Development of an Expert System for Power Transformers Fault Diagnosis Using Random Forest Technique

Thumbnail Image
Arar, Ghazi
Journal Title
Journal ISSN
Volume Title
جامعة النجاح الوطنية
This research will contribute to the field of power system protection. As traditional protection has failed to overcome its limitations to classify and discriminate different statuses of transformer, a need has risen to find new techniques to solve the problem. In this thesis, ensemble techniques are used to solve this issue. Hence, the differential protection constructed by using ensemble techniques to provide protection element via a trip and no trip actions. And further, conditional monitoring functions are used to distinguish five different statuses of transformer including normal, inrush, over-excitation, current transformer-saturation and internal fault. By capturing practical transformer rating models for 20 different transformers with 5 different operating cases, 100 examples were provided as a data set to train and test models with 1600 observations. The 100 original and raw data were used to train random forest, then it has been validated with internal measures including out-of-bag error, margin, confusion matrix, and outliers. Afterward, an updated and weighted data set was generated to be used in training and testing random forest. OOB error and margin were captured for weighted examples to be compared with raw examples. Different train to test, which are 80-to-20 and 60-to-40, have been used to validate system strength and reliability. Moreover, a faster version of random forest models constructed with different sizes of data window included ¾, ½, and ¼ cycles, resulted in an accurate protection and high accurate conditional monitor. Besides, two different versions of random forest in terms of individual trees depth have been tested concerning the greedy and limited size. Boosting technique was also applied to both, original data set and weighted data set with different train to test ratio including 80-to-20 and 60-to-40 to validate the model. And yet, the model has been tested conditioned with optimum number of trees by using out-of-bag error and cross validation folds. Due to that, the variable importance was achieved by using the optimum number of trees. It is worth noting that the variable importance was captured by using ensemble techniques, and therefore conclusion for signal importance at different instances investigated. In conclusion, random forest and boosting have shown promising results and showed the ability to classify the suggested problem. Thus, it provides accurate, fast, and reliable results.