Regional optimization of new-energy bus charging stations using a hybrid model integrating graph convolutional networks and simulated annealing

Electrification of public bus fleets poses significant challenges for urban transit agencies, particularly in the strategic placement and capacity expansion of charging infrastructure to meet growing operational demands. Traditional optimization methods often neglect the complex spatial interdependencies within transit networks, resulting in uneven load distribution and unmet demand. This study introduces a hybrid framework that integrates Graph Convolutional Networks (GCNs) with a Simulated Annealing (SA) metaheuristic to optimize the placement of electric bus charging stations at a regional scale. Using Suzhou, China, as a case study, we construct a graph of 617 community zones—characterized by features such as bus route density, stop density, population, land use, and POI data—and train a two-layer GCN to rank zones by charging-suitability score. The top 100 candidates are then passed to an SA algorithm, which selects five additional station locations to augment the existing network of 148 sites. Graph-theoretic evaluation demonstrates that this approach reduces under-supplied stations by over 70 %, lowers the standard deviation of station load from 0.8743 to 0.7258, and enhances network connectivity and resilience. The modular design of the GCN + SA pipeline allows for straightforward retraining and re-optimization as service frequencies, demographic patterns, or demand profiles evolve, offering a flexible, data-driven tool for iterative infrastructure planning in diverse urban contexts.