面向在线地图的GCN-LSTM神经网络速度预测

为从路网速度中完整提取路段速度的时空特征，实现高精度路段速度预测，通过调用在线地图的路径规划应用程序接口，采集路段的在线地图速度；利用图卷积神经网络(GCN)提取空间特征，利用长短期记忆(LSTM)神经网络提取时间特征，建立面向在线地图的GCN-LSTM神经网络，提取路段速度的时空特征，预测路段速度；为测试面向在线地图的GCN-LSTM神经网络表现，并评价在线地图下GCN-LSTM神经网络的优势与面向检测器速度预测模型的可替代性，以局部路网为例分析模型表现，并对比在线地图下不同模型的表现与不同数据源下近似模型的表现。研究结果表明:GCN-LSTM神经网络在训练集和测试集上的平均绝对误差(MAE)均低于5，均方根误差(RMSE)均低于6，平均绝对百分比误差(MAPE)均低于30%，训练误差和测试误差均处于较低水平，总体表现良好；GCN-LSTM神经网络的路段MAPE服从Gumbel分布，均值均落在19%±4%之间，85%分位点均落在34%±5%之间，2项指标均处于较低水平，个体表现良好；在面向在线地图的速度预测模型中，GCN-LSTM神经网络的MAE、RMSE、MAPE以及MAPE拟合曲线均值、85%分位点最低，总体和个体表现均为最佳，在面向在线地图的速度预测中具有一定优势；在近似模型中，GCN-LSTM神经网络的MAE、RMSE、MAPE以及MAPE拟合曲线均值、85%分位点最低，总体和个体表现均为最佳，则面向在线地图速度预测的可靠性高，可代替面向检测器的速度预测。 

Speed prediction by online map-based GCN-LSTM neural network

Online map speeds of roads were collected by calling the path-planning application programming interface of the online map to completely extract the spatio-temporal features of the road speed from road network speed and then achieve high-precision road speed prediction. The spatial features were extracted using a graph convolutional network (GCN), and the temporal features were extracted using a long short-term memory (LSTM) neural network. An online map-based GCN-LSTM neural network was established, the spatio-temporal features of the road speed were extracted, and the road speed was predicted. The performance of the online map-based GCN-LSTM neural network was assessed, and the advantages of the online map-based GCN-LSTM neural network and the substitutability of the detector-based speed prediction model were evaluated. By using the local road network as an example, the performance of the model was analyzed, and the performances of different online map-based models and similar models with different data sources were compared. Analysis results show that the mean absolute errors(MAEs) of the GCN-LSTM neural network are lower than 5, the root mean square errors (RMSEs) are lower than 6, and the mean absolute percentage errors (MAPEs) are lower than 30% in the training and testing sets. Hence, the training and testing errors are low, indicating good comprehensive performance. The MAPE of the GCN-LSTM neural network of the roads follows a Gumbel distribution, whose mean ranges between 19%±4%, and the 85% quantile ranges between 34%±5%. Hence, both indexes are low, indicating good individual performance. Among the online map-based speed prediction models, the MAE, RMSE, MAPE, mean, and 85% quantile of the MAPE fitting curve of the GCN-LSTM neural network have the lowest values. Hence, its comprehensive and individual performances are the best, and it exhibits advantages in online map-based speed prediction. Among the similar models, the MAE, RMSE, MAPE, mean, and 85% quantile of the MAPE fitting curve of the GCN-LSTM neural network have the lowest values. Hence, its comprehensive and individual performances are the best. Furthermore, the reliability of online map-based speed prediction is high, so that it can be used as a substitute for detector-based speed prediction. 4 tabs, 13 figs, 30 refs.