Haijian Bai
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Haijian Bai has a total of 24 co-authors. Here are the top co-authors by number of shared publications:
Weihua Zhang
School of Automotive and traffic Engineering, Hefei University of Technology, China
6 shared publications- Heng Ding
Hefei Univ. of Technology
4 shared publications Xingyu Chen
School of Automotive and traffic Engineering, Hefei University of Technology, China
3 shared publicationsHeng Ding
School of Automotive and traffic Engineering, Hefei University of Technology, China
2 shared publicationsWenjuan Huang
School of Automotive and Transportation Engineering, Hefei University of Technology, Hefei, China
2 shared publications- Xingyu Chen
School of Civil Engineering, Hefei University of Technology, Hefei, China
2 shared publications
Latest Publications
- 20 November 2026
- Published by SAGE Publications in Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering
A car-following model based on two dimensional optimal velocity function(OVM-2D): Designed to improve ability to focus on multiple lateral moving targets for autonomous vehicles
Classical single-lane car-following models (e.g. the Optimal Velocity Model (OVM)) often exhibit limitations in representing the smooth and adaptive behavior required for cross-following scenarios involving dynamic target switching, such as adjacent vehicle cut-in, temporary lane intrusion, and lead vehicle cut-out. These limitations may increase potential safety risks during vehicle interactions. To enhance the continuity and adaptability of autonomous vehicle motion control, this paper proposes a multi-target attentive car-following model, termed OVM-2D, built upon the classical OVM. The proposed model introduces a Flexible Transition Function (FTF) to characterize the capability of attending to multiple targets during lateral interactions, enabling the ego vehicle to dynamically identify the most influential car-following target among multiple leading vehicles distributed across different lanes and achieve smooth, adaptive target switching in cross-following scenarios. This mechanism provides a simple, interpretable, and behaviorally consistent framework for modeling dynamic target switching. Linear stability analysis and phase-plane evolution under large perturbations suggest an expanded stability region under the considered conditions. Cross-validation using the CitySim and TOD_VT datasets demonstrates stable trajectory-fitting performance and satisfactory cross-dataset transferability under the evaluated scenarios. Comprehensive evaluations across four representative cross-following scenarios indicate that, compared with the selected baseline models, the proposed model achieves lower trajectory-prediction errors while producing smoother longitudinal responses during target transitions, thereby supporting the effectiveness of the proposed target-switching mechanism. Meanwhile, it exhibits advantages in reducing jerk, estimated fuel consumption, and emissions while mitigating acceleration fluctuations caused by minor longitudinal and lateral disturbances, providing an effective framework for dynamic target switching in complex multi-lane traffic.
- 20 November 2026
- Published by SAGE Publications in Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering
A car-following model based on two dimensional optimal velocity function(OVM-2D): Designed to improve ability to focus on multiple lateral moving targets for autonomous vehicles
Classical single-lane car-following models (e.g. the Optimal Velocity Model (OVM)) often exhibit limitations in representing the smooth and adaptive behavior required for cross-following scenarios involving dynamic target switching, such as adjacent vehicle cut-in, temporary lane intrusion, and lead vehicle cut-out. These limitations may increase potential safety risks during vehicle interactions. To enhance the continuity and adaptability of autonomous vehicle motion control, this paper proposes a multi-target attentive car-following model, termed OVM-2D, built upon the classical OVM. The proposed model introduces a Flexible Transition Function (FTF) to characterize the capability of attending to multiple targets during lateral interactions, enabling the ego vehicle to dynamically identify the most influential car-following target among multiple leading vehicles distributed across different lanes and achieve smooth, adaptive target switching in cross-following scenarios. This mechanism provides a simple, interpretable, and behaviorally consistent framework for modeling dynamic target switching. Linear stability analysis and phase-plane evolution under large perturbations suggest an expanded stability region under the considered conditions. Cross-validation using the CitySim and TOD_VT datasets demonstrates stable trajectory-fitting performance and satisfactory cross-dataset transferability under the evaluated scenarios. Comprehensive evaluations across four representative cross-following scenarios indicate that, compared with the selected baseline models, the proposed model achieves lower trajectory-prediction errors while producing smoother longitudinal responses during target transitions, thereby supporting the effectiveness of the proposed target-switching mechanism. Meanwhile, it exhibits advantages in reducing jerk, estimated fuel consumption, and emissions while mitigating acceleration fluctuations caused by minor longitudinal and lateral disturbances, providing an effective framework for dynamic target switching in complex multi-lane traffic.
