融合路段传输模型和深度学习的城市路网短时交通流状态预测

城市路网短时交通流预测是实现智慧城市的关键技术，随着人工智能的发展，越来越多的深度学习算法被应用于城市道路交通状态估计和预测研究。但是深度学习因缺少对交通流演化机理的刻画导致其可解释性不强，而交通流解析模型常因预测精度问题导致其应用效果受到限制。为了取长补短，首先对路段传输模型（Link Transmission Model，LTM）进行改进，提出了可以利用真实数据实时校准仿真网络从而提高预测精度的数据驱动型路段传输模型（Data-driven Link Transmission Model，D²LTM），并在此基础上引入时空深度张量神经网络模型（Spatial-temporal Deep Tensor Neural Networks，ST-DTNN）来捕获网络交通流数据中的时间维、空间维和深度维特征信息，形成融合路段传输模型和深度学习的城市路网短时交通流预测模型D²LTM-STDTNN。该混合模型一方面通过D²LTM机理模型来揭示交通流演化的基本规律，发挥其对城市路网交通流状态时空演化过程的精细刻画能力，增强混合模型机理的可解释性；另一方面利用ST-DTNN模型强大的高维数据挖掘能力和动态特征学习能力，提高城市级路网交通流的短时预测精度。该模型还考虑了交叉口不同转向的短时预测问题，具有更细的空间粒度和时间粒度，因此也具有更大的预测难度。实测结果表明：D²LTM-STDTNN混合模型相对于基准模型预测精度更高，且具备模拟演化机理方面的优势，提升了城市路网短时交通流状态预测能力，揭示了路段间的交通流动态演化规律，可为网络交通流模拟推演和主动管控提供了技术支撑。

Short-term Traffic-state Prediction of Urban Road Networks Based on the Fusion of a Link-transmission Model and Deep Learning

Short-term traffic prediction of urban road networks is a key technology for the development of smart cities. With the development of artificial intelligence, an increasing number of deep-learning algorithms have been applied to traffic estimation and the prediction of urban road networks. However, deep learning cannot describe the evolution mechanism of a traffic flow, which leads to poor interpretability, and the application of analytical transportation models is often limited owing to low prediction accuracy. In this study, the link-transmission model (LTM) was improved, and a data-driven link-transmission model (D²LTM) was proposed. D²LTM uses real-time data to calibrate the simulation network, which improves the prediction accuracy. Accordingly, a spatio-temporal deep tensor neural network (ST-DTNN) was introduced to capture the time-, space-, and depth-dimension feature information in the network-wide traffic data to form a hybrid model, D²LTM-STDTNN, which is a short-term traffic-prediction model for urban road networks. The hybrid model utilized the D²LTM model to reveal the fundamental laws of traffic evolution and could finely describe the temporal and spatial evolution of the urban road network-wide traffic state, which enhanced its interpretability. Furthermore, it employed the ST-DTNN model's capabilities to mine high-dimensional data and dynamic features, improving the accuracy of short-term traffic prediction. The hybrid model also considered the short-term prediction of different turns at intersections, which have finer spatial and temporal granularities and greater prediction difficulties. The actual measurement results show that the D²LTM-STDTNN hybrid model has a higher prediction accuracy than the benchmark model and has the advantage of simulating the evolution mechanism. The model improves the short-term traffic-prediction ability of urban road networks and reveals the dynamic evolution rules of traffic flow between road sections. This model can facilitate network-wide traffic simulation and help realize active transportation management and control.