The problem of multiparameter, nonlinearity and complex mechanism in process industry limits the performance of shallow algorithm, and therefore the deep learning is introduced into the processing industry predictive modeling. However, feature mining is insufficient for diversity data by single deep network, so an improved stacked autoencoder is proposed. Firstly, attribute of input is divided into several classes by clustering algorithm, and then the parallel sparse autoencoder is entered to detect feature locally. The parallel output is integrated into the following deep networks to extract feature layer by layer and get the fitting results. To overcome predictive error due to the overfitting, the “dropout” technique is introduced. In the prediction modeling of hydrocracking, the presented algorithm has better prediction level and generalization ability.