Energy Digest

Researchers have developed a machine learning approach to design observers for nonlinear systems that are robust against measurement noise and learning errors. By incorporating contraction conditions into the training loss function, they can determine both the correction term and the contraction metric without solving complex matrix partial differential inequalities. The proposed observers show exponential input-to-state stability and computable bounds on learning errors.

Quantification and Surrogate Model Estimation of Spatial-Temporal Carbon Intensity Factors involves a novel approach to calculate local carbon intensity factors on an urban district level, with results ranging from 0 to 316 g/kWh. A transferable surrogate model is developed using limited input parameters, achieving highly accurate estimations in densely built areas and supporting sustainable city operations. Neglecting local variations can lead to emission estimation errors of up to 9%.

A human-AI teaming framework is presented to enhance safety and robustness in autonomous voltage regulation of distribution networks. The framework combines a Soft Actor-Critic agent with an adaptive Lagrange constraint mechanism and a human-guidance module for sensitivity-based corrections, improving policy internalization of safe control behavior. This approach achieves lower voltage-violation severity and reduced power losses compared to baseline methods.

A human-on-the-loop (HOTL) resilient control architecture has been proposed to manage inverter-based resources under actuator degradation, addressing limitations of traditional fault-tolerant control and adaptive control strategies. The framework incorporates human supervisory judgment to detect subtle off-nominal behavior, adjust operational objectives, and maintain system operability. It achieves superior results compared to conventional methods, preserving control reserves and preventing actuator saturation in grid-connected inverter simulations.

Large-scale AI training turns computing facilities into periodic loads on the grid, but whether multiple independent jobs share the same power envelope or stay synchronized depends on how the power-management stack handles load-dependent throttling and shared cooling. The coupling between jobs is repulsive to leading order, becoming attractive only when there's a phase lag in control loop frequency; protection requires rate diversity. This synchronization can be influenced by phase-scattering scheduling.

Quantum-resistant distributed optimization is necessary for multi-region unit commitment due to vulnerable encrypted data flow against retrospective decryption. A customized Benders decomposition-based approach with global summation structure enables secure aggregation, incorporating additive masking, variable transformation hiding, and reveal-bound lattice-based zero-knowledge proofs. The proposed method achieves low suboptimality and lightweight computational overhead while recovering significant system cost via inter-regional reserve sharing.

Optimization models and algorithms are used to determine the optimal strategy for meeting electricity demand at minimum cost by committing power generation units at each point in time, solving a combinatorial problem with long solution time requirements. A proposed decomposition method using alternative models from the EGRET library achieves significant computational speed ups, outperforming benchmarking systems. The approach provides a solution to the challenging unit commitment problem.

A novel framework introduces human-in-the-loop distributed consensus control for demand response participation in multiple buildings, where a facility manager serves as the leader to balance energy use with occupant comfort. A nonlinear observer is proposed to address the challenge of the leader-follower system's lack of direct access to the decision-making process. The approach effectively achieves consensus and maintains system performance, validating its effectiveness in managing demand response events.

A new extension for the Ex-Fuzzy library has been developed to enable interpretable fuzzy regression with scalar consequents learned directly from data, introducing a target-aware partition initialisation strategy based on Fuzzy C-Means clustering. The proposed method achieves high predictive accuracy in regression tasks and produces compact rule bases of 10-15 human-readable rules. It offers a transparent and competitive alternative to black-box models, supporting practical interpretability with competitive predictive performance.