- Open Access
- 21 March 2026
- Published by Elsevier in SSRN Electronic Journal
A state-degradation-risk-embedded sliding mode perimeter control for regional road networks under cyberattacks
With the gradual popularization of connected and automated vehicle (CAV) technologies, urban traffic systems increasingly rely on communication networks to achieve vehicle-infrastructure cooperation and cloud-based control. However, when communication networks suffer from cyberattacks, non-linear changes in the macroscopic traffic flow characteristics are induced, which result in traffic network capacity attenuation and throughput degradation. To resist the adverse effects of cyberattacks on network traffic states, this paper proposes a state-degradation-risk-embedded sliding mode perimeter (SDRE-SMP) control approach based on the macroscopic fundamental diagram (MFD) theory. First, the network traffic flow is divided into five states through fuzzy inference based on MFD characteristics. On this basis, the "state risk flow" is introduced to accurately quantify the disruption caused by cyberattacks to the supply-demand balance. Second, according to the vehicle accumulation ratio and risk flow intensity, a dynamic state transition probability matrix is constructed to predict the risk probability of the network state transitioning to a deteriorated state. A comprehensive state degradation risk index is then formulated by assigning differentiated weights. Third, target vehicle accumulation is dynamically adjusted based on the network state degradation condition, and a regional sliding mode perimeter controller incorporating the comprehensive degradation risk index is developed. Finally, in a scenario of a 9×9 grid road network with a central 4×4 grid region under four different cyberattack modes, four control methods are tested: no control, PI control, model predictive control (MPC) with dynamic target vehicle accumulation, and SDRE-SMP control. Comparison results demonstrate that SDRE-SMP control can effectively suppress the regional MFD capacity attenuation and traffic congestion caused by cyberattacks. These results suggest that the proposed approach can not only improve traffic flow operational efficiency but also significantly reduce total delay and lower carbon emissions.
- 21 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA state-degradation-risk-embedded sliding mode perimeter control for regional road networks under cyberattacks
With the gradual popularization of connected and automated vehicle (CAV) technologies, urban traffic systems increasingly rely on communication networks to achieve vehicle-infrastructure cooperation and cloud-based control. However, when communication networks suffer from cyberattacks, non-linear changes in the macroscopic traffic flow characteristics are induced, which result in traffic network capacity attenuation and throughput degradation. To resist the adverse effects of cyberattacks on network traffic states, this paper proposes a state-degradation-risk-embedded sliding mode perimeter (SDRE-SMP) control approach based on the macroscopic fundamental diagram (MFD) theory. First, the network traffic flow is divided into five states through fuzzy inference based on MFD characteristics. On this basis, the "state risk flow" is introduced to accurately quantify the disruption caused by cyberattacks to the supply-demand balance. Second, according to the vehicle accumulation ratio and risk flow intensity, a dynamic state transition probability matrix is constructed to predict the risk probability of the network state transitioning to a deteriorated state. A comprehensive state degradation risk index is then formulated by assigning differentiated weights. Third, target vehicle accumulation is dynamically adjusted based on the network state degradation condition, and a regional sliding mode perimeter controller incorporating the comprehensive degradation risk index is developed. Finally, in a scenario of a 9×9 grid road network with a central 4×4 grid region under four different cyberattack modes, four control methods are tested: no control, PI control, model predictive control (MPC) with dynamic target vehicle accumulation, and SDRE-SMP control. Comparison results demonstrate that SDRE-SMP control can effectively suppress the regional MFD capacity attenuation and traffic congestion caused by cyberattacks. These results suggest that the proposed approach can not only improve traffic flow operational efficiency but also significantly reduce total delay and lower carbon emissions.
