A Meta-BAGRU-Based Spectrum Prediction Scheme for Low-Altitude Intelligent Networks

The high mobility of unmanned aerial vehicles (UAVs) and interference between nodes in Low-Altitude Intelligent Networks (LAINs) cause spectrum conditions to change quickly. The sparse data and dynamic spectrum environments in LAINs pose significant challenges to spectrum prediction. To address these issues, we propose a spectrum prediction scheme, which integrates Gated Recurrent Unit (GRU) Network with Bahdanau Attention in a Meta-Learning Framework, namely Meta-BAGRU scheme. As the base learner, the hybrid GRU-Bahdanau Attention (BAGRU) model integrates the sequential modeling capability of GRU with the attention mechanism to capture long-term dependencies in spectrum data. Moreover, the proposed scheme leverages the few-shot learning capability and parameter initialization optimization capability of meta-learning, enabling accurate cross-band predictions with insufficient data under dynamic spectrum environments. Simulations results show that the proposed Meta-BAGRU outperforms traditional methods, achieving higher accuracy and superior robustness with sparse data conditions.