With the development of computer and network technologies, the network traffic has come to an explosive growth. Considering the non-linearity of network data, we choose echo state network (ESN) to predict the data. ESN is a novel kind of recurrent neural networks, where a reservoir is generated randomly and only output weight matrix is adaptable. In our paper, we propose an improved network of ESN considering its weekness on selection of matrix initialization and activation function. The improved network defines the scope of matrix initialization and replaces the activation function of middle layer with wavelet function. And result shows that our improved network is more effective compared to original ESN. Our evaluation index is normalized mean square error (NMSE), and it drops from 0.7435 to 0.5852 by making improvements. network predicts well. Numerous experiments show that ESN has better adaptability to nonlinear time series.The rest of this paper is organized as follows. Section 2 describes previous related work about prediction of network traffic. In Section 3, we make an analysis of the data acquired from telecom operator. In Section 4, we introduce the structure of echo state network (ESN) and improvement of model. Finally, we give simulation results and conclusion in section 6. Related workBoth short-term and long-term prediction of network traffic are feasible directions for research. With historical traffic data and suitable method, we can have a general idea of the size of future traffic. With the prediction, operators can adjust the distribution of network resource in time. Also, equipment manufacturers can replace network devices with insufficient load capacity in advance.In recent years, there have been many technologies applied in the filed of traffic prediction. Pang et al [4] establised an adpative fuzzy traffic predicator based on the theory of fuzzy system [5] . They used nearest neighborhood clustering learning algrothim to present the fuzzy traffic predicator, and the results showed that the predicator was accurate and flexible applied in ATM networks. Sang et al [6] predicted network traffic using two stationary traffic models, including the Auto-Regressive Moving Average(ARMA) [7] model and the Markov-Modulated Possion Process(MMPP) [8] model. These two models mainly use mathematical analysis and experiments to do the predictability analysis of network traffic. Besides, wavelet analysis has been one of the most effective method to deal with non-stationary time series for traffic prediction. Wang et al [9] proposed a novel method of combining wavelet and RLS to forecast the Internet traffic and results showed this method achieved extraodinary accuracy comapred to other models. As we all know, Artificial Intelligence (AI) has been applied in many areas widely, such as expert system, fuzzy reasoning and fuzzy neural. Specially, there are many techniques about neural networks(NNs) used in the field of prediction. Neural networks are suitable for learing more complex nonlinear relationships. Mo...