Assessing the power transformer insulation health condition using a feature-reduced predictor mode

Condition monitoring (CM) techniques in the form of electrical, chemical and physical tests are conducted to assess the ageing of transformers on the basis of its oil-paper insulation strength. A thorough examination of the transformer insulation condition can only be done through a complete set of CM standard tests. Some of the conducted tests are considered expensive in terms of the required facilities and experienced personnel. The concept of the health index (HI) was developed to assess the transformer's health condition and effective remnant age as a function of all CM tests. The main components involved in the HI computation are related to the transformers' insulation condition, service record and design. Several computational methods have been developed to calculate the transformer HI. Each method has a different interpretation of how the tests are correlated with the HI value. The drawback of these methods is the associated cost of the required monitoring test inputs for an accurate HI computation. Thus, alternative methods are needed to predict the HI value with a smaller number of CM tests. Artificial intelligent (AI) methods, such as artificial neural networks (ANN), can learn how to map the response output (HI) to a given correlated set of inputs (monitoring tests). The presented work introduces a general cost-effective ANN based HI predictor model that achieved 95% prediction accuracy using only a subset of the required input features. Moreover, the developed model has been tested on a set of transformer database from another utility company and achieved 89% prediction accuracy.