With the deepening global focus on sustainable development and carbon emission reduction, electric vehicles (EVs) have become a key component of clean energy transportation. To address the challenges of computational inefficiency in largescale node scenarios and the difficulty of synchronous spatiotemporal feature extraction in traditional graph neural networks, this paper proposes a Mixture-ofExperts Spatiotemporal Synchronous Graph Kolmogorov-Arnold Network based on hierarchical clustering. First, the global charging pile network is partitioned into 4 spatiotemporally correlated subgraphs using an agglomerative hierarchical clustering algorithm, and temporal patching strategy is employed for feature reconstruction. Second, a Spatiotemporal Synchronous Graph KolmogorovArnold Network is designed as expert models, where Fourier coefficients replace traditional spline functions to achieve frequency-domain fusion of spatiotemporal features. Finally, a gating network integrates expert outputs through transfer learning. Validation on real-world data from 24,761 charging piles in 247 Shenzhen districts demonstrates: For 15/30/45-minute prediction tasks, the proposed model achieves MAEs of 1.63/3.02/3.69, outperforming state-of-the-art baseline PAG by 4.7%/4.4%/7.3% respectively. The training time is reduced to 20.43 minutes. Experimental results confirm the method effectively balances prediction accuracy and computational efficiency in large-scale charging infrastructure scenarios.
