The development of multi-mode transportation systems, e.g., bus, metro, taxi, and bike-sharing, presents a fundamental challenge in forecasting demand across heterogeneous, noisy, and complexly interacting data streams. From a feature modeling perspective, this requires a shift from simple data fusion to a more principled approach. This paper introduces a novel end-to-end framework, Multi-mode Spatiotemporal Adaptive Fusion Network (MSTAFN), that systematically addresses this challenge through a two-stage process: 1) unsupervised shared feature selection, and 2) dynamic asymmetric feature interaction modeling. For the first stage, we design an Infomax module that employs an information-theoretic principle to obtain a clean low-dimensional shared latent representation from cross-mode data. This representation captures the underlying semantic drivers of demand, such as latent commuting patterns, while mitigating noise and redundancy. For the second stage, we propose a Multi-Flashback module to explicitly model the complex asymmetric interactions between heterogeneous features, particularly across different temporal granularities. Its Cross-Flashback mechanism is designed to allow low-frequency modes (e.g., metro) to be informed by the latest fine-grained dynamics of high-frequency modes (e.g., bike-sharing). Experiments on a large-scale real-world dataset from New York City demonstrate that our two-stage paradigm outperforms state-of-the-art baselines, especially on the mode with the coarsest time granularity. This validates the superiority of our proposed modeling framework, supporting the improvement of operational efficiency for multi-mode transportation systems.

While considerable research has been dedicated to the prediction of area trip demand, relatively little attention has been paid to the area trip attraction prediction problem. A primary challenge in short-term trip attraction prediction stems from the partial observability of destination information, as trips are either still en route or have recently arrived but not yet confirmed as completed, leading to delayed and incomplete measurement of attraction dynamics. This paper proposes the field theory-guided area trip attraction prediction fusion network (FG-ATAPFN), which leverages potential energy fields (PEFs) to model underlying physical principles of urban dynamics. Furthermore, it establishes a connection between trip attraction and traffic flow through the continuity equation derived from field theory. FG-ATAPFN consists of three pivotal components: spatial-temporal Transformer to model spatial-temporal dependencies in historical area trip attraction sequences effectively, an OD flow PEF prediction network to derive precise trip destination insights, and an edge flow PEF prediction network to track and predict real-time human mobility flows dynamically, all leveraging both historical and real-time data to enable a comprehensive and precise analysis of travel behaviors within urban transportation systems. The proposed model is validated on an open-sourced real-world dataset comprising human movement trajectories from 100,000 users in Nagoya, Japan, and benchmarked against state-of-the-art prediction models. Experimental results indicate that FG-ATAPFN achieves a 15.5% reduction in mean absolute percentage error, surpassing baseline models in predicting trip attraction. We demonstrate that field theory effectively enhances the accuracy of trip attraction prediction, which is critical for optimizing transportation infrastructure design and guiding urban policy.

Addressing traffic congestion is significant in enhancing urban mobility. Traditional navigation systems use real-time traffic status to push the temporally shortest path to drivers, forming selfish-routing, which guide traffic flow to the same roadway, thereby causing imbalance traffic flow distribution and navigation-induced congestion. Besides, navigation based on real-time detected traffic metrics may cause the congestion oscillation during periods with high travel demand variations. This paper proposes the Traffic Prediction and Reinforcement Learning-based Navigation Guidance (TP-RLNG) approach for active control of traffic flow on urban expressways. The TP-RLNG introduces a differentiated guidance approach with multiple origin-destination (OD) pairs, replacing the conventional all-or-nothing route guidance strategies. To enhance the stability, the TP-RLNG integrates traffic prediction and reinforcement learning (RL)-based dynamic optimization to prospectively harmonize traffic supply and demand overall the expressway system. In contrast to the reinforcement learning navigation guidance (RLNG) (no prediction) approach, the integration of traffic prediction model in TP-RLNG enhances the anticipation of critical nodes and enables proactive traffic management. We use SUMO-based traffic simulation to examine the efficacy of TP-RLNG and alternative approaches in controlling traffic flow within the Hangzhou Liushi Expressway network under varying demand patterns. The results underscore the ability of TP-RLNG to enhance road network efficiency and mitigate urban expressway congestion. The findings indicate that, relative to RLNG approaches, our methodology achieves a reduction in average travel time of 9.3%.

