Robust Power System State Estimation using Physics-Informed Neural Networks

Solon Falas, Markos Asprou, Charalambos Konstantinou, , Maria K. Michael · 2025

Read Robust Power System State Estimation using Physics-Informed Neural Networks by Solon Falas, Markos Asprou, Charalambos Konstantinou, , Maria K. Michael (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-05874.txt.

Introduction

Falas and co-authors use physics-informed neural networks to estimate power grid state with more grit under faults and attacks. Grid operators need fast, exact state reads to run the system safely. Their hybrid model bakes physical laws into the network's design. Tests covered normal runs, faulty readings, and data tampering attacks. Against plain learning models, the physics-guided one won clearly. It scored up to 83% higher on held-out slices of its training data.

On wholly new datasets it still led by 65%. Under an attack on a critical bus, it kept its footing better than rivals. The work argues physics is a shield, not just a prior.

Catalog record

CREATORSolon Falas, Markos Asprou, Charalambos Konstantinou, , Maria K. Michael
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-05874

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Robust Power System State Estimation using Physics-Informed Neural Networks Solon Falas, Markos Asprou, Charalambos Konstantinou, , Maria K. Michael Abstract Modern power systems face significant challenges in state estimation and real-time monitoring, particularly regarding response speed and accuracy under faulty conditions or cyber-attacks. This paper proposes a hybrid approach using physics-informed neural networks (PINNs) to enhance the accuracy and robustness, of power system state estimation. By embedding physical laws into the neural network architecture, PINNs improve estimation accuracy for transmission grid applications under both normal and faulty conditions, while also showing potential in addressing security concerns such as data manipulation attacks. Experimental results show that the proposed approach outperforms traditional machine learning models, achieving up to ∼ \sim 83% higher accuracy on unseen subsets of the training dataset and ∼ \sim 65% better performance on entirely new, unrelated datasets. Experiments also show that during a data manipulation attack against a critical bus in a system, the PINN can be up to ∼ \sim 93% more accurate than an equivalent neural network. † † aftertitle: I Introduction The escalating global electricity demand, driven by rapid urbanization, transportation electrification, and digital technology proliferation, has underscored the need for robust and stable power systems. Ensuring the stability and security of critical infrastructures, particularly power transmission networks, is essential for economic stability and public safety. However, the growing complexity of modern grids, driven by renewable energy integration, adoption of smart grid technologies, and interconnected networks, presents significant challenges in monitoring, control, and system resilience [ 1 , 2 ] . In particular, addressing challenges related to real-time data management and stability has become increasingly critical, necessitating advanced monitoring schemes to ensure system stability and integrity. It is therefore important to have advanced monitoring schemes to ensure the system stability and integrity. Traditional state estimation methods, such as Weighted Least Squares (WLS), rely on Supervisory Control and Data Acquisition (SCADA) systems and iterative processes, but they are limited by insufficient real-time capabilities, high computational demands, and reliance on sparse data [ 3 , 4 ] . These limitations are especially problematic with intermittent renewable energy sources like solar and wind, which cause rapid fluctuations and transient errors. Phasor Measurement Units (PMUs) enhance real-time monitoring with high-precision synchronized data, but conventional methods still fall short in addressing the complexity of modern systems [ 5 ] . Compounding these technical challenges, power grids have simultaneously become prime targets for cyber-attacks, such as Denial of Service (DoS) and False Data Injection (FDI) attacks [ 6 , 7 ] . Such sensor manipulation attacks can compromise the integrity of grid monitoring, leading to potential outages, and severe physical damage [ 8 , 9 ] . Therefore, a more cyber-physical approach to state estimation is imperative to enhance the resilience of power grids against such threats. To address these limitations, machine learning (ML) techniques, including neural networks and support vector machines, have been explored to improve state estimation by leveraging statistical patterns and historical data [ 10 ] . While these data-driven approaches enhance accuracy and efficiency, they often neglect the physical constraints inherent to power systems [ 11 , 12 ] , limiting their reliability under varying operational