Yongxing Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Yongxing Wang has a total of 32 co-authors. Here are the top co-authors by number of shared publications:
- Jun Bi
Beijing Jiaotong University
9 shared publications - Xiaolong Zuo
School of Traffic and Transportation Beijing Jiaotong University Beijing China
4 shared publications - Chaoru Lu
Department of Built Environment, Oslo Metropolitan University, 0166 Oslo, Norway
3 shared publications Jun Bi
3 shared publicationsQiuyue Sai
3 shared publicationsZixian Zhang
3 shared publications
Latest Publications
- Open Access
- 25 August 2026
- Published by Elsevier in SSRN Electronic Journal
A Sequence Aware Feasibility Constrained Hierarchical Reinforcement Learning Framework for Electric Vehicle Charging Navigation under Dynamic Uncertainty
Range anxiety remains a critical barrier to electric vehicle (EV) adoption, calling for charging navigation methods that can provide real time guidance under dynamic network conditions. The task couples two decisions, charging station selection and route planning, each affected by time varying queueing delays, charging prices, travel times, and energy consumption. Existing methods either reduce these uncertainties to deterministic surrogates or merge the two decisions into a single value function, which leaves the station route coupling insufficiently represented. This paper develops a feasibility constrained hierarchical reinforcement learning framework for EV charging navigation. The framework combines a high level Double Deep Q Network (DDQN) for station selection with a low level Long Short Term Memory (LSTM) augmented Advantage Actor Critic (A2C) for sequential routing. The station selector handles stochastic queueing delays and charging prices, while the route planner treats routing as a trajectory dependent policy learning problem. To support feasible policy learning, action masks are applied at both decision levels. The high level mask removes unreachable charging stations. The low level mask removes routing actions that violate road connectivity, residual battery, or route acyclicity constraints. Experiments on a representative urban network show that the proposed framework reduces the total weighted navigation cost by 32.6%, 30.3%, 29.3%, and 7.9% compared with Dijkstra, a genetic algorithm, single layer A2C, and a value based hierarchical variant, respectively. It also achieves successful completion in all evaluated trips. These findings demonstrate the potential of sequence aware hierarchical reinforcement learning as a soft computing approach for constrained, dynamic, and coupled transportation energy decision problems.
- 25 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Sequence Aware Feasibility Constrained Hierarchical Reinforcement Learning Framework for Electric Vehicle Charging Navigation under Dynamic Uncertainty
Range anxiety remains a critical barrier to electric vehicle (EV) adoption, calling for charging navigation methods that can provide real time guidance under dynamic network conditions. The task couples two decisions, charging station selection and route planning, each affected by time varying queueing delays, charging prices, travel times, and energy consumption. Existing methods either reduce these uncertainties to deterministic surrogates or merge the two decisions into a single value function, which leaves the station route coupling insufficiently represented. This paper develops a feasibility constrained hierarchical reinforcement learning framework for EV charging navigation. The framework combines a high level Double Deep Q Network (DDQN) for station selection with a low level Long Short Term Memory (LSTM) augmented Advantage Actor Critic (A2C) for sequential routing. The station selector handles stochastic queueing delays and charging prices, while the route planner treats routing as a trajectory dependent policy learning problem. To support feasible policy learning, action masks are applied at both decision levels. The high level mask removes unreachable charging stations. The low level mask removes routing actions that violate road connectivity, residual battery, or route acyclicity constraints. Experiments on a representative urban network show that the proposed framework reduces the total weighted navigation cost by 32.6%, 30.3%, 29.3%, and 7.9% compared with Dijkstra, a genetic algorithm, single layer A2C, and a value based hierarchical variant, respectively. It also achieves successful completion in all evaluated trips. These findings demonstrate the potential of sequence aware hierarchical reinforcement learning as a soft computing approach for constrained, dynamic, and coupled transportation energy decision problems.
- Open Access
- 17 May 2026
- Published by MDPI AG in World Electric Vehicle Journal
- Vol. 17 (5) , 268
The Contribution of Electric Vehicles to the Realization of the Dual-Carbon Goal
The transportation sector is the main source of environmental pollution in cities, fueling the energy crisis
- 17 May 2026
- Published by MDPI AG in World Electric Vehicle Journal
- Vol. 17 (5) , 268
Open AccessThe Contribution of Electric Vehicles to the Realization of the Dual-Carbon Goal
The transportation sector is the main source of environmental pollution in cities, fueling the energy crisis
- Open Access
- 26 March 2026
- Published by MDPI AG in World Electric Vehicle Journal
- Vol. 17 (4) , 174
Optimal Scheduling of Electric Bus Fleets Considering Battery Degradation Effects
During routine operation, the power batteries of electric buses (EBs) gradually age due to the combined effects of numerous factors, including charging/discharging cycles, load fluctuations, and ambient temperature. This paper focuses specifically on the problem of battery aging in the context of actual urban electric bus system operations. It explores how to comprehensively incorporate the battery degradation effect into optimization schemes for EB fleet scheduling. This paper proposes an integrated optimization methodology that combines a capacity degradation model, a scheduling optimization model, and a genetic algorithm. A comprehensive scheduling optimization model is constructed, incorporating vehicle procurement costs, operational charging costs, off-peak charging costs, and battery capacity degradation costs, subject to rigorously defined constraints. Subsequently, an improved genetic algorithm framework is developed. Finally, the constructed model is validated using operational data from the Chongqing bus system. An analysis of the optimization mechanisms is provided, and a sensitivity analysis is conducted on vehicle procurement costs and battery capacity degradation costs. Based on the results, the model can reduce the total cost to 92% of the original level, proving that it is effective to a certain extent in reducing the operating costs of electric buses.
