Abstract: To address the insufficient accuracy and robustness of high-frequency communication MUF prediction under complex ionospheric conditions and highly variable communication scenarios, this study proposes a high-frequency communication frequency prediction method based on multi-feature, multi-scale fusion. First, an adaptive k-value graph convolution layer is constructed, in which spectral entropy is computed to dynamically determine the order of the Chebyshev graph convolution, enabling accurate extraction and efficient fusion of multi-feature data and thereby capturing complex dependencies among heterogeneous features. Second, a cross-scale attention fusion layer is designed to capture, in parallel, temporal evolution patterns at different scales; attention scores are generated to perform weighted integration across scales, fully exploiting temporal regularities over multiple time horizons and improving prediction accuracy. Finally, a gated sparse feature integration layer is introduced to control model redundancy by selecting core feature information while producing an integrated output, which enhances prediction robustness. Experimental results on the Hangzhou–Guangzhou link and Xinxiang–Wuhan link show that the MAE of the algorithm proposed in this paper reaches 0.32 MHz and 0.38 MHz, respectively, which are reduced by 8.57% and 11.63% compared with the ASTGCN algorithm. Therefore, the proposed algorithm has high prediction accuracy and strong robustness. This algorithm helps to optimize frequency planning and resource allocation in short-wave communication systems and improve the overall efficiency and reliability of the communication system.