Power transformer condition assessment using an immune neural network approach to Dissolved Gas Analysis

One commonly used engineering tool for condition assessment of power transformers is Dissolved Gas Analysis (DGA) which can detect internal and incipient faults and can be done without disrupting the operation of the transformer. The drawback of DGA is that the conventional methods that are used to interpret DGA test results have limitations. To address the limitations of the conventional methods, a combined Artificial Immune System (AIS) and Artificial Neural Network (ANN), called an Immune Neural Network, is used in this paper as an alternative approach for condition assessment of transformers. Radial Basis Function Neural Network (RBFNN) is used for nonlinear mapping of DGA data inputs to different transformer health conditions such as normal condition and faulty conditions involving internal arcing, localized overheating, partial discharge activity, or multiple faults. DGA data inputs include concentrations of five dissolved gases (hydrogen, methane, ethane, ethylene, and acetylene) in transformer oil, gas generation rate in ppm/day, and gas ratios. An immune system-inspired model known as the aiNet model is used to determine the centers of the RBFNN. The aiNet is compared to random selection and k-means clustering in determining the RBFNN hidden centers. It is proven in the study that the aiNet has better training convergence and has an advantage over k-means due to non-empty clusters results. The study also showed that unlike conventional methods, the Immune Neural Network approach always gives a definite diagnosis, and it has better diagnosis accuracy for normal, single-fault, and multiple-fault transformer conditions.