Spectral graph neural networks (GNNs) can leverage frequency domain properties to effectively analyze graph-structured data. However, existing dynamic GNNs typically model temporal graph evolution as a holistic process, overlooking the inherent diversity of spectral frequencies within dynamic graphs: low-frequency components that capture stable, gradually evolving structures, and high-frequency components that reflect abrupt, localized changes. To address this gap, we propose DynSpectral, a multi-channel temporal spectral GNN that explicitly decomposes and models these co-existing frequency signals. Our approach introduces a novel frequency-aware graph decomposition strategy. For each temporal snapshot, we construct three complementary views: (1) a Union Graph that aggregates edges across consecutive snapshots to preserve low-frequency structural stability; (2) a Difference Graph that isolates newly appearing and disappearing edges to capture high-frequency dynamics; and (3) the Original Graph as a full-spectrum reference. Each view is then processed by a specialized temporal encoder tailored to its unique evolutionary characteristics, effectively modeling both smooth, long-term trends and abrupt, short-term changes. To ensure coherent and robust representations, we employ a contrastive learning framework that aligns these multi-frequency views while preserving their distinct information. Extensive experiments on dynamic link prediction benchmarks validate the efficacy of our frequency decomposition approach, demonstrating its capability to provide a more comprehensive modeling of temporal graph dynamics while achieving promising and competitive performance.