Researchers Detect Faults, Improving Power System Protection

Fault detection in electrical distribution systems presents a continuing challenge for reliable power delivery. Sidharthenee Nayak and Victor Sam Moses Babu, both from ABB Ability Innovation Center, Asea Brown Boveri Company, alongside Chandrashekhar Narayan Bhende of the School of Electrical Sciences, Indian Institute of Technology Bhubaneswar, and Pratyush Chakraborty from the Department of Electrical and Electronics Engineering, BITS Pilani Hyderabad Campus, with further contributions from Mayukha Pal at ABB Ability Innovation Center, Asea Brown Boveri Company, present a novel approach utilising autoencoders to address this critical issue. This research is significant because it offers an anomaly-based method capable of achieving high accuracy, 97.62% on simulated data and 99.92% using publicly available datasets, while also reducing training time through the implementation of Convolutional Autoencoders. This collaborative work between ABB Ability Innovation Center, Asea Brown Boveri Company, Indian Institute of Technology Bhubaneswar, and BITS Pilani Hyderabad Campus provides a promising advancement in intelligent fault detection techniques for electrical power systems.

Scientists are tackling the persistent problem of electrical grid failures with a novel artificial intelligence technique. Protecting our power supply demands rapid and accurate identification of faults, yet current methods struggle with real-world complexity. This new approach promises to improve reliability by swiftly pinpointing issues before they escalate into widespread outages.

Scientists have demonstrated considerable interest in fault detection within electrical power systems, attracting attention from both academic researchers and industry professionals. Despite the development of numerous fault detection methods over the past decade, their practical application remains highly challenging. Given the probabilistic nature of fault occurrences, certain decision-making tasks could be approached from a probabilistic standpoint.

Protective systems are tasked with the detection, classification, and localization of faulty line magnitudes, culminating in the activation of circuit breakers to isolate the faulty line. Obtaining reliable data for training and testing is essential for designing effective fault detection systems, which is often scarce. Leveraging deep learning techniques, particularly the capabilities of pattern classifiers in learning, generalizing, and parallel processing, offers promising avenues for intelligent fault detection.

This paper proposes an anomaly-based approach for fault detection in electrical power systems, employing deep autoencoders and utilising Convolutional Autoencoders (CAE) for dimensionality reduction, which requires less training time compared to conventional autoencoders. The electric power grid plays a vital role in modern society, reliably supplying electricity to residential, commercial, and industrial sectors.

As our dependence on electricity grows, there is a corresponding increase in the need for robust and effective electrical distribution systems (EDSs). Guaranteeing the safety and dependability of these systems requires mitigating risks and ensuring uninterrupted power delivery, making sophisticated fault detection and classification methods crucial for optimising performance and strengthening grid resilience.

The electrical power system, comprising various dynamic elements, is susceptible to disturbances and faults, necessitating swift fault detection and protection operation to maintain stability. It is important to swiftly detect and classify faults on transmission lines, with protection systems initiating relays to prevent outages. Effective fault detection and classification, ensuring rapid restoration of the power system, are imperative for service reliability and minimising outages.

Protection schemes must promptly detect and remove affected segments during a fault incident to minimise its impact. However, the expansion of modern power networks poses challenges for protection systems, requiring integrated schemes capable of monitoring different grid layers. Wide Area Protection (WAP) using phasor measurements from Phasor Measurement Units (PMUs) has been proposed, yet challenges remain in interpreting data and identifying faulty components.

Existing fault detection algorithms for transmission networks often rely on iterative solutions or require numerous PMUs, while distribution networks face issues due to distributed generation impacting fault levels and relay operation. Synchrophasor measurements offer a more reliable alternative but are currently limited to distribution networks, highlighting the need for an integrated scheme applicable to both distribution and transmission networks.

Fault diagnosis is categorized into two main types: model-based and process history-based approaches. Model-based methods involve analysing faults by representing a system or process using quantitative or qualitative models. Process history-based techniques rely on empirical data, establishing connections between inputs and desired outputs without prior mathematical modelling.

