⚡ electrical engineering

Physics-Informed Neural Networks for Accelerating Power System State Estimation

This paper proposes a physics-informed neural network (PINN) approach that significantly accelerates power system state estimation by integrating physical laws as prior knowledge, achieving improved accuracy, reduced result variability, and faster convergence compared to traditional iterative methods on the IEEE 14-bus system.

Original authors: Solon Falas, Markos Asprou, Charalambos Konstantinou, Maria K. Michael

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Solon Falas, Markos Asprou, Charalambos Konstantinou, Maria K. Michael

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Guessing the Weather vs. Knowing the Laws of Physics

Imagine you are trying to guess the weather in a city.

The Old Way (Traditional State Estimation):
Imagine you have a team of detectives trying to figure out the temperature, wind speed, and humidity. They only have a few old, broken thermometers scattered around the city (these are the "SCADA" sensors). To get a good guess, they have to run complex math simulations over and over again, adjusting their guesses until everything "sort of" fits. It's slow, it's computationally heavy, and if the data is messy, they might get the wrong answer.

The New Way (Machine Learning):
Now, imagine you train a super-smart AI student on thousands of past weather reports. This student can look at the few broken thermometers and instantly guess the weather. It's fast! But there's a catch: the student is just a "pattern matcher." If it sees a weird situation it hasn't studied before (like a sudden storm), it might guess that it's snowing in the desert because it saw that pattern once in its training data. It doesn't actually know how weather works; it just knows what usually happens.

The Paper's Solution (Physics-Informed Neural Networks - PINNs):
This paper introduces a "Super-Student" who is both a pattern matcher and a physics expert.

The Analogy: Imagine training a student to guess the weather, but you also give them a copy of the Laws of Thermodynamics and tell them, "You can guess based on the data, but your answer must obey these laws. If you guess it's 100°F but the humidity says it's impossible, you get a penalty."
The Result: The student learns faster, makes fewer silly mistakes, and doesn't need to memorize as many past weather reports to be accurate.

How It Works in the Power Grid

In a power grid, "State Estimation" is like taking a snapshot of the entire electrical system to see if it's healthy. We need to know the Voltage (pressure) and Angle (timing) at every single substation (bus).

The Problem: Traditional methods are like trying to solve a giant, tangled knot of wires by pulling on one end. It takes a long time. If the grid is huge, it takes forever.
The Flaw of Standard AI: If you just use a standard AI, it might learn the "shape" of the knot but forget the rules of how electricity actually flows. If the grid changes suddenly (like a generator shutting down), the AI might get confused.
The PINN Fix: The authors built a Neural Network that has the Laws of Electricity baked right into its brain.
- The "Data" Part: It looks at the measurements (Power flowing in/out).
- The "Physics" Part: It constantly checks itself against the rules of electricity (Kirchhoff's laws, Ohm's law). It asks, "Does my guess make sense physically?"
- The "Loss Function": Think of this as a teacher grading the student.
  - Grade 1: How close is your guess to the real sensor data?
  - Grade 2: Did you break the laws of physics?
  - The AI tries to get a perfect score on both.

The "Secret Sauce": The Training Recipe

The authors didn't just turn the physics on immediately. They used a clever training schedule:

Early Training: They let the AI look mostly at the data to learn the general shape of the problem (like learning the alphabet).
Later Training: They gradually turned up the volume on the "Physics Rules." This forced the AI to refine its answers to be physically perfect.

It's like teaching a child to drive: first, you let them steer the car in an empty parking lot (learning from data), then you slowly introduce traffic rules and physics (speed limits, friction) to make them a safe, expert driver.

What Did They Find? (The Results)

They tested this on a standard model of a power grid (the IEEE 14-bus system) under two conditions: normal operation and a sudden emergency (a generator shutting down).

Here is what happened compared to the "Plain AI":

More Accurate: The PINN was about 11% more accurate. It made fewer mistakes.
More Stable: The results were much more consistent. If you ran the test 5 times, the Plain AI gave 5 slightly different answers. The PINN gave almost the same answer every time. (Think of it as a dart thrower who hits the bullseye every time vs. one who hits the board but bounces around).
Faster Learning: The PINN reached its peak performance 30% faster. It didn't need to study as long to become an expert.

Why Does This Matter?

Power grids are getting smarter and faster. We are installing new sensors (PMUs) that send data hundreds of times a second.

The Old Way: The math is too slow to keep up with this flood of data.
The PINN Way: It's fast enough to handle real-time data, and because it understands the physics, it won't get tricked by noisy or missing data.

Summary

This paper proposes a smarter way to monitor power grids. Instead of just using a "black box" AI that guesses based on history, they built an AI that knows the rules of physics. This makes the system faster, more accurate, and more reliable, especially when the grid is under stress or changing quickly. It's the difference between a student who memorizes answers and a student who truly understands the subject.

