⚡ electrical engineering

Predicting Fault-Ride-Through Probability of Inverter-Dominated Power Grids using Machine Learning

This study demonstrates that machine learning models can accurately predict the fault-through probability of inverter-dominated power grids using a synthetic dataset and successfully generalize these predictions to the IEEE-96 test system, thereby offering a computationally efficient alternative to conventional dynamic simulations for risk assessment.

Original authors: Christian Nauck, Anna Büttner, Sebastian Liemann, Frank Hellmann, Michael Lindner

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Christian Nauck, Anna Büttner, Sebastian Liemann, Frank Hellmann, Michael Lindner

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

Imagine the power grid as a vast, complex dance floor where electricity flows between various partners. In the past, this dance was led by heavy, rotating machines (conventional power plants) that acted like a solid rhythm section, keeping everyone in step. But today, we are replacing these heavy machines with solar panels and wind turbines. These new partners are lighter and more flexible, but they lack the same heavy "inertia" to keep the dance stable if someone stumbles.

This article is about figuring out how likely it is that this new, inverter-heavy dance floor will stay on its feet if someone accidentally knocks over a chair (a "fault") and then quickly removes it.

Here is the breakdown of the research using simple analogies:

1. The Problem: Too Many Scenarios to Check

Imagine you are a dance instructor trying to predict whether a particular group of dancers will stay balanced if a chair is kicked out from under their feet.

The old way: You could simulate every single possible way a chair could be knocked over, how hard it is pushed, and where the dancers are standing. But with thousands of dancers and millions of possible scenarios, this takes so much time and computing power that you can only check a tiny handful. You might miss that one dangerous scenario that causes the whole group to fall.
The new idea: The authors wanted to use Machine Learning (ML) – imagine this as a super-smart student who watches thousands of simulations and learns the "rules of balance" without having to run every single simulation again.

2. The Solution: Training a "Super-Student"

To teach this student, the researchers had to create a huge library of practice scenarios.

Synthetic dance floors: Instead of examining just one real city power grid, they built 1,000 different "fake" power grids on the computer. These grids varied in size and structure, just like real cities.
The test: For each node (a connection point in the grid), they simulated a "fault" (like a sudden voltage drop) and then observed what happened. They calculated a "fault survival probability."
- Analogy: Imagine a score from 0 to 1. A score of 1 means the dancer always stays balanced, no matter how the chair is knocked over. A score of 0 means they always fall. A score of 0.5 means they stay balanced half the time.

3. The Secret Ingredient: The Map is Crucial

The researchers discovered something critical: You cannot just look at the dancer; you must look at the entire dance floor.

Local vs. Global: Some methods look only at the specific dancer who was pushed (local features). But the article shows whether a dancer falls often depends on who their neighbors are and how the floor is connected (topology).
The tool: They used a special type of AI, a Graph Neural Network (GNN).
- Analogy: If normal AI is like a teacher looking at students one by one, a GNN is like a teacher who can see the entire classroom, the desks, and how the students are sitting next to each other, all at once. This allowed the AI to understand that a "push" in one corner could trigger a chain reaction in another corner.

4. The Results: The Student Passed the Test

The researchers trained their AI with the 1,000 fake grids and then tested it on a completely different, real test grid (the IEEE-96 system) that the AI had never seen before.

The result: The AI was surprisingly good at predicting the stability scores.
The "generalization" success: Although the test grid was different from the training grids, the AI still recognized the rules of stability. This is like teaching a student on a practice gymnastics floor, and they successfully predict how a gymnast would perform on a completely different, real Olympic stage.
The winner: The AI that looked at the entire "map" (the GNN) performed better than the AI that looked only at individual points.

5. Why This Matters

The article concludes that we no longer need to run expensive, slow simulations for every single possible scenario.

The advantage: Once the AI is trained, it can predict the stability of a grid instantly. This helps grid operators identify "weak links" (components likely to fail) much faster.
The goal: This enables better planning for future power grids and ensures that the lights stay on even when something goes wrong as we transition to 100% renewable energy.

In short: The article shows that by training an intelligent AI on thousands of fake power grids, we can quickly and accurately predict how real power grids will handle emergencies, saving time and preventing blackouts.

Technical Summary: Predicting Fault Ride-Through Probability in Grid-Forming Inverter-Dominated Power Systems Using Machine Learning

Problem Statement

As power grids transition to 100% renewable energy sources (RES), they face significant stability challenges characterized by reduced inertia, decentralized generation, and volatile production. These systems rely heavily on power electronic inverters, necessitating a shift from traditional synchronous generation to grid-forming inverters. A critical challenge in operating and planning these future grids is the need for comprehensive risk assessments covering a wide range of fault scenarios.

While traditional dynamic simulations are accurate, they are computationally intensive. The costs of probabilistic stability analysis scale at least quadratically with the number of nodes, making it impossible to analyze all relevant configurations for large-scale real-world grids. Consequently, grid operators often rely on conservative assumptions or analyze only a limited set of fault scenarios, potentially overlooking critical failure modes. This contribution addresses the need for efficient methods to predict the fault ride-through probability ( $p_{frt}$ ) – the probability that a grid component remains within defined ride-through characteristics after fault clearance – without incurring the prohibitive computational costs of exhaustive dynamic simulations.

