融合时序关联动态图与常微分方程的区域间出租车需求预测

Inter-regional taxi demand forecasting based on time series correlationdynamic graph and ordinary differential equation

为解决出租车行业中的高空驶率和不均衡的需求分布问题,通过对出租车出行的区域进行深入研究,提出了一个融合时序关联动态图与常微分方程的需求预测模型TCG-ODE(temporal correlation graphs-ordinary differential equations)。首先,模型使用ODE(ordinary differential equation)对图卷积神经网络(graph convolutional network,GCN)进行微分方程建模,将堆叠局部特征信息抽象为动态图,以节点的时序特性来推进局部节点状态;然后,设计了一种基于注意力分数调整采样策略的蒸馏方案,提高对多层稀疏图的适应效果,以更稳定地表征复杂时空特征,最终实现对区域间出租车需求量的预测。在真实的出租车订单数据集上进行实验,研究结果表明,TCG-ODE模型的预测效果均优于对照模型和改进前的模型。通过精准预测不同区域之间的出租车需求量,可以为出租车司机和乘客出行提供决策支持信息,从而优化供需关系。

ed the stacked local feature information into a dynamic graph.It advanced the local node state based on the timing characteristics of nodes.Then,it designed a distillation scheme based on the attention score adjustment sampling strategy to improve the adaptation effect to the multi-layer sparse graph,so as to more stably represent the complex spatio-temporal characteristics,and finally realized the prediction of inter-regional taxi demand.Experimental results conducted on real taxi order datasets demonstrate that the TCG-ODE model outperforms both benchmark models and the pre-improvement model in terms of demand prediction accuracy.By accurately forecasting taxi demand among different regions,this model provides decision support information for taxi drivers and passengers,thereby optimizing the supply-demand relationship.