1. Problem Statement

Power system state estimation (SE) is critical for grid monitoring and control, providing the real-time operating condition of the system. However, traditional SE methods face significant challenges:

Computational Intensity: Conventional techniques (e.g., Gauss-Newton, Weighted Least Squares) rely on iterative algorithms that are computationally expensive, especially for large-scale systems.
Data Limitations: Traditional methods often rely on sparse, error-prone SCADA data, leading to delays and reduced accuracy.
Machine Learning Drawbacks: While standard Machine Learning (ML) approaches offer speed, they often fail to capture complex physical dynamics and constraints. They rely heavily on large historical datasets and can produce physically implausible results when data is noisy or scarce.
Real-Time Requirements: The increasing integration of Phasor Measurement Units (PMUs) necessitates faster, real-time estimation to capture short-duration transients, which traditional iterative solvers struggle to meet efficiently.

2. Methodology

The authors propose a Physics-Informed Neural Network (PINN) framework that integrates domain-specific physical laws directly into the neural network's training process to overcome the limitations of pure data-driven models.

A. Core Architecture

Input: Active ( $P_i$ ) and Reactive ( $Q_i$ ) power injections at each bus.
Output: Voltage magnitude ( $V_{mag}$ ) and voltage angle ( $V_{ang}$ ) for each bus.
Network Structure: A feedforward neural network with 32 neurons in a single hidden layer, using the Hyperbolic Tangent (tanh) activation function and the Adam optimizer.

B. Physics-Informed Loss Function

The novelty lies in the augmentation of the loss function to enforce physical consistency. The total loss ( $L$ ) is a weighted sum of a data-driven term ( $u$ ) and a physics-driven term ( $f$ ):
$Loss = \lambda_1 \cdot u + \lambda_2 \cdot f$

Data-Driven Term ( $u$ ):
- Measures the Mean Squared Error (MSE) between the predicted voltage ( $V_{pred}$ ) and the ground truth voltage ( $V_{true}$ ).
- Ensures the model fits the observed data.
Physics-Driven Term ( $f$ ):
- Enforces the power system's physical laws, specifically the relationship between voltage and current injections defined by the admittance matrix ( $Y$ ).
- The network predicts voltages, which are then used to calculate predicted current injections ( $I_{pred} = Y \cdot V_{pred}$ ).
- This term measures the Mean Absolute Error (MAE) between $I_{pred}$ and the ground truth current injections ( $I_{true}$ ), derived from the input power data.
- Key Advantage: This allows the model to learn the system topology and constraints (Kirchhoff's laws) encoded in the admittance matrix without requiring explicit time-series modeling or massive datasets.
Training Strategy:
- Dynamic Weighting: The weights $\lambda_1$ and $\lambda_2$ are adjusted during training. The strategy starts with a focus on data fitting ( $\lambda_1$ high) and progressively increases the influence of physics constraints ( $\lambda_2$ high) to fine-tune the model.
- Normalization: Both loss components are normalized to the range $[0, 1]$ to prevent scaling issues.

3. Experimental Setup

Test System: IEEE 14-bus benchmark system.
Datasets: Two scenarios generated using PowerWorld software:
1. Steady-State: Varying load conditions (192 time instances).
2. Transient: A 20-second period involving a generator shutdown at Bus 2 and recovery (2000 time instances).
Validation: 5-fold cross-validation to assess stability and generalization.
Noise: Measurement noise was added to simulate real-world SCADA/PMU conditions.
Baseline: A standard "plain" Neural Network (NN) trained only on the data-driven loss ( $u$ ).

4. Key Results

The proposed PINN was compared against the benchmark plain NN across both datasets.

Accuracy Improvement:
- Steady-State: PINN achieved a 6.7% average improvement in accuracy.
- Transient (Generator Shutdown): PINN achieved a 4.6% average improvement.
- Note: The abstract mentions up to an 11% increase in accuracy, likely referring to specific optimal configurations or peak performance points.
Algorithmic Stability:
- The standard deviation of the validation error across the 5 folds was significantly reduced, indicating more consistent performance.
- Steady-State: Standard deviation reduced by 57.8%.
- Transient: Standard deviation reduced by 56.6%.
- This demonstrates that PINNs are less sensitive to variations in training data splits compared to plain NNs.
Convergence Speed:
- The PINN reached its peak performance (best validation error) significantly faster than the plain NN.
- Steady-State: Converged 18.1% faster (fewer epochs required).
- Transient: Converged 24.6% faster.
- This implies reduced computational effort and training time.

5. Significance and Contributions

Hybrid Approach: Successfully bridges the gap between data-driven ML and physics-based modeling, ensuring predictions are not only statistically accurate but also physically valid.
Data Efficiency: By leveraging physical laws as regularization, the method reduces reliance on massive datasets, making it suitable for scenarios with limited or noisy data.
Topology Independence: The approach does not require system-specific architectural designs based on topology; the physics constraints (via the admittance matrix) inherently encode the system structure.
Real-Time Viability: The reduction in training time and the ability to handle transient events make PINNs a viable solution for next-generation power grids with high PMU observability, where rapid state estimation is critical for stability.

Conclusion

The paper demonstrates that integrating physical laws into neural network training via PINNs significantly enhances the accuracy, stability, and convergence speed of power system state estimation. This approach offers a robust alternative to traditional iterative solvers and pure data-driven models, particularly in the era of high-speed PMU data. Future work will explore limited observability scenarios and fault detection applications.