Spectrum prediction via long short term memory

Abstract

Spectrum prediction algorithm has always been a research hot spot of communication because it can reduce considerable time and energy consumption when the receiver senses channel states. Users only sense those channels whose prediction result of channel state in the next time slot will be idle, thus reducing the number of perceptual processes. Many spectrum prediction algorithms have achieved good performance, and with the rise of deep learning it will be a good innovation research in the application of spectrum prediction. A prediction model composed of Long Short Time Memory (LSTM) layers is constructed in this paper and then is trained through supervised learning before prediction. The output of LSTM network is transformed into 0 or 1 which indicates idle state or occupied compared to the threshold. Therefore, the maximum accuracy can be achieved under certain threshold settings. Influence of different LSTM network depth and width on the prediction accuracy is also studied in this paper together with Back Propagation (BP) neural network performance. Results show that LSTM network has better performance than BP network under the condition of same number of hidden layers and neurons. When there are something wrong with spectrum sensing, channel states with error and correct channel states both can be used as learning labels. The simulation results indicate that prediction accuracy for the latter one is about 4% better than the previous one.