Individual next location prediction plays a crucial role in location-based applications, such as route navigation and service recommendation. Although the existing research based on deep learning effectively captures users' spatiotemporal travel preferences, there are challenges in the interpretability of location prediction, heavily relying on large-scale historical travel data for model training. Drawing inspiration from the powerful reasoning capabilities of large language models (LLMs), this study proposes a novel multiround continuous dialogue mechanism and candidate set enhancement method, leveraging LLMs for next location prediction through step-by-step reasoning. In the first round of dialogue, we introduce activity prediction as an auxiliary task to narrow down the candidate locations. Subsequently, we establish an activity-aware prompt to enable LLM to achieve accurate location prediction and provide corresponding reasoning. Finally, we incorporate a third round of dialogue to prompt LLM to make necessary corrections by integrating the prediction results of deep learning models. To address the issues of LLMs being affected by element ranking within the candidate set, we propose a new candidate set enhancement method based on the entropy-weighted Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS). Our model can understand user travel preferences by fusing location, activity, and time information through natural language. Extensive experiments are conducted on two public datasets of user check-ins, and the results show that our model achieves prediction performance comparable to deep learning models in full-sample prediction and outperforms them in the few-shot settings. Our model provides logical and explainable reasoning, offering insightful guidance for downstream application tasks.

Data-driven methods have demonstrated high trajectory prediction accuracy with advancements in trajectory monitoring technology and deep learning algorithms. Human driving behavior shows significant heterogeneity, and aggressive driving behavior poses a significant threat to traffic safety. However, the aggressive trajectory data are usually scarce and few-shot, which makes it difficult to model. A cross-in domain transfer learning model (CIDTL) tailored for vehicular trajectory prediction under few-shot conditions of aggressive driving behavior is proposed to tackle the above challenges. CIDTL aligns cross-domain features from different road intersections and in-domain features of three types of driving behaviors, which can improve the prediction performance on few-shot samples in the target domain by leveraging sufficient data from source domains. Specifically, we develop a simple yet efficient method for driving behavior clustering, dividing trajectory data into three distinct styles. Then, CIDTL is proposed to align cross-domain features and in-domain features simultaneously. To evaluate the effectiveness of our proposed model, we conduct numerical experiments on the inD dataset. The results show that accurate predictions can be achieved even with limited aggressive driving behavior data, which indicates that CIDTL can successfully reduce the distribution difference of trajectory data between ordinary driving behavior and aggressive driving behavior, as well as the difference between source domains and target domain, by learning common driving knowledge and domain-invariant knowledge. The proposed model is robust and has good generalization, which expands the application scope of traditional trajectory prediction models in reality.

This paper revisits the Reactive Dynamic User-Equilibrium (RDUE) model for dynamic traffic assignment (DTA) of macroscopic traffic flow in two-dimensional continuum space, focusing on the Eikonal equation—a crucial partial differential equation (PDE) with specific boundary conditions. Traditionally, solving Eikonal equations has relied on iterative numerical methods through the discretization of the continuum space. However, this discretization compromises the precision of numerical solutions and could lead to non-convergence issues during iterative processes. This study refers to Physics-Informed Neural Networks (PINNs) and develops the Distance-Informed Neural Eikonal Solver (NES-DI) for solving Reactive Dynamic User-Equilibrium models. While the previously proposed Neural Eikonal Solver (NES) performs badly in a strong heterogeneous cost field with large cost differences, NES-DI explicitly considers the influence of solid boundaries during the factorization process by incorporating accurate distance information. Numerical examples of RDUE at both the static and dynamic levels are presented to illustrate the performance and applications of the NES-DI framework. The results demonstrate that NES-DI greatly outperforms both NES and the fast sweeping method. Moreover, NES-DI overcomes the limitations of discretization, enabling predictions of solutions at arbitrary locations within the computational domain. At the dynamic level, transfer learning is employed to leverage historical solutions to solve RDUE problems more efficiently. Overall, NES-DI shows the potential of solving reactive dynamic problems with strong heterogeneity, which offers a promising alternative to discretization-reliant numerical methods.

