Spatio-temporal graph data have been widely applied in traffic flow prediction tasks. Traditional methods often combine graph convolutional networks with recurrent neural networks based on the original graph structure. However, these methods typically struggle with issues such as incomplete sensor distributions and unresolved potential dependencies between different parallel intersections. Moreover, the model structure often fails to capture the noise inherent in traffic predictions, which becomes a significant barrier to accurate forecasting. To address these challenges, we proposed a novel approach that diverges from traditional traffic flow models, which rely on predefined graph structures. Instead, our method uncovers hidden edge connectivity and integrates the data into a unified framework, enabling the model to better capture the underlying spatial relationships. Specifically, we introduced a new framework that combines neural control differential equations with stochastic differential equations to enhance spatio-temporal graph modeling. This integration improves the model's capacity to capture complex spatio-temporal patterns, thereby enhancing the accuracy of prediction tasks. Extensive experiments conducted on benchmark datasets validate the effectiveness of our method.