Predicting individual mobility is essential for transport planning and sustainability strategies, as it helps anticipate travel demand. However, most studies of student mobility remain descriptive and cannot be applied to predict trip attributes, despite such prediction being essential for proposing personalised mobility recommendations that account for students' contextual constraints, such as academic schedules. This paper introduces CAH2RL-Move, a two-stage method for predicting student trip features, including origin, destination, travel mode, and departure time. The approach uses recurrent encoders with attention mechanisms to capture spatial routines and variations in travel mode, then applies deep reinforcement learning to model the influence of academic schedules on departure times. The model is validated using a dataset constructed from real student mobility preferences and actual academic schedules. Compared with baseline approaches, CAH2RL-Move achieves 80.74%-88.61% departure time prediction accuracy within 30-minute windows across four student mobility cases, enabling informed mobility recommendations.
Ride-sharing is an emerging mode that can effectively alleviate urban transport pressure. However, real-world transport systems may frequently face various traffic conditions, such as congestion. This paper proposes a responsive ride-sharing dispatching model, in which vehicles are allowed to divert from their current routes to respond to sudden and unforeseen events. The assumptions and formulation are established with consideration of real-world road networks. A variable neighbourhood search with intelligent search reduction (VNS-ISR) is developed to solve the problem efficiently, particularly under time-critical dispatching tasks. The proposed model and algorithm are validated using real-world trip data from New York, US. Results show that the model improves solutions by 7.88%-23.47% under traffic events. Moreover, a Manhattan case simulation highlights the expected environmental benefit equivalent to a 93,462 L reduction in gasoline consumption in annual operations.
To accommodate fluctuating public transport (PT) passenger demand and recharging management of battery-electric vehicles, a multi-depot multi-type vehicle scheduling problem is modelled to consider not only heterogeneous batteries and load capacities, but also timetabling as an emerging study for heterogeneous fleets. The objective is to minimize the cost of the number of chargers, recharging queue times, and fleet size. The network-level PT operations planning is extended to include arcs to represent deadheading trips. A mixed integer linear programming model is derived by means of linearization techniques. The complex multidimensional computation problem is decomposed by a Lagrangian relaxation-based algorithm. Three algorithmic enhancements, including domain reduction, valid cuts, and multiple chains' identification, are embedded into the tailored heuristic to facilitate the relaxed subproblems' convergence. We use the deficit function to track the dynamic number of chargers. The results show a significant total cost reduction of 35.37% with a 36.11% decrease in the total PT fleet size compared with the benchmark scheme.
Navigation systems improve efficiency but expose traffic networks to route guidance attacks that manipulate routing information. We develop an analytical framework that integrates the generalized bathtub model with three driver behaviours (perfect rationality, stochastic logit, bounded rationality) to quantify network-level resilience under falsified travel times. The framework maps misinformation to traffic redistribution and evaluates impacts using total system travel time, a resilience index, and cumulative performance loss. Across attack intensities in our single-reservoir setting, we found that perfectly rational users produce the largest delays, whereas boundedly rational users preserve baseline performance for small attacks (below the indifference band), and logit users reduce losses relative to perfectly rational users across all attack levels. We identify a critical threshold aligned with the drivers' indifference band: beyond this point, widespread rerouting triggers congestion growth, a Braess-like inefficiency under misperceived costs. The results provide quantitative evidence that heterogeneity in driver behaviour modulates RGA impacts at the network level and inform detection thresholds as well as dispersion-based mitigation strategies.
Bus fleet electrification is a critical approach for reducing transportation emissions and promoting sustainable development. However, effectively scheduling mixed fleets of diesel buses (DBs) and electric buses (EBs), while assessing the associated economic and environmental impacts remains an operational challenge. To address this issue, we develop a mixed integer programming model that assigns DBs and EBs to scheduled trips and manages EB charging activities, aiming to minimize operational and emission costs. Building on the lower operating expenses and emissions of EBs, we propose an EB priority scheduling (EBPS) algorithm, which assigns EBs to as many trips as possible with DBs covering any remaining trips. Numerical experiments on synthetic datasets demonstrate that EBPS yields near-optimal solutions comparable to those produced by Gurobi, while requiring substantially less computational time. A real-world case study in Dalian emphasizes the importance of comprehensive scheduling approaches to capture economic and environmental benefits of bus electrification.
