Energy Digest

Developed a method to manage energy resources during normal grid-connected operation to preserve islanded readiness, using a robust viability kernel to identify sufficient battery energy, fuel, and load-shifting capability to sustain critical loads. The method successfully sustained critical loads during 14-day islanding periods at 0.999 availability, and enforcing the resource-dependent requirement preserved readiness at a modest increase in operating cost. The framework helps commanders understand how readiness changes with operating conditions.

Compact LC-filter design and inner-loop control must be coordinated for stable grid-forming operation in virtual-admittance-based inverter control, with a nominal resonance-placement limit near one-third of the sampling frequency and gain-dependent contraction of the positive-resistance region. A 3-kVA prototype demonstrates an 83.3% LC-product reduction, while output-impedance analysis and grid-connected tests identify additional grid-dependent gain constraints. Delayed inverter-current feedback can oppose VFD-assisted damping above one-sixth of the sampling frequency.

A dual-agent reinforcement learning framework is developed for electric-hydrogen-transport systems, combining vehicle scheduling and multi-energy dispatch. The framework achieves the lowest vehicle-service delay and energy cost among 11 evaluated methods, demonstrating effective coordination between these two processes. The dual-agent framework also reveals that updates to the upstream scheduling policy can shift the demand distribution experienced by the downstream dispatcher.

A novel model is presented for large-scale Hydrogen Electrolyzers, specifically Proton Exchange Membrane-based systems, which can provide grid ancillary services and enhance power system stability. The model, called Generalized ESS Model (GEM), is extended to PEM HEs for the first time and includes a method for scaling matrices across different capacities. The model is validated through comprehensive simulations on the WSCC system, demonstrating its accuracy and effectiveness.

Distributionally robust optimization (DRO) can hedge against errors in the distribution estimated from a finite record by using a Kullback-Leibler (KL) ball centered at a learned nominal density model. This approach allows the adversary to up-weight costly unobserved realizations, but can accommodate only a few scenarios. Centering the ball at the learned nominal reduces cost under distribution shift, particularly when the record is scarce and the model has captured the underlying structure.

Modeling choices for wind and solar power have a limited impact on predicting variable renewable energy (VRE) droughts and the overall design of a net-zero European power system. These choices can alter the amount of capacity needed to withstand system stress, but do not significantly change the frequency or severity of droughts. The actual effect of these choices on the system is comparable to the spread of electricity delivered during the most critical weather years.

This work develops a physics-informed multimodal deep learning framework for 10- to 110-minute-ahead probabilistic photovoltaic power forecasting, which outperforms 14 baselines with a forecast skill score of 0.2950, achieving significant improvements over the next-best model. The proposed model incorporates multiple branches, including plant-level temporal history and inverter-level spatial cloud dynamics, to provide accurate forecasts in challenging conditions. The model retains substantial skill in overcast and variable sky conditions, with FSS values of 0.198 and 0.255.

GradientSHAP is introduced as a method to provide interpretability for multivariate, multistep wind power forecasting, offering local and global explanations of relationships between environmental factors, site-specific conditions, and temporal dynamics. This approach can help address the opacity of AI models in the energy sector, supporting more informed decision-making in day-ahead grid management. GradientSHAP's complexity poses challenges for practical implementation, and future research should focus on streamlining these interpretability methods for routine deployment.