Traffic behavior in large-scale networks is a rather complex nonlinear system, and there is no mature method to study it at home and abroad. Considering the nonlinear characteristics of network traffic, this paper decomposes the traffic time series into trend component, periodic component, abrupt component and random component through different mathematical models. According to the decomposition, four relatively simple sub-components are modeled with corresponding mathematical tools to simulate complex traffic. The decomposition model is used to analyze the long-term traffic behavior of CERNET backbone network and NSFNET backbone network, and the analysis results are compared with the traditional ARIMA seasonal model. By comparing the simulation autocorrelation function with the prediction error, it is found that the decomposition model has the advantages of simplicity and high accuracy in describing the macro-behavior of traffic. Fractal or self-similar model can describe the long-term dependence of network traffic process well, but it can't really describe the singularity of network traffic process on a small scale. From another angle, this paper studies and establishes a waterfall model based on waterfall process, and simulates and analyzes the real network traffic data. It is found that waterfall model can explain the scale characteristics of actual network traffic to some extent and accurately describe the singularity of network traffic on a small scale. Through the scale analysis of simulated data, it is proved that waterfall model has the ability to describe the multifractal characteristics of real traffic data.