Urban public transportation systems are critical for mass mobility, but unexpected disruptions can cause delays and passenger congestion, challenging operational reliability and safety. This paper investigates the integrated optimization of train rescheduling, rolling stock circulation, and passenger flow control under metro disruptions. We develop a mixed-integer nonlinear programming model to minimize timetable deviation, passenger waiting time, and operating costs associated with skip-stop operations and short-turning strategies. To handle uncertain passenger demand, we adopt a scenario-based robust optimization model that remains feasible across all scenarios. To reduce the computational complexity introduced by bilinear terms in passenger waiting time, we apply a piecewise McCormick envelope method combined with a bound contraction algorithm, enabling rapid generation of high-quality solutions. The proposed approach is validated using data on the Beijing Yizhuang Line, demonstrating its robustness and practical applicability for punctuality and safety in disrupted metro operations.
This study develops a dual-dimension dynamic day-to-day evolution model for mixed traffic with autonomous vehicles (AVs) and regular vehicles (RVs), capturing commuters' transition from user equilibrium (UE) to system optimum (SO). Distinct AV and RV travel-cost functions are specified, and evolution processes are formulated under dynamic user equilibrium (DUE) and dynamic system optimum (DSO) principles. Analytical results and numerical illustrations show that equilibrium states under both principles converge and remain stable. Under DUE, an efficiency-loss traffic paradox occurs, with AVs experiencing longer queueing times than RVs. Under DSO, congestion charges and subsidies regulate departure-time choices, balance AV arrivals, and reduce queueing times close to zero in the benchmark case, thereby mitigating the paradox and moving the system toward SO. Sensitivity analysis indicates that DUE optima exist within a range, whereas the DSO optimum concentrates near a fixed point in the tested setting.
This paper proposes a customized bus network design using modular vehicles and in-vehicle transfers. Four test scenarios (Modular-with-Transfer, Non-Modular-without-Transfer, Modular-without-Transfer, and Non-Modular-with-Transfer) are established. A mixed-integer nonlinear programming (MINLP) model with exclusive transfer constraints is formulated and a linearization strategy transforms it into tractable mixed-integer linear programming (MILP). Standard particle swarm optimization (PSO) is revised into a route-sequence-aware variant for better convergence on large networks. A synergy quantification framework quantifies combined benefits of modularity and transfers Validated on Sioux Fallsand Beijing transit data, the method cuts system cost by up to 20.82%, shortens travel distance by 26.89%-42.86% and lifts load factor by 8.24%-37.42%. For Beijing's real network, cost drops 1.21%-2.64%, load factor rises 22.72%-29.27%, andtravel time falls 0.2%-1.23%. Findings reveal route overlap dominates passenger choices, and coordinated expansion realizes synergistic load balancing.
Pedestrian safety at unsignalized crosswalks remains a major urban challenge, and V2X technology offers promising mitigation. This study examines the effects of pedestrian crossing facilities (no crosswalk, standard crosswalk, crosswalk with yellow flashing lights), warning lead times (4.5, 5.5, 6.5 s), and warning modalities (image, voice prompt, voice command, and combined image-voice) on driver yielding behaviour using driving simulation experiments. An explainable machine learning framework (XGBoost with SHAP) modeled drivers' decisions to stop or decelerate when yielding. Results show that crosswalks with yellow flashing lights significantly promote earlier and smoother braking and increase full-stop yielding. Multimodal warnings were most effective, particularly at longer lead times, whereas short lead times often induced emergency braking. Interaction effects indicate that multimodal warnings are critical on roads without crosswalks, while visual warnings alone are sufficient at flashing-light crosswalks. Key thresholds distinguishing yielding strategies were identified through braking and eye-movement indicators.
This article proposes and evaluates predictive models for quantifying, assessing, and managing the impact of road or intersection closures on the performance of an urban road network. We measure the network's performance loss using resilience-based metrics and apply rerouting to mitigate the adverse effects of road closures. The simulations are conducted in the extended Athens city centre. The findings indicate that the optimal area for implementing rerouting strategies varies depending on whether a single or multiple corridors are closed. By integrating this information with other factors, such as the number of closed corridors and traffic flow, we develop a supervised machine learning model to predict performance declines in the network resulting from road closures.
Transit network design has been studied for decades, but practical transit route planning remains manual and iterative due to limited travel demand data, lack of transit route mapping, and simplified transit assignment. This study presents a multimodal transit network design framework that addresses these challenges by integrating high-resolution travel demand data, precise route mapping, and an advanced multimodal assignment model. Unlike traditional approaches, we utilise real-world cellular signalling data from 8.8 million users to capture dynamic demand patterns, reconstruct transit routes using GPS data and web mapping APIs, and develop an assignment model that accounts for strict capacity constraints, a boarding priority mechanism, and the common line problem. The problem is solved efficiently using a parallel genetic algorithm, allowing application to large-scale networks. The implementation of this framework in Foshan, China, demonstrates that the optimized network outperforms the existing network in terms of network structure, operation, and ridership.
