Application of Self-Organising Map Algorithm for Analysis and Interpretation of Dissolved Gases in Power Transformers K.F. Thang R.K. Aggarvval University of Bath University of Bath Bath, BA27AY, UK Bath, BA27AY, UK Abstracf Onset of incipient faults in power transformers can degrade the mineral oil and cellulose insulation, leading to the formation of dissolved gases. The process from oil sampling to quantification of gases is known as dissolved gas analysis (DGA). Despite the availability of DGA interpretation schemes and artificial intelligence (AI) methods for transformer condition monitoring (CM) based on DGA data, it is pointed out in this paper that these approaches are less than ideal and practical in implementation, In view of that, this paper illustrates a novel approach for analysis and interpretation of DGA data, which leads to a more credible CM of power transformers. The proposed approach, which is based on the self-organising map (SOM) algorithm, has been validated using real fault-cases and thereby is proven to be more reliable in portraying the current condition of power transformers. Keywords: Incipient faults; Power transformer; Dissolved gas analysis; Artificial intelligence; Condition monitoring; Self- organising map. I. INTRODUCTIC)N Mineral oil in power transformcz-s consists of many different hydrocarbon molecules, Electrical and thermal faults can break-up bonds linking these molecules, which lead to the formation of gases. For a slowly developing or incipient fault, the gases formed will dissolve in oil, with only a small proportion diffusing from the oil into any gas phase above it,, This kind of fault can therefore be monitored using dissolved gas analysis (DGA). DGA requires the sampling of seven key gases as stated in IEC 599 Standard [1], viz. carbon dioxide (C02), carbon monoxide (CO), hydrogen (Ha), methane (CH4), ethane (CJHJ, ethylene (C2HJ) and acetylene (CJH2), As described in IEC 599 Standard, partial discharge (PD) occurs in the case of low-level energy, such as breakdown in gas-filled cavities resulting from incomplete oil-impregnation. In this case, the major gas produced is H2. In other types of fault, the decomposition of oil is mainly caused by heat. Decomposition occurs at normal operating temperature, producing mainly Hz and CH~. Higher decomposition temperature, resulting from thermal fault (TF) such as hot- spot and overheating, produces mainly CH4. With firther increases in temperature, an increasing amount of C2H6and CZH4 will be rsleased. In the case of a much higher temperature resulting from disruptive faults such as electrical discharges (ED), the production of C2H2becomes A. J. McGrail D.G. Esp The National Grid Company plc The National Grid Company plc Featherhead, KT227ST, UK Wokingham,RG31 5BN, UK significant. On the other hand, if cellulose materials such as paper, pressboard etc. are involved at the location of the fault, further gases, principally COZ and CO will also be generated. Owing to the credible relationship between the relative composition of dissolved gases and the type and severity of faults, several DGA interpretation schemes have been established, such as the well-known Dornenburg Ratios [2], Rogers Ratios [3], IEC Ratios [1] and Duval Triangle [4]. The implementation of these schemes, however, is less than ideal since different diagnosis of fault is often result. Attempts have been made to employ AI methods [5-9] for improving fault diagnosis and condition monitoring (CM) of power transformers. Although some of these approaches are still dependent on DGA schemes for their development, improvement over conventional schemes have been reported. A novel approach for analysis and interpretation of DGA data is proposed in this paper, Unlike the above stated methods, the proposed approach does not depend on DGA interpretation schemes for its development; only measured DGA data is needed. A brief description and comparison of DGA interpretation schemes and currently available AI methods for CM of power transformers is presented in first part of the paper. The second part of the paper concentrates on introducing the self-organizing map (SOM) algorithm, outlining the proposed approach, and presenting the performance and validation of the proposed approach. II. REVIEW OF DGA SCHEMES AND AI APPROACHES FOR CM OF POWER TRANSFORMERS A. DGA Interpretation Schemes Several well-known DGA schemes are, for example, Dornenburg Ratios [2], Rogers Ratios [3], IEC Ratios [1] and Duval Triangle [4]. These schemes have been implemented, either in improvised or modified format, by various power utilities throughout the world. The use of these schemes requires the computation of several key-gas ratios, as listed in Table 1. Fault diagnosis is achieved by associating these ratios with a set of pre-defined faults or conditions. 0-7803-7031-7/01/$10.00 (C) 2001 IEEE 0-7803-7173-9/01/$10.00 © 2001 IEEE 1881