Abstract: Methods to achieve the smallest sized network which can learn the training data within a given error bound are analyzed.Rough sets theory is applied to construct neural networks.A pruning algorithm for RBF(radial basis function) neural network based on rough sets is proposed,and this algorithm is compared with the existing methods.The proposed algorithm is applied to build the dynamic model of the superheated steam temperature in thermal process.Simulation is made,and the results show that the neural model based on this algorithm is of high approximation accuracy and good generalization ability.