Transformer fault diagnosis using CatBoost optimized with LASSO and PLO

Abstract

To address the issues of high-dimensional redundancy in features and the difficulty in optimizing diagnostic model parameters in traditional transformer fault diagnosis techniques, this paper proposes a new method for transformer fault diagnosis based on the Least Absolute Shrinkage and Selection Operator (LASSO) and the Polar Light Optimization (PLO) algorithm to optimize CatBoost. First, five typical fault characteristic gases are extracted using Dissolved Gas Analysis (DGA) technology and expanded to 20 dimensions using the ratio method; subsequently, the LASSO algorithm is employed for sparse dimensionality reduction to streamline the fault feature set. Second, a CatBoost-based diagnostic model is constructed, and the PLO algorithm is used to perform adaptive global optimization of its core hyperparameters, establishing a PLO-CatBoost joint fault diagnosis model. Experimental results demonstrate that the proposed method achieves a comprehensive fault diagnosis accuracy of up to 99.44% for transformers, representing improvements of 6.66%, 1.11%, and 2.22% compared to the original CatBoost, WOA-CatBoost, and SSA-CatBoost models, respectively. This study provides an efficient and reliable solution for the complex multi-class fault diagnosis of oil-immersed transformers and holds strong potential for engineering applications.