Abstract: There is a significant bias when imbalanced fault data are dealt with by transformer fault identification methods. To solve this problem, we have built a hybrid ensemble model based on an improved lightweight gradient boosting machine. First, we propose a gradient harmonizing mechanism loss and cross entropy loss improved light gradient boosting machine (GCLightGBM). This approach enhances the model's attention to minority samples in the dataset. Then, to address the issue of parameter specificity affecting the model's recognition capability in GCLightGBM, we propose a hybrid ensemble model based on GCLightGBM. This model further improves accuracy while ensuring good performance on real-world varied and imbalanced datasets. The experimental results show that GCLightGBM can effectively solve the problem of low accuracy of minority samples, and the overall accuracy is as high as 0.911. In addition, for other variable and unbalanced datasets, the fault identification method based on GCLightGBM hybrid ensemble model has an average accuracy of 0.988.