Firefly algorithm, whose convergence speed and precision is improved by chaotic optimization theory and adaptive variable step size mechanism, is introduced to achieve fault diagnosis of power transformer efficiently. A power transformer fault diagnosis method based on improved firefly algorithm (IFA) and multi-classification support vector machine (SVM) is proposed. The method uses IFA to optimize the parameters of SVM. Multi-class SVM is constructed based on binary tree method to identify power transformer fault types in the method. The simulation results of power transformer fault diagnosis examples show that the convergence and optimization ability of IFA are better than those of FA and particle swarm optimization (PSO), and the optimized power transformer fault diagnosis model has higher accuracy.