Fault detection in power distribution systems presents a continuing challenge to reliable power supply. With further contributions from Sidharthenee Nayak and Victor Sam Moses Babu of the ABB Ability Innovation Center, Chandrashekhar Narayan Bende of the Department of Electrical Sciences, Indian Institute of Technology Bhubaneswar, Pratyush Chakraborty of the Department of Electrical and Electronic Engineering, BITS Pilani Hyderabad Campus, and Mayukha Pal of the ABB Ability Innovation Center, Asea Brown Boveri Company presented a new approach utilizing autoencoders to address this critical issue. This study is important because it provides an anomaly-based method that can achieve high accuracy of 97.62% on simulated data and 99.92% using publicly available datasets, while reducing training time through the implementation of a convolutional autoencoder. This collaboration between ABB Ability Innovation Center, Asea Brown Boveri Company, Indian Institute of Technology Bhubaneswar, and BITS Pilani-Hyderabad Campus brings promising advances in intelligent fault detection technology for power systems.
Scientists are using new artificial intelligence techniques to tackle the persistent problem of power grid failures. To protect power supplies, faults must be identified quickly and accurately, but current methods struggle with real-world complexities. This new approach promises to improve reliability by quickly identifying problems before they develop into widespread failures.
Scientists have shown significant interest in fault detection within power systems, attracting attention from both academic researchers and industry professionals. Although many fault detection methods have been developed in the past decade, their practical implementation remains very difficult. Given the probabilistic nature of failure occurrence, certain decision-making tasks can potentially be approached from a probabilistic perspective.
The protection system is responsible for detecting, categorizing, and locating faulty circuits, ultimately tripping circuit breakers to isolate the faulty circuit. Obtaining reliable data for training and testing is essential to designing effective fault detection systems, but data is often lacking. Exploiting deep learning techniques, especially the capabilities of pattern classifiers in learning, generalization, and parallel processing, provides a promising avenue for intelligent fault detection.
In this paper, we propose an anomaly-based approach for power system fault detection. This approach employs a deep autoencoder and utilizes convolutional autoencoder (CAE) for dimensionality reduction. This reduces training time compared to traditional autoencoders. Electric power grids play an important role in modern society by providing a reliable supply of electricity to households, commercial and industrial sectors.
As reliance on electricity increases, so does the need for robust and effective electrical distribution systems (EDS). Ensuring the safety and reliability of these systems requires mitigating risks and ensuring uninterrupted power supply, and advanced fault detection and classification methods will be essential to optimize performance and enhance grid resiliency.
Power systems are composed of various dynamic elements and are susceptible to disturbances and faults, requiring rapid fault detection and protective actions to maintain stability. It is important that protection systems initiate relays to prevent outages and quickly detect and classify faults in transmission lines. Effective fault detection and classification is essential to ensure rapid recovery of power systems and to ensure service reliability and minimize outages.
The protection scheme must quickly detect and remove affected segments in the event of a failure to minimize its impact. However, the expansion of modern power grids poses challenges to protection systems and requires integrated schemes that can monitor different grid layers. Wide area protection (WAP) using phasor measurements from phasor measurement units (PMUs) has been proposed, but challenges remain in interpreting the data and identifying faulty components.
Existing fault detection algorithms for power transmission networks often rely on iterative solutions or require a large number of PMUs. Meanwhile, power distribution networks face challenges due to distributed generation, which affects disturbance levels and relay operation. Synchrophasor measurements offer a more reliable alternative, but are currently limited to power distribution networks, highlighting the need for integrated schemes that are applicable to both power distribution and transmission networks.
Fault diagnosis is classified into two main types: model-based approaches and process history-based approaches. Model-based methods involve analyzing failures by representing systems or processes using quantitative or qualitative models. Process history-based methods rely on empirical data and establish connections between inputs and desired outputs without prior mathematical modeling.
Feature extraction is very important in process history-based methods as it helps in obtaining important information from empirical data for pattern recognition. With advances in signal processing and a deeper understanding of power systems, a variety of techniques have emerged for direct measurement and transformation, allowing the extraction of unique fault characteristics.
Commonly used methods include wavelet transform and Fourier transform, which effectively separate fault-related characteristics with robustness and accuracy. However, these classical methods can yield inaccurate results due to assumptions about line parameters. Artificial neural networks (ANN) and support vector machines (SVM) are robust pattern recognition techniques that can efficiently generalize dynamic parameters using both supervised and unsupervised learning approaches.
Recently, machine learning algorithms have been widely used to combine signal processing approaches to quickly and accurately identify faults. Signal processing techniques extract features from the initially acquired voltage and current signals to determine the occurrence and type of fault. Improving the fault detection accuracy of EDS has emerged as an important research focus.
Existing fault detection techniques are often supervised approaches and require prior labeling, creating challenges for real-time applications, and online fault detection and clustering remains difficult to achieve with high accuracy. Recently, autoencoders have emerged as an interesting option for time series anomaly detection, as they need to be trained only on regular data.
Researchers have used deep autoencoders for anomaly detection in wireless communication networks and videos. This method uses a deep convolutional autoencoder model to detect faults in power distribution and transmission systems. First, the model is trained with regular time series data of current. During training, the autoencoder learns to reconstruct regular time series data, and the maximum reconstruction error is selected as the threshold.
During testing, the model is given current signals containing different types of faults. If the reconstruction error exceeds a threshold, those segments of the signal are identified as faulty segments. The core of this work lies in the ability of convolutional autoencoders to reduce dimensionality while minimizing training time compared to traditional autoencoders due to fewer parameters.
This efficient dimensionality reduction is critical for processing the complex time series data inherent in power system monitoring. The success of this model relies on an unsupervised learning approach that requires only regular data for training, avoiding the need for pre-labeled fault examples. This is a huge advantage for real-time applications. Moreover, the performance of this system outperforms other traditional machine learning models in accurately detecting faults.
Deep convolutional autoencoders allow you to directly extract relevant features from voltage and current signals, eliminating the need to select arbitrary features across different frequency ranges. This automated feature extraction contributes to the consistency and reliability of the fault detection process. Reconstruction error metrics provide a quantifiable measure of anomalies and allow accurate identification of faulty segments within the current signal.
Deep convolutional autoencoder enables highly accurate power system fault detection
The proposed anomaly-based fault detection system achieves 97.62% accuracy on simulated datasets and 99.92% accuracy on publicly available datasets, demonstrating excellent performance in identifying faults in power systems. This level of accuracy was achieved by implementing a deep convolutional autoencoder model trained only on regular current time series data.
The autoencoder learns to reconstruct typical current patterns and establishes a maximum reconstruction error that serves as a failure threshold. During testing, current signals containing different types of faults are processed, and reconstruction errors exceeding this threshold are flagged as faulty segments.
Fault detection using a compressed representation of normal system behavior
Convolutional autoencoder (CAE) is a type of neural network designed for efficient data compression and reconstruction, which underpins the fault detection method adopted in this work. We chose autoencoders because they can learn complex nonlinear relationships in data without explicit programming and are well-suited for identifying subtle anomalies that indicate failures.
The core principles include training the CAE under normal operating conditions, allowing it to develop a robust internal representation of normal system behavior. Deviations from this learned representation then indicate a potential failure. To facilitate this, voltage and current data representing normal system operation were first dimensionally reduced using CAE.
This process reduces the number of input variables, reduces computational load, and reduces the risk of overfitting during training. Unlike traditional autoencoders, implementing convolutional layers within a CAE architecture allows you to directly extract relevant features from the input data.
