基于全局时空特性的城市路网交通速度预测模型

城市路网交通速度预测是智能交通系统中的重要组成部分，其可为出行者提供实时的交通信息，对提升道路通行能力具有重要意义。现有基于图卷积网络的预测模型一定程度上加强了对一阶相邻路段间空间关联程度的挖掘，但在非一阶相邻路段关联度大于一阶相邻路段关联度的情况下，如果仍输入原始的邻接矩阵，会遗失一些相对重要的路段空间信息，无法得到较好的预测结果。为准确挖掘城市路网中的时空特性，提出一种基于全局图卷积和门控循环单元的城市路网交通速度预测模型G-GCGRU。考虑全局路网下非一阶相邻路段间的空间影响程度，利用相关性分析方法计算得到路段间的关联度矩阵，并作为新的卷积方式进一步加深对空间特征的挖掘，在此基础上，采用门控循环单元方法提取路网时间特征。使用深圳市罗湖区城市路网车速数据进行实验，结果表明，该模型预测性能优于图卷积网络（GCN）、门控循环单元（GRU）和GCN-GRU混合模型，以均方根误差为评价指标，预测精度分别提高25.3%、4.7%和2.1%。

Urban Road Network Traffic Speed Prediction Model Based on Global Spatio-Temporal Characteristics

Urban road network traffic speed prediction is an important aspect of Intelligent Transportation System(ITS) and is of great significance in improving road capacity by providing real-time traffic information for travelers.The existing prediction model based on the Graph Convolution Network(GCN) strengthens the mining of the spatial correlation degree between first-order adjacent sections to a certain extent.However, if the correlation degree of some non-first-order adjacent sections is greater than that of some first-order adjacent sections, when the original adjacency matrix is input, some relatively important section spatial information may be lost, precluding the realization of better prediction results.To accurately mine the temporal and spatial characteristics of urban road networks, a traffic speed prediction model of an urban road network based on global GCN and Gated Recurrent Unit(GRU), G-GCGRU, is proposed.Considering the degree of spatial influence between non-first-order adjacent road sections under the global road network, the correlation degree matrix between road sections is calculated using the correlation analysis method, which is a new convolution method to further deepen the mining of spatial features.Based on this, the gated cycle unit method is used to extract the time features of the road network.The results show that the performance of the model is better than those of the GCN, Gated Recurrent Unit (GRU), and GCN-GRU hybrid models.Using the root mean square error as the evaluation index, the prediction accuracy is improved by 25.3%, 4.7%, and 2.1%, respectively.