Fault Detection in Electrical Distribution System using Autoencoders

Abstract: In recent times, there has been considerable interest in fault detection within electrical power systems, garnering attention from both academic researchers and industry professionals. Despite the development of numerous fault detection methods and their adaptations over the past decade, their practical application remains highly challenging. Given the probabilistic nature of fault occurrences and parameters, certain decision-making tasks could be approached from a probabilistic standpoint. Protective systems are tasked with the detection, classification, and localization of faulty voltage and current line magnitudes, culminating in the activation of circuit breakers to isolate the faulty line. An essential aspect of designing effective fault detection systems lies in obtaining reliable data for training and testing, which is often scarce. Leveraging deep learning techniques, particularly the powerful capabilities of pattern classifiers in learning, generalizing, and parallel processing, offers promising avenues for intelligent fault detection. To address this, our paper proposes an anomaly-based approach for fault detection in electrical power systems, employing deep autoencoders. Additionally, we utilize Convolutional Autoencoders (CAE) for dimensionality reduction, which, due to its fewer parameters, requires less training time compared to conventional autoencoders. The proposed method demonstrates superior performance and accuracy compared to alternative detection approaches by achieving an accuracy of 97.62% and 99.92% on simulated and publicly available datasets.

• The paper proposes a convolutional autoencoder trained only on normal three-phase current signals, using reconstruction-error thresholds to detect previously unlabeled faults.
• The method achieves 97.62% accuracy and 91.35% recall on a PV-integrated simulated feeder, and 99.92% accuracy with 100% recall on a public benchmark dataset.
• The results suggest autoencoders can reduce dependence on labeled fault data, but testing across fault locations, impedances, noise levels, changing grid conditions, and real-time deployment is still needed.

Overview and motivation

This paper proposes an unsupervised, anomaly-based approach to fault detection in electrical distribution systems (EDSs) using a convolutional autoencoder (CAE). The authors—affiliated with ABB Ability Innovation Center, IIT Bhubaneswar, and BITS Pilani Hyderabad—motivate the work by noting that existing fault detection schemes are predominantly supervised, requiring labeled fault data that is scarce in practice, and often depend on signal-processing feature extraction (Wavelet/Fourier transforms) whose frequency-band selection is arbitrary. The core idea is to train the autoencoder exclusively on normal three-phase current signals; because the model learns only nominal operating patterns, faulted segments yield elevated reconstruction errors and can be flagged without any fault labels.

Methodology

The autoencoder formulation follows the standard encoder–decoder setup: an encoder $A: \mathbb{R}^n \rightarrow \mathbb{R}^p$ maps input signals to a latent representation, and a decoder $B: \mathbb{R}^p \rightarrow \mathbb{R}^n$ reconstructs the input, with parameters optimized by minimizing the expected reconstruction loss under the $\ell_2$-norm. The authors position the autoencoder as a nonlinear generalization of PCA: with linear $A$ and $B$, the model recovers the PCA subspace, whereas the nonlinear variant learns a manifold better suited to capturing complex current-signal structure.

The specific architecture uses CNN-based encoder and decoder blocks. Input signals are segmented using overlapping windows of fixed length $T$ over a signal of $N$ points, producing $N - T + 1$ samples. Training proceeds on phase-A current under no-fault conditions; the maximum reconstruction error observed on training data is set as the detection threshold $\alpha$. At inference, per-phase current signals are reconstructed, and any segment whose mean absolute error exceeds $\alpha$ is classified as faulty. Evaluation is performed at the data-point level via confusion-matrix-derived metrics: accuracy, precision, recall, specificity, and F1-score.

A notable design choice is that the model is trained on a single phase's normal current yet applied to all three phases at test time, exploiting the fact that fault signatures manifest as deviations from learned normal behavior regardless of which phase is affected.

Datasets

Two data sources are used:

• Simulated dataset: A MATLAB/Simulink model comprising a 400 kW solar PV farm (four 100 kW arrays) connected to a distribution network with two feeders (8 km and 14 km), three loads, and three buses. Three-phase voltage and current measurements are taken at Bus 2. Four fault types are injected—single line-to-ground (AG), double line-to-ground (ABG), triple line-to-ground (ABCG), and line-to-line (AB)—at 1 km distance with 0.01 Ω fault resistance, sampled at $B: \mathbb{R}^p \rightarrow \mathbb{R}^n$0 s (320 samples per 60 Hz cycle), each fault lasting 2000 samples.
• Public dataset: A Kaggle dataset (~7600 data points) generated from a MATLAB power system simulation, used primarily for benchmarking against traditional machine learning baselines.

Results

On the simulated dataset, the model achieves 97.62% accuracy, with precision of 96.85%, recall of 91.35%, specificity of 99.23%, and F1-score of 94.02%. On the public dataset, performance is markedly stronger: 99.92% accuracy, precision of 99.89%, perfect recall (1.00), specificity of 99.74%, and F1-score of 99.94%.

Metric	Simulated data	Public data
Accuracy	97.62%	99.92%
Precision	96.85%	99.89%
Recall	91.35%	1.00
Specificity	99.23%	99.74%
F1-score	94.02%	99.94%

Against supervised baselines on the public dataset, the proposed model outperforms logistic regression (60.11%), support vector machines (99.66%), and K-nearest neighbors (99.34%). The claim that an unsupervised autoencoder surpasses these supervised classifiers is the paper's strongest empirical assertion, though it rests on comparison against third-party Kaggle results rather than controlled reimplementation.

The lower recall on simulated data (91.35%) relative to the public dataset suggests the PV-integrated distribution scenario—with distributed generation altering fault levels—is more challenging than the benchmark data, a point the paper does not analyze in depth.

Limitations and open questions

Several constraints qualify the reported results. First, the simulation covers only four fault types at a single location (1 km) with a fixed low fault resistance (0.01 Ω); sensitivity to fault impedance, distance, and high-impedance faults—which are notoriously difficult to detect—is not evaluated. Second, the threshold $B: \mathbb{R}^p \rightarrow \mathbb{R}^n$1 is defined as the maximum reconstruction error on training data, making it sensitive to outliers or noise in the normal training signal; no robustness analysis of this thresholding rule is provided. Third, the comparison against SVM and KNN relies on externally published Kaggle accuracies rather than experiments run under identical conditions. Fourth, the method is demonstrated offline; real-time deployment, computational latency, and adaptive retraining under changing grid topology or renewable output variability remain unaddressed. Finally, the discrepancy between simulated-data recall (91.35%) and public-data recall (100%) is left unexplained.

Conclusion

This paper demonstrates that a CNN-based autoencoder trained solely on normal current signals can detect faults in distribution systems without labeled fault data, achieving 97.62% accuracy on a PV-integrated simulated feeder and 99.92% on a public benchmark, exceeding standard supervised classifiers. The approach offers a practical path around the label scarcity problem in protection engineering, though its robustness across fault impedances, locations, noise conditions, and real-time operation remains to be established.