Vehicle trajectory prediction is a key technology for enabling safe decision-making in autonomous driving systems. To address the issues of insufficient feature extraction and low prediction accuracy in existing methods under complex traffic environments, this paper proposes a vehicle trajectory prediction network called TFRPNet, based on time-frequency feature fusion and reliable trajectory propagation. This network employs multi-scale temporal convolutional networks and implicit neural representations to enhance the temporal features of trajectories and the frequency-domain features of maps, respectively. By leveraging an attention mechanism, cross-modal feature fusion is achieved, overcoming the limitations of traditional single-dimensional feature representations. Additionally, a reliable trajectory propagation mechanism is utilized to generate initial reliable trajectories that guide the final trajectory generation, effectively improving prediction accuracy by addressing the lack of correction mechanisms in single-stage prediction. Experiments on the Argoverse dataset demonstrate that TFRPNet outperforms existing benchmark models across multiple performance metrics. Compared to the SIMPL model, which performs well among benchmark models, TFRPNet improves ADE by 5.6%, reduces FDE by 8.5%, and decreases MR by 3.7%. This study provides valuable insights for vehicle trajectory prediction.