A Stairway Statistical Neural Model for DGA Analysis
Abstract
This paper proposes a new approach for power transformers dissolved gas analysis (DGA) using Statistical Machine Learning Techniques and Neural Networks to compose a stairway model which performs analysis in three levels in order to check the existence of faults and which type it most probably is. The proposed approach shortcuts the problem of lacking reliable data related to the type of fault creating a model with three levels of analysis. The first one uses real data from an energy company and from IEC TC 10 data to classify the DGA samples as faulty or normal. After that, a second one based just on IEC TC 10 takes place to classify three possible types of the fault. The third level is used to classify 5 types of fault in a more detailed analysis. The proposed levels of the model achieved an accuracy in the test set of 100 %, 94 % and 92 % respectively.