Properties of 5G video conferencing traffic and its prediction by artificial intelligence methods

The objectives and prospects of short- and long-term forecasting of data traffic in 5G networks are outlined. An overview of current traffic forecasting methods is presented. A large number of works indicate a variety of ap-proaches to the analysis and forecasting of mobile traffic. These approaches include both traditional statistical methods and deep learning methods, which is an important factor for the development of network technologies and optimization of radio access networks. Forecasting and management plays an important role both at the level of the entire 5G network and at the level of its individual components. With the increase in the volume of big data, traffic forecasting becomes a difficult task due to mobility and different user behavior. In connection with this circumstance, using the example of mobile traffic data from Google Meet, MS Teams and Zoom video conferencing systems, the features of this type of traffic are analyzed and it is concluded that the distribution of time intervals between packet arrivals corresponds to a stable distribution. Based on the results of studying some machine learning methods designed to predict time series, a comparative analysis of the effectiveness of these methods for short-term forecasting of traffic intensity was performed. The following models were used for analysis and prediction: naive seasonal, exponential smoothing, linear regression, ARIMA, Theta and NBEATS. It was found that linear regression provides the best quality indicators for Google Meet traffic, the exponential smoothing method shows the best quality indicators for MS Teams, and the Theta method has the best indicators for Zoom.