With respect to urban flooding frequency, insufficient mitigation resources cannot cover all inundation sites, making efficient drainage strategy development crucial for improving traffic performance. Evaluating candidate strategies requires traffic simulations to capture the nonlinear impacts of flooding and drainage strategies on urban mobility, but these simulations are computationally expensive and non-differentiable. Conventional methods typically treat interrelated scenarios induced by drainage strategies independently, limiting the potential for knowledge transfer. To overcome these challenges, this study proposes a Simulation-Based Optimization framework based on Multi-Task Representation Learning (MTRL-SBO). A shared encoder is employed to extract task-invariant features, and Conditional Batch Normalization (CBN) is introduced to model task-specific variations. In addition, task-specific surrogate models are trained to accelerate strategy evaluation and guide the optimization search. A case study in central Guangzhou shows that the optimal drainage strategy reduces the average traffic delay by 10.3%, demonstrating superior optimization capability and practical value for urban flood mitigation planning.
In mixed-traffic environments, autonomous vehicles must interact with human drivers exhibiting strategic but bounded-rational behaviors, which introduces significant uncertainty for decision-making. To address this challenge, we propose CHSAC (Cognition-aware Hierarchical Soft Actor-Critic), a novel framework that integrates cognitive modeling with deep reinforcement learning for safe highway lane-change planning. CHSAC unifies risk assessment, real-time inference of surrounding vehicles' cognitive levels, and belief-augmented policy optimization, enabling the vehicle to anticipate human intentions and adjust its policy foresightedly. Extensive simulations demonstrate that CHSAC significantly outperforms competitive reinforcement learning and game-theoretic baselines in safety and efficiency, achieving near-zero collisions while maintaining smooth and efficient driving. These results highlight the effectiveness of incorporating cognitive belief inference into deep policy learning, offering an adaptive and robust decision-making paradigm for autonomous driving in complex mixed-traffic scenarios.
Physics-informed Gaussian processes can improve traffic prediction, but their performance depends strongly on where physical constraints are imposed. This paper proposes a robust acquisition strategy for adaptive virtual-point selection. The method combines predictive uncertainty with the severity of physical inconsistency and accounts for uncertainty in the traffic fundamental diagram, so that virtual points are placed where they are most informative for both data fitting and physical consistency. A sequential training procedure then updates the model and adds new virtual points during learning. Experiments on synthetic and real-world corridor benchmarks show clear gains over random virtual-point placement in both traffic-state reconstruction and travel-time prediction. Across the tested corridor scenarios, the proposed method reduces travel-time prediction error by about 11% to 25% while using 60% fewer virtual points. It is especially effective under non-recurrent incident conditions such as lane closures.
With the growing adoption of electric vehicles (EVs), charging behavior has become intertwined with daily activities, leading to clustered charging demand at destinations such as workplaces and shopping districts. EV users need to consider both road congestion and queuing delays at charging stations, resulting in spatiotemporally coupled travel and charging decisions and potentially a mismatch between demand and limited infrastructure capacity. Given the dynamic evolution of charging processes, delays and congestion tend to propagate across time and space, posing challenges for traditional static analytical frameworks. This paper develops an integrated dynamic traffic assignment framework that combines charging demand-supply dynamics with travel behavior analysis. Waiting times and charging relocation incentives are embedded into a dynamic user equilibrium (DUE) model, formulated as a variational inequality and solved using a column-generation-based algorithm. Numerical experiments demonstrate that the charging relocation incentive can effectively improve charging supply utilization and the demand service rate.
The potential spread of infectious diseases, e.g. Influenza A, SARS, and COVID-19, has brought challenges to railway operations. One issue is how to improve railway utilization while implementing social distancing strategies. This study proposes to jointly optimize train stop planning, scheduling, and seat allocation when incorporating social distancing strategies (TSPS-SA-SD). The proposed approach incorporates heterogeneous social distancing requirements for passengers from areas associated with different levels of risks in epidemics/pandemics. To tackle the TSPS-SA-SD problem, an integer nonlinear programming (INLP) model is developed, which minimizes travel time and maximizes rail revenue while ensuring that social distancing strategies are respected. The proposed INLP is reformulated into an integer linear programming (ILP) model. Then, a decomposition-based heuristics solution (DHS) algorithm, which leverages the adaptive large neighborhood search (ALNS) and GUROBI solver, has been introduced to solve large instances. The proposed approach is evaluated on a small network and two real instances.
