Combining conventional and artificial intelligence DGA interpretation methods using optimized weighting factor

Abstract

Rogers Ratio Method (RRM), Duval Triangle Method (DTM), IEC Ratio Method (IRM) (Basic Gas Ratios Method) and GB/T 7252 (National Standard of the People's Republic of China) are popular conventional methods for interpreting the possible faults indicator of transformer. Each of them has its own characteristics and these methods may have different interpreting results for the same data. This research proposes combined methods by combining four conventional Dissolved Gas Analysis (DGA) methods (RRM, DTM, IRM and GB/T) and two artificial intelligence methods based on Restricted Boltzmann Machine (RBM) and Back Propagation Neural Network (BNN) using weighting factors. The values of weighting factors that used to generate the final results optimized by using Tabu Search (TS) optimization algorithm based DGA practical data. From 707 DGA practical data which the real faults are known, the combined methods can have higher accuracy than single methods. The highest accuracy of single methods is 91.80% from RBM based method and the highest accuracy of combined methods is 93.08%.