Dissolved gas analysis (DGA) is significant for early warning and diagnosis of transformer faults. To enhance the accuracy and reliability of transformer fault diagnosis, a transformer fault diagnosis method is proposed based on the tree-structured parzen estimator (TPE) algorithm to optimize the light gradient boosting machine (LightGBM). Firstly, a 16-dimensional DGA feature set including gas ratios and encodings in oil is established, and the least absolute shrinkage and selection operator (LASSO) algorithm is used to select effective feature quantities for transformer fault diagnosis. Secondly, a transformer fault diagnosis method based on LightGBM is constructed, and the TPE algorithm is introduced to optimize the parameters of the LightGBM diagnosis model, forming an optimal fault diagnosis model. Finally, evaluation metrics such as accuracy, recall, and F1 score are selected to assess the performance of the proposed diagnosis model. The research results indicate that the average accuracy of TPE-LightGBM is 90.23%, and its diagnostic accuracy and robustness are superior to algorithms such as RF and XGBoost. At the same time, compared with the commonly used three-ratio method in practice, the proposed method shows significantly improved accuracy and reliability. This method can effectively enhance the level of intelligent operation and maintenance of power transformers.