Robust Power System State Estimation using Physics-Informed Neural Networks

1. Problem Statement

Modern power systems face increasing complexity due to renewable energy integration, smart grid technologies, and interconnected networks. This complexity creates significant challenges for State Estimation (SE), which is critical for real-time monitoring and grid stability.

• Limitations of Traditional Methods: Conventional techniques like Weighted Least Squares (WLS) rely on iterative processes and sparse SCADA data. They suffer from high computational demands, slow response times, and difficulty handling dynamic conditions or transient errors caused by intermittent renewables.
• Limitations of Pure Data-Driven ML: While Neural Networks (NNs) offer speed and handle non-linearities well, they often ignore fundamental physical laws. This makes them prone to overfitting, unreliable under sparse data, and vulnerable to cyber-attacks (e.g., False Data Injection) where manipulated sensor data leads to incorrect state estimates.
• The Gap: There is a need for a hybrid approach that combines the speed of machine learning with the reliability and constraint adherence of physical laws to ensure robustness against faults and cyber threats.

2. Methodology

The authors propose a Physics-Informed Neural Network (PINN) framework specifically designed for power system state estimation.

A. Core Concept

Unlike standard NNs that minimize only the error between predicted and observed data, PINNs embed physical laws directly into the loss function. The model is trained to satisfy both observed data points and the governing equations of the power system.

B. Loss Function Formulation

The total loss function ($Loss$) is a weighted sum of three components:
$$Loss = \lambda_d \cdot d + \lambda_p \cdot p + \lambda_c \cdot c$$
Where:

1. Data Term ($d$): Mean Squared Error (MSE) between the estimated complex voltages ($V_{est}$) and ground-truth voltages ($V_{true}$).
2. Physics Term ($p$): MSE enforcing Kirchhoff's Current Law and network topology. It calculates the difference between the actual injected currents ($I_{true} = Y \cdot V_{true}$) and the estimated currents ($I_{est} = Y \cdot V_{est}$), where $Y$ is the known admittance matrix. This forces the network to respect the system's topology.
3. Constants Term ($c$): MSE penalizing deviations from known system constants (e.g., voltage magnitudes at generator buses or the slack bus angle during steady state).

C. Input/Output Architecture

• Inputs: Active ($P$) and Reactive ($Q$) power injections at each bus.
• Outputs: Estimated complex voltages (magnitude and phase angle) for all buses.
• Mechanism: By learning the correlation between power injections and voltages while minimizing the physics-based error, the network implicitly learns the system topology without explicit architectural modifications.

D. Optimization Strategy

The authors employ a Systematic Hyperparameter Optimization approach using a Tree-Structured Parzen Estimator (TPE) meta-optimizer.

• They systematically search for optimal weights ($\lambda_d, \lambda_p, \lambda_c$) that balance the loss components.
• They optimize architectural parameters (layers, neurons, learning rate, batch size) to ensure the model generalizes well across different operating conditions.

3. Key Contributions

1. Hybrid Framework: Proposes a novel PINN architecture that integrates physical laws (power flow equations, admittance matrix constraints) directly into the neural network training process for SE.
2. Robustness to Anomalies: Demonstrates that embedding physical constraints significantly improves resilience against data manipulation attacks and system faults (e.g., three-phase faults, generator shutdowns) compared to purely data-driven models.
3. Systematic Optimization: Introduces a rigorous hyperparameter tuning procedure to determine the optimal balance between data fidelity and physical consistency, addressing a common limitation in existing PINN literature.
4. Scalability Validation: Validates the approach on two standard IEEE benchmark systems (14-bus and 118-bus), proving scalability from small to large grids.

4. Experimental Results

The model was tested on IEEE 14-bus and IEEE 118-bus systems under various scenarios, including steady-state, generator outages, transient faults, and cyber-attacks.

• Accuracy Improvements:
  • Unseen Data: PINN achieved ~83% higher accuracy on unseen subsets of the training data and ~65% better performance on entirely new, unrelated datasets compared to standard NNs.
  • Steady State: On the IEEE 14-bus system, PINN reduced Mean Absolute Error (MAE) by ~57.3% compared to an optimized NN. On the IEEE 118-bus system, the improvement reached ~82.91%.
• Robustness to Faults:
  • During generator shutdowns and load spikes, PINN maintained stable MAE, whereas standard NN performance deteriorated.
  • In 3-phase fault scenarios, PINN showed ~20.53% (14-bus) and ~37.67% (118-bus) better accuracy than NNs.
• Cyber-Attack Resilience:
  • Under data manipulation attacks (False Data Injection) targeting specific buses, the PINN acted as a filter, leveraging physical constraints to reject erroneous data.
  • Result: PINN was up to ~93% more accurate than an equivalent NN in estimating the system state during attacks on the 14-bus system and ~76% more accurate on the 118-bus system.
• Computational Efficiency:
  • Training times were reasonable (e.g., ~34s for 14-bus, ~40s for 118-bus).
  • Inference times were extremely fast (~0.035–0.040 ms), making the approach suitable for real-time applications.

5. Significance and Conclusion

This paper establishes that Physics-Informed Neural Networks are a superior alternative to both traditional iterative methods and pure data-driven machine learning for power system state estimation.

• Security: By anchoring predictions to physical laws, PINNs inherently detect and mitigate the impact of cyber-attacks that violate these laws, offering a critical layer of defense for grid security.
• Generalization: The approach requires less data to achieve high accuracy and generalizes better to unseen fault conditions and system topologies.
• Future Impact: The study suggests that integrating physical constraints is essential for the next generation of resilient, cyber-physical power systems. Future work may explore multi-PINN architectures for localized clustering and transfer learning to further enhance anomaly detection capabilities.