As a fundamental technology in Intelligent Transportation Systems, accurate traffic prediction has become a critical challenge for real-time applications. Traffic series typically exhibit both seasonal fluctuations and long-term trends, and the complex interaction between these components presents significant forecasting challenges. However, existing methods often either model the raw time series directly, thereby overlooking the entangled effects or perform decomposition without fully exploiting the interactions between the decomposed components. Moreover, many methods rely on a single or predefined graph structure to characterize spatial associations, limiting their ability to capture dynamic spatial dependencies effectively. To this end, we propose a Seasonal-Trend Dynamic Interactive Integration Network, namely STDIINet, for accurate traffic prediction. Specifically, the trend and seasonal parts of the time series data, are decoupled, modeling each according to its distinct characteristics. For the seasonal component, we employ a dynamic spatial learning module based on a graph construction method, combined with a temporal convolution module, to capture seasonal features. For the trend component, we utilize a multilayer perceptron to extract long-term trend information. Furthermore, we design a dynamic interactive integration framework that integrates these components while supporting bidirectional dynamic fusion between trend and seasonal features, thereby enhancing the model’s capability to learn complex spatio-temporal representations. Experiments on three real-world traffic datasets demonstrate that STDIINet outperforms the state-of-the-art methods.