Power Transformer Diagnosis Based on Dissolved Gases Analysis and Copula Function

Xiaoqin Zhang, Hongbin Zhu, Bo Li, Ruihan Wu, Jun JiangEnergies 2022 · DOI

TL;DR

This paper introduces a novel approach to diagnosing power transformer faults using dissolved gas analysis (DGA) by employing copula functions to model the complex dependencies between different fault gases. This method offers a more robust and accurate diagnosis than traditional ratio-based techniques, particularly in identifying healthy versus defective transformer states.

What Problem Does This Paper Address?

Power transformers are critical components in electrical grids, and their reliable operation is paramount. Dissolved Gas Analysis (DGA) of the insulating oil is a common method for detecting internal transformer faults. However, traditional DGA methods often rely on simplified gas ratio codes and fail to adequately capture the intricate, non-linear dependencies between various fault-characteristic gases. This can lead to misclassifications and missed diagnoses. Existing advanced methods like neural networks and support vector machines, while powerful, can be data-hungry and require extensive prior knowledge for optimal configuration. This paper aims to bridge this gap by developing a more interpretable and data-efficient method that explicitly models the interdependencies of dissolved gases for improved fault diagnosis.

Methods

The core of this research lies in the application of copula functions for modeling the joint distribution of dissolved gases. Copula functions are statistical tools that allow for the separation of marginal distributions from their dependence structure, making them ideal for analyzing complex relationships between variables.

The methodology involves several key steps:

  • Variable Selection: Identifying the most relevant fault characteristic gases. The paper explores factor analysis for this purpose but ultimately settles on using pairs of gases due to limitations in factor analysis's effectiveness with the given data. Methane and ethane were chosen as the input variables due to their strong correlation.
  • Marginal Distribution Fitting: Fitting the individual distributions of the selected gases (methane and ethane) using generalized extreme value and generalized Pareto distributions.
  • Copula Function Selection and Fitting: Evaluating various bivariate copula functions (Frank, Gaussian, Gumbel, T, and Clayton) to model the dependence between the selected gases. Bayesian estimation is used to fit the parameters of these copula functions.
  • Goodness-of-Fit Evaluation: Using Root Mean Square Error (RMSE) and Nash-Sutcliffe Efficiency (NSE) to select the best-performing copula function.
  • Transformer State Diagnosis: Utilizing the fitted copula function to establish a relationship between the cumulative distribution probability (CDF) of the dissolved gases and the transformer's operational state (healthy, sub-health, or failure).

Datasets

The study utilizes field DGA data from power transformers. Two datasets were combined:

  • Dataset 1: Live detection data from medium and high-temperature overheat fault transformers, comprising 2073 data points from 27 transformers.
  • Dataset 2: Normal transformer live detection data, comprising 4691 data points from 76 transformers.

In total, data from 103 transformers were used, with most transformers meeting IEC 60296-2003 standards.

Key Insights and Ideas

  • Copulas as a Superior Dependence Model: The research highlights that copula functions are significantly better at capturing the complex, non-linear dependencies between dissolved gases compared to traditional ratio-based methods or simple correlation coefficients.
  • Hydrocarbon Gas as a Key Indicator: The study found that using hydrocarbon gases (specifically methane and ethane in their analysis) as the marginal variables in the copula function effectively delineates between healthy and defective transformer states.
  • CDF Threshold for Diagnosis: A critical insight is that a Cumulative Distribution Probability (CDF) value close to 0.8 in the copula function is a sensitive indicator. Fluctuations around this threshold can lead to sharp changes in the probability of a transformer being diagnosed as faulty or healthy, suggesting this region is crucial for early detection.
  • Data-Driven, Low-Prior-Knowledge Approach: Copula functions offer a powerful alternative to complex machine learning models that often require extensive domain expertise for architecture design and parameter tuning.

Key Results

  • The Frank copula function was identified as the best fit for modeling the joint distribution of methane and ethane, achieving the highest NSE (0.9897) and lowest RMSE (0.7690) among the tested copula types.
  • The study demonstrates that the copula function's CDF boundary can effectively separate healthy and defective transformer states.
  • A CDF value around 0.8 was found to be a critical point where small changes in gas concentration lead to significant shifts in the probability of a fault diagnosis.
  • The proposed copula-based method provides a more robust diagnostic capability, especially when compared to relying solely on CDF ranges, where errors can occur.

Limitations and Future Work

  • Factor Analysis Limitations: The paper notes that factor analysis, initially considered for variable selection, did not yield satisfactory results for the given DGA data, leading to the use of pairwise combinations of gases. This suggests that further exploration into more advanced feature selection techniques might be beneficial.
  • Binary Copula Focus: The study primarily focuses on binary copula functions (two variables). While effective, extending this to multivariate copulas to incorporate more gases simultaneously could potentially offer even greater diagnostic power.
  • Dataset Specificity: While the study uses field data, the effectiveness of the model might vary with different datasets and transformer types. Further validation on a broader range of data is always valuable.
  • Future Work: The authors suggest evaluating their approach with an increasing scale of DGA data and exploring its application in field scenarios for practical implementation. They also hint at the potential for integrating this method with other diagnostic techniques for a more comprehensive system.

Summary updated

Abstract

The traditional DGA (Dissolved Gas Analysis) diagnosis method does not consider the dependence between fault characteristic gases and uses the relationship between gas ratio coding and fault type to make the decision. As a tool of the dependence mechanism between variables, a copula function can effectively analyze the correlation between variables when it cannot determine whether the linear correlation coefficient can correctly measure the correlation between variable relationships. In this paper, the edge variable of a copula function is selected from the fault characteristic gas of a transformer, and the distribution type of the edge variable is fitted at the same time. Then, Bayesian estimation with the Gaussian residual likelihood function is used to fit the parameters of a copula function and a copula function is selected to describe the optimal dependence of the fault characteristic gas of transformer. The relationship between a copula function and the state of transformer is studied. The results show that the copula function boundary with hydrocarbon gas as edge variable can divide the transformer as healthy or defective state. When the cumulative distribution probability (CDF) value of the dissolved gas in the oil in the copula function is close to 0.8, the fluctuation of its gas concentration leads to a sharp change in the probability. Therefore, the analysis of dissolved gas in oil based on a copula function can be used as a powerful technical solution for oil-immersed power transformer fault diagnosis.
BibTeX
@inproceedings{zhang2022power,
  author = {Xiaoqin Zhang and Hongbin Zhu and Bo Li and Ruihan Wu and Jun Jiang},
  title = {{Power Transformer Diagnosis Based on Dissolved Gases Analysis and Copula Function}},
  booktitle = {Energies},
  year = {2022},
  doi = {10.3390/en15124192},
}
Figure