[Purposes] Traffic flow time series possess complex nonlinearity, significant spatio-temporal dependence, and strong uncertainty. Long-term prediction is more valuable in practical traffic management applications. This study aims to construct a probabilistic prediction model that integrates a graph multi-head attention network (GMAN) with a t-distributed generalized autoregressive conditional heteroscedasticity (GARCH-t) model. This provides decision support tools for intelligent traffic management that combine high-precision long-term mean prediction with reliable interval estimation. [Methods] A deep coupling model integrating GMAN and GARCH-t (the GMAN-GARCH-t model) was proposed. First, a multi-head spatio-temporal attention mechanism was applied to dynamically capture the nonlinear spatio-temporal correlations among road network nodes, accurately extracting the core features of traffic flow evolution. Second, the GARCH-t model was adopted to accurately quantify and predict the heteroscedasticity of the residuals. Finally, a dual-layer output structure enabling collaborative optimization of "point prediction-interval estimation" was constructed. The model was validated on measured data from 105 road sections in Suzhou City for predictions spanning 1 to 12 time steps. [Findings] In the mean prediction, the mean absolute error (MAE) of the 12-step prediction reaches 17.22 vehicles/5 min. In interval estimation, the coverage rate of the 95% confidence interval reaches 94.6%, and the interval width can adapt to changes in traffic conditions. [Conclusions] Through the organic integration of spatio-temporal attention mechanism and thick-tail volatility modeling, the GMAN-GARCH-t coupling model simultaneously improves the accuracy and interpretability of traffic flow probability prediction, providing a reliable quantitative decision-making basis for dynamic traffic control.