Accurately predicting human mobility is crucial for various applications, e.g., transportation services, epidemic control, and advertisement recommendation. Although numerous sequential modeling based methods (e.g., recurrent neural networks) have been proposed for human mobility prediction, accurately modeling individuals’ high-order travel preferences and the influence of social neighbors on their travel decisions remains challenging. In this paper, we construct a novel multi-context aware model for next location prediction, which aggregates multi-dimensional contextual features, including individual preferences, social relations, and activity-location associations. First, we define activity prediction as an auxiliary task and propose an activity-location association pruning method to mitigate the impact of data sparsity on model prediction. Second, we present a novel motif-preserving individual travel preference learning method that leverages a motif-induced hypergraph convolutional network to capture high-order travel preference features explicitly. Third, we identify virtual social neighbors with similar preferences based on individual travel preference learning results, and design a new social gated fusion structure to model the influence of social neighbors on individual travel choices. Finally, experimental results on two real-world travel datasets demonstrate the superiority of the proposed model over baseline models. Our proposed universal method can be seamlessly integrated with other sequential prediction models to improve the accuracy and stability of human mobility prediction.

Recent research has shown that some network traffic flow patterns are similar across multiple traffic regions. Identifying and transferring these domain-invariant features can significantly boost model accuracy and robustness, providing new insights into dealing with modeling issues like traffic data insufficiency and dataset shift. However, how to acquire transferable network traffic flow patterns from multiple traffic regions and adapt such knowledge to downstream prediction tasks of target regions remains challenging. To realize domain-invariant traffic flow pattern transfer and provide more robust prediction under insufficient data conditions, we propose a macroscopic fundamental diagram (MFD) guided transfer learning method, namely physics-guided multi-source domain adversarial network (PG-MDAN). First, an MFD similarity measure is proposed to determine what traffic flow patterns are transferable and to what extent they can be transferred. PG-MDAN embeds this physics-informed transferability measure in domain adversarial pre-training for better adaptation ability. Numerical experiments based on two real-world urban network traffic datasets show that PG-MDAN can successfully transfer recurrent and non-recurrent network traffic flow patterns from multiple regions to provide more robust and responsive prediction performance. Finally, extensive sensitivity analysis is conducted, and the results validate that applying such physical regularization can effectively avoid negative transfer and provide a flexible tool to initiate traffic flow pattern transfer in practice.

Metro bears a large number of passenger flows in urban transportation systems. Short-term metro origin-destination (OD) passenger flow prediction is an essential component of intelligent transportation systems (ITS), which allows operators to better monitor the metro system and improve the level of service for passengers. In this paper, we exploit a novel data structure, hypergraph, to represent the complex correlation between OD pairs, and propose an elaborately designed Spatio-Temporal Dynamic Attentive Multi-HyperGraph Network (ST-DAMHGN) to tackle the short-term OD passenger flow prediction problem. In the proposed framework, we construct multiple hypergraphs to model the relationship between OD pairs and adopt the perceptual field to realize efficient and effective vertex feature extraction. Then, we utilize the attention mechanism to adaptively and dynamically synthesize information from multiple hypergraphs and make a trade-off between exploration and exploitation. A case study is conducted on the metro system of Hangzhou, China. The results of extensive experiments show that ST-DAMHGN outperforms baseline models. The efficiency is validated for the multi-hypergraph model, perceptual field, spatial feature extraction, and attention mechanism. The hypergraph structure used in our model is verified suitable for modeling traffic data without physical road networks to map directly, e.g., OD passenger flow. ST-DAMHGN can be widely implemented by defining proper correlations for model relationships.

