The reliable operation of power transmission networks is critical, and timely fault detection and localization are essential for preventing power outages and equipment damage. Traditional fault diagnosis methods are often time-consuming and limited in accuracy due to the complexity of the networks and variability of fault conditions. Machine learning (ML) and deep learning (DL) algorithms offer a promising alternative, but their effectiveness heavily relies on the quality and quantity of training data. Acquiring sufficient labeled data in real-world power systems is challenging due to the infrequent and unpredictable nature of faults. This research addresses this limitation by proposing a novel approach that leverages variational autoencoders (VAEs) to generate synthetic data to augment real-world datasets, thereby improving the performance of ML algorithms for fault classification and localization in power transmission networks. The increasing complexity and demand for electric power necessitate more sophisticated methods for fault diagnosis and this paper proposes a solution for the accurate and efficient detection of faults in high-voltage transmission networks.

Existing literature explores various techniques for fault diagnosis in power transmission networks, including wavelet analysis, genetic algorithms (GAs), phasor measurement units (PMUs), and multi-information-based techniques. Traditional rule-based or model-based approaches require detailed knowledge of network topology and fault characteristics, which can be challenging to obtain in complex systems. Recent studies have investigated the use of generative models, such as GANs and VAEs, to create synthetic data for enhancing ML model performance in data-scarce scenarios. Previous work has explored VAE-based synthetic data generation for fault diagnosis in specific applications, such as wind turbines. This paper builds upon this prior work by applying VAEs to generate synthetic data for power transmission line fault diagnosis, using a diverse set of machine learning algorithms for both classification and localization, and focusing on three-phase high-voltage networks.

The proposed methodology involves several key steps: 1. **Data Acquisition:** Real-world data representing various types of faults (line-to-ground, line-to-line, double line-to-ground, three-phase) were obtained from simulations using the Aspen One-liner tool. 2. **Synthetic Data Generation:** A Variational Autoencoder (VAE) was trained on the real-world data to generate synthetic fault data. The VAE architecture learns the underlying probability distribution of the data and generates new data points that follow the same distribution. The objective function used balances reconstruction loss and KL divergence loss for optimal performance. 3. **Data Preprocessing and Feature Selection:** The combined real and synthetic datasets were preprocessed to remove noise and irrelevant information. Forward Feature Selection (FFS) was applied to select the most relevant features for efficient model training. 4. **Model Training:** Five machine learning algorithms were trained on the preprocessed dataset: CatBoost, SVM, Decision Trees, Random Forest, and K-Nearest Neighbors. Hyperparameter tuning was performed using grid search to optimize model performance. 5. **Model Evaluation:** The trained models were evaluated using a test dataset, with performance metrics including accuracy, precision, recall, F1 score, MAE, and ROC curves. Confusion matrices were used to analyze the classification performance of each model. The methodology employs stratified cross-validation to handle imbalanced data and to enhance the robustness of the proposed models.

The results demonstrate that the proposed method, using VAE-generated synthetic data and a combination of machine learning algorithms, significantly improves fault classification and localization accuracy compared to baseline methods. The CatBoost algorithm showed superior performance achieving 99% accuracy in fault classification. The MAE for fault localization was 0.2. Other algorithms like SVM, Decision Trees, Random Forest, and KNN also exhibited high accuracy, demonstrating the robustness of the approach. The use of synthetic data generated by the VAE was crucial for achieving these high accuracy rates, as it addressed the limitation of insufficient labeled real-world data. The scatter plots of the synthetic dataset showed a clear separation between different fault types and locations, demonstrating the effectiveness of the VAE in generating realistic and representative data. Detailed confusion matrices and ROC curves provided further insights into the classifier performance. The regression plots show the high accuracy of fault localization and the calculated absolute errors were low, highlighting the effectiveness of the regression models.

The findings address the research question by demonstrating the feasibility and effectiveness of using VAE-generated synthetic data to enhance fault classification and localization in power transmission networks. The high accuracy rates achieved (99% for classification and MAE of 0.2 for localization) indicate that this approach outperforms existing methods and offers a significant improvement in the reliability and efficiency of fault diagnosis. The success of the approach highlights the potential of using synthetic data generation to overcome data scarcity issues in other domains where acquiring labeled data is challenging. The results suggest that the combination of VAEs and robust machine learning algorithms provides a powerful tool for improving the safety and reliability of power transmission systems. Future research could explore more advanced generative models and investigate the impact of different data augmentation techniques on model performance.

This study successfully demonstrated a novel approach to fault classification and localization in power transmission networks using VAE-generated synthetic data and a suite of machine learning algorithms. The method achieved exceptional accuracy, surpassing state-of-the-art techniques. The use of synthetic data effectively mitigated the challenges of limited real-world data. Future work could investigate the application of this method to more complex network topologies and explore the integration of this approach with real-time monitoring systems for predictive maintenance.

While the proposed method demonstrates high accuracy, some limitations exist. The accuracy is dependent on the quality of the synthetic data generated by the VAE. The model's performance might be affected by variations in the real-world data distribution not fully captured by the training data. The computational cost associated with training the VAE and multiple ML models can be high. Further research is needed to assess the generalizability of the method to different network configurations and fault types.