Multi-Scale Spatio-Temporal Representation-Based 3-D Long-Term Spectrum Prediction

The recent advancements in space-air-ground-sea integrated networks necessitate accurate forecasting of future evolution patterns in the three-dimensional (3D) spatio-temporal spectrum. However, it is difficult to obtain accurate 3D spectrum situation due to the complex spectrum features in the 3D space. In this work, we present a 3D long-term spatio-temporal spectrum prediction network (LTST-PreNet) that incorporates a spatio-temporal spectrum graph convolution module (STSGC) and a time-frequency attention module (TFA). The STSGC is designed to capture cross-dimensional spatio-temporal correlations across different timeslots and locations. Meanwhile, the TFA is capable of jointly representing multi-scale temporal features in the spectrum, accounting for variations across hourly, daily, and weekly timescales in both the frequency and time domains. In addition, we have designed a complementary completion network and a graph sampling structure (GS) to enhance the prediction accuracy and generalizability of the proposed network. The completion network improves the integrity of the prediction input data by recovering spatially missing data from samples collected by unmanned aerial vehicles (UAVs). Based on our proposed 3D spatio-temporal spectrum simulator, the GS enhances the training efficiency of the graph network and the generalizability of our method across diverse scenes. The experimental results demonstrate that the proposed method outperforms existing spectrum prediction schemes in terms of accuracy and generalizability.