针对现有交通流预测研究中存在的未能充分融合复杂时空相关性和时空异质性的问题，该文设计了一种基于栅格数据的交通流预测网络——时空异质化两阶段融合网络（Spatiotemporal Heterogeneous Two-Stage Fusion Neural Network，ST_HTFNN）。该网络使用分阶段、层次化的时空特征提取架构，采用静态和动态特征提取阶段串行的新模式，在静态特征提取阶段引入新颖的类曼巴线性注意力（Mamba-Like Linear Attention，MLLA）块作为静态异质化融合单元，实现空间上的相关性和异质性融合挖掘，在动态特征提取阶段设计了简单高效的动态异质化融合单元，通过膨胀卷积和门控机制的结合来自适应融合捕捉全局和局部的时空相关性和异质性。同时，针对细致到道路级的交通流特征，设计了道路特征增强模块来重建和增强道路信息，以解决深度卷积过程中道路特征平滑的问题。最后，设计了外部扰动特征融合模块来融合外部扰动特征对交通流预测结果的影响。在3个现实世界的交通数据集BikeNYC、TaxiCQ和TaxiBJ上进行的模型实验表明，ST_HTFNN模型展现出了超越现有基线模型的卓越性能，相应的预测精度平均绝对误差分别降低了6.13%、0.8%和7.01%。

In response to the existing traffic flow prediction studies that fail to fully integrate complex spatiotemporal correlations and heterogeneities, this paper designs a traffic flow prediction network based on grid data, namely the spatiotemporal heterogeneous two-stage fusion neural network marked as ST_HTFNN. This network employs a phased and hierarchical spatiotemporal feature extraction architecture, and adopts a new model where the static and dynamic feature extraction stages are serialized. In the static feature extraction stage, a novel Mamba-like linear attention (MLLA) block is introduced as a static heterogeneous fusion unit to achieve spatial correlation and heterogeneity fusion mining. In the dynamic feature extraction stage, a simple and efficient dynamic heterogeneous fusion unit is designed, and dilated convolution is combined with gating mechanisms to adaptively fuse and capture global and local spatiotemporal correlations and heterogeneities. Furthermore, to address the smoothing of road features during the deep convolution process for road-level traffic flow characteristics, a road feature enhancement module is designed to reconstruct and enhance road information. Finally, an external disturbance feature fusion module is designed to integrate the impact of external disturbance features on traffic flow prediction results. Experimental results on three real-world traffic datasets, namely BikeNYC, TaxiCQ and TaxiBJ, demonstrate that the ST_HTFNN model outperforms the existing benchmark methods, respectively with a decrease of 6.13%, 0.8% and 7.01% in the mean absolute error of prediction accuracy.