基于多图时空图卷积模型的城市交通流长时预测

Long Term Prediction on Urban Traffic Flow Based on Multi-source Spatio-temporal Graph Convolutional Neural Network Model

交通流预测是智能交通系统中的重要组成部分,准确、及时和有效的预测信息对于城市道路的交通控制、诱导具有重要意义。然而,由于城市路网交通流时刻受到用地性质、天气变化等多种外部因素的影响,其预测面临着巨大挑战。为了有效预测城市路网交通流而提出了一种融合多源数据的多图时空图卷积模型。将影响城市路网交通流的外部因素分为静态和动态两类,并提供了一种明确和有结构的分类依据来理解影响交通流的各种外部因素。再将静态因素编码为多图,具体为距离矩阵、功能相似性矩阵和连通性矩阵,使用多图组成三通道输入时空相关性建模模块。该模块使用图卷积网络对交通流的空间相关性进行建模,学习节点特征和邻接信息;使用门控循环单元对交通流的时间相关性进行建模,捕捉交通流数据的动态变化和周期性规律。最后,使用融合层将多通道输出与动态因素进行融合作为最后预测输出。为了验证模型的有效性,使用SZ-TAXI数据集,与7种基准预测模型进行对比试验,结果表明,融合了多源外部因素的多图时空图卷积模型在评价指标MAE和RMSE上都比基准模型预测值更具准确性。并设计消融试验分析,处理静态因素的多图方法以及融合动态因素的方法均有效提高了城市路网交通流长时预测性能。

Traffic flow prediction is an important part of intelligent transport system.Accurate,timely and effective prediction information are of great significance for urban traffic control and guidance.However,due to the fact that urban road network traffic flow is affected by various external factors such as land use properties and weather changes,the traffic flow prediction faces enormous challenges.In order to predict the traffic flow of urban road networks effectively,multi-source spatio-temporal graph convolutional neural network model is proposed.The external factors affecting urban traffic flow are divided into static and dynamic categories.A clear and structured classification basis to understand the various external factors affecting traffic flow is provided.Then,the static factors are coded into multiple graphs,specifically distance matrix,functional similarity matrix,and connectivity matrix.The three-channel input spatio-temporal correlation modeling module is composed of multiple graphs.The graph convolution network is used to model the spatial correlation of traffic flow,learning node features,and adjacency information.The door control cycle single element is used to model the temporal correlation of traffic flow for capturing dynamic changes and periodic patterns of traffic flow data.Finally,the fusion layer is used to fuse the multi-channel output with dynamic factors as the final predicted output.In order to verify the model effectiveness,SZ-TAXI dataset is used for comparing with 7 benchmark models.The result shows that the multi-source spatio-temporal graph convolutional neural network model integrated multiple external factors achieves optimal performance than benchmark model in the evaluation indicators of MAE and RMSE.The ablation experiment shows that the multi-source method for handling static factors and the method for integrating dynamic factors effectively improve the long-term prediction performance of urban road network traffic flow.