Summary
Dissolved gas analysis in transformer oil (DGA) is one of the most informative methods for diagnosing the condition of power transformers. The use of online gas analyzers enhances the capabilities of DGA, but algorithms based on machine learning methods are required to process large data sets. This paper presents a comparative study of several classes of machine learning models for classifying transformer defects based on data from online gas analyzers. The models were trained on labeled data obtained from online gas analyzers installed on operating power transformers.
Read more Read lessThe results of offline testing showed that multi-layer perceptron (MLP) models provide the highest classification accuracy (over 99%). However, experimental testing under conditions of streaming data processing from 40 transformers and limited computing resources showed that only the probabilistic neural network (PNN) model provides stable real-time operation.
Additionally, it has been demonstrated that the use of time series analysis models in combination with classification algorithms allows for the early detection of the development of defective conditions in transformers.
The prospects for applying machine learning methods in the absence of labeled data, in the presence of server or cloud infrastructure enabling the use of more complex forecasting and classification models, as well as for the diagnosis of transformers filled with complex esterbased dielectric fluids, are discussed.
Additional informations
| Publication type | Session Materials |
|---|---|
| Reference | D1_11225_2026 |
| Publication year | |
| Publisher | CIGRE |
| Country | Russian Federation |
| Study committees | |
| File size | 843 KB |
| Price for non member | 30 € |
| Price for member | 30 € |
Authors
GARIFULLIN Marsel - Kazan State Power Engineering University; RAKHMANKULOV Shamil - Kazan State Power Engineering University; GALIEV Ilgiz - Kazan State Power Engineering University
Keywords
Transformer, Diagnosis, Dissolved Gas Analysis, Monitoring, Machine Learning, Probabilistic Neural Network, Multilayer Perceptron, Time-Series, Real Time Processing