- 27 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (4) , 7132-7151
A Path-Following-Based CAV Motion Planning Framework for Unified Description of Car-Following and Lane-Changing
In motion planning for connected and automated vehicles (CAVs), separate modeling of car-following (CF) and lane-changing (LC) behaviors increases decision-making complexity, while scenario switching further compounds motion description challenges. To address this gap, this article proposes a path-following-based CAV motion planning framework (PFF) that unifies the modeling of both driving behaviors. The PFF integrates two core components: 1) a smooth and time-independent nonlinear path is dynamically generated and coupled with a velocity Gaussian constraint strategy to ensure lateral safety, significantly enhancing the computational efficiency of the path generation algorithm and 2) the dual-target collision avoidance model (DT-CAM) has been developed, utilizing a soft transition mechanism for following targets and adaptive velocity control, with real-time monitoring of vehicle states on either side of the road to ensure smooth and stable driving in the path. Calibration and simulation tests in three typical scenarios (“Original,” “Overtaking,” and “Being overtaken”) of the NGSIM dataset show that the PFF can effectively simulate human driving behavioral characteristics with a millisecond update of dynamic planning. Compared to four advanced benchmark models, PFF not only exceeds the safety-focused risk field-based behavior model (RFM) and comfort-oriented model predictive control (MPC) in their respective key metrics, but also demonstrates superior human-like behavior over all benchmarks, including the two-dimensional following lane changing (2DF-LC) and deep reinforcement learning (DRL). Furthermore, the PFF’s strong transferability is confirmed on the CitySim dataset, while its effectiveness in handling complex interactive scenarios is verified through CarSim simulation. Therefore, PFF not only reduces decision-making complexity during driving but also shows better overall performance, thereby contributing to the advancement of driving automation.
- 27 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (4) , 7132-7151
A Path-Following-Based CAV Motion Planning Framework for Unified Description of Car-Following and Lane-Changing
In motion planning for connected and automated vehicles (CAVs), separate modeling of car-following (CF) and lane-changing (LC) behaviors increases decision-making complexity, while scenario switching further compounds motion description challenges. To address this gap, this article proposes a path-following-based CAV motion planning framework (PFF) that unifies the modeling of both driving behaviors. The PFF integrates two core components: 1) a smooth and time-independent nonlinear path is dynamically generated and coupled with a velocity Gaussian constraint strategy to ensure lateral safety, significantly enhancing the computational efficiency of the path generation algorithm and 2) the dual-target collision avoidance model (DT-CAM) has been developed, utilizing a soft transition mechanism for following targets and adaptive velocity control, with real-time monitoring of vehicle states on either side of the road to ensure smooth and stable driving in the path. Calibration and simulation tests in three typical scenarios (“Original,” “Overtaking,” and “Being overtaken”) of the NGSIM dataset show that the PFF can effectively simulate human driving behavioral characteristics with a millisecond update of dynamic planning. Compared to four advanced benchmark models, PFF not only exceeds the safety-focused risk field-based behavior model (RFM) and comfort-oriented model predictive control (MPC) in their respective key metrics, but also demonstrates superior human-like behavior over all benchmarks, including the two-dimensional following lane changing (2DF-LC) and deep reinforcement learning (DRL). Furthermore, the PFF’s strong transferability is confirmed on the CitySim dataset, while its effectiveness in handling complex interactive scenarios is verified through CarSim simulation. Therefore, PFF not only reduces decision-making complexity during driving but also shows better overall performance, thereby contributing to the advancement of driving automation.