Understanding human drivers’ intentions and predicting their future motions are significant to connected and autonomous vehicles and traffic safety and surveillance systems. Predicting multimodal vehicular trajectories at urban unsignalized intersections remains challenging due to dynamic traffic flow and uncertainty of human drivers’ maneuvers. In this paper, we propose a comprehensive trajectory prediction framework that combines a multimodal trajectory generation network with inverse reinforcement learning (IRL) and risk aversion (RA) modules. Specifically, we first construct a multimodal spatial-temporal Transformer network (mmSTTN) to generate multiple trajectory candidates, using trajectory coordinates as inputs. Accounting for spatio-temporal features, we formulate the IRL reward function for evaluating all candidate trajectories. The optimal trajectory is then selected based on the computed rewards, a process that mimics human drivers’ decision-making. We further develop the RA module based on the driving risk field for optimal risk-averse trajectory prediction. We conduct experiments and ablation studies using the inD dataset at an urban unsignalized intersection, demonstrating impressive human trajectory alignment, prediction accuracy, and the ability to generate risk-averse trajectories. Our proposed framework reduces prediction errors and driving risks by 25% and 30% compared to baseline methods. Results validate vehicles’ human-like risk-averse diverging-and-concentrating behavior as they traverse the intersection. The proposed framework presents a novel approach for forecasting multimodal vehicular trajectories by imitating human drivers and incorporating physics-based risk information derived from the driving field. This research offers a promising direction for enhancing the safety and efficiency of connected and autonomous vehicles navigating urban environments.

Simultaneously and accurately predicting trajectories of multiple heterogeneous agents is crucial for intelligent transportation systems (ITS) applications, e.g., connected and autonomous vehicles. Existing model-based and data-driven methods can achieve good prediction accuracy, but most of them neglect the domain shift issue and prevalent imperfect data problems, i.e., few-shot learning and zero-shot learning issues. To address these issues, we propose a multi-source transfer learning (TL) framework, transferable hierarchical Siamese Transformer network (T-HSTN), for trajectory prediction of multiple heterogeneous agents, e.g., vehicles, bicycles, and pedestrians, at urban unsignalized intersections under small data conditions. Specifically, by extending the self-attention mechanism and exploring feature representations of traffic scenes, a Transformer-based network that hierarchically extracts temporal/spatial features and map features is introduced as the basic prediction model. Moreover, a TL framework with adaptive learning and feature alignment modules is built to explore the feature representations of unfixed traffic scenes and align both statistical and deep features to learn domain-invariant knowledge. More challenging trajectory prediction experiments are designed, corresponding to newly-built or badly-instrumented intersections under real-world scenarios. Experimental results verify the proposed method’s high accuracy, transferability, and generability. Our work fills the gap in solutions and benchmarks for TL tasks in trajectory prediction for heterogeneous agents. The conducted TL experiments provide a more practical setting of considering imperfect data problems in trajectory prediction.

Forecasting of traffic conditions plays a significant role in smart traffic management systems. With the prevalent use of massive vehicle trajectory data, agencies inevitably encounter missing data issues that hinder traffic flow forecasting in an urban road network. This paper studies the urban network-wide short-term forecasting of traffic speed with consideration to missing link speed data via (i) a data recovery algorithm to impute missing speed data for the segment network with nonlinear spatial and temporal correlations; and (ii) forecasting of spatially heterogeneous traffic speed within the road network using the GraphSAGE model. The influences of partially missing data and recovered data on the traffic speed forecasting are investigated. A case study of the urban area in Hangzhou, China, is presented, and it is found that the proposed recovery algorithm has the best performance in terms of traffic speed information reconstruction compared to benchmark methods. The case study also shows that using the recovered data acquires higher accuracy and efficiency in the short-term speed forecasting, compared to the case of using the original data without recovery. The proposed methods tackle missing traffic data issues and forecasting problems in the presence of missing data in an urban road network.

Real-time traffic speed prediction is an essential component of intelligent transportation systems applications on large-scale urban networks, e.g., proactive traffic management, advanced information provision, and prompt incident response. The family of traffic prediction models (e.g., convolutional neural networks) based on multi-detector speed diagrams in the time-space plane has been one of the most frequently used approaches for individual roads and the entire network. However, the predefined stacking sequence of traffic detectors along the spatial dimension of the speed diagram has a significant influence on the prediction performance, which makes network-wide speed prediction more challenging. To tackle the above challenge and better capture complicated traffic dynamics, we propose a novel speed prediction approach, named spatial-temporal deep tensor neural networks (ST-DTNN), for a large-scale urban network with mixed road types. Spatial and temporal dependencies of different road segments are simultaneously taken into account to improve the network-wide prediction accuracy. A scalable deep tensor is constructed for the ST-DTNN to eliminate the potentially negative impact caused by the manually stacking sequence of speed time series collected at different locations. Multi-step ahead traffic speeds can be simultaneously predicted based on probe data for a real-world large-scale urban network with hundreds of detectors installed on freeways, highways, and major/minor arterials. The results demonstrate the capability and effectiveness of the proposed ST-DTNN approach. Compared with the benchmark models, the ST-DTNN performs higher prediction accuracy during either peak or off-peak periods within an acceptable training time and has more stable prediction performance on the spatial scale. The proposed approach can be extended to develop network-wide traffic state monitoring, optimize routing in navigation services, and support congestion mitigation.