- 26 March 2026
- Published by MDPI AG in World Electric Vehicle Journal
- Vol. 17 (4) , 174
Open AccessOptimal Scheduling of Electric Bus Fleets Considering Battery Degradation Effects
During routine operation, the power batteries of electric buses (EBs) gradually age due to the combined effects of numerous factors, including charging/discharging cycles, load fluctuations, and ambient temperature. This paper focuses specifically on the problem of battery aging in the context of actual urban electric bus system operations. It explores how to comprehensively incorporate the battery degradation effect into optimization schemes for EB fleet scheduling. This paper proposes an integrated optimization methodology that combines a capacity degradation model, a scheduling optimization model, and a genetic algorithm. A comprehensive scheduling optimization model is constructed, incorporating vehicle procurement costs, operational charging costs, off-peak charging costs, and battery capacity degradation costs, subject to rigorously defined constraints. Subsequently, an improved genetic algorithm framework is developed. Finally, the constructed model is validated using operational data from the Chongqing bus system. An analysis of the optimization mechanisms is provided, and a sensitivity analysis is conducted on vehicle procurement costs and battery capacity degradation costs. Based on the results, the model can reduce the total cost to 92% of the original level, proving that it is effective to a certain extent in reducing the operating costs of electric buses.
- Open Access
- 9 March 2026
- Published by Elsevier in SSRN Electronic Journal
Hierarchical Reinforcement Learning for Electric vehicle Charging Navigation in Dynamic Transportation Networks
Range anxiety remains a significant barrier to sustainable electric vehicle (EV) adoption. Existing navigation methods often decouple station selection from route planning, struggling to adapt to real-world uncertainties. To address this, we propose a real-time, bi-level stochastic optimization framework capturing the spatiotemporal dynamics of coupled transportation and charging networks. We develop a novel Hierarchical Reinforcement Learning approach (HE-DDQN-A2C) to solve this system. The upper-level Double Deep Q-Network optimizes station selection to minimize queuing and charging costs, while the lower-level Advantage Actor-Critic algorithm executes sequential route planning to minimize energy consumption and travel time. Validated on an urban network, HE-DDQN-A2C significantly outperforms traditional routing (Dijkstra, Genetic Algorithms) and single-layer reinforcement learning baselines. By achieving a 100% success rate and substantially reducing total navigation costs, our framework offers a robust solution for alleviating range anxiety and enhancing sustainable EV mobility.
- 9 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHierarchical Reinforcement Learning for Electric vehicle Charging Navigation in Dynamic Transportation Networks
Range anxiety remains a significant barrier to sustainable electric vehicle (EV) adoption. Existing navigation methods often decouple station selection from route planning, struggling to adapt to real-world uncertainties. To address this, we propose a real-time, bi-level stochastic optimization framework capturing the spatiotemporal dynamics of coupled transportation and charging networks. We develop a novel Hierarchical Reinforcement Learning approach (HE-DDQN-A2C) to solve this system. The upper-level Double Deep Q-Network optimizes station selection to minimize queuing and charging costs, while the lower-level Advantage Actor-Critic algorithm executes sequential route planning to minimize energy consumption and travel time. Validated on an urban network, HE-DDQN-A2C significantly outperforms traditional routing (Dijkstra, Genetic Algorithms) and single-layer reinforcement learning baselines. By achieving a 100% success rate and substantially reducing total navigation costs, our framework offers a robust solution for alleviating range anxiety and enhancing sustainable EV mobility.
- Open Access
- 7 March 2026
- Published by Elsevier in SSRN Electronic Journal
Hierarchical Reinforcement Learning for Electric vehicle Charging Navigation in Dynamic Transportation Networks
Range anxiety remains a significant barrier to sustainable electric vehicle (EV) adoption. Existing navigation methods often decouple station selection from route planning, struggling to adapt to real-world uncertainties. To address this, we propose a real-time, bi-level stochastic optimization framework capturing the spatiotemporal dynamics of coupled transportation and charging networks. We develop a novel Hierarchical Reinforcement Learning approach (HE-DDQN-A2C) to solve this system. The upper-level Double Deep Q-Network optimizes station selection to minimize queuing and charging costs, while the lower-level Advantage Actor-Critic algorithm executes sequential route planning to minimize energy consumption and travel time. Validated on an urban network, HE-DDQN-A2C significantly outperforms traditional routing (Dijkstra, Genetic Algorithms) and single-layer reinforcement learning baselines. By achieving a 100% success rate and substantially reducing total navigation costs, our framework offers a robust solution for alleviating range anxiety and enhancing sustainable EV mobility.
