DATA ANALYTICS FOR TRANSFORMER DISSOLVED GAS ANALYSIS TO AID ASSET MANAGEMENT

Abstract

A power transformer is one of the most expensive assets owned and operated by a power utility. Hence, asset managers are keen to know the health status of each unit in a large fleet of transformers in order to take proper and timely maintenance actions and avoid catastrophic failures. In this respect, Dissolved Gas Analysis (DGA) is an effective transformer diagnostic technique with proven reliability. This work analysed DGA databases gathered by a major power utility in the UK, including around 60,000 DGA records from over 700 transmission power transformers. Through graphical and statistical analyses, long-term DGA patterns concerning new transformers up to over 60 years in service were examined. Hydrogen, Methane, Ethane, and Ethylene showed no influence of transformer service age on the gas level, but an increasing trend was clearly observed in Carbon Oxide gases. In addition, stray gassing of Hydrogen and Ethane in hydro-treated uninhibited oil and elevated Ethylene and Carbon Oxide gas records were observed in certain transformer families commissioned before the 1970s. These observations support to incorporate the service age and design families in DGA interpretations. Estimating the rate of change in gas level (or gassing trends) causes difficulties, mainly due to the uncertainty of the DGA data produced by gas level fluctuations and changes in DGA sampling frequency. Therefore, this study developed an automated data analytical technique to determine gas trends. This technique can be used as an asset management tool to assess gassing trends in a large fleet of transformers confidently. In extending the study with this technique, population analysis of gassing trends was evaluated using the in-service transformers and three scrapped transformer groups, which were categorised as transformers with faults of dielectric (D), overheating of insulating paper (TP) and overheating of insulating oil (TO). The results showed trend characteristics in the scrapped categories D and TO significantly different from the scrapped category TP and in-service transformers. The scrapped category D showed the generation of five gases, Acetylene, Ethylene, Methane, Ethane and Hydrogen, together beyond a set of thresholds. The same gases except Acetylene were observed in the scrapped category TO. However, the scrapped category TP and in-service transformers showed mostly none, occasionally with one to two simultaneous gas generations. Therefore, this study observed that it is challenging to identify DGA-related symptoms relating to the overheating of paper insulation in free-breathing transformers. Furthermore, an anomaly detection technique was introduced to identify anomalies in the DGA data. This technique enables asset managers to prioritise transformers, which need close monitoring. In addition, two numerical methods to examine statistical validations for abnormal gas behaviours within large databases were presented based on two field experiences. One case showed an abnormal rise of Hydrogen and Carbon monoxide after the passivation in less than 10% of the passivated transformers. The other case showed an unusual Ethylene production in 66% of oil-reclaimed transformers. Therefore, these abnormal observations must be considered to avoid misleading interpretations of DGA in facilitating asset management decisions after oil treatments. Finally, the feasibility of machine learning models in DGA status (acceptable or significant gas production) classification was investigated. Accordingly, the Random Forrest model, with a novel feature extraction method, including the gas trends, was built, and its diagnostic accuracy reached 97.5%. Especially, the obtained accuracy was an increment of around 25% over the conventional feature set concerning only the gas levels. This developed machine learning model helps autonomous DGA monitoring programs where the necessary human expertise is lacking.

Date of Award	1 Aug 2023
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Zhongdong Wang (Supervisor) & Qiang Liu (Supervisor)