Source-linked AI summary
A Systematic Evaluation of Machine Learning Methods for Fault Detection and Line Identification in Electrical Power Grids
Julian Oelhaf, Georg Kordowich, Paula Andrea Pérez-Toro, Tomás Arias-Vergara, Andreas Maier, Johann Jäger, Siming Bayer
TL;DR
Renewable integration complicates fault protection, while differing simulation and analysis setups limit comparisons among ML approaches. The paper systematically evaluates classifiers for three-phase fault detection and faulty-line identification across 10–50 ms windows and runtime. Top models achieve F1 scores up to 0.99 for detection and 0.98 for line identification at 10 ms, with competitive runtimes reported for MLP, Gradient Boosting, and Stacking.
Problem
Renewable integration challenges conventional protection, and differing grid configurations, preprocessing methods, and analysis windows limit consistent comparison of ML approaches.
Method
The study systematically evaluates 14 ML classifiers for binary fault detection and four-class line identification using simulated multivariate voltage-current windows of 10–50 ms.
Results
At 10 ms, top models achieve mean F1 scores up to 0.99 for fault detection and 0.98 for fault-line identification, while MLP, Gradient Boosting, and Stacking combine high scores with competitive runtimes.
Takeaways & Limitations
Most evaluated models perform strongly in the specified three-phase short-circuit scenarios using a 10 ms measurement window.
Takeaways & Limitations
The study is limited to a single grid and does not yet establish transferability across grids, topologies, fault types, operating conditions, or noisy or incomplete data.
Abstract
from arXiv · showhide
The integration of renewable energy sources into the electrical grid introduces complex challenges in fault detection and coordination of grid recovery mechanisms. Traditional relay protection systems, which operate based on static rules and predefined thresholds, are inadequate for addressing these challenges, particularly in detecting and isolating faults such as short circuits. Consequently, the conventional methodologies applied to electrical network protection frequently fail to achieve optimal performance in fault detection, especially in terms of adherence to safety standards and the selective limitation of damage. Recent research indicates that machine learning (ML)-based approaches can effectively tackle these issues; however, variations in grid configurations and analysis windows have impeded consistent comparative assessments. In this study, we assess the efficacy of various ML models in detecting electrical faults and pinpointing defective transmission lines within a 10 ms measurement interval - a critical time-frame for real-time operational viability, for the first time. The most effective model attained an F1 score of 0.991 +/- 0.018 and demonstrated a processing time of 0.342ms +/- 0.509ms.
I. INTRODUCTION
Renewable integration increases grid complexity and challenges conventional protection, while existing ML studies remain difficult to compare and have not evaluated the 10 ms real-time boundary. This study systematically evaluates ML models for three-phase fault detection and transmission-line identification across window lengths and runtimes.
- Motivation: Renewable and distributed energy integration increases grid complexity through meshing and AC/DC hybrid arrangements, challenging traditional protection methods.Renewable-source fault currents can differ from traditional-source currents, affecting short-circuit characteristics and conventional protection behavior.
- Motivation: Traditional protection relies on fixed rules and thresholds for high currents, but struggles with nonlinear classification and changing renewable-source fault currents.These challenges affect fault-detection performance, sensitivity, and selectivity.
- Scope: Fault detection recognizes anomalous conditions, whereas fault-line identification identifies the affected grid section; fault localization is a separate task excluded here.The study therefore focuses on detecting three-phase short circuits and identifying their faulty transmission lines.
- Real-time constraint: 10 ms is a critical protection timescale because prompt fault clearing supports reliable and safe operation, while relay guidance specifies detection windows from 10 ms to 50 ms.VDEW requirements cited in the paper specify maximum detection times of 25 ms, 30 ms, and 40 ms for extra-high-, high-, and medium-voltage networks, respectively.
- Research gap: ML fault-protection studies are difficult to compare because they use different grid simulations, preprocessing methods, target metrics, and realism assumptions.The paper also notes that published methods had not been evaluated at the 10 ms lower bound of real-time operation.
- Study objective: The study applies domain randomization and compares fault-detection performance over 10–50 ms context windows while analyzing model runtime.The evaluation targets three-phase short circuits in transmission lines and seeks generalization from synthetic data toward later real-world data.
A. Grid Topology and Data Generation
The study generates physics-based, randomized simulations on a four-line Double-Line grid and records instantaneous relay voltage and current measurements with labels for fault timing and affected line.
- Simulation setup: Physics-based PowerFactory simulations generate one-second episodes with varied grid and fault parameters on the Double-Line topology.Fault initiation ranges from 0.2 s to 0.5 s, fault duration from 5 ms to 600 ms, and fault resistance from 0.1 Ω to 10 Ω.