Feature extraction is crucial in process history-based methods as it helps capture essential information from empirical data for pattern recognition. With advancements in signal processing and a deeper understanding of power systems, various techniques have emerged for direct measurement and transformation, enabling the extraction of inherent fault characteristics.

Commonly utilised methods include Wavelet and Fourier transforms, which effectively isolate fault-related characteristics with robustness and precision. However, these classical methods may yield inaccurate results due to assumptions about line parameters. Artificial neural networks (ANNs) and support vector machines (SVMs) are robust pattern recognition methodologies capable of efficiently generalizing dynamic parameters using both supervised and unsupervised learning approaches.

Recently, machine learning algorithms have been widely used by combining signal processing approaches to rapidly and accurately identify faults. Signal processing techniques extract features from initially captured voltage and current signals to determine fault occurrence and their types. Enhancing the fault detection accuracy for EDSs has emerged as a significant research focus.

Existing fault detection techniques are often supervised approaches, posing challenges for real-time applications due to the requirement of prior labelling, and achieving online fault detection and clustering with a high degree of accuracy remains elusive. Recently autoencoders have emerged as an interesting option for anomaly detection in a time series because they need to be trained only on normal data.

Researchers have used deep autoencoders for anomaly detection in wireless communication networks and videos. In this method, a deep convolutional autoencoder model is used to detect faults in the distribution and transmission system. At first, the model is trained on normal time-series data of current. During training, the autoencoder learns to reconstruct the normal time series data and the maximum reconstruction error is chosen as the threshold.

While testing, current signals with various types of faults are given to the model. If the reconstruction error is more than the threshold, those segments of the signal are identified as fault segments. The core of this work lies in the convolutional autoencoder’s ability to reduce dimensionality while minimising training time compared to conventional autoencoders due to its fewer parameters.

This efficient dimensionality reduction is crucial for processing the complex time-series data inherent in power system monitoring. The model’s success hinges on its unsupervised learning approach, requiring only normal data for training and circumventing the need for pre-labelled fault examples, a significant advantage for real-time applications. Furthermore, the system’s performance surpasses that of other traditional machine learning models in accurately detecting faults.

The use of a deep convolutional autoencoder allows for the extraction of relevant features directly from the voltage and current signals, eliminating the need for arbitrary feature selection across different frequency ranges. This automated feature extraction contributes to the consistency and reliability of the fault detection process. The reconstruction error metric provides a quantifiable measure of anomaly, enabling precise identification of fault segments within the current signals.

Deep convolutional autoencoders enable high accuracy power system fault detection

Achieving 97.62% accuracy on a simulated dataset and 99.92% on a publicly available dataset, the proposed anomaly-based fault detection system demonstrates substantial performance in identifying electrical power system faults. This level of accuracy was attained through the implementation of a deep convolutional autoencoder model trained exclusively on normal current time-series data.

The autoencoder learns to reconstruct typical current patterns, establishing a maximum reconstruction error that serves as a fault threshold. During testing, current signals containing various fault types are processed, and any reconstruction error exceeding this threshold is flagged as a fault segment.

Fault detection via compressed representation of normal system behaviour

Convolutional Autoencoders (CAE), a type of neural network designed for efficient data compression and reconstruction, underpin the fault detection methodology employed in this work. We selected autoencoders due to their capacity to learn complex, non-linear relationships within data without explicit programming, making them well-suited to identifying subtle anomalies indicative of faults.

The core principle involves training the CAE on normal operating conditions, enabling it to develop a robust internal representation of healthy system behaviour. Subsequently, deviations from this learned representation signal potential faults. To facilitate this, voltage and current data representing normal system operation were initially subjected to dimensionality reduction using the CAE.

This process reduces the number of input variables, decreasing computational burden and mitigating the risk of overfitting during training. Unlike traditional autoencoders, the implementation of convolutional layers within the CAE architecture allows for the extraction of relevant features directly from the input data.