Speakers
Description
Insulating systems of high-voltage electrical devices used in the power industry are exposed to various types of exposure during their operation. Overvoltages in the nature of lightning surges, resulting from the direct or indirect impact of atmospheric discharges - lightning, constitute a certain group of these exposures. For this reason, the need to perform tests and design and acceptance tests of insulating systems of devices that may be subject to such impacts is justified. Insulating materials, insulators, insulating systems of switches, transformers, rotating machines, cables, etc. are tested. The aim is to verify the structural correctness of a given type of device or a specific product by checking the electrical strength of its insulating system with a test voltage of the appropriate shape and value according to standards.
The paper focuses on exploring the potential application of neural networks for the classification of voltage surges compliance with the norm. Therefore, the training dataset comprising 269 voltage surges - 134 correct ones and 135 with incorrect parameters was generated. Three potential neural network architectures were considered for the task - a convolutional neural network (further referred to as CNN), a model combining convolutional and LSTM layers (CNN+LSTM) and a transformer model.
Due to the small dataset size, network models have relatively few layers. For the same reason, data were preprocessed using a sub-sampling of series from 40000 to 1000 points. In the next step, z-score normalization was applied to highlight the differences between the individual time series. The data set was randomly split in an 80:20 ratio to training and test datasets, respectively. Twenty per cent of training data created a validation set.
All three network models were trained by 42 epochs with a batch equal to 32, with an early stopping criterion applied if validation loss did not decrease for the following ten epochs. The best results were achieved by the CNN+LSTM model (accuracy of 87% on the test dataset), followed by the simple transformer model (accuracy: 81.5%), and CNN (accuracy: 72%).
The preliminary results described above show that neural networks can be successfully applied to the task of voltage surge validation and allow for a quick assessment of whether the generated impulse meets the appropriate standard.
In future work, the authors aim to expand the dataset by collecting additional time samples and generating extra data through artificial augmentation. They also intend to improve the current model by hyperparameters tuning and considering other types of networks, such as those treating the input as a two-dimensional image.
BIBLIOGRAPHY
[1] Flisowski Z.: Technika wysokich napięć, WNT, Warszawa, 1992
[2] IEC 60060-1. High-Voltage Test Techniques. Part 1: General Definitions and Test Requirements
[3] IEC 60060-1. High-Voltage Test Techniques. Part 2: Measurements Systems
[4] L. Tong, Y. Liu, Y. Chen, S. Su and P. Liang, "A CVT Based Lightning Impulse Wave Measuring Method Using Convolutional Neural Network," 2021 IEEE 4th International Electrical and Energy Conference (CIEEC), Wuhan, China, 2021, pp. 1-6
[5] Figoń P., Analiza numeryczna przebiegów udarowych — algorytmy obliczeniowe, Przegląd Elektrotechniczny, no 10, 2015, pp. 201-205
[6] Yutthagowith P. Non-Iterative Technique for Determination of Full Lightning Impulse Voltage Parameters. Energies. 2022; 15(12):4199
[7] Tuethong P, Kitwattana K, Yutthagowith P, Kunakorn A. An Algorithm for Circuit Parameter Identification in Lightning Impulse Voltage Generation for Low-Inductance Loads. Energies. 2020; 13(15):3913
[8] Keras: Deep Learning for Humans: https://keras.io/
[9] TensorFlow: https://www.tensorflow.org/
