Parallel Multi-model Fusion Spectrum Prediction Based on Multi-channel Feature Extraction

Radio spectrum prediction is crucial for mining spectrum behavior patterns in complex environments, managing spectrum usage, and improving the probability of cognitive radio access to the spectrum. In this paper, we propose a parallel multi-channel multi-model fusion network (PM2FN) for feature extraction of complex electromagnetic data to achieve accurate spectrum prediction based on the obvious time-frequency correlation exhibited by the spectrum data. However, the accurate point prediction method ignores the random characteristics of the complex electromagnetic environment when predicting data with high volatility, and the traditional deterministic prediction can hardly eliminate the prediction error. Therefore, this paper combines the quantile regression and parallel multi-channel multi-model fusion network (QPM2FN) for probabilistic prediction of the electromagnetic spectrum to effectively quantify the prediction uncertainty. In this paper, a large number of comparative experiments are conducted on the Aachen dataset, and the experimental results show that the proposed model has higher prediction accuracy and effectiveness than the baseline models in terms of accurate prediction and probabilistic prediction.