Robust Long-Term Spectrum Prediction With Missing Values and Sparse Anomalies

Due to an increasingly instrumental role in dynamic spectrum access, spectrum prediction causes extensive concern. Recently, the long-term spectrum prediction scheme based on tensor completion (LSP-TC) was proposed, which performs prediction over a time-frequency-day spectrum tensor model. Nevertheless, LSP-TC suffers from missing data and sparse anomalies. In this paper, we show that the low-rank property of the tensor model may be destroyed by sparse anomalies, resulting in an invalid tensor completion. Therefore, robust tensor recovery (RTR) is introduced to fill the missing data and separate anomalies from the observation data. On the other hand, the regularity of the data sequences may also be severely affected by anomalies and missing data. To this end, a new prediction scheme named robust long-term spectrum prediction (RLSP) is proposed. Specifically, it alternatively performs prediction on partial data sequences (also referred to as pre-fill operation) and RTR in an iterative way, which not only reduces the damage to data sequence regularity brought by anomalies and missing data but also improves the accuracy of spectrum prediction. Finally, simulations based on the synthesis of data and real-world satellite spectrum data are given to show the superiority of our proposed RLSP over LSP-TC.