Accurate predicted the downlink traffic contributes to traffic load balancing and information security management in educational resources grid. Wavelet neural network is suitable for modeling and nonlinear prediction in grid downlink traffic which has the randomness and uncertainty characteristic. General wavelet neural network prediction model had some defects such as convergence slower, larger error and poor stability. In order to eliminate or improve the existing defects, a momentum was added in the scheme which was used to adjust the network weights and parameters based on gradient descent algorithm, meanwhile, an improved algorithm with random sample replacement mechanism in temporarily prediction results was proposed. Experimental results show that the proposed algorithm can reduce the convergence time in network training and improve the prediction accuracy and stability.