The rapid development of information technology leads to the fact that the number of users in telecommunication networks is constantly growing along with the need to improve the quality of service. Network traffic prediction is an important task in this area, as it underlies network diagnostics and efficient use of its resources. Various linear and nonlinear methods such as neural networks are actively used to analyze temporal and spatial relationships and forecast network traffic. However, most of them cannot accurately describe the dynamics of the network, as the importance of different nodes changes over time, which complicates the topology. This article proposes to modify autoregressive and gradient boosting models to detect spatial features and work with data with network structure in order to solve the above problem. Experimental results on three publicly available datasets with network traffic show that the proposed methods are superior to their one-dimensional counterparts and can compete with the most modern solutions. Additionally, it was found that the logarithmic transformation significantly increases the accuracy of the forecast, and models based on decision trees are superior to autoregressive ones. Also, increasing the size of the training sample does not always improve the accuracy of the forecast. Moreover, singular spectrum analysis is superior to exponential smoothing and moving average for Internet traffic. The performance of the proposed models achieved MAPE values of 4.4% and 8.9% on the PeMSD7 dataset for gradient boosting and autoregression methods, respectively. Thus, using these models as a forecasting tool will help optimize complex network systems and improve the quality of their service.