Traffic flow prediction is critical for intelligent transportation systems, facing challenges in modeling complex networks and extracting dynamic traffic patterns. As traffic networks scale, computational complexity and large hidden state spaces of models become issues, potentially leading to high computational costs and weak generalization, which can limit practical use. To tackle these challenges, in this paper, we propose a Spatio-temporal Dual Graph Network with Learnable Bases (STLB-GN), an efficient and effective model for complex spatio-temporal correlations: (i) Considering the evident periodicity in traffic flow data, a dynamic periodicity embedding learning approach is adopted for modeling the traffic network structure, which can better explore the data’s latent periodicity; (ii) We introduce learnable bases to replace Chebyshev polynomials in spectral graph convolution, thereby enhancing the model’s generalization ability; (iii) A dual graph convolution is proposed to automatically identify specific patterns for individual nodes and shared patterns between nodes, helping the model in capturing complex spatio-temporal features within traffic flow data; (iv) An improved multi-scale gated temporal convolution is proposed to grasp the temporal dynamics of traffic flow data and solve the temporal inconsistency problem in previous studies; (v) We performed a series of experiments and evaluations on real public transportation datasets. The outcomes show that STLB-GN performs more competitively than state-of-the-art models. Furthermore, STLB-GN shows high computational efficiency, which outperforms other high-performance models.