A Private Strategy for Workload Forecasting on Large-Scale Wireless Networks

Abstract

The growing convergence of various services characterizes wireless access networks. Therefore, there is a high demand for provisioning the spectrum to serve simultaneous users demanding high throughput rates. The load prediction at each access point is mandatory to allocate resources and to assist sophisticated network designs. However, the load at each access point varies according to the number of connected devices and traffic characteristics. In this paper, we propose a load estimation strategy based on a Markov’s Chain to predict the number of devices connected to each access point on the wireless network, and we apply an unsupervised machine learning model to identify traffic profiles. The main goals are to determine traffic patterns and overload projections in the wireless network, efficiently scale the network, and provide a knowledge base for security tools. We evaluate the proposal in a large-scale university network, with 670 access points spread over a wide area. The collected data is de-identified, and data processing occurs in the cloud. The evaluation results show that the proposal predicts the number of connected devices with 90% accuracy and discriminates five different user-traffic profiles on the load of the wireless network.