Dynamic vehicular origin–destination (OD) flow is a fundamental component of traffic network modeling and its estimation has long been studied. Although ideal observing conditions and behavioral assumptions are often indispensable for estimation, day-to-day traffic recurrences and variations are seldom utilized to improve the estimation performance. In this paper, we propose a new method to recover day-to-day dynamic OD flows using both connected vehicle (CV) trajectories and automatic vehicle identification (AVI) observations. The method involves two modules: the first module provides reliable prior OD flows given limited observations, while the second module seeks the optimal estimates based on the prior OD flows. In the first module, linear projection is extended to consider temporal and spatial variation of the CV penetration rate, and non-negative Tucker decomposition (NTD) is adopted to address the data sparsity issue caused by the low CV penetration rate. In the second module, a self-supervised learning model called the latency-constrained autoencoder (LCAE) is established to search for the optimal OD flows according to the priors with given robust latent features. To avoid local minima and ensure consistency between estimates, a novel algorithm called adaptive sub-sample correction (ASC) is proposed and integrated into the optimization process of LCAE, which can iteratively correct the most inconsistent samples based on the day-to-day traffic flow characteristics. The proposed method is examined on an empirical urban arterial network, a calibrated simulation network, and a synthetic large-scale grid network. Our results indicated that the proposed method requires very few AVI detectors and CV trajectories to achieve competitive estimation performance against two benchmark models. Furthermore, general robustness to several factors with respect to observing conditions and data quality was investigated, and satisfactory scalability was also demonstrated in terms of both estimation accuracy and computational cost.

动态车辆起点-终点 (OD) 流是交通网络建模的基本组成部分，长期以来对其估计进行了研究。尽管理想的观测条件和行为假设对于估计通常是必不可少的，但很少利用日常交通重现和变化来提高估计性能。在本文中，我们提出了一种使用联网车辆 (CV) 轨迹和自动车辆识别 (AVI) 观察来恢复日常动态 OD 流的新方法。该方法涉及两个模块：第一个模块在给定有限观察的情况下提供可靠的先验 OD 流量，而第二个模块基于先验 OD 流量寻求最佳估计。在第一个模块中，扩展了线性投影以考虑 CV 渗透率的时空变化，采用非负塔克分解（NTD）来解决低CV渗透率导致的数据稀疏问题。在第二个模块中，建立了称为延迟约束自动编码器（LCAE）的自监督学习模型，以根据具有给定稳健潜在特征的先验搜索最佳 OD 流。为了避免局部最小值并确保估计之间的一致性，提出了一种称为自适应子样本校正（ASC）的新算法并将其集成到 LCAE 的优化过程中，该算法可以根据日常流量迭代校正最不一致的样本流动特性。所提出的方法在经验城市主干网络、校准模拟网络和合成大规模网格网络上进行了检查。我们的结果表明，所提出的方法需要很少的 AVI 检测器和 CV 轨迹来实现对两个基准模型的竞争性估计性能。此外，研究了与观测条件和数据质量有关的几个因素的总体稳健性，并且在估计精度和计算成本方面也证明了令人满意的可扩展性。