Traditional transformer fault diagnosis has some problems, such as magnanimous monitor data and poor classification of small example. Aiming at these problems above, this paper proposed a hybrid fault diagnosis method based on variable prediction model (VPMCD) and stacked denoising autoencoder (SDAE). Firstly, the chromatographic data of transformer oil was collected and normalized. Secondly, SDAE was trained layer by layer to achieve the high-level feature representation and the parameters of network structure; Then, the optimal model type and model parameters of fault type were obtained by training four mathematical models of VPMCD. Finally, the model was fined-tuned by a small amount of labeled data to obtain the optimital network parameters and complete the fault diagnosis. The results showed that the hybrid model has higher accuracy, stronger robustness and scalability.