At present, China's urban rail transit is developing rapidly. Short-term traffic flow prediction plays an important role on the metro safety,efficiency and cost. Many existing approaches with shallow architecture failed to provide favorable results, because short-term traffic flow are highly random, cyclical, correlative and non-linear. Therefore, we propose a prediction model with deep architecture---DBN-P/GSVM based on the deep believe network(DBN) and support vector machine(SVM). The parameters of the SVM are obtained based on the genetic algorithm(GA) and the particle swarm optimization(PSO). Abundant experiments are conducted on the Chengdu Metro North Railway Station. The results show that the proposed DBN-P/GSVM model performs better than such shallow architecture models as the GA-SVM, PSO-SVM and back propagation neural network(BPNN) and such deep architecture models as the long short-term memory(LSTM) and LSTM-Softmax in terms of mean squared error, root mean square error, mean absolute error and mean absolute percentage error.