To address the limitations of existing models in jointly modeling group interactions and temporal evolutionary trends, this paper proposes a spatiotemporal trend-aware network for multi-agent trajectory prediction. First, initial node and edge features are extracted through a trajectory embedding layer. Second, we innovatively design a triple dynamic denormalization-based encoder architecture: A Pairwise Interaction Transformer(DyTPRT) captures fine-grained agent interactions, a Hyperedge Relation Transformer(DyTHRT) models group-level dependencies and a Temporal Trend-Aware Transformer (DyTTaT) learns contextual time-series patterns. Finally, a multi-head decoder fuses spatiotemporal features for trajectory forecasting. Extensive experiments on three benchmark datasets(NBA, SDD, and ETH-UCY) demonstrate significant superiority over state-of-the-art (SOTA) methods.