model power systems

This study addresses the degradation of accuracy observed when physics-informed surrogate models—despite performing well in isolated evaluations—are integrated into dynamic power system simulators, particularly under high-stress operating conditions. Framing this challenge as a verification and validation (V&V) problem, the work proposes a theoretical framework that combines algebraic coupling sensitivity, dynamic error amplification mechanisms, and finite-time simulation horizons to bound prediction errors. The approach integrates model-based verification with data-driven conformal calibration. Using a physics-informed neural network as a surrogate for synchronous machine dynamics, the authors conduct residual analysis and generate conformal predictions within a differential-algebraic equation simulation framework. Their findings reveal that small residuals in governing equations do not necessarily guarantee small trajectory errors, thereby underscoring the critical need—and a new paradigm—for validating surrogate models in coupled dynamical environments.

Under energy transition, high renewable penetration and cross-border interconnections intensify grid uncertainty, rendering conventional power flow solvers inadequate for real-time operation, while purely data-driven models lack physical consistency. This paper proposes a physics-informed neural network (PINN) framework for power flow simulation, integrating Kirchhoff’s laws and other physical priors via hybrid modeling—combining MLP/GNN architectures with physics-constrained regularization and unsupervised loss. We introduce a novel four-dimensional evaluation framework (accuracy, physical consistency, industrial deployability, out-of-distribution generalization) and the LIPS benchmark platform. Systematic ablation studies demonstrate that explicit graph-structured modeling and direct optimization of physical equations are critical to reliability. The resulting model achieves high accuracy, strong physical consistency, robust out-of-distribution generalization, and practical industrial deployment potential. Code is fully open-sourced.

Interdisciplinary modeling and large-scale simulation of smart grids face challenges including model heterogeneity, complex inter-domain couplings, and poor computational scalability. To address these, this paper proposes a systematic, multi-domain co-modeling framework that integrates power system dynamics, energy market mechanisms, and demand-side response behaviors into a unified backbone model. It introduces an innovative distributed subsystem optimization architecture to support flexible and scalable prosumer-coordinated scheduling. By unifying system-level modeling, distributed optimization, and cross-domain simulation techniques, the framework enables integrated modeling and co-simulation of heterogeneous, multi-source resources. The developed simulation tool efficiently validates diverse grid evolution scenarios, enabling system-level hypothesis testing at human timescales. Empirical evaluation demonstrates a 37% improvement in modeling efficiency and a 22% reduction in error for critical scenarios.

Power system domain experts often lack programming proficiency, hindering efficient modeling and analysis. Method: This paper proposes a symbolic, model-driven solution framework built upon an open model architecture. It enables intuitive definition of system components (e.g., AVRs, LTCs) via mathematical expressions—including complex-valued variables—without requiring code implementation. The framework integrates continuous power flow computation and equality-constrained Gauss–Newton state estimation, and ensures compatibility with standard datasets through a MATPOWER data converter. Contribution/Results: Compared to conventional tools, the framework substantially lowers the entry barrier for non-programmers, enhances modeling flexibility and accessibility, and supports rapid prototyping. Its design facilitates seamless adoption in power system education, research, and engineering practice, while preserving numerical rigor and interoperability with established simulation ecosystems.

This work addresses the lack of standardized, publicly available fault waveform datasets in high-voltage protection research, which hinders reproducible evaluation of data-driven methods. To bridge this gap, the authors introduce PROTECT-90, a benchmark dataset generated through electromagnetic transient (EMT) simulations on a standard 90 kV double-circuit transmission line topology. By systematically randomizing grid operating points, line parameters, and fault conditions, they produce 9,022 physically consistent and scenariowise balanced short-circuit simulation samples. Three-phase voltage and current waveforms, along with structured metadata, are synchronously recorded at eight measurement points. PROTECT-90 represents the first openly shared high-voltage fault dataset, with fully documented modeling assumptions and generation procedures, thereby establishing a standardized and reproducible benchmark for evaluating signal processing techniques and learning-based protection algorithms.

This study addresses the limited engineering applicability of machine learning surrogate models in power system dynamic simulation due to their lack of physical interpretability. For the first time, neural tangent kernel (NTK) theory is introduced into this domain, integrated with small-signal eigenvalue analysis to establish a formal connection between system physical stiffness and neural network optimization stiffness. This linkage reveals the evolution mechanism of error modes during training, enabling the development of an adaptive loss weighting strategy. The proposed approach not only provides modal-level physical explanations for performance differences across network architectures—such as ActNet—but also significantly enhances training convergence and reliability. Consequently, this work lays a theoretical foundation for designing interpretable, structure-aware surrogate models tailored to power system dynamics.

This study addresses the operational complexities and lack of standardized benchmarks in power systems arising from renewable energy integration by systematically reviewing Graph Machine Learning (GML) applications in forecasting and control. Synthesizing 800 publications, we establish a catalog of ML-ready benchmark requirements and advocate for open data dissemination. The review elucidates GML’s advantages in efficiency and generalization derived from topological inductive biases while identifying critical bottlenecks in deployment and interpretability. Furthermore, standardized benchmark specifications are formulated to bridge existing gaps. This work provides a pivotal framework for enhancing research reproducibility and guiding future data-driven modeling in modern power grids, effectively addressing the current absence of unified evaluation standards in this rapidly evolving domain.

This work addresses the limitations of existing synthetic power grid scenarios, which often lack AC power flow feasibility and operational robustness, thereby hindering their utility in real-world system analysis. The authors propose a feasibility-aware hierarchical diffusion generative framework that embeds AC power flow equations and operational constraints directly into the generation process, enabling joint modeling of network topology, branch parameters, and time-varying load within a physically feasible distribution. By integrating domain-specific engineering knowledge through a three-stage architecture, the method significantly enhances power flow convergence, contingency robustness, and statistical fidelity of the generated scenarios—without requiring post-hoc optimization—thus achieving efficient, high-dimensional, and physically consistent grid scenario synthesis.

This study addresses the limitations of conventional computing reliant on specialized hardware by introducing, for the first time, a “power-grid-native computing” paradigm that leverages existing electrical grids as fixed physical computational operators. Methodologically, computation is executed through voltage perturbations and current responses, integrating time-domain simulation, Kirchhoff’s laws, power electronic interfaces, and surrogate models to construct a comprehensive encoding–decoding framework. This approach substantially reduces trainable parameters while supporting spatial concurrency and temporal multiplexing. Experimental evaluations demonstrate classification accuracies of 91.5% and 82.25% on MNIST and Fashion-MNIST, respectively. These results validate the critical influence of grid topology and signal representation on computational utility, establishing a promising foundation for physics-based analog computing.

This study addresses the physical inconsistency in neural AC power flow solvers, where accurate voltage predictions coexist with large power residuals. We reveal that this discrepancy originates from output errors deviating from the dominant solution subspace. By analyzing the geometric structure of these errors via singular value decomposition, we propose a calibration subspace projection method that relies solely on training data to effectively suppress outlying error components. This work is the first to demonstrate that output error geometry is the critical factor determining physical consistency. Experiments across four mainstream neural foundation models show that the proposed method reduces average power balance residuals by up to 68.9% while simultaneously improving voltage magnitude accuracy.