Fault diagnosis of power transformers based on dissolved gas analysis and improved LightGBM hybrid integrated model with dual‐branch structure

Aiming at the fault diagnosis problems of imbalanced data and insufficient mapping of characteristic information in fault samples collected by transformers at present, which lead to low accuracy and large diagnostic deviation in actual applications, a power transformer fault diagnosis method based on dissolved gas analysis and an improved LightGBM hybrid integrated model with a dual‐branch structure (DIL‐DS) is proposed. Firstly, multi‐characteristic dissolved gas ratio analysis is used to construct multi‐dimensional supplementary feature vectors, which enrich the characterisation features of transformers and facilitate efficient diagnosis of classification models. Secondly, a dual‐branch structure combining focal‐gradient harmonic loss and cross‐entropy loss is introduced to improve the attention and recognition ability of the model to a few categories in the dataset and alleviate the influence of data imbalance on the diagnostic results. Then, an improved grey wolf optimisation (GWO) is designed to improve LightGBM and realise the iterative optimisation of hyperparameters. At the same time, the Jacobian regularisation method is introduced to denoise LightGBM to solve the problem that the model is sensitive to noise. Finally, the LightGBM hybrid integrated model is developed to ensure the accuracy and stability of model diagnosis under the changeable and imbalanced dataset. Experiments show that the proposed DIL‐DS can effectively solve the limitation of class imbalance, improve the overall fault diagnosis performance, and is suitable for transformer fault identification. The proposed power transformer fault diagnosis method DIL‐DS, based on DGA and the improved LightGBM hybrid integrated model with a dual‐branch structure, can be mainly divided into the following parts: construction of supplementary feature vectors, design of dual‐branch structure, improvement of the LightGBM model, and development of specific ensemble learning, which can improve the overall fault diagnosis performance of the model for imbalanced data of power transformers.