Methodology

1. Dataset Generation

To overcome the lack of training data for dynamic stability, the authors generated a new, large-scale dataset of synthetic power grids:

Grid Topology: 1,000 synthetic power grids were generated using a random growth algorithm, with sizes ranging from 70 to 80 nodes. This size was chosen to capture topological effects while remaining computationally manageable.
Component Modeling: The grids model 100% RES integration.
- Grid-Forming Inverters: Modeled using the "Normal Form" (a technology-agnostic dynamic model) with three different parameter sets (NF1, NF2, NF3), representing different levels of virtual inertia.
- Grid-Following Inverters: Modeled as PQ nodes.
- Transmission Lines: Modeled using Pi-models with parameters based on the German 380-kV transmission system (dena standards).
Fault Simulation: For each node in every grid, 1,000 post-fault clearance states were generated using Sobol sequences (Quasi-Monte Carlo sampling) and an ambient forcing algorithm to ensure algebraic constraints were satisfied.
Target Calculation: Dynamic simulations were performed to determine whether system transients remained within defined voltage and frequency ride-through characteristics. The target variable $p_{frt}$ was calculated as the ratio of successful rides-through to the total number of examined states for each node.
Validation Set: An updated IEEE-96-RTS test system, modified to include grid-forming capabilities and updated line parameters, was used as a held-out test set to evaluate generalization.

2. Machine Learning Approach

The study compares two classes of Machine Learning (ML) models for predicting $p_{frt}$ :

Non-Graphical ML Methods: Methods relying solely on node features (e.g., active/reactive power, local line parameters, node type) without considering the global grid topology. Tested models include Linear Regression (linreg) and Gradient Boosted Trees (GBT).
Graph Neural Networks (GNNs): Models processing the complete graph structure (node features, line features, and topology). Specifically, the authors used Topological Adaptive Graph (TAG) convolutional layers, which aggregate information from larger neighborhoods than standard Graph Convolutional Networks (GCNs).

3. Training and Evaluation

Features: Input features included node power setpoints, sums of line conductance/susceptance, node categories (Slack, Load, grid-forming), and the real part of the Normal Form parameter $R(B_x)$ .
Metrics: Performance was evaluated using the coefficient of determination ( $R^2$ ) and the Spearman rank correlation coefficient ( $\rho$ ).
Regularization: To address overfitting and improve generalization, a regularized version of the TAG model (TAGreg) using dropout was tested.

Key Results

1. Dynamics and Topology

Numerical analysis of the simulations revealed that stability outcomes are highly nonlinear. Scenarios with similar fault magnitudes (power deviations $\Delta P, \Delta Q$ ) led to diverging stability outcomes. Crucially, the study showed that $p_{frt}$ is strongly influenced by grid topology, not just local component parameters. Nodes with identical machine parameters exhibited significantly different stability probabilities depending on their position within the grid.

2. ML Performance on Synthetic Grids

On the synthetic dataset, the GNN-based TAG model significantly outperformed non-graphical methods:

TAG: Achieved an $R^2$ of 90.36% and a Spearman correlation ( $\rho$ ) of 95.05%.
Non-Graphical Models: GBT and Linear Regression achieved $R^2$ values of approximately 62–63%.
This confirms that incorporating topological information is essential for accurately predicting dynamic stability in grid-forming inverter-dominated grids.

3. Generalization to the IEEE Test System

The models were tested on the IEEE-96-RTS system, which was not part of the training data and exhibited different structural properties:

Initial Generalization: The standard TAG model showed reduced performance on the IEEE test case ( $R^2 \approx 60.46\%$ ) compared to the synthetic data, indicating difficulties with distribution shifts.
Regularized Model: The regularized TAG model (TAGreg) improved generalization, achieving an $R^2$ of 77.47% and a $\rho$ of 87.32% on the IEEE test case.
Comparison: Although GBT performed slightly better than the unregularized TAG on the IEEE test case, the regularized GNN (TAGreg) outperformed GBT on both the synthetic and IEEE datasets. This suggests that GNNs with appropriate regularization can learn causal, topology-independent patterns that generalize to unseen grid structures.

Significance and Claims

The contribution claims that machine learning, particularly Graph Neural Networks, offers a viable path to overcome the computational bottlenecks of probabilistic dynamic stability analysis in future power grids.

Efficiency: ML models can predict fault ride-through probabilities with high accuracy in near-instantaneous time, enabling the analysis of thousands of configurations that would be infeasible with dynamic simulations.
Generalization: The ability of GNNs trained on synthetic data to generalize to a real-world test case (IEEE 96-RTS) suggests these models can be deployed to investigate probabilistic stability in actual power grids without requiring exhaustive simulations for every new configuration.
Operational Impact: By identifying critical components vulnerable to failure across a broad spectrum of probabilistic scenarios, these tools can help grid operators move away from conservative, static planning. This enables the early integration of dynamic effects into grid expansion planning, potentially avoiding costly compensation measures and improving the overall reliability of inverter-dominated systems.

The authors emphasize that while the study used synthetic data, the successful generalization to the IEEE test case underscores the potential to apply these methods to grid-scale power systems where probabilistic stability assessments are currently computationally prohibitive.