融合语义的时空相关多步乘车需求预测方法

Multi-step Passenger Demand Prediction Based on Spatiotemporal Correlation Incorporating Semantic Information

为了准确预测乘车需求,提高车辆利用率,缓解供需不平衡和交通拥堵,提出一种融合语义的时空相关多步乘车需求预测方法。在空间相关性建模方面,计算区域历史需求数据的互信息以对空间历史需求模式相关性建模;同时考虑基于兴趣点的城市功能区对乘车需求的影响,利用反映城市功能特征的语义信息,借鉴词频-逆文档频率方法对区域兴趣点赋予权重,继而计算区域功能相似度并对空间功能相关性建模。在时间相关性建模方面,利用深度学习网络Transformer捕获数据中潜在的长时依赖性。最后,在真实数据集上验证所提方法的有效性。结果表明:单步预测时,与基线方法的均值相比,所提方法的均方根误差、平均绝对误差和平均绝对百分比误差3个指标分别降低了38.77%、38.79%和54.11%;多步预测时,所提方法的预测精度也有较大提升,在预测步长为6时,与基线方法的均值相比,所提方法的3个指标分别降低了21.35%、21.98%和15.72%。该方法能够更准确预测乘车需求时空分布,合理配置运力资源。

In order to accurately predict passenger demand,improve vehicle utilization,alleviate the supply-demand imbalance,and mitigate traffic congestion,a novel method for predicting multi-step passenger demand based on spatiotemporal correlation incorporating semantic information was proposed.In terms of spatial correlation,the mutual information of regional historical demand data was calculated to model the correlation of spatial historical demand patterns among regions.At the same time,the impact of urban function based on points of interest on passenger demand was considered.The term frequency-inverse document frequency method was referred to calculate the weight of each category of points of interest.Then the regional functional similarity was further calculated to model spatial functional correlation.In terms of temporal correlation,the deep learning network Transformer was utilized to capture potential long-term dependence in data.Finally,the effectiveness of the proposed method was verified on real datasets.For single-step prediction,compared with the average results of the baseline methods,the proposed method reduces the mean squared error,mean absolute error and mean absolute percentage error by 38.77%,38.79%and 54.11%,respectively.For multi-step prediction,the proposed method also improves the accuracy.When the prediction step is 6,compared with the average results of the baseline methods,the proposed method reduces the three indicators by 21.35%,21.98%and 15.72%,respectively.The proposed method can predict the spatiotemporal distribution of passenger demand accurately,which helps to allocate transport capacity resources rationally.