Existing urban traffic flow prediction methods still face numerous challenges when dealing with nonlinear, multivariate, and long-time series data. To address these challenges, this paper introduces STVMamba, a spatio-temporal collaborative prediction framework that combines selective state space modeling with a variable-aware mechanism to improve prediction accuracy and stability. In spatial modeling, the model utilizes graph convolutions to extract the topological structure and spatial correlation features of the traffic network. In temporal modeling, a time-location encoding mechanism is introduced that integrates dynamic information and learnable embeddings to mitigate issues such as periodic misalignment and weakening of temporal features. By constructing a graph structure based on node distances and implementing a dynamic reordering mechanism, the model enhances its ability to model interactions among multiple variables. Combined with Mamba and LSTM, it achieves multi-scale collaborative modeling of long-term and short-term temporal features. Experiments on four publicly available PeMS datasets and real-world traffic flow data from Baotou City demonstrate that STVMamba significantly outperforms state-of-the-art methods in multi-step prediction tasks, exhibiting superior performance in both prediction accuracy and generalization capabilities.