Deep Learning for Spectrum Prediction From Spatial–Temporal–Spectral Data
Xi Li, Zhicheng Liu, Guojun Chen, Yinfei Xu, Tiecheng Song
Abstract
Spectrum prediction is challenging owing to its complex inherent dependency and heterogeneity among the spectrum data. In this letter, we propose a novel end-to-end deep-learning-based model, entitled spatial-temporal-spectral prediction network (STS-PredNet), to collectively predict the states of various frequency bands in all locations of interest at the same time. More specifically, the predictive recurrent neural network (PredRNN) is trained to capture the spatial-temporal-spectral dependencies of spectrum data. Three components of PredRNN units are employed to model the three kinds of temporal properties in spectrum data, i.e. closeness, daily period, and weekly trend. The final prediction is then performed in a dynamically aggregated way. Extensive experiments are conducted based on a real-world spectrum measurement dataset, which illustrate the superiority of the proposed STS-PredNet over the state-of-the-art baselines.