Learnings

Outcomes

The aim of this project was to research and develop techniques suitable for transformer characterisation and objective interpretation of Frequency Response Analysis results. Mechanical winding faults are difficult to detect by conventional test techniques at 50Hz, i.e. leakage impedance. FRA was developed as an alternative and effective way to detect the mechanical integrity of windings, taking advantage of a wide range of frequencies and allowing geometric dimension of the winding to be reflected by winding impedance across a few Hz to a few MHz. The effectiveness of FRA heavily relies on the correct interpretation of results, based on the understanding of the relationship between winding geometry, transformer equivalent circuit and corresponding FRA fingerprints. The project has furthered our understanding of what may be achieved using different modelling techniques to identify transformer design features where they are not already known and to pick out defects inside the transformer, whilst being unmoved by changes in the results arising from minor differences in test measurement or construction.

The following three areas of research summarise the key outcomes:

• A mathematical expression was developed to depict FRA measurement results in a common format, i.e. a transfer function;
• Both supervised and un-supervised machine learning algorithms were applied to categorise winding construction types
• A “grey box” modelling approach was developed with the aim of producing a fault fingerprint library for an objective FRA modelling based interpretation, without the necessity of transformer design data. 

Looking at these in more detail, specific technical conclusions were as follows: 

Two methods for transfer function estimation were developed. The first method is called Feature Extraction Method. The key information - complex zeros and complex poles - are extracted from each frequency region and combined to form a Feature Transfer Function. Then a Difference Transfer Function is used to correct the deviation between the Feature Transfer Function and the measured data. This method was applied on different winding types and a satisfactory match for both magnitude response and phase response was achieved. The second method is called Extreme Points Iteration Method. The height and frequency location of the resonance and antiresonance were used to estimate the complex parameters. Iterations were conducted to eliminate the mutual influences between the complex parameters. This method has been successfully applied on the FRA traces from a Multiple Layer winding and a Plain Disc winding.

Machine learning methods were applied to identify transformer winding construction types. An SVM-based supervised method was proposed for transformer winding type recognition using FRA data. The model was built with FRA traces of 400/275/13kV transformers, and applied to 51 FRA traces with unknown winding type. The prediction result was validated with the classification made by expert experience. The unsupervised machine learning method Hierarchical Clustering was also used for the identification of winding types according to the distances between FRA traces. It was found that tertiary, common, and series windings can be roughly clustered together. It is possible to identify frequency responses from the tertiary windings, common windings and series winding, using 5 Hz to 2 kHz frequency response. Windings with high series capacitance and windings with low series capacitance, can be clustered together, using the 20 kHz to 200 kHz frequency response.

To produce a transformer “grey box” model without known design data, the GA optimisation methodology of constructing an equivalent circuit network was developed. A comparison with a corresponding “white box” model was made to verify the method. The work so far clearly indicates that the technical roadmap is feasible to conduct objective and modelling based FRA interpretation using the grey box model and a deformation fingerprint library developed based on the model.

Recommendations for further work

The work conducted for this project could be continues in the broad area of “digital twin” for transformers taking advantage of the development of AI techniques and optimisation algorithms. Working together with a transformer manufacturer willing to provide design data could result in the development of a digital twin covering electrical, mechanical and thermal aspects of a transformer over its lifetime.

Lessons Learnt

A structured approach to the collection and storage of FRA results is required to allow the use of pattern recognition and other computer analysis tools. The use of IEC 60076-18 standard measurement and data reporting would make the development and analysis of an FRA database significantly simpler in the future. While the diagnostic outcomes of different measurements made in the same way is consistent, measurements made in different configurations cannot be directly compared, which highlights the need for standardisation.

Dissemination

1. Xiaozhou Mao, Zhongdong Wang, Zanji Wang, and Paul Jarman, “Accurate Estimating Algorithm of Transfer Function for Transformer FRA Diagnosis”, 2018 IEEE Power & Energy Society General Meeting, Boston, July 2018. https://doi.org/10.1109/PESGM.2018.8586633
2. Bozhi Cheng, Peter Crossley, Zhongdong Wang, Paul Jarman and Andrew Fieldsend-Roxborough, “Interpretation of FRA Results through Low Frequency Transformer Modelling”, IEEE International Conference of Electrical Materials and Power Equipment, Guanzhou, April 2019. https://doi.org/10.1109/ICEMPE.2019.8727328
3. Xiaozhou Mao, Zhongdong Wang, Paul Jarman, and Andrew Fieldsend-Roxborough, “Winding Type Recognition through Supervised Machine Learning using Frequency Response Analysis (FRA) Data”, IEEE International Conference of Electrical Materials and Power Equipment, Guanzhou, April 2019. https://doi.org/10.1109/ICEMPE.2019.8727354
4. Xiaozhou Mao, Shuntao Ji, Zhongdong Wang, Paul Jarman, Andrew Fieldsend-Roxborough and Gordon Wilson, “Applying Unsupervised Machine Learning Method on FRA Data to Classify Winding Types”, 21st International Symposium on High Voltage Engineering, Budapest, Hungary, August 26-30, 2019. https://doi.org/10.1007/978-3-030-31676-1_91
5. Bozhi Cheng, Peter Crossley, Zhongdong Wang, Paul Jarman, Andrew Fieldsend-Roxborough, and Gordon Wilson, “Interpreting First Anti-resonance of FRA Responses through Low Frequency Transformer Modelling”, 21st International Symposium on High Voltage Engineering, Budapest, Hungary, August 26-30, 2019. https://doi.org/10.1007/978-3-030-31676-1_92
6. X. Mao, Z.D. Wang, P. Crossley, P. Jarman, A. Fieldsend-Roxborough, G.Wilson: ‘Winding Type Recognition through Applying Support Vector Machine on Transformer Frequency Response Analysis Data’, IET High Voltage Volume 5, Issue 6, December 2020 pp 704-715 https://doi.org/10.1049/hve.2019.0294 (open access)
7. Z.D. Wang, B.Z. Cheng, P.A. Crossley, D.M. Sofian J. Sanchez, ‘Fundamental understanding of Frequency Response Analysis ‘U shape’ through transformer modelling’, CIGRE India Colloquium, 2019.
8. Bozhi Cheng, Peter Crossley, Zhongdong Wang “Using Lumped Element Equivalent Network Model to Derive Analytical Equations for Interpretation of Transformer Frequency Responses” IEEE Access Vol. 8 https://doi.org/10.1109/ACCESS.2020.3027798 (open access)
9. B. Z. Cheng, X. Z. Mao, Y. X. Yang, Z. D. Wang, P. A. Crossley, and A. Fieldsend-Roxborough, “Factors dominating low frequency ‘V, ∩, U’shape features in transformer FRA spectra,’’ in Proc. IEEE Int. Conf. High Voltage Eng. Appl., Beijing, China, Sep. 2020, Paper 10840.