Energy Digest

Oracle is a multi-objective reinforcement learning-based framework for optimizing analog circuit design that replaces scalar reward optimization with vector-valued learning and preference-aware conditioning. This approach enables a single trained model to generate designs across diverse trade-off settings without retraining, addressing limitations of existing methods. Oracle achieves significant performance improvements over state-of-the-art approaches, reducing runtime by 20.4x-104.4x and meeting 99.9% of target specifications.

Synchronized generating units have lower availability probabilities when needed compared to non-synchronized units, as they are already synchronized to the grid, whereas non-synchronized units must first start and synchronize. The difference in availability probabilities is quantified by proposed failure probabilities SynFORd and NonSynFORd, with non-synchronized units exhibiting higher and more dispersed failure rates. Case studies of generating-unit data from the New England system show a significant disparity between synchronized and non-synchronized unit failures.

Grid-forming inverters' limited overcurrent capability makes current limiting essential. An analytical framework predicts grid-voltage boundaries for CCL activation, determining whether the operating equilibrium persists as a saturated stable equilibrium point (satSEP) or is lost. The predictions show that CCL activation can cause immediate equilibrium loss or allow satSEPs to persist until later losses.

A new framework for load frequency control is proposed that integrates contingency detection with predictive control to improve resilience against power system dynamics disruptions. The approach models contingencies as stochastic discrete events and uses a disturbance-aware residual formulation to identify active modes and estimate unknown disturbances. This leads to more accurate contingency detection and improved closed-loop performance under multiple scenarios and unknown disturbances.

The article proposes a real-time security assessment of distribution grid security using data from smart meters to detect power quality challenges. An iterative algorithm is designed to adaptively identify a subset of smart meters that can certify all voltage magnitudes are within bounds. The algorithm consistently identifies nodes with high and low voltage magnitude in numerical tests on the IEEE 123 distribution feeder.

DiffAPQP is a solver-flexible framework that accelerates decision-focused learning for power systems by automating the solution and differentiation of large optimization problems during training. This approach achieves significant speedups over existing methods, with up to 3.91x increase in closed-loop training speed, while reducing peak memory usage by approximately 50%. DiffAPQP demonstrates its effectiveness on a realistic power network, such as the IEEE 118-bus system, and outperforms other backends like CvxpyLayers.

Short-term load forecasting plays a crucial role in safety-critical environments like the German transmission grid, where determinism, reproducibility, and auditability are essential engineering requirements. A 41-day live challenge evaluated a pipeline that implemented EU-AI Act Requirements, which achieved better forecast accuracy than the official ENTSO-E day-ahead forecast. The open-source Python library spotforecast2-safe demonstrates competitive performance with large pre-trained models while being transparent, low-cost, and auditable.