Long-Term Spectrum State Prediction: An Image Inference Perspective

Abstract

Spectrum prediction techniques have drawn much attention for enabling the dynamic spectrum access. As new algorithms emerge endlessly, most of them can only predict the future spectrum states in a slot-by-slot manner. A new thought to realize the long-term and comprehensive spectrum state prediction efficiently is deserving our exploration. In this paper, we formulate the spectrum situation of multiple frequency points or bands in a whole day with multiple time slots as an “image” and propose an idea of image inference to predict the spectrum situation of a whole day in the future based on multiple “images” composed of historical spectrum data. First, we model a new kind of three-order spectrum tensor and convert the spectrum prediction problem to a tensor completion problem. We analyze the impacts of prefilling proportion and the parameter $m$ of the third dimension on the prediction performance via an illustrative example of predicting a mosaic image. Then, a new long-term spectrum prediction scheme based on tensor completion (LSP-TC) is developed. Experiments with real-world satellite spectrum data demonstrates that the proposed LSP-TC is superior to the benchmark scheme in both the accuracy and the runtime overhead of prediction.