Transformer oil is one of the main insulating material in power transformers. The density index of oil is closely related to the safe operation of transformers. On the basis of the principle of multi-frequency ultrasound, genetic algorithm GA and back propagation neural network BPNN, a prediction study of density of transformer oil is proposed. Taking 110 sets of transformer oil belonged to China southern power grid as an example, 100 of which are training sets and 10 are forecast sets, a prediction model of density of transformer oil is established based on BPNN, with the 242 dimensional multi-frequency ultrasonic data of oil sample as the input and density as the output, through the experimental method to determine the BP neural network hidden layer neurons number, the nonlinear mapping relationship, and use genetic algorithm GA to optimize the BP neural network connection weights and threshold of every layer. By adjusting the number of hidden layer neurons, the network is trained. Moreover, the genetic algorithm GA is introduced to optimize the network parameters. All results show that compared with the traditional standard BPNN model, the output value of density of transformer oil with the GA-BPNN model is much close to the real value with small errors, which lays a solid foundation to test transformer oil other parameters with tell multi-frequency ultrasonic technology.