Purpose/Significance To explore the potential of rs-fMRI-based functional connectivity (FC) as a predictive biomarker for autism spectrum disorder (ASD), so as to address the challenge of identifying reliable biological markers in the diagnosis of neuropsychiatric disorders. Method/Process A hierarchical multi-dimensional graph convolutional network (HMGNet) is proposed, which incorporates a time-series encoder to extract temporal features with long-term dependencies and enhance FC matrix modeling. The model integrates a graph attention mechanism to dynamically adjust edge weights, enabling the identification of critical inter-regional brain interactions. Additionally, residual learning is employed to construct a deeper graph neural network architecture, thereby improving the modeling performance for complex brain functional connectivity. Result/Conclusion Empirical results demonstrate that HMGNet achieves a classification accuracy of 74.4% on two variants of the ABIDE dataset, outperforming most existing approaches. The neuroimaging biomarkers identified by the model show strong concordance with established medical knowledge, offering a promising and clinically viable pathway for the diagnosis of ASD.