The development of the data analysis method of deep learning provides a new solution and technical means for the state prediction of power transformers. This article proposes a prediction method of dissolved gas concentration in transformer oil based on modified deep belief networks (DBNs). First, the multivariate ReLu–DBN model is established with the multivariate dissolved gas as the input of networks. Then, the networks are pretrained with a CD-k algorithm and the network structure parameters are fine tuned with an ReLu function to weaken redundant information and speed up network training. Finally, the example data are put into the prediction model to verify the validity of the algorithm. The results show that the model proposed in this article can excavate the inherent law of the data itself and overcome the drawbacks of low stability in the traditional methods.