- 1 November 2025
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 69
The revolution of roundabouts in the autonomous driving era: A lane-free self-organized motion planning framework
- 1 November 2025
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 69
The revolution of roundabouts in the autonomous driving era: A lane-free self-organized motion planning framework
- 6 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 25 (8) , 9039-9057
A Sigmoid-Based Car-Following Model to Improve Acceleration Stability in Traffic Oscillation and Following Failure in Free Flow
This paper presents an improved Intelligent Driving Model (Sigmoid-IDM) to address the issues of excessive acceleration in traffic oscillation and following failure in free flow. The Sigmoid-IDM utilizes a Sigmoid function to enhance the starting-following characteristics, improve the output strategy of the spacing term, and stabilize the steady-state velocity in free flow. Furthermore, the model’s asymmetry is enhanced by introducing cautious following distance, caution driving factor, and segmentation function. The anti-interference ability of the Sigmoid-IDM is demonstrated through local stability and string stability analyses. The model parameters were calibrated using the Hefei dataset and High D data across various traffic scenarios: start-up, stop-go, and free-flow. The Sigmoid-IDM outperforms the IDM by significantly reducing errors and enhancing performance metrics. Specifically, in start-up and stop-go scenarios, the Sigmoid-IDM achieves a 28.57% and 19.04% reduction in Root Mean Square Error (RMSE) for acceleration, respectively. Comfort error during start-up is also lowered by 18.1%. In the free-flow scenario, the RMSE for spacing and velocity decreases by 15.64% and 16.36%, respectively. Furthermore, the Sigmoid-IDM demonstrates a more pronounced asymmetric behavior than the IDM, offering a more accurate representation of human drivers’ following patterns. The model’s efficacy was further validated through circular road simulation and Simulink-Carsim co-simulation, confirming its ability to accurately simulate the transition from synchronized flow to wide moving jams under variable parameters, as well as the traceability of its trajectory planning.
- 6 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 25 (8) , 9039-9057
A Sigmoid-Based Car-Following Model to Improve Acceleration Stability in Traffic Oscillation and Following Failure in Free Flow
This paper presents an improved Intelligent Driving Model (Sigmoid-IDM) to address the issues of excessive acceleration in traffic oscillation and following failure in free flow. The Sigmoid-IDM utilizes a Sigmoid function to enhance the starting-following characteristics, improve the output strategy of the spacing term, and stabilize the steady-state velocity in free flow. Furthermore, the model’s asymmetry is enhanced by introducing cautious following distance, caution driving factor, and segmentation function. The anti-interference ability of the Sigmoid-IDM is demonstrated through local stability and string stability analyses. The model parameters were calibrated using the Hefei dataset and High D data across various traffic scenarios: start-up, stop-go, and free-flow. The Sigmoid-IDM outperforms the IDM by significantly reducing errors and enhancing performance metrics. Specifically, in start-up and stop-go scenarios, the Sigmoid-IDM achieves a 28.57% and 19.04% reduction in Root Mean Square Error (RMSE) for acceleration, respectively. Comfort error during start-up is also lowered by 18.1%. In the free-flow scenario, the RMSE for spacing and velocity decreases by 15.64% and 16.36%, respectively. Furthermore, the Sigmoid-IDM demonstrates a more pronounced asymmetric behavior than the IDM, offering a more accurate representation of human drivers’ following patterns. The model’s efficacy was further validated through circular road simulation and Simulink-Carsim co-simulation, confirming its ability to accurately simulate the transition from synchronized flow to wide moving jams under variable parameters, as well as the traceability of its trajectory planning.
- Open Access
- 1 January 2024
- Published by Wiley in Journal of Advanced Transportation
- Vol. 2024 (1)
2D‐Action Asynchronous Cooperative Lane Change Trajectory Planning Method for Connected and Automated Vehicles
The ability to change lanes safely, efficiently, and comfortably is an important prerequisite for the application of Connected‐Automated Vehicles (CAVs). Based on the five‐order polynomial trajectory planning for CAVs, the 2D‐Action Asynchronous Lane Change (AALC) trajectory planning model is constructed by further considering the longitudinal and lateral driving action execution time parameters. This is done to improve the applicability of the lane change model and increase the CAV lane change success rate. The continuous collision space algorithm is constructed by determining the continuity condition of collision trajectory parameter solution space through the monotonicity of trajectory curve parameters and collision form classification. AALC trajectory safety judgment is realized through this algorithm. A cooperative lane change trajectory evaluation objective function is constructed, considering multivehicle comfort and efficiency. Finally, the AALC model is solved in the continuous collision space according to the optimal objective function, and the lane change is divided into free, cooperative, and refused according to the optimization. The results indicate that the AALC model achieves the transfer of collision space between lanes through asynchronous process of behavior execution time window, thereby reducing the possibility of vehicle collision. The AALC model reduces the degree of change of cooperative lane change parameters by asynchronous process of behavior, increasing the number of free lane change trajectories by about 17%, effectively reducing the occurrence of lane change refusal, improving the successful rate of lane change, and enhancing the overall evaluation of the lane change. The AALC model realizes the reallocation of collision space between different lanes through asynchronous process, making it more suitable for environments with large differences in vehicle gaps such as ramp merging. The collision‐based trajectory optimization algorithm can quickly obtain the corresponding safety space and optimal trajectory. The maximum calculation time for a single cooperative lane change is 0.073 s, thus enabling real‐time trajectory planning.
