大城市多种交通方式交织形成了彼此互通的客流网络，而跨交通方式客流之间的时空关联关系错综复杂，需要客流联合预测来解析整体出行规律，这对实现跨交通方式客流无缝衔接出行至关重要。针对跨交通方式客流网络，引入跨交通方式超图关联矩阵刻画公交和地铁的客流超图网络之间关联关系，提出一种基于双模时空超图卷积网络（BSTHCN）的跨交通方式客流联合预测模型。具体来说，模型由3部分组成：输入模块、时空卷积模块（包括时间卷积和空间卷积）和输出模块，可同时捕捉公交、地铁站点和线路的客流特征，以及两种客流网络之间换乘客流特征。最终，模型可识别提取重要信息特征，并进行特征聚合和分配。实证分析结果表明，相对于经典预测模型而言，BSTHCN具有更优的预测精度，以预测1h后的客流为例，在公交和地铁数据集上的平均绝对误差（MAE）指标至少分别减小了8.93%和8.10%，均方根误差（RMSE）指标至少分别降低了10.64%、7.47%。且BSTHCN的参数量和模型运行时间均在合理范围内，相比于时空卷积网络（S-TGCN）和扩散卷积递归神经网络（DCRNN），BSTHCN在实现更精准预测的同时，运行时间仅增加了4.82%。综合来看，BSTHCN显示出较强的竞争力。消融实验结果则进一步表明，BSTHCN在引入超图并考虑跨交通方式关联后，可更好地反映客流网络中的局部与整体特征，提升客流预测的精度。

The traffic modes in metropolis is interwoven to form an interconnected passenger flow network, and the spatio-temporal relationship between cross-transportation modes is complicated, which requires the joint prediction of passenger flow to analyze the overall travel law. For the cross-transportation passenger flow network, the supergraph correlation matrix of cross-transportation modes is introduced to describe the correlation between the passenger flow supergraph network of bus and subway, and a joint prediction model based on dual-mode spatial-temporal supergraph convolution network (BSTHCN) is proposed. Specifically, the model consists of three parts: an input module, a spatio-temporal convolution module (including temporal and spatial convolutions), and an output module, which can simultaneously capture the passenger flow network characteristics of both buses’ and metros’ stations and routes, as well as the transfer passenger flow characteristics between the two different passenger flow networks. The proposed model can identify and extract important information features, and perform feature aggregation and allocation. The experimental results show that the proposed model has better prediction accuracy compared to classical prediction models. The proposed model reduces the mean absolute error (MAE) by 8.93% and 8.10% on the bus and metro datasets, respectively, while the RMSE decreased by 10.64% and 7.47%. Moreover, the parameter volume and model runtime of proposed model are within a reasonable range. Compared to S-TGCN and DCRNN, proposed model achieves more accurate predictions with only a 4.82% increase in runtime. On the whole, proposed model demonstrates strong competitiveness. The ablation experimental results further demonstrate that after incorporating hypergraphs and considering multi-modal transportation correlations, the proposed model can better reflect both local and global characteristics in passenger flow networks, thereby improving the accuracy of passenger flow prediction.