As urbanization and intelligence progress, the demand for intelligent transportation is rising continuously. The fundamental challenge of traffic flow prediction consists in how to effectively model the complex spatiotemporal dependencies in traffic data. While many models with superior performance do exist, there are a number of remaining concerns. The majority of models utilize the topology between nodes to construct static graphs, which restricts the capability of learning dynamic urban traffic patterns; traffic data frequently contains important cyclical information, yet current methods have generally ignored frequency-domain modeling and only extract features from the time domain, neglecting the fact that there are time delays in the propagation of traffic conditions between different locations in the transportation system. To address these problems, we propose a dual-frequency spatiotemporal graph neural network, namely DFST-Net, for accurate traffic flow prediction. It builds a dual-stream graph: a static graph based on temporal embeddings to model long-term dependencies, and a dynamic graph constructed via FFT-based phase similarity to capture short-term interactions. A learnable gating module fuses these graphs into a time-varying adjacency tensor. Beyond the propose mentioned above, a Physically-motivated Frequency Enhancement Module applies spectral filtering and sparsification to enhance key frequency patterns while suppressing noise. We adopt dilated convolutions for temporal modeling and design a gated spatiotemporal fusion mechanism to integrate both spatial and temporal signals. Experimental results on multiple benchmark datasets show that DFST-Net outperforms state-of-the-art models in terms of accuracy, and robustness.