- 1 January 2024
- Published by Wiley in Journal of Advanced Transportation
- Vol. 2024 (1)
Open Access2D‐Action Asynchronous Cooperative Lane Change Trajectory Planning Method for Connected and Automated Vehicles
The ability to change lanes safely, efficiently, and comfortably is an important prerequisite for the application of Connected‐Automated Vehicles (CAVs). Based on the five‐order polynomial trajectory planning for CAVs, the 2D‐Action Asynchronous Lane Change (AALC) trajectory planning model is constructed by further considering the longitudinal and lateral driving action execution time parameters. This is done to improve the applicability of the lane change model and increase the CAV lane change success rate. The continuous collision space algorithm is constructed by determining the continuity condition of collision trajectory parameter solution space through the monotonicity of trajectory curve parameters and collision form classification. AALC trajectory safety judgment is realized through this algorithm. A cooperative lane change trajectory evaluation objective function is constructed, considering multivehicle comfort and efficiency. Finally, the AALC model is solved in the continuous collision space according to the optimal objective function, and the lane change is divided into free, cooperative, and refused according to the optimization. The results indicate that the AALC model achieves the transfer of collision space between lanes through asynchronous process of behavior execution time window, thereby reducing the possibility of vehicle collision. The AALC model reduces the degree of change of cooperative lane change parameters by asynchronous process of behavior, increasing the number of free lane change trajectories by about 17%, effectively reducing the occurrence of lane change refusal, improving the successful rate of lane change, and enhancing the overall evaluation of the lane change. The AALC model realizes the reallocation of collision space between different lanes through asynchronous process, making it more suitable for environments with large differences in vehicle gaps such as ramp merging. The collision‐based trajectory optimization algorithm can quickly obtain the corresponding safety space and optimal trajectory. The maximum calculation time for a single cooperative lane change is 0.073 s, thus enabling real‐time trajectory planning.
- 13 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Vehicles
- Vol. 9 (1) , 427-445
Two-Dimensional Following Lane-Changing (2DF-LC): A Framework for Dynamic Decision-Making and Rapid Behavior Planning
Lane changes require dynamic decision-making and rapid behavior planning, which are challenging for traffic modeling. We propose a two-dimensional following lane-changing framework (2DF-LC) that exploits the benefits of car-following (CF) models for computational efficiency, collision avoidance, and human-like behavior. This framework uses a sigmoid-based intelligent driver model (SIDM) with both longitudinal and lateral following. To avoid excessive acceleration at start-up, we develop an SIDM that ensures a smooth start-up. In the longitudinal plane, we introduce a transition function to create a double-target car-following model (DT-SIDM) that can handle sudden acceleration changes due to target switching, thereby guaranteeing stable longitudinal motion and dynamic collision avoidance. In the lateral plane, we develop a lateral movement car-following model (LM-SIDM) inspired by a social force model. The LM-SIDM defines both lane and gap forces, resulting in effective lateral motion and collision avoidance during lane changes. Simulations and tests in three typical scenarios show that 2DF-LC has high computational efficiency: it completes calculations within milliseconds. Compared with the widely used hierarchical motion planning system (HMPS) and integrated model and learning combined algorithm (IMLC) methods, 2DF-LC based on real trajectories reduces the errors by 49.5% and 16.1%, respectively, and achieves a 28.63% lower time-integrated anticipated collision time (TI-ACT) than the original trajectories, indicating improved safety. Moreover, 2DF-LC produces a smooth acceleration curve, with an average jerk value of 0.358 m/s 3 . The lane-change trajectory generated by 2DF-LC can also be followed and executed effectively in CarSim tests.
