Estimating the Transformer Health Index Using Artificial Intelligence Techniques

Abstract

Transformer Asset Management (TAM) is concerned with the strategic activities that monitor and manage the transformer asset in the power system. The outcomes of TAM aim at setting proper monitoring methods and maintenance plans, with minimal cost of time and money. Monitoring methods in the form of electrical, chemical and physical tests are conducted to assess the transformer operational condition. The main part, which is directly related to the ageing of the transformer, is the oil-paper insulation system. The standard practiced monitoring test methods used by TAM companies are considered highly effective and useful. However, a full feedback of the transformer's condition requires a number of monitoring tests to be conducted. Such an exercise is considered expensive and difficult to implement for some of the tests. Moreover, the individual conducted tests cannot provide a comprehensive understanding of the transformer condition based on a single factor. Thus, the concept of the Health Index (HI) was developed to accurately assess the transformer's condition and effective remnant age. The main components involved in the HI computation are related to the transformers' insulation condition, service record and design. Finding the transformer HI is normally done through using several industry computational methods. The drawback of these methods is the large number of tests required to achieve high level of condition assessment accuracy. Thus, alternative Artificially Intelligent (AI) methods should be used to design the HI model. AI methods, such as Artificial Neural Networks (ANN), can learn the pattern of the response output (HI), based on a given set of input (monitoring tests). The use of feature selection technique such as stepwise regression, can lead to an effective reduction of redundant tests in the presence of more significant ones. The presented work produces a general cost-effective AI based HI predictor model that can be used by different utility companies. Such a predictor would be able to produce a HI output value with a 95% prediction accuracy using only a subset of the required input features. Furthermore, the model can produce the same prediction accuracy with a predicted costly feature as one of the input features.