Traffic prediction is a critical component of intelligent transportation systems. In the traffic prediction task, utilizing graph models to capture dependencies in spatial-temporal data is a promising strategy for improving prediction accuracy. However, current graph models face the following several challenges: i) Many methods neglect fine-grained temporal dependency modeling when capturing temporal patterns, struggling to adapt to multi-scale temporal evolution in traffic data. ii) Most methods fail to account for the noise in the initial data, leading to suboptimal performance when handling non-stationary data. To address these challenges, we propose a novel approach called the Spatial-Temporal Multi-Scale Time Difference (STMSTD) model, which incorporates two core modules: the multi-scale time difference module and the spatial-temporal feature fusion module. The multi-scale time difference module first employs Empirical Mode Decomposition (EMD) to decompose raw time series into Intrinsic Mode Functions (IMFs) for signal denoising, then establishes temporal differencing mechanism to capture fine-grained multi-scale temporal dynamics. And the spatial-temporal feature fusion module integrates temporal attention, spatial attention, and Graph Convolution Network (GCN) to extract dynamic features comprehensively. Finally, experimental results demonstrate that our method significantly enhances the accuracy of traffic prediction.