基于ATT-LSTM模型的高速公路交通事件持续时长预测

为揭示交通事件对高速公路运行状态持续时间的影响规律，研究了高速公路交通事件持续时长预测方法。考虑高速公路交通事件时间序列特性，基于循环神经网络理论，从时间序列数据中提取交通事件时间依赖关系；通过引入长短时记忆网络，结合特征、时序注意力层挖掘历史时刻信息和当前时刻数据间的相关性，构建基于注意力机制-长短时记忆网络的高速公路交通事件持续时长预测模型。以2018年西安绕城高速公路交通监测数据集为例，开展了高速公路交通事件持续时长预测模型验证，对比了所提模型与反向传播神经网络、随机森林、支持向量机、长短时记忆网络模型这4种典型算法的预测精度，并分析了事件类型、天气条件、车辆类型、交通量等不同影响因素对持续时长的影响程度。结果表明:使用同一数据集，注意力机制-长短时记忆网络预测模型的预测结果平均绝对误差为24.43，平均绝对百分比误差为25.24%，均方根误差为21.17，预测精度优于其他4种预测方法。在模型的各影响因素权重中，事件类型所占权重最大为0.375，其次分别为车道数、车辆类型、天气等；采用立交出入口小时交通量作为修正参数，可以进一步提升预测精度，预测结果的绝对误差、平均绝对百分比误差和均方根误差可分别降低21.3%、7.5%和16.9%。研究结果能进一步提高高速公路交通事件持续时长预测的精度，为公路安全高效运行提供技术支持。 

Prediction of the Duration of Freeway Traffic Incidents Based on an ATT-LSTM Model

In order to study the impacts of traffic incidents on freeway operation, a method for predicting the duration of freeway traffic incidents is studied. Time-dependent characteristics of traffic incidents on freeways are extracted from time series data based on the recurrent neural network (RNN) theory. The feature and the temporal attention layer of a long short-term memory (LSTM) network are combined to study the correlation between historical and current moment data. Based on attention (ATT) mechanism and the LSTM, a model for predicting the duration of traffic incidentson freeways is developed. Validation of the model is carried out based on traffic monitoring dataset collected in 2018 along the Xi'an Ring Freeway. The prediction accuracy of the proposed model is compared with the following models: back propagation neural network (BP), random forest (RF), support vector machine (SVM), and long short-term memory (LSTM). The impacts of different factors, including the types of events, weather conditions, types of vehicles, and traffic volume, on the duration is also analyzed. Study results indicate that under the condition of the same dataset, the mean absolute error (MAE), the mean absolute percentage error (MAPE), and the root mean square error (RMSE) of the ATT-LSTM model is 24.43, 25.24%, and 21.17, respectively, which is better than that of other models. The "type of events" has the maximum weight of 0.375 among all of factors considered within the model, followed by the "number of lanes" "vehicle type" and "weather". By using the hourly traffic volume at the entrances and exits of interchanges as the correction parameter, the prediction accuracy is improved, and the MAE, MAPE, and RMSE of the model is decreased by 21.3%, 7.5%, and 16.9%, respectively. This study improves the prediction accuracy of the duration of traffic incidents on freeways and provides technical support for their safe and efficient operation.