Improvement Of Dissolved Gas Analysis Interpretation By Combining Duval Triangle Method With Machine Learning Techniques

Power transformer is one of the main components in any power transmission and distribution network. Transformers have complicated winding structures and are subjected to electrical, thermal and mechanical stresses. During the last few years, there has been a trend of continuous increasing of transformer failures. It is therefore vital to diagnose the incipient faults for safety and reliability of an electrical network. Thus, these transformers are needed to be routinely maintained. Due to the large number of transformers with different manufactures and capacities, routine maintenance and diagnosis of such transformers are difficult since different transformers exhibit different characteristics and problems. By means of dissolved gas analysis (DGA), it is possible to distinguish faults such as partial discharge (corona), overheating (pyrolysis) and arcing in oil-filled equipment. Dissolved gas analysis is one of the most effective tools for power transformer condition monitoring. There are several interpretation techniques for DGA results including Key Gas, Doernenburg, IEC Ratio, Roger’s Ratio and Duval Triangle. However, DGA interpretation is still a challenge issue as most of the techniques are relying on personnel experience more than standard mathematical formulation. As a result, various interpretation techniques do not necessarily lead to the same conclusion for the same oil sample. Furthermore, significant number of DGA results fall outside the proposed codes of the current based-ratio interpretation techniques and cannot be diagnosed by these methods. Moreover, ratio methods fail to diagnose multiple fault conditions due to the mixing up of produced gases. To overcome these limitations, this thesis proposes a new Artificial Intelligence (AI) approach to reduce dependency on expert personnel and to aid in standardizing DGA interpretation techniques. The approach relies on incorporating the Duval triangle method (DTM) with two machine learning classifiers named decision tree (DT) and random forest (RF), which the final interpretation will apply the voting combination method to get the final prediction of the incipient fault inside the power transformer. DGA results of oil samples where the real fault already known that were collected from different published papers were used to train and test the classifiers. The results demonstrate that combining the conventional method with artificial intelligence based on DGA interpretation methods gives reliable diagnosis of the incipient fault in the power transformer