Diverse-Feature Spatial-Temporal Graph Attention Network for Long-Term Interference Prediction

With the explosive growth of communication networks, the unmet demand for spectrum has created escalating interference issues, necessitating high-precision multi-step interference prediction methods. However, collecting interference data can impose significant sensor overhead. To address this challenge, we propose a framework that leverages physical resource block (PRB) traffic for interference prediction. Given the intricate spatial-temporal characteristics of PRB traffic and interference intensity, capturing these intrinsic features accurately poses a considerable challenge. In response, we introduce a novel Diverse Spatial-Temporal Graph Attention Network (DSTGAT) designed to precisely predict interference intensity from PRB traffic. The DSTGAT comprises two main components: first, a multi-feature extraction module that integrates four adjacency matrices to characterize the spatial features of PRB traffic; second, a time-encoding temporal self-attention module crafted to uncover temporal correlations. Experimental results substantiate the superior performance of our proposed model, especially in managing multi-step prediction scenarios.