the convergence rate of the networked dynamics, which can be improved through tuning of controller parameters. The relationship between the convergence rate and connectivity of the network is also analyzed.

The scenario approach is a data-driven design framework that uses data for both design and certification without separate test datasets. A new framework introduces baseline and post-design appropriateness criteria, providing upper and lower bounds on the risk of failing to meet post-design appropriateness using distribution-free methods. This method can be used to infer comprehensive knowledge of performance indexes from available datasets.

A capacity-constrained incentive-based demand response approach for residential smart grids uses deep reinforcement learning to adjust hourly incentive rates based on wholesale electricity prices and aggregated residential load, aiming to reduce peak demand by up to 22.82%. The proposed framework considers both service provider and end-user financial interests, modeling heterogeneous user preferences through appliance-level home energy management systems. It achieves a smoother aggregated load profile and effective reduction in peak-to-average ratio.

A proposed framework for coordinating energy storage peak shaving and stacked services using non-parametric kernel regression models constructs state-of-charge trajectory bounds from historical data, while a second stage utilizes remaining capacity for energy arbitrage via transfer learning. The method achieves 1.3 times improvement in performance over the state-of-the-art forecast-based method, resulting in cost savings and effective peak management. It effectively reduces electricity costs and extends battery lifetime in commercial buildings without relying on predictions.

The MARLEM framework is an open-source multi-agent reinforcement learning simulation environment for studying implicit cooperation in decentralized local energy markets, featuring a modular market platform and physically constrained agent models. It enables agents to learn strategies that benefit the entire system through enhanced observations and rewards, facilitating emergent coordination and improving market efficiency. The framework has been applied to various case studies demonstrating its potential to strengthen grid stability.

The proposed methodology optimizes the placement and sizing of PV-based DG units in a distribution network by identifying candidate nodes based on their active power loading capacity, and then using the Monte Carlo method to determine optimal locations and sizes that minimize voltage deviation and reduce active power losses. The results show significant reductions in network active power losses (50.37%, 58.62%, and 65.16%) with improved voltage profiles for different numbers of DG units. This approach allows for larger DG capacities while maintaining better voltage profiles compared to existing studies.