Proactive Resource Allocation in Spectrum Sharing with Radar Systems via ML-based Wireless Network Time Series Forecasting

The prediction of wireless network parameters can help to facilitate proactive resource allocation solutions not only in cellular networks but also in enterprise networks. In this thesis, temporal and spatial analysis of network traffic using real traffic data of an enterprise network are presented, where we use three different machine learning (ML-based) methods and two different classical methods for temporal forecasting traffic usage. The results show that no universal best forecasting method can predict the traffic usage of every AP in an enterprise wireless network. The ML-based combined models are also used for spatio-temporal forecasting, improving the forecasting performance of single AP traffic usage. However, analyzing and predicting only network layer traffic data is not enough to execute decisions for resource allocation since the channels utilized in enterprise networks are shared unlicensed channels. Therefore, analyzing and forecasting the physical layer data of a channel are also investigated. In this thesis, we also improved the performance of conventional ML-based methods for time series forecasting by proposing a features-like grid training data structure which uses historical data as features.

After proving that physical layer data has more predictive power in the time series forecasting aspect with all forecasting models, physical layer data transmit power (TP) and interference prediction are applied in spectrum sharing with a radar system to improve the secure radar protection and efficient transmission of APs. In particular, weather radar operating in the 5.6 GHz band is considered a primary system and the secondary system is an enterprise network consisting of APs in a university campus. First, the transmit power time series of APs in the campus are modeled with multinomial distribution based on real collected data. Then, the aggregated interference due to the transmissions from the APs at the radar is predicted using LSTM with Monte Carlo (MC) dropout by considering model uncertainty. Finally, an efficient sharing and radar protection (ESRP) system based on two algorithms is proposed by using an averaged and predicted upper limit interference time series. The results show that the proposed ESRP system with an upper limit of interference prediction ensures radar protection with better throughput than the conventional radar protection systems. Moreover, better radar protection can be achieved with a small trade-off for the throughput of the secondary users by adjusting the MC dropout value used in the ESRP system.