城市交通时间预测的混合神经网络模型

Urban traffic time prediction model based on hybrid neural networks

针对基于模型和直接匹配的城市交通时间预测方法很难有效整合影响预测的多重因素问题,提出一种基于一维卷积神经网络(Conv1d)-长短期记忆单元(LSTM)-残差网络(ResNet)的混合神经网络预测模型CLRTT。模型利用CNN和LSTM网络提取轨迹的空间和时间相关性,将影响交通时间的外部特征转化为低维向量,级联到时间预测组件的输入,通过在损失函数中引入权重系数的方法结合轨迹局部和整体预测结果,通过3层残差全连接网络得到整段路径的预测时间。针对原始轨迹的路网匹配修正能够有效提升模型预测精度,误差平均减小11%;不同时段和不同长度的轨迹预测实验结果表明,与传统的AVG和KNN类算法的模型相比,CLRTT模型预测误差MAPE在不同测度平均降低10%以上;CLRTT模型具有较好的平稳性,MAPE振幅小于15%,对较长轨迹时间预测精度提升明显。

Aiming at the problem that traditional model-based and direct matching methods are difficult to effectively integrate multiple factors affecting urban traffic time prediction,a hybrid neural network prediction model CLRTT based on one-dimensional convolutional neural network(CONV1d)-long short term memory unit(LSTM)-residual network(RESNET)is proposed.In this model,CNN and LSTM networks are used to extract the spatio-temporal correlation of trajectories,and the external features that affect the traffic time are transformed into low dimensional vectors,which are cascaded to the input of the time prediction component.Finally,the weight coefficient is introduced into the loss function to determine the prediction time of the whole path on the local and overal prediction results of trajectories and through the three-layer residual fully connected network.The experimental results on the data set of the actual track of 25 day taxi in Chengdu show that the road network matching correction for the original track can effectively improve the prediction accuracy of the model,with an average error of 11%decrease.Prediction experiment results in different periods and at different lengths of trajectories show that the CLRTT model has higher accuracy and prediction error MAPE is reduced by more than 10%in different measures.In addition,CLRTT model has good stability,and MAPE amplitude is less than 15%,especially for long track time prediction accuracy.