Dwarf nova (DN) is a special and rare class of semi-contiguous binary star. To discovery more DNs is significant for the further study of matter transfer theory. It also has been profound for understanding the evolution of close binary stars. It is a research hot spot to extract features of celestial spectra and then classify them by deep learning. Traditional auto-encoder is a classical neural network model with only one hidden layer. However, its coding ability is limited and data representation learning ability is insufficient. Broadening the depth of the neural network with modularity can make the network learn features of the celestial spectrum successively. High-level features can be obtained through gradual abstract learning of underlying features so as to improve the spectral classification accuracy. In this paper, a deep feedforward stack network is constructed consisting of an input layer, several hidden layers and an output layer on the basis of auto-encoder. This network with multi-layer perceptron architecture is utilized to process massive spectral data sets. It excavates the depth structure features hidden in the spectra and realizes the accurate classification of DN spectra. Parameters set for the network with deep architecture will seriously affect the performance of the constructed network. In this paper, the optimization of network parameters is divided into two processes: hierarchical training and inverse propagation. The preprocessed spectral data first enter the network from the input layer, and then the network parameters are trained layer by layer with the auto-encoder algorithm and weight sharing policy. In the reverse propagation stage, the initial sample data are input into the network again, and the network is initialized with the weights obtained from the hierarchical training process. Then the local optimization training results of each layer are fused and the network parameters are adjusted according to the set output error cost function. Hierarchical training and inverse propagation are operated repeatedly until the global optimal network parameters are obtained. Finally, the last hidden layer is adopted as the reconstruction layer to connect the support vector machine classifier, and the feature extraction and classification of DNs are realized. In the process of network parameter optimization, the idea of mean network is utilized to make the output of network hidden layer unit attenuate according to dropout coefficient. The reverse propagation algorithm is adopted to fine-tune the entire network to prevent depth overfitting in the network. Such operation can reduce to extract duplicate feature caused by mutual moderation of hidden layer neurons, and improve the generalization ability. The distributed multi-layer architecture of the network can provide effective data abstraction and representational learning. The feature detection layer can learn the depth structure features implicitly from the unlabeled data, and effectively characterize the nonlinearity and random fluctuation of spectral data, thus avoiding the explicit extraction of spectral features. The network shows strong data fitting and generalization ability. Weight sharing between different layers can reduce the interference of redundant information and effectively resolve the risk that the traditional multi-layer architecture network is prone to fall into the local minimization of weight. Experiments show that that of the accuracy of the deep architecture network in DNs classification is 95.81%, higher than the classical LM-BP network.