This paper proposes a hybrid Bat-BP approach based on dissolved gas-in-oil data set (DGA) to optimize the structure of back propagation neural network (BPNN). BPNN is a multilayer feed forward neural network. The rule of local decline that BPNN used is easy to fall into local optimum. Bat algorithm is a metaheuristic bionic algorithm with great local performance, which is adopted to optimize the initial value of BPNN. The recommended Bat-BP method has been employed in power transformer fault diagnosis for the first time. To prove the proposed method has better ability of power transformer fault diagnosis, this paper compares the fitness of Bat-BP with BPNN and other optimized approaches including PSO-BP, GA-BP based on the same DGA data set. The mean squared error (MSE) is used in this paper to evaluate the performance of the total four methods. The experimental results show the Bat-BP has increased the fault diagnosis accuracy from 75.68% to 95.22%, which is higher than those optimized models.