Spatial-temporal traffic state estimation and the prediction of urban expressways is a vital component of traffic management and information systems. The adaptive smoothing method is one of the most frequently used approaches to estimate traffic states. However, the fixed filter parameters used in existing approaches sometimes fail to characterize traffic dynamics well. To better capture generation, propagation, and mitigation dynamics of traffic congestion, we propose an adaptive rolling smoothing (ARS) approach by dynamically tuning the filter parameters in a rolling horizon scheme for online applications. The fusion of heterogeneous traffic data combines aggregate traffic measurements (e.g., traffic flow rate, time occupancy, and speed collected by remote microwave sensors) and disaggregate information (e.g., timestamps of individual vehicles detected by license plate recognition cameras). A nonlinear traffic flow filter based on the virtual trajectory algorithm is established to reconstruct the spatial-temporal traffic state and estimate experienced travel times of individual vehicles. The results demonstrate the capability and effectiveness of the proposed ARS approach in the historical traffic state estimation and short-term traffic flow prediction. Complicated traffic states of weaving, merging, and diverging segments can be well distinguished by reconstructing time-space speed diagrams. The proposed approach can be extended to develop efficient missing data imputation algorithms and hierarchical control strategies for heterogeneously congested urban expressways.

Ride-sourcing services are becoming an increasingly popular transportation mode in cities all over the world. With real-time information from both drivers and passengers, the ride-sourcing platform can reduce matching frictions and improve efficiencies by surge pricing, optimal vehicle-trip assignment, and proactive ridesplitting strategies. An important foundation of these strategies is the short-term supply-demand forecasting. In this paper, we tackle the problem of predicting the short-term supply-demand gap of ride-sourcing services. In contrast to the previous studies that partitioned a city area into numerous square lattices, we partition the city area into various regular hexagon lattices, which is motivated by the fact that hexagonal segmentation has an unambiguous neighborhood definition, smaller edge-to-area ratio, and isotropy. To capture the spatio-temporal characteristics in a hexagonal manner, we propose three hexagon-based convolutional neural networks (H-CNN), both the input and output of which are numerous local hexagon maps. Moreover, a hexagon-based ensemble mechanism is developed to enhance the prediction performance. Validated by a 3-week real-world ride-sourcing dataset in Guangzhou, China, the H-CNN models are found to significantly outperform the benchmark algorithms in terms of accuracy and robustness. Our approaches can be further extended to a broad range of spatio-temporal forecasting problems in the domain of shared mobility and urban computing.

Data fusion is an approach that combines multiple data sources for a more efficient statistical purpose. There have been some explorations on the application of data fusion for short-term traffic predictions. Unlike the previous work, this paper attempts to propose a probabilistic data fusion approach. This approach regards different data sources as random variables with some empirical distributions, and it attempts to fuse the data sources with the consideration of their probability distributions so as to improve probabilistic inference and hypothesis test. The density ratio model is introduced and utilized for this probabilistic data fusion approach, which estimates a fused probability distribution with different data sources. Real-world case studies are conducted to investigate the goodness-of-fit of the probabilistic data fusion and its impact on traffic predictions. This paper finds that probabilistic data fusion can improve the prediction accuracy when the fused probability distribution contains “incomplete” characteristics of the empirical distribution.

Network design problems (NDPs) have long been regarded as one of the most challenging problems in the field of transportation planning due to the intrinsic non-convexity of their bi-level programming form. Furthermore, a mixture of continuous/discrete decision variables makes the mixed network design problem (MNDP) more complicated and difficult to solve. We adopt a surrogate-based optimization (SBO) framework to solve three featured categories of NDPs (continuous, discrete, and mixed-integer). We prove that the method is asymptotically completely convergent when solving continuous NDPs, guaranteeing a global optimum with probability one through an indefinitely long run. To demonstrate the practical performance of the proposed framework, numerical examples are provided to compare SBO with some existing solving algorithms and other heuristics in the literature for NDP. The results show that SBO is one of the best algorithms in terms of both accuracy and efficiency, and it is efficient for solving large-scale problems with more than 20 decision variables. The SBO approach presented in this paper is a general algorithm of solving other optimization problems in the transportation field.

