Spectrum occupancy prediction using deep learning techniques for next generation wireless communications

Due to the rapidly evolving needs of future wireless communication technologies and the increasing number of connected devices, there is an unprecedented level of demand for radio frequency spectrum resources and efficient communication infrastructure. As spectrum is a limited resource conventionally managed by fixed allocation strategies, it becomes challenging to meet these demands efficiently. Dynamic spectrum management assisted with spectrum prediction has been proposed in the literature to formulate a solution for this challenge and has developed into an important research area. In dynamic spectrum access (DSA) solutions, different spectrum management functions are performed by cognitive radios (CR) in order to gain awareness about the radio environment and take actions or decisions to allocate spectrum to unlicensed users in an opportunistic manner. These functions introduce non-idealities such as delays and reduce spectrum utilization efficiency. To solve these inefficiencies, spectrum occupancy prediction algorithms are adopted to enable proactive spectrum management. These algorithms estimate future spectrum characteristics by utilizing the patterns in the spectrum observations captured by the CRs. Based on the predicted outcomes, informed decisions can be made to manage the spectrum proactively and efficiently. This research centres around spectrum occupancy prediction using deep learning algorithms. The key objective of spectrum prediction is to enable unlicensed users in a DSA scenario to proactively operate over unused spectrum with minimal interference to the licensed users. Such a proactive spectrum management approach can offer many advantages, including improved spectrum utilization and energy efficiency, higher system throughput, etc. As a part of this research, we also extensively review the spectrum management techniques and the spectrum prediction methods such as conventional and machine learning (ML) based algorithms. The first contribution of this research aims at improving the performance of a deep learning based spectrum occupancy prediction model. In this contribution, we perform spectrum prediction in the temporal domain using deep learning due to its many advantages. To tackle the computational challenge of training the deep learning model, we develop an initialization methodology and validate its ability to improve the training overhead of the prediction algorithm. We analyse and model the parameters of the spectrum prediction algorithm based on Monte Carlo simulations. Based on the statistical analysis of the trained model parameters and the improved initialization method, we propose a spectrum prediction framework to achieve fast learning in deep learning based spectrum prediction applications. Our results indicate that the proposed methodology and framework improves the training convergence over simulated and real-world spectrum datasets, while achieving similar prediction performance as compared to conventional initialization methods. In the second part of this research, we focus on the performance analysis of an ML based spectrum prediction model in order to gain insight on how the prediction algorithm will perform in a realistic environment. Therefore, in the second contribution, we analyse the performance of the prediction algorithm theoretically and numerically by conducting statistical analysis and characterization of the outcomes. We also propose closed form expressions for different performance measures of the algorithm. Our results indicate that the proposed theoretical model and closed form expressions are able to accurately characterize the performance of the spectrum prediction model over simulated and measured spectrum datasets. In the third contribution, our aim is to extend the deep learning spectrum prediction algorithm to a multi-user cooperative radio environment. We apply an ML based model to perform local spectrum prediction and propose fusion rules to soft combine the local prediction outcomes to enhance the performance in different radio environment conditions. We also extract performance characteristic curves of the proposed fusion rules compared to local and hard fusion combining under different input model cases and channel conditions. Our results indicate that the proposed soft fusion rules can provide improvements in the prediction performance compared to local prediction and hard fusion combining in different radio environment conditions. The fourth contribution deals with multi-time step ahead spectrum prediction and performance analysis of different ML based prediction models. Our aim is to understand the performance abilities of the proposed models in predicting the spectrum occupancy status well ahead into the future. We analyse the prediction performance of the proposed prediction model for a range of predictive time steps, input model conditions, and probability of spectrum occupancy. We also obtain performance characteristic curves of different ML models to understand the impact of spectrum occupancy and compare them with that of Bayesian prediction. Our results indicate that the prediction performance deteriorates as the prediction window extends more into the future. However, for increasing value of spectrum occupancy, the prediction error shows a decreasing trend.