Prediction-Based Spectrum Access Optimization in Cognitive Radio Networks

Abstract

Cognitive radio (CR) has received wide attention for enhancing the spectrum utilization. Spectrum prediction can be fully utilized in both time and frequency domains for cognitive access. In this paper, Long Short-Term Memory (LSTM) networks method is adopted for spectrum prediction. Based on prediction of the power of future time slots, the method can help to achieve spectrum utilization flexibility by calculating throughput of shared channels. A new spectrum access strategy, which integrates optimal spectrum sensing interval in time domain and channel selection based on the presented LSTM networks method in frequency domain, is proposed. In particular, for multiple channels of the shared spectrum, the optimal sensing interval of each channel is calculated, which is then adopted as the reference output length of the LSTM networks method, with predicted results, a feedback which consists of a preferred channel list will be conducted for spectrum access. With both real-life and generated data, the proposed LSTM networks method is verified, which outperforms commonly used Neural Network (NN) and Hidden Markov Model (HMM) methods. Numerical results also show that the proposed spectrum access strategy is capable of increasing throughput of the cognitive system and reducing energy consumption of spectrum sensing significantly.