Freeway traffic state estimation is a vital component of traffic management and information systems. Macroscopic-model-based traffic state estimation methods are widely used in this field and have gained significant achievements. However, tests show that the inherent randomness of traffic flow and uncertainties in the initial conditions of models, model parameters, and model structures all influence traffic state estimations. To improve the estimation accuracy, this paper presents an ensemble learning framework to appropriately combine estimation results from multiple macroscopic traffic flow models. This framework first assumes that any models existing are imperfect and have their own strengths/weaknesses. It then estimates the online traffic states in a rolling horizon scheme. This framework automatically ensembles the information from each individual estimation model based on their performance during the selected regression horizon. In particular, we discuss three weighting algorithms, namely, least square regression, ridge regression, and lasso, which represent different presumptions of model capabilities. A field test based on real freeway measurements indicates that lasso ensemble best handles various uncertainties and improves estimation accuracy significantly. It should be also pointed out that the proposed framework is a flexible tool to assemble nonmodel-based traffic estimation algorithms. This framework can be also extended for many other applications, including traffic flow prediction and travel-time prediction.

Travel time serves as a fundamental measurement for transportation systems and becomes increasingly important to both drivers and traffic operators. Existing speed interpolation algorithms use the average speed time series collected from upstream and downstream detectors to estimate the travel time of a road link. Such approaches often result in inaccurate estimations or even systematic bias, particularly when the real travel times quickly vary. To get rid of this problem, Coifman proposed a creative interpolation algorithm based on kinetic-wave models. This algorithm reconstructs vehicle trajectories according to the velocities and the headways of vehicles. However, it sometimes gives significant biased estimation, particularly when jams emerge from somewhere between the upstream and downstream detectors. To make an amendment, we design a new algorithm based on the temporal-spatial queueing model to describe the fast travel-time variations using only the speed and headway time series that is measured at upstream and downstream detectors. Numerical studies show that this new interpolation algorithm could better utilize the dynamic traffic flow information that is embedded in the speed/headway time series in some special cases.

In this paper, we propose a simple temporal-spatial queueing model to quantitatively address some typical congestion patterns that were observed around on/off-ramps. In particular, we examine three prime factors that play important roles in ramping traffic scenarios: the time τinfor a vehicle to join a jam queue, the time τoutfor this vehicle to depart from this jam queue, and the time intervalTfor the ramping vehicle to merge into the mainline. Based on Newell's simplified car-following model, we show how τinchanges with the main road flow rateqmain. Meanwhile,Tis the reciprocal of the ramping road flow rateqramp. Thus, we analytically derive the macroscopic phase diagram plotted on theqmain-versus-qrampplane and τin-versus-Tplane based on the proposed model. Further study shows that the new queueing model not only reserves the merits of Newell's model on the microscopic level but helps quantify the contributions of these parameters in characterizing macroscopic congestion patterns as well. Previous approaches distinguished phases merely through simulations, but our model could derive analytical boundaries for the phases. The phase transition conditions obtained by this model agree well with simulations and empirical observations. These findings help reveal the origins of some well-known phenomena during traffic congestion.

In this paper, we link two research directions of road traffic-the mesoscopic headway distribution model and the microscopic vehicle interaction model-together to account for the empirical headway/spacing distributions. A unified car-following model is proposed to simulate different driving scenarios, including traffic on highways and at intersections. Unlike our previous approaches, the parameters of this model are directly estimated from the Next Generation Simulation (NGSIM) Trajectory Data. In this model, empirical headway/spacing distributions are viewed as the outcomes of stochastic car-following behaviors and the reflections of the unconscious and inaccurate perceptions of space and/or time intervals that people may have. This explanation can be viewed as a natural extension of the well-known psychological car-following model (the action point model). Furthermore, the fast simulation speed of this model will benefit transportation planning and surrogate testing of traffic signals.