Real-Time Faulted Line Localization and PMU Placement in Power Systems Through Convolutional Neural Networks

Wenting Li, Deepjyoti Deka, Michael Chertkov, Meng Wang

TL;DR

This paper introduces a robust, real-time method for locating faulted power lines using Convolutional Neural Networks (CNNs) and limited sensor data. By combining physically interpretable feature extraction with a novel "model-aware" sensor placement strategy, the authors achieve high localization accuracy even when only 7% to 15% of the grid's buses are monitored.

What Problem Does This Paper Address?

When a fault occurs in a power grid (like a short circuit), operators must identify and isolate the affected line immediately to prevent cascading failures or massive blackouts. However, modern grids are massive, and installing high-speed sensors—Phasor Measurement Units (PMUs)—at every single bus is prohibitively expensive.

Existing localization techniques often fall into two traps: they either require a "fully observed" system (sensors everywhere) or rely on extremely high sampling rates (1kHz to 1MHz) that exceed standard PMU capabilities. Furthermore, many data-driven methods treat the grid as a "black box," ignoring the underlying physics, which makes them fragile when load conditions change or measurements become noisy.

Methods

The authors propose a two-pronged approach:

  1. Physically-Informed Feature Extraction: Instead of feeding raw voltage data into a neural network, they use Substitution Theory to derive a feature vector ($\psi$). By multiplying the pre-fault admittance matrix ($Y^0$) by the change in bus voltages ($\Delta U$), they create a vector that physically represents unbalanced currents. In a perfect system, this vector is "sparse," meaning it only has non-zero values at the buses directly connected to the fault.
  2. CNN-Based Classification: They employ a 4-layer CNN (inspired by AlexNet) to classify which line is faulted. CNNs are particularly well-suited for this because their "convolutional" nature naturally mimics the local connectivity and topology of a power grid.
  3. Greedy PMU Placement: To decide where to put a limited number of sensors, the authors developed a greedy algorithm. This algorithm selects bus locations that minimize the CNN’s training loss while prioritizing "high-degree" buses (hubs with many connections), ensuring the AI gets the most informative "view" of the network.

Datasets

The methodology was validated using two standard benchmarks in power systems research:

  • IEEE 39-bus system: A simplified model of the New England power grid.
  • IEEE 68-bus system: A more complex model representing the New York/New England interconnected systems.

Data was generated using the Power System Toolbox (PST), simulating four types of faults: Three-Phase (TP), Line-to-Ground (LG), Double-Line-to-Ground (DLG), and Line-to-Line (LL), under varying load conditions and fault impedances.

Key Insights and Ideas

  • Physics as a Pre-processor: The "aha moment" is that using the admittance matrix to transform raw data into physically meaningful features makes the machine learning task significantly easier. It reduces the complexity of the CNN and makes it more robust to "noise" in the grid.
  • Local Features for Local Faults: Because faults primarily affect their immediate electrical neighborhood, the sparse connectivity of CNNs is a much better architectural fit for power grids than traditional "fully-connected" neural networks.
  • Graceful Degradation: Even when the system has very low observability (e.g., only 7% of buses have sensors), the algorithm doesn't just guess randomly. It consistently identifies the correct line within its "Top 3" or "Top 5" predictions, effectively narrowing the search area for repair crews.
  • Topology-Aware AI: The paper demonstrates that PMU placement shouldn't just be about covering the map; it should be optimized specifically to help the underlying AI model learn the grid's behavior.

Key Results

  • High Accuracy with Low Coverage: In the 68-bus system, the CNN achieved over 90% accuracy with only 15% PMU coverage.
  • Superiority over Baselines: The CNN consistently outperformed Support Vector Machines (SVM) and standard Neural Networks (NN), especially as the number of sensors decreased.
  • Robustness: The model maintained high performance despite:
  • Noise: Signal-to-Noise Ratios (SNR) as low as 60dB.
  • Delays: Measurement synchronization delays of up to 40ms.
  • Uncertainty: Significant random fluctuations in power loads.
  • Efficiency: Once trained, the system can localize a fault in just 4 milliseconds, making it truly "real-time."

Limitations and Future Work

  • Simulation vs. Reality: The study relies on synthetic data generated by the Power System Toolbox. The authors acknowledge that testing on real-world utility data is a necessary next step.
  • Distribution Grids: The current work focuses on transmission systems. Extending this to distribution grids—which are more complex, have more branches, and less instrumentation—remains a challenge.
  • Pinpointing the Exact Spot: While the method identifies the faulted line, it doesn't yet pinpoint the exact meter-mark along that line. The authors suggest this could be handled by a second-stage algorithm once the line is identified.

Summary updated

Abstract

Diverse fault types, fast reclosures, and complicated transient states after a fault event make real-time fault location in power grids challenging. Existing localization techniques in this area rely on simplistic assumptions, such as static loads, or require much higher sampling rates or total measurement availability. This paper proposes a faulted line localization method based on a convolutional neural network (CNN) classifier using bus voltages. Unlike prior data-driven methods, the proposed classifier is based on features with physical interpretations that improve the robustness of the location performance. The accuracy of our CNN-based localization tool is demonstrably superior to other machine learning classifiers in the literature. To further improve the location performance, a joint phasor measurement units placement strategy are proposed and validated against other methods. A significant aspect of our methodology is that under very low observability (7% of buses), the algorithm is still able to localize the faulted line to a small neighborhood with high probability. The performance of our scheme is validated through simulations of faults of various types in the IEEE 39-bus and 68-bus power systems under varying uncertain conditions, system observability, and measurement quality.
BibTeX
@article{li2019real,
  author = {Wenting Li and Deepjyoti Deka and Michael Chertkov and Meng Wang},
  title = {{Real-Time Faulted Line Localization and PMU Placement in Power Systems Through Convolutional Neural Networks}},
  journal = {IEEE Transactions on Power Systems},
  year = {2019},
  doi = {10.1109/tpwrs.2019.2917794},
}

Loading related content...

Figure