<p>Knowledge hypergraphs, extending knowledge graphs, inherently exhibit heterogeneity from diverse entity types, relationships, and higher-order associations. Leveraging this enables nuanced representation of complex semantics and structures. However, existing knowledge hypergraph link prediction methods face critical challenges: inadequate modeling of high-order semantic relations results in (1) suboptimal utilization of hypergraph structural information, and (2) limited reasoning capabilities. Consequently, these methods fail to improve prediction performance. To address this, we introduce a novel heterogeneous graph structure learning framework. Specifically, the hypergraph is transformed into a heterogeneous graph with five node types and twelve edge types to fully exploit structural information. We further propose a hypergraph attention layer to capture structural features of nodes and edges, and a Transformer layer to model these features for predicting missing entities and relationships. Experiments on public datasets show that our method significantly outperforms baseline methods, confirming its effectiveness and robustness.</p>