- 13 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Vehicles
- Vol. 9 (1) , 427-445
Two-Dimensional Following Lane-Changing (2DF-LC): A Framework for Dynamic Decision-Making and Rapid Behavior Planning
Lane changes require dynamic decision-making and rapid behavior planning, which are challenging for traffic modeling. We propose a two-dimensional following lane-changing framework (2DF-LC) that exploits the benefits of car-following (CF) models for computational efficiency, collision avoidance, and human-like behavior. This framework uses a sigmoid-based intelligent driver model (SIDM) with both longitudinal and lateral following. To avoid excessive acceleration at start-up, we develop an SIDM that ensures a smooth start-up. In the longitudinal plane, we introduce a transition function to create a double-target car-following model (DT-SIDM) that can handle sudden acceleration changes due to target switching, thereby guaranteeing stable longitudinal motion and dynamic collision avoidance. In the lateral plane, we develop a lateral movement car-following model (LM-SIDM) inspired by a social force model. The LM-SIDM defines both lane and gap forces, resulting in effective lateral motion and collision avoidance during lane changes. Simulations and tests in three typical scenarios show that 2DF-LC has high computational efficiency: it completes calculations within milliseconds. Compared with the widely used hierarchical motion planning system (HMPS) and integrated model and learning combined algorithm (IMLC) methods, 2DF-LC based on real trajectories reduces the errors by 49.5% and 16.1%, respectively, and achieves a 28.63% lower time-integrated anticipated collision time (TI-ACT) than the original trajectories, indicating improved safety. Moreover, 2DF-LC produces a smooth acceleration curve, with an average jerk value of 0.358 m/s 3 . The lane-change trajectory generated by 2DF-LC can also be followed and executed effectively in CarSim tests.
- 18 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 24 (11) , 12736-12753
A Superposition Assessment Framework of Multi-Source Traffic Risks for Mega-Events Using Risk Field Model and Time-Series Generative Adversarial Networks
In this study, a novel traffic risk assessment framework of mega-events that integrate risk field and deep learning is proposed. Considering the inherent difference of different traffic risks, the risk quantification and standardization is conducted first. Then several risk field models are constructed to quantify the impacts of multi-source traffic risk superposition on mega-events. Then a time-series generative adversarial networks (TimeGAN) is used to predict the evolution of superposition risk. We select 2022 Beijing Winter Olympics as a case to explore the superposition effects of different traffic risks on the convoy entrance of this mega-event. The results illustrate the superposition risks are significantly associated with the strength of each traffic risk, and the distance from the traffic risk location to the convoy entrance. Furthermore, the temporal evolutions for different traffic risks and their superposition are forecasted using TimeGAN. The results show the unexpected traffic congestion risk presents the highest predictive performance (i.e., the average error for RMSE, MAE, and MSE is 0.135%) and the superposition traffic risks present the lowest predictive performance (the average error is 0.536%). Comparison between different methods demonstrates TimeGAN outperforms other methods in predicting both single traffic risks and superposition risks. The research findings could be potentially referenced in multi-source traffic risk management for mega-events.
- 18 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 24 (11) , 12736-12753
A Superposition Assessment Framework of Multi-Source Traffic Risks for Mega-Events Using Risk Field Model and Time-Series Generative Adversarial Networks
In this study, a novel traffic risk assessment framework of mega-events that integrate risk field and deep learning is proposed. Considering the inherent difference of different traffic risks, the risk quantification and standardization is conducted first. Then several risk field models are constructed to quantify the impacts of multi-source traffic risk superposition on mega-events. Then a time-series generative adversarial networks (TimeGAN) is used to predict the evolution of superposition risk. We select 2022 Beijing Winter Olympics as a case to explore the superposition effects of different traffic risks on the convoy entrance of this mega-event. The results illustrate the superposition risks are significantly associated with the strength of each traffic risk, and the distance from the traffic risk location to the convoy entrance. Furthermore, the temporal evolutions for different traffic risks and their superposition are forecasted using TimeGAN. The results show the unexpected traffic congestion risk presents the highest predictive performance (i.e., the average error for RMSE, MAE, and MSE is 0.135%) and the superposition traffic risks present the lowest predictive performance (the average error is 0.536%). Comparison between different methods demonstrates TimeGAN outperforms other methods in predicting both single traffic risks and superposition risks. The research findings could be potentially referenced in multi-source traffic risk management for mega-events.