The exponential growth of mobile network traffic makes accurate traffic prediction essential for network optimization. However, this remains challenging due to complex spatial-temporal dependencies and diverse traffic patterns across base stations. Existing methods often rely on single-factor modeling or fail to effectively handle this multi-pattern heterogeneity. We propose the Meta-Optimized Multi-Period Spatial-Temporal Transformer (MMSTT), a novel framework integrating dynamic graph-based spatial modeling and multi-period temporal fusion within a Transformer architecture. To address the multi-pattern nature of traffic data, we incorporate a clustering algorithm and a meta-learning optimization process. This process captures shared features across patterns during meta-training and adapts to pattern-specific characteristics during meta-testing. Experimental results on a real-world dataset demonstrate that MMSTT significantly outperforms existing baselines, reducing the mean absolute error (MAE) to 28.39, root mean square error (RMSE) to 42.83, and mean absolute percentage error (MAPE) to 0.37. Compared to the standard Transformer and GCN baselines, MMSTT achieves improvements of 38.71% in MAE, 33.57% in RMSE, and 60.22% in MAPE.