Identifying Faults in Power Transformers Based on Machine‐Learning Algorithms Compared With Other Techniques

ABSTRACT This paper implemented a comprehensive variety of modern machine‐learning techniques, which were demonstrated to be effective in handling complex tabular data, generating accurate predictions, and ensuring high dependability in fault classification of oil‐immersed power transformers. The primary objective of utilizing the techniques is to improve computing efficiency and generalization by using neural networks, gradient boosting models, and hybrid ensemble techniques. Light Gradient Boosting Machine (LGBM) was selected due to its rapidity and memory efficiency algorithms, integrating histogram‐based learning and leafwise growth strategies to allow the effective handling of large‐scale data sets. In parallel, Categorical Boosting was implemented because of its powerful ability to manage categorical data through adaptive categorical encoding and ordered boosting. These methods contribute to mitigating biases and overfitting. According to study comparisons, LGBM performs more accurately than several other boosting algorithms in a variety of real‐life situations. Additionally, a Multilayer Perceptron was applied to capture complex nonlinear interactions that tree‐based models frequently miss. On the basis of utilizing 700 samples reported by the International Technical Committee 10, Central Chemical Laboratories at Egyptian Electricity Holding Company, and related databases, it is concluded that the hybrid LGBM + Extra Trees Classifier model achieved the best performance, with accuracy values ranging between 93% and 98%. Furthermore, the accuracy of the hybrid LGBM + ET has been compared with the accuracy of different conventional and artificial intelligence techniques and has proven to be more accurate.