The quality and efficiency of public transportation are crucial for mitigating environmental pollution and traffic congestion, with bus transit playing a dominant role. Consequently, optimizing bus transit services has been extensively studied. Passenger flow prediction is essential for optimizing bus transit systems, especially with decreasing public funding and the need for improved operational efficiency. Existing methods often struggle to fully capture the complex interplay of spatial and temporal dependencies, as well as the influence of complex exogenous factors. We propose a novel deep learning framework, Multi-scale Temporal Decomposition Spatiotemporal Network (MTDTSN), for accurate regional bus passenger flow prediction. It leverages multi-scale temporal decomposition, integrating the decomposed components with exogenous factors using specialized modules to capture both temporal dynamics at various granularities and the complex spatial relationships between bus stations. Furthermore, a Time-Station Cross-Attention mechanism is utilized to supplement the interaction between time and space, further enhancing the model's accuracy. We evaluate MTDTSN against state-of-the-art baselines, including MLP-based, GNN-based, and Transformer-based models, using a real-world dataset from Chongqing. MTDTSN achieves superior performance, demonstrating average improvements of 16.79% and 6.86% in MAE and RMSE, respectively, compared to these baselines.