In a recent publication in the journal Scientific Reports, researchers provided a comprehensive comparison of different machine learning (ML) and deep learning (DL) methods for three-dimensional (3D) localization of partial discharge (PD) sources within power transformer tanks using single-sensor electric field measurements.
Architecture of BPNN model. The first layer is the input layer, each fully connected layer has 512 units, and the output layer estimates the x, y and z coordinates of the PD source. Image Credit: https://www.nature.com/articles/s41598-024-62527-9
Background
Power transformers are essential components of the electric power system, responsible for converting voltage levels and transmitting electricity over long distances. However, they are susceptible to failures due to various factors such as aging, overloading, and environmental conditions.
One of the primary causes is PD, which is an electrical breakdown within the transformer's insulation system, including oil, paper, or bushings. Over time, PDs can degrade the insulation, leading to a complete breakdown and costly repairs/replacements. Therefore, early detection and localization of PDs are crucial for enabling preventive maintenance and ensuring the reliable and efficient operation of the power grid.
About the Research
In this article, the authors examined various ML and DL techniques, including convolutional neural networks (CNN), support vector regression (SVR), support vector machines (SVM), backpropagation neural networks (BPNN), k-nearest neighbors (KNN), multilayer perceptron’s (MLP), and extreme gradient boosting (XGBoost) for 3D localization of partial discharge (PD) sources within power transformer tanks.
The study focused on multiple case studies each with varying attributes such as sensor positioning, the number of sensors, the frequency content of the PD signal, and the size of the transformer tank. The primary objective was to determine the PD location using single-sensor electric field measurements, with one exception involving three sensors.
Five key steps were involved in data preprocessing: cut-off, normalization, resampling, label shifting, and train-test dataset splitting. The authors utilized a laboratory-scale model of a power transformer tank with a single electric field sensor placed inside. They generated PD signals using a needle-plate electrode system at different locations within the tank.
Electric field signals were recorded, and features such as peak value, rise time, pulse width, and energy were extracted. The data were simulated using computer simulation technology microwave studio (CST-MWS) software, ensuring consistency and accuracy in the input signals. ML models were then trained and tested using preprocessed data, with grid search employed for hyperparameter optimization.
Furthermore, the researchers evaluated the methods' performance using two metrics: the correlation coefficient (CC) and the root mean square error (RMSE) between the predicted and actual locations. They also explored transfer learning techniques to improve model accuracy, particularly when using multiple sensors.
Research Findings
The authors demonstrated that using ML and DL techniques significantly improved the accuracy of PD localization. They presented detailed results for various case studies, highlighting the performance of each method. For instance, in case study 1, a single monopole antenna was used to receive the PD signal, and 600 Monte Carlo simulations were conducted to evaluate the model's performance. The outcomes showed that the CNN model outperformed the other methods in terms of location accuracy and robustness, with an average CC of 0.99 and an average RMSE of 0.02 m for the testing data. The other methods had lower CC and higher RMSE values.
The CNN model also demonstrated a high degree of generalization and adaptability to different scenarios, effectively handling various sensor positions, PD signal frequency contents, and transformer tank sizes without significant loss of accuracy. This is because the CNN model can learn the complex relationship between the electric field signal features and the PD location, capturing the spatial and temporal patterns of the PD signals. The model uses a combination of convolutional, pooling, and fully connected layers to process the input data and output the location coordinates.
Moreover, when CNN was combined with transfer learning techniques, it provided the highest accuracy in localizing PD sources. In scenarios involving multiple sensors, the accuracy was further improved. For example, in case study 4, which uses a larger transformer tank and three sensors, the RMSE error decreased significantly when using CNN models with transfer learning. The paper also showed that increasing the number of sensors and employing advanced preprocessing methods could enhance the model's performance.
Applications
The study offers numerous valuable applications for the power industry and research community. It has the potential to simplify the measurement system by reducing the number of sensors required, while simultaneously enhancing the reliability and accuracy of localization results. The developed methods can be further extended and refined to address more complex and realistic scenarios, including multiple PD sources, noisy environments, and various transformer types and configurations.
Furthermore, the findings can significantly assist power utilities and operators in monitoring and maintaining their transformers, thereby preventing costly and hazardous consequences. It can also be applied to other fields that require precise source localization, such as medical diagnostics, acoustic monitoring, and electromagnetic studies.
Conclusion
In summary, the paper demonstrated the significant potential of ML techniques, particularly the CNN model, for accurate and reliable PD localization. The researchers showed that the CNN model could effectively handle various attributes and scenarios without compromising performance, making it a valuable tool for power transformer maintenance and safety.
Moving forward, future research could focus on integrating these techniques into real-time monitoring systems, which would further enhance the reliability and efficiency of electrical power networks. By leveraging the power of ML and DL, the industry can proactively address potential issues and minimize the risk of costly and hazardous failures.
Machine Learning Powering Breakthroughs in Climate Forecasting and Modeling