- Parameter variation: Domain randomization varies grid parameters within stable operating ranges to represent real-world operational diversity and support simulation-to-reality domain shift.The simulations enforce load-flow convergence, numerical stability, and R/X ratios between 0.05 and 0.5.
- Grid topology: The Double-Line model contains one external grid, two loads, three buses, and four transmission lines used for grid-protection simulations.Both b lines are occasionally deactivated to improve robustness across topology configurations.
- Measurements: Each transmission line has two protection relays measuring instantaneous three-phase voltage and current, producing 48 measurements per time step across the four lines.The measurements form a multivariate time series and are provided directly to models without phasor transformation.
- Labels and fault scope: Labels encode fault start time for detection and a categorical fault-line variable for identifying the affected transmission line.The study focuses exclusively on three-phase faults, which the paper describes as severe but less frequent.
B. Data Preprocessing
Raw episodes are trimmed around fault onset and converted into overlapping windows of five candidate lengths to model real-time protection decisions.
- Episode trimming: Episodes are trimmed to ±80 ms around fault start to capture critical events immediately before and after the fault.This preprocessing centers the retained data on the event used for detection and line identification.
- Windowing: Sliding windows advance in 5 ms steps and overlap at least two segments, approximating real-time relay processing.The overlap is maintained across all evaluated context lengths.
- Window lengths: The evaluation compares 10 ms, 20 ms, 30 ms, 40 ms, and 50 ms windows, with corresponding timestep, window-count, fault-window, and feature-count details reported in Table II.Window length determines the number of time steps and features supplied to the models.
C. Machine Learning Models for Fault Detection and Line Identification
Fault detection is evaluated as binary classification and line identification as four-class classification using multivariate voltage-current windows, a broad classifier set, cross-validation, and runtime measurement.
- Task formulation: Fault detection is binary classification, while line identification is four-class classification using concatenated simulated voltage and current time series.The feature count varies with the selected window length.
- Labeling: Fault labels are assigned when a fault starts fully inside a window, using a 5 µs margin from both window boundaries.The condition uses the window start and end timestamps to ensure the event is entirely contained.
- Model set: The evaluation includes linear, neighborhood, neural, margin-based, tree, boosting, bagging, stacking, and voting classifiers.Specific models include Logistic Regression, Ridge, SGD, KNN, MLP, SVM, AdaBoost, Bagging, ExtraTrees, Gradient Boosting, Random Forest, Stacking, and Voting.
- Evaluation protocol: Ten-fold cross-validation uses a 9:1 training-to-test split, with standardized features and accuracy, precision, recall, sensitivity, specificity, and F1 metrics.The F1 score is reported because it balances relevant classification dimensions.
- Evaluation protocol: Fault detection uses all extracted windows, whereas line identification uses only windows containing faults because identification follows fault detection.Runtime is measured for scaling and predicting one window over 5000 iterations.
III. EXPERIMENTS AND RESULTS
Across fault detection and fault-line identification, the evaluation compares 14 ML models using F1 scores and runtime measurements. Several models perform strongly, while a small subset underperforms or incurs substantially higher runtime.
- Up to 0.99 F1 was achieved by ET, GB, MLP, RF, Stacking, and SVM across all fault-detection window lengths.
- F1 scores below 0.69 occurred for LG, Ridge, and SDG in fault detection, while LG, Ridge, SDG, and AdaBoost reached 0.58 or lower in line identification.
- Up to 0.97 mean F1 was achieved by ET, GB, MLP, RF, and Stacking for fault-line identification.
- KNN had the slowest measured runtime at 88.6±5.8 ms, making it a clear outlier among the tested models.
IV. DISCUSSION
The discussion identifies 10 ms as a viable short measurement window, with strong task performance from several models. It also highlights a trade-off between predictive quality and runtime across model families.
- Most models remained effective regardless of window length, supporting the use of the shortest evaluated window for real-time-oriented analysis.
- At 10 ms, top models achieved mean F1 scores up to 0.99 for fault detection and 0.98 for fault-line identification.
- MLP, GB, and Stacking combined the highest F1 scores across both tasks with competitive runtimes of 0.34 ms, 1.40 ms, and 2.18 ms, respectively.
- The three fastest models performed worst in both tasks, indicating that their simplicity was insufficient for fault detection or line identification.
V. CONCLUSION
The study evaluates ML models for three-phase short-circuit detection and transmission-line identification, finding strong performance with 10 ms windows in the specified scenarios. Its conclusions remain bounded by the single-grid evaluation and hardware-dependent runtime measurements.
- The study systematically evaluates ML models for detecting three-phase short circuits and identifying faulty transmission lines in electrical power grids.
- Most evaluated models achieved exceptional scores with a 10 ms window in the specified scenarios.
- The study is limited to a single grid, while runtime discrepancies depend on the hardware used.