Energy Digest

A staged physics-informed neural network (PINN) is proposed for estimating the parameters of a DC buck converter from time-domain data, with the five-stage schedule recovering inductance and capacitance to within 5% and series ESR to 10%. A three-stage variant of the PINN achieves the same accuracy 11% faster. The PINN remains accurate for simultaneous perturbations of all four components, achieving a waveform RMSE of 2.6 x 10^-2 over 5 ms.

Historical blackout events show that interconnected power systems can split into multiple areas with distinct dynamic behavior under extreme events. Conventional two-cluster representations are insufficient under these events, as multiple unstable modes can excite and lead to multiple coherent areas. Simulations on three test systems demonstrate the limitations of these representations.

A new framework for calibrating dynamic models of AI data center loads is proposed, using an identifiability-guided calibration method to improve accuracy, particularly for cooling-component recovery, with an aggregate active-power R-squared of 0.948 in evaluation. The method involves a grey-box estimator trained with branch-level supervision, yielding improved results compared to an LSTM baseline. The cooling composition and protection behavior can affect retained-load states and voltage responses near protection thresholds.

The article develops a framework to quantify operational flexibility in electric distribution systems under uncertainty, considering both active and reactive loads. It introduces a method to measure flexibility as the largest scaling of a prescribed uncertainty set that can be accommodated while satisfying network and actuator limits. The framework is applied to a case study on the Iowa-240 feeder, demonstrating the impact of control authority and communication structure on achievable flexibility margin.

Topology-Adaptive Certified Quantile Estimation (TACQE) is proposed to estimate the maximum voltage unbalance factor (VUF) in distribution networks with finite-sample guarantee. The method updates cross-couplings among power-flow sequences following a local network change, certifying candidate power-flow solutions against updated nonlinear equations. TACQE provides rigorous bounds on the maximum VUF and a uniform sampling bound, allowing for simultaneous upper-tail probabilities and quantiles.

A severity-aware minimum security set framework is proposed to reduce the computational burden of preventive AC security-constrained optimal power flow (PSCOPF) using machine learning, focusing on contingencies and constraints affecting preventive control decisions. The framework identifies a compact minimum security set containing non-dominated contingencies and associated post-contingency constraints that impact the preventive operating point. The proposed framework reduces problem size and computation time while controlling residual post-contingency severity.

Data centers are fundamentally reshaping the magnitude, spatial concentration, and electrical characteristics of power demand due to the proliferation of cloud computing and artificial intelligence workloads. Reliable and sustainable grid integration requires a comprehensive assessment of hourly and locational supply, network deliverability, marginal emissions, and other factors beyond peak demand and annual energy matching. Current methods are insufficient, and a more holistic approach that considers various aspects of data center power management is needed.

A pretrained tabular foundation model (TFM) was found to outperform or match task-specific neural models on three power grid steady-state analysis tasks: power flow prediction, N-k contingency screening, and voltage stability margin prediction. The TFM achieved strong results on 118- and 2000-bus systems, including accurate predictions for voltage magnitudes and angles, and high accuracy for N-1 and N-2 contingency screening. The model's strong performance was achieved without retraining or fine-tuning, indicating its potential as a general-purpose neural solver for power grid analysis.