In order to solve the problem of low accuracy of fault diagnosis of oil-immersed transformers, a transformer fault diagnosis method of kernel principal component analysis (KPCA) with improved pelican optimization algorithm (IPOA) optimized least squares support vector machine (LSSVM) is proposed. Firstly, KPCA is used to extract features from multidimensional transformer fault data, reducing computational complexity. Secondly, logistic chaotic mapping, adaptive weight strategy, and lens imaging reverse learning strategy are introduced to improve the pelican optimization algorithm (POA). Finally, the KPCA-IPOA-LSSVM fault diagnostic model is established, and the diagnostic accuracy is 94.24%. Compared with the PCA-IPOA-SVM, KPCA-IPOA-SVM, KPCA-WOA-LSSVM, and KPCA-POA-LSSVM fault diagnostic models, the accuracy is improved respectively by 18.31%, 11.53%, 11.87%, 7.46%. The results show that the transformer fault diagnosis model proposed in this paper effectively improves the accuracy of fault diagnosis, proving that the diagnostic model has certain significance in theoretical research and practical engineering application.