Energy Digest

The Input-Convex Encoder-only Transformer (IC-EoT) is a novel architecture that combines parallel processing with guaranteed tractability of input convexity, addressing limitations of recurrent models in learning-based Model Predictive Control. The IC-EoT is structurally immune to gradient instability and substantially reduces MPC solver times, being 2.7-8.3 times faster than its recurrent counterpart across horizons from one to eight hours. This leap in computational efficiency enables effective real-time MPC for building energy management.

BOOST-RPF, a novel method for power flow analysis, reformulates voltage prediction into a sequential path-based learning problem using gradient-boosted decision trees. It achieves state-of-the-art results with its Parent Residual variant, outperforming baselines in accuracy and generalization tasks, while maintaining high precision across unseen feeders. The framework displays linear $O(N)$ computational scaling and increased sample efficiency through per-edge supervision.

A novel power management framework has been proposed to manage hybrid energy storage systems in low-carbon industrial microgrids, eliminating traditional state of charge constraints to improve dispatch flexibility. The framework uses a full-timescale hierarchical model predictive control architecture with an adaptive feedback mechanism based on micro trajectory inverse projection (MTIP) to optimize energy release and cycle efficiency. The approach has been validated through experiments using real-world data from 14 months, achieving significant improvements in load smoothing rate and cycle efficiency.

The article introduces IF-CPS, a modular framework for diagnosing and attributing failures in cyber-physical systems (CPS), which improves upon existing methods by considering CPS-specific properties like closed-loop dynamics and safety constraints. It proposes three variants of influence functions tailored to CPS: safety, trajectory, and propagated influence, which outperform standard methods in most evaluation settings. The framework demonstrates promising results in diagnosis, curation, and safety attribution tasks on six benchmarks.

Power flow equations define a smooth bijection between nodal voltage phasors and active/reactive power injections within a feasible region meeting practical stability requirements. A data-based evaluation method using differential geometry and analytic functions can imply the associated power flow manifold from limited data points around a single operating point. The proposed method reduces computational complexity, allowing for efficient evaluation of the entire power flow manifold with just a few local measurements.

A novel heuristic approach combines optimal line switching and substation reconfiguration with congestion management through bus splitting to minimize dispatch costs, demonstrating improved economic gains over existing methods through simulations on nine IEEE test systems. The approach establishes a unified sensitivity framework for both line switching and bus splitting, enabling more effective transmission system operation. Incorporating bus splitting achieves greater cost savings than line switching alone in these simulations.

Grid-forming voltage source converters play a crucial role in power systems with large amounts of converter-based generation and are critical to transient stability. Three active-power control strategies were investigated to enhance transient stability: a wide-area control strategy, a local transient damping method, and a novel local control strategy that achieved the best performance without requiring communication infrastructure. These strategies improved critical clearing time in short-circuit simulations on a two-area test system with 100% GFM-VSC generators.

The article proposes a hybrid physics-informed neural network (PINN) combined with finite control set MPC for controlling DC-DC boost converters. This approach embeds physical laws into neural training to provide accurate state predictions while ensuring constraint satisfaction and multi-objective optimization. The method offers improved transient response, reduced voltage ripple, and robust operation across conduction modes in experimental results on a commercial boost module.

A proposed portfolio-level optimization framework helps hydrogen-centric companies manage electricity, hydrogen, and green certificate markets, co-optimizing asset scheduling and market decisions across multiple sites to reduce costs and increase hydrogen production. The model supports participation in various markets and is applied to three operational scenarios to evaluate its economic and operational impacts. Centralized control can unlock up to 2.42-fold increase in hydrogen production and 9.4% reduction in daily operational costs while satisfying all company policy constraints.

The proposed End-to-End Differentiable Predictive Control (E2E-DPC) framework addresses the limitations of conventional DPC by utilizing an Encoder-Only Transformer to model complex system dynamics, jointly training the model and control policy with a performance-oriented loss, and providing theoretical guarantees for recursive feasibility and constraint satisfaction. The framework achieves near-perfect constraint satisfaction while minimizing electricity expenditure in high-fidelity EnergyPlus simulations. This work establishes a deployable, performance-driven control solution for building energy management.