The increasing integration of distributed energy resources (DERs) poses significant challenges for real-time and scalable solutions to the optimal power flow (OPF) problem. Traditional OPF solvers, while mathematically rigorous, often suffer from heavy computational burdens when faced with high-dimensional distribution networks and frequent topology reconfigurations. Recent advances in graph neural networks (GNNs) offer data-driven power flow surrogate models that leverage the underlying graph structure of power systems. However, existing GNN-based approaches face two critical limitations: (i) a lack of convexity guarantees, which limits the integration of GNN-based power flow models and the downstream optimization tasks; and (ii) insufficient generalization to switching configurations not represented during training, which restricts their practical deployment in distribution networks with frequent topology changes. To address these challenges, this paper proposes a novel ConvexGraphSAGE model that integrates convex power flow mapping into the inductive graph neural network architecture, thereby forming a power flow surrogate that guarantees a convex mapping and merits excellent generalization ability. To balance training effort with generalization of the ConvexGraphSAGE model, a composite clustering metric is proposed to systematically select representative switching configurations for training, thereby improving the model's zero-shot generalization to unseen switching configurations. To further enhance the feasibility of OPF solutions, a risk-sample-mining-based iterative data enhancement strategy is designed to expand training datasets to effectively support subsequent retraining. Numerical experiments demonstrate that the proposed approach significantly reduces computational effort while guaranteeing competitive accuracy across seen and unseen switching configurations of the test feeder.
The increasing participation of residential load aggregators (LAs) in power system ancillary services presents new challenges due to demand response uncertainties and the massive scale of small-capacity LAs. These factors complicate the coordination of LAs by virtual power plants (VPPs) during secondary frequency regulation, often leading to prohibitive computational delays in real-time power distribution. To address these issues, a two-stage equal incremental rate principle for power distribution algorithm (TEIRP-PDA) is proposed in this paper. First, a power distribution model based on linear Distflow is constructed to minimize the VPP's operational costs while accounting for response uncertainties. Then, by decomposing the equal incremental rate principle into offline preparation and real-time execution stages, TEIRP-PDA analytically precomputes structural transitions. This decomposition significantly reduces the real-time computational complexity, enabling the power distribution problem to be solved with ultra-low latency. Numerical experiments demonstrate that, compared to the interior point method implemented by commercial solvers, TEIRP-PDA accelerates real-time dispatch speeds by about two orders of magnitude without sacrificing optimality, thereby effectively enhancing power system frequency stability.
Large-scale integration of renewable energy resources has increased operational volatility and transmission congestion, creating a need for reduced-order models that preserve essential network characteristics while remaining computationally tractable. This paper proposes a three-stage methodology for constructing an AC reduced-order zonal power system model. A new bus feature derived from the Zbus matrix is introduced to quantify voltage sensitivity associated with highly volatile resources, and a modified torque clustering algorithm is applied to identify zones without prior specification of the number of clusters. A hybrid aggregation–elimination framework is then developed to preserve inter-zonal tie-lines while reducing intra zonal topologies through zonal Kron reduction. A new equivalent generator scheme is further introduced at boundary buses to represent the aggregated reactive capability of internal resources, ensuring consistent voltage profiles under varying operating conditions. Theoretical aspects of the applied methodology are analytically developed. Comprehensive numerical studies conducted on the IEEE 39-bus, the real-world Nordic 74-bus, and the large-scale Texas 2000-bus power systems demonstrate that the proposed reduced-order models achieve high fidelity relative to their full-order counterparts while significantly reducing computational complexity.
Transient stability preventive control (TSPC) requires repeated time-domain simulations (TDS) within the optimization loop, which makes numerical approaches computationally prohibitive. To address this challenge, a gradient-aware surrogate model (GASM) is introduced to replace TDS in the optimization loop. This enables fast and accurate evaluations of the transient stability index (TSI) and its gradient information, thereby accelerating stability-constrained optimization. An active-learning refinement (ALR) mechanism is further incorporated to continuously assess surrogate uncertainty and to trigger on-demand numerical corrections when informative, out-of-distribution (OOD) operating points (OPs) are detected. This adaptation mechanism prevents gradient drift and ensures the effectiveness of the optimization process without requiring extensive retraining. Case studies on the IEEE 39- bus System and a 2,000- bus provincial scale power grid demonstrate that the proposed method accelerates TSPC by more than one order of magnitude compared with full simulation-based optimization, while maintaining the feasibility and accuracy of the resulting control actions.
This letter proposes a physics-informed Transformer (PIformer) for day-ahead load forecasting (DALF) on holidays. The proposed approach embeds sector-guided physical priors into attention scores, encoding historical loads into feature maps associated with industrial production and ambient temperature. These feature maps are then decoded into industrial and temperature-sensitive load components under physical constraints, which are aggregated to produce DALF outputs. By incorporating the dominant drivers, i.e., the industrial operating rhythm and temperature, PIformer improves holiday forecasting accuracy while providing physical interpretability. Experiments on real-world load data from China validate its effectiveness.
