As an essential part of the modern intelligent traffic management system, traffic speed prediction is a challenging task. In recent studies, deep neural networks (LSTM and WaveNet) and graph neural networks (GCN and GNN) have been extensively investigated on traffic networks evaluation, which is better than statistical-based models (MA and ARIMA). However, the demerits existing in these deep learning forecasting process include (1) carry out vehicle speed as an individual input and insufficient ability to handle the other related factors, such as the number of equivalent lanes, accident occurrence, and toll data; (2) inadequate capability of considering both linear and nonlinear components as a whole; (3) unstable performance on forecasting task given various heterogeneous series. Therefore, we propose a hybrid end-to-end model to combine both spatio-temporal features and other effective features. First, a heterogeneous graph attention network approach (HetGAT) was proposed, and a temporal dilated convolution architecture (TCN) was adopted to simulate the impacts on traffic flow of the multi-scale context of temporal factors. Then, a weighted graph attention network (GAT) encodes input temporal features, and a decoder predicts the output speed sequence via a freeway network structure. Based on the end-to-end architecture, we integrate multiple Spatio-temporal factors effectively for the prediction. To validate the efficiency of the proposed model, three sets of field-captured data were employed to run the test. Compared with conventional sequence analysis models and deep prediction models, experimental results demonstrated the superiority of HetGAT for all cases with regards to MAE, MAPE, and RMSE.

作为现代智能交通管理系统的重要组成部分，交通速度预测是一项艰巨的任务。在最近的研究中，深度神经网络（LSTM和WaveNet）和图神经网络（GCN和GNN）已经在交通网络评估中得到了广泛研究，这比基于统计的模型（MA和ARIMA）更好。但是，这些深度学习预测过程中存在的缺点包括：（1）将车速作为个人输入来执行，并且处理其他相关因素（例如等效车道数，事故发生和通行费数据）的能力不足；（2）整体考虑线性和非线性成分的能力不足；（3）在给定各种异构序列的情况下，预测任务的性能不稳定。因此，我们提出了一种混合的端到端模型，以结合时空特征和其他有效特征。首先，提出了一种异构图注意力网络方法（HetGAT），并采用了时间膨胀卷积架构（TCN）来模拟时间因素多尺度上下文对交通流的影响。然后，加权图注意力网络（GAT）对输入的时间特征进行编码，并且解码器通过高速公路网络结构预测输出速度序列。基于端到端的体系结构，我们有效地集成了多个时空因素进行预测。为了验证所提出模型的效率，使用了三组现场捕获数据来运行测试。与传统的序列分析模型和深度预测模型相比，