Spectrum prediction is of crucial importance for realizing the cognitive Internet of Things to tackle the spectrum scarcity problem. Deep-learning-based spectrum prediction methods have attracted extensive attention due to their superior accuracy. However, the training speed of deep networks is low and the architecture of traditional networks is uninterpretable. In order to tackle these problems, a radio frequency machine-learning-driven spectrum prediction scheme is proposed by exploiting a novel model-enabled autoregressive (AR) network. A cell with only two parameters is exploited in each layer of the AR, which accelerates the network training. Moreover, the domain knowledge of the AR structure enables our proposed scheme to be explainable. Simulation results show that our proposed scheme has the best prediction accuracy than the long short-term memory (LSTM)-based scheme and the AR scheme. It is also shown that its convergence speed is higher than that of the LSTM-based scheme.

中文翻译：

---

RFML-Driven Spectrum Prediction：一种新型的模型启用自回归网络

频谱预测对于实现认知物联网解决频谱稀缺问题至关重要。基于深度学习的频谱预测方法由于其卓越的准确性而引起了广泛的关注。然而，深度网络的训练速度较低，传统网络的架构难以解释。为了解决这些问题，提出了一种射频机器学习驱动的频谱预测方案，该方案利用了一种新的模型支持的自回归 (AR) 网络。在 AR 的每一层中都使用了一个只有两个参数的单元，从而加速了网络训练。此外，AR 结构的领域知识使我们提出的方案可以解释。仿真结果表明，我们提出的方案比基于长短期记忆 (LSTM) 的方案和 AR 方案具有最好的预测精度。还表明其收敛速度高于基于 LSTM 的方案。