conditions or sparse data [ 13 ] . PINNs [ 14 ] offer a transformative approach by embedding physical laws, constraints, and boundary conditions directly within the ML framework. By integrating principles such as conservation laws and system constants, PINNs align predictions with the physical realities of power systems [ 15 ] . This not only improves model accuracy but also reduces data requirements, both in quantity and quality, compared to purely data-driven methods. PINNs’ attributes make them particularly suited for applications requiring robustness, accurate fault diagnosis, and rapid responses during disruptions [ 16 , 17 ] . This work presents a hybrid framework for power system state estimation that leverages the capabilities of PINNs. The proposed approach seeks to improve both the accuracy and robustness of state estimation, particularly in the presence of disturbances such as cyberattacks and operational anomalies. The primary contributions of this work are: 1. Proposing a hybrid model that incorporates physical laws into neural networks to improve the accuracy of state estimation in power systems. 2. Conducting a comprehensive benchmark against state-of-the-art data-driven neural networks through an extensive set of experimental evaluations. 3. Demonstrating significant improvements in accuracy (up to \sim 93\% ) and robustness under various challenging conditions, including three-phase faults and data manipulation attacks. The proposed PINN-based approach highlights the potential for enhancing the stability, integrity, and security of power grids by utilizing advanced ML models that integrate physical principles to enable accurate and timely state estimation. The paper is organized as follows: Section reviews existing research on physics-enhanced neural networks for state estimation. Section II explains the methodology and training process of the proposed model. Section presents experimental results, and Section III summarizes the study’s key insights and future directions. II Methodology II-A Background The foundational concept behind PINNs [ 14 ] is to incorporate physical laws directly into the training process of neural networks by embedding these laws as constraints in the loss function. In general, PINNs can approximate a target function f by training a neural network \hat{f}(\mathbf{z};\theta) , parameterized by weights \theta , where \mathbf{z} represents input variables. Instead of relying purely on data, PINNs include physics-based constraints that govern the system’s behavior, ensuring that the model output conforms to known physical laws. The PINN loss function combines a data-driven term, L_{\text{data}} , which minimizes prediction error on observed data, and a physics-informed term, L_{\text{physics}} , which enforces consistency with the underlying physical laws: L_{\text{data}}=\sum_{i=1}^{N}\|\hat{f}(\mathbf{z}_{i};\theta)-f_{i}\|^{2}, (1) where \{(\mathbf{z}_{i},f_{i})\}_{i=1}^{N} represents observed data points, and L_{\text{physics}}=\sum_{j=1}^{M}\|G(\hat{f}(\mathbf{z}_{j};\theta),\mathbf{z}_{j})\|^{2}. (2) Here, G(\hat{f},\mathbf{z})=0 denotes a general physical law or system constraint that should hold at each point \mathbf{z}_{j} in the domain, represented by \{\mathbf{z}_{j}\}_{j=1}^{M} . The combined loss is expressed as: L=L_{\text{data}}+\lambda L_{\text{physics}}, (3) where \lambda is a weighting factor that balances the data-driven and physics-informed components of the loss. During training, minimizing L guides the model \hat{f} to align with observed data while satisfying the physical constraints, allowing PINNs to model complex behaviors accurately and consistently with the underlying physical system. This approach enhances model generalization and reduces dependence on extensive data by leveraging embedded physics, making PINNs particularly effective for applications in scientific and engineering domains. Fig. 1: Model training procedure for the power system state estimation use case. II-B Enhancing Loss Functions with Physical Constraints In this study, we build upon the foundational principles of PINNs by systematically introducing additional terms to the loss function that impose physics-based constraints relevant to power system state estimation. These terms are derived from established relationships and laws governing the system, such as power flow equations, Kirchhoff’s laws, and the admittance matrix. By embedding these constraints directly into the loss function, we ensure that the neural network’s predictions align with the physical laws governing power systems. Consequently, during the training process, the network aims to minimize not only the discrepancy between predicted and actual values but also the violation of these physics-based constraints. This approach enhances the model’s accuracy and reliability, making it particularly suitable for the complexities inherent in power system state estimation. In the proposed

Read the complete text