基于GRU-GCN-RDrop模型的交通速度预测  

Traffic Speed Prediction Based on GRU-GCN-RDrop Model

作　　者：赵嘉雨 段亚茹 何立明[1] ZHAO Jia-yu;DUAN Ya-ru;HE Li-ming(School of Information Engineering,Chang’an University,Xi’an 710064,China)

机构地区：[1]长安大学信息工程学院,陕西西安710064

出　　处：《计算机技术与发展》2023年第4期120-125,139,共7页Computer Technology and Development

基　　金：国家自然科学基金项目(51308058)。

摘　　要：准确并实时地预测交通速度是实现智能交通管控和建设智慧交通系统必不可少的一环,然而现有的预测方法无法准确地挖掘其潜在的时空相关性。为了进一步挖掘数据的时空特性以及提高预测精度,设计了基于门控循环神经网络(GRU),图卷积网络(GCN)和正则化Dropout(R-Drop)结合的GRU-GCN-RDrop组合模型。GCN用于学习复杂的拓扑结构来捕获空间依赖性,GRU用于学习交通数据的动态变化来捕获时间依赖性。GCN和GRU相结合后使用R-Drop方法提高模型泛化能力。以SZ-taxi数据集为例进行预测分析,GRU-GCN-RDrop模型预测了未来在15分钟、30分钟、45分钟和60分钟的交通速度,并得出对应的均方根误差、平均绝对误差、精度、判定系数和解释方差。对比GCN、GRU单个模型,GRU-GCN-RDrop模型有效解决了误差的迅速累积问题。对比大多数现有基准模型,GRU-GCN-RDrop模型对于交通速度序列的特征提取及解释能力较为优秀。对比T-GCN模型和ST-AGTCN模型,GRU-GCN-RDrop模型泛化能力更强。由此说明了GRU-GCN-RDrop模型预测的交通速度时间序列具有较高的精度和稳定性。Accurate and real-time prediction of traffic speed is an essential part of intelligent traffic control and construction of smart transportation system.However,the existing prediction methods cannot mine the potential spatio-temporal correlation in the data accurately.In order to further mine the spatio-temporal characteristics of data and improve the prediction accuracy,a GRU-GCN-RDrop combined model based on gated recurrent unit(GRU),graph convolutional network(GCN)and regularized dropout(R-Drop)is designed.The GCN is used to learn complex topological structures to capture spatial dependence and the gated recurrent unit is used to learn dynamic changes of traffic data to capture temporal dependence.After the combination of GCN and GRU,the R-Drop method is used to improve the generalization ability of the model.SZ-taxi data set is taken as an example for prediction and analysis.GRU-GCN-RDrop model predicts the future traffic speed in 15 minutes,30 minutes,45 minutes and 60 minutes,and obtains the corresponding RMSE,MAE,Accuracy,R2 and Var.Compared with GCN and GRU models,GRU-GCN-RDrop model effectively solves the problem of rapid accumulation of errors.Compared with other benchmark models,GRU-GCN-RDrop model is superior in feature extraction and interpretation of traffic speed series.Compared with T-GCN model and ST-AGTCN model,GRU-GCN-RDrop model has stronger generalization ability.It is showed that the timeseries of traffic speed predicted by GRU-GCN-RDrop model has high accuracy and stability.

关 键 词：智慧交通系统 交通速度预测 图卷积网络 门控循环单元 正则化Dropout 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论] U491.2[自动化与计算机技术—计算机科学与技术]