Combined with the support vector regression machine (support vector regression, the SVR) and Particle Swarm Optimization algorithm, (Particle Swarm Optimization, PSO), the paper proposes PSO-SVR model which is suitable for the learning of small samples. The paper selects the data of coal railway freight volumes and influence factors from 1995 to 2011 as the learning samples and then uses the particle swarm algorithm to optimize model parameters. The SVR model with good learning and generalization ability is established through training and testing．Setting up the BP neural network model and comparing the predicted value of both, the results show that the prediction accuracy of PSO-SVR model is superior to the BP neural network model.