- 7 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHierarchical Reinforcement Learning for Electric vehicle Charging Navigation in Dynamic Transportation Networks
Range anxiety remains a significant barrier to sustainable electric vehicle (EV) adoption. Existing navigation methods often decouple station selection from route planning, struggling to adapt to real-world uncertainties. To address this, we propose a real-time, bi-level stochastic optimization framework capturing the spatiotemporal dynamics of coupled transportation and charging networks. We develop a novel Hierarchical Reinforcement Learning approach (HE-DDQN-A2C) to solve this system. The upper-level Double Deep Q-Network optimizes station selection to minimize queuing and charging costs, while the lower-level Advantage Actor-Critic algorithm executes sequential route planning to minimize energy consumption and travel time. Validated on an urban network, HE-DDQN-A2C significantly outperforms traditional routing (Dijkstra, Genetic Algorithms) and single-layer reinforcement learning baselines. By achieving a 100% success rate and substantially reducing total navigation costs, our framework offers a robust solution for alleviating range anxiety and enhancing sustainable EV mobility.
- 31 December 2025
- Published by Elsevier in Measurement
- Vol. 264
Driving range estimation for electric buses considering battery degradation levels: A self-attention deep learning approach
- 31 December 2025
- Published by Elsevier in Measurement
- Vol. 264
Driving range estimation for electric buses considering battery degradation levels: A self-attention deep learning approach
- Open Access
- 1 January 2025
- Published by Maximum Academic Press in Digital Transportation and Safety
- Vol. 4 (3) , 207-214
Estimating energy consumption rates for battery electric truck based on real-world data
Battery electric trucks (BETs) have been regarded as an environmentally friendly mode of road freight transportation. However, since BETs have limited driving ranges, the freight transportation companies experience the challenge of establishing reasonable freight transport schemes, where the operators of freight transportation companies need to understand the energy consumption rate (ECR) of BETs and ensure that the transport tasks can be finished without breaking down due to energy depletion. The accurate estimation of ECR is an effective method to alleviate the negative effects of limited driving range. To this end, this study utilizes the BET operational data collected from real-world situations to investigate the estimation of ECR. A data-driven approach is adopted to establish the nonlinear regression models for estimating the ECR of BETs, accompanied by its mathematical expression. The estimated performance of the models is evaluated, and consequently, the cubic function model is selected by considering the complexity and robustness. Based on the established estimation model, the economic driving speed of BETs is investigated. The results indicate that the economic driving speed of BETs is 52.02 km/h, and the corresponding ECR is 0.53 kWh/km.
- 1 January 2025
- Published by Maximum Academic Press in Digital Transportation and Safety
- Vol. 4 (3) , 207-214
Open AccessEstimating energy consumption rates for battery electric truck based on real-world data
Battery electric trucks (BETs) have been regarded as an environmentally friendly mode of road freight transportation. However, since BETs have limited driving ranges, the freight transportation companies experience the challenge of establishing reasonable freight transport schemes, where the operators of freight transportation companies need to understand the energy consumption rate (ECR) of BETs and ensure that the transport tasks can be finished without breaking down due to energy depletion. The accurate estimation of ECR is an effective method to alleviate the negative effects of limited driving range. To this end, this study utilizes the BET operational data collected from real-world situations to investigate the estimation of ECR. A data-driven approach is adopted to establish the nonlinear regression models for estimating the ECR of BETs, accompanied by its mathematical expression. The estimated performance of the models is evaluated, and consequently, the cubic function model is selected by considering the complexity and robustness. Based on the established estimation model, the economic driving speed of BETs is investigated. The results indicate that the economic driving speed of BETs is 52.02 km/h, and the corresponding ECR is 0.53 kWh/km.
- 11 October 2024
- Published by Elsevier in Renewable Energy
- Vol. 237
Optimal battery electric bus system planning considering heterogeneous vehicles, opportunity charging, and battery degradation
- 11 October 2024
- Published by Elsevier in Renewable Energy
- Vol. 237
Optimal battery electric bus system planning considering heterogeneous vehicles, opportunity charging, and battery degradation
- Open Access
- 27 June 2023
- Published by MDPI AG in World Electric Vehicle Journal
- Vol. 14 (7) , 171
A Charging Guidance Optimization Model for Electric Vehicle Travel by Considering Multi-Dimensional Preferences of Users
The dual-carbon strategy advocates a green, environmentally friendly, and low-carbon lifestyle. In the field of transportation, electric vehicles (EVs) have been regarded as an effective solution to reduce carbon emissions and to conserve energy. Developing a reasonable charging guidance scheme for users is a feasible way to solve problems, such as the range anxiety of EV users, and has a great application value for the promotion of EVs in the future. In practical situations, how to develop charging induction schemes for users that better meet their needs according to the type of user and their multi-dimensional preferences is the focus of this paper. To this end, this study utilized charging behavioral data to investigate the multi-dimensional charging preference of users based on the collaborative filtering algorithm. Then, a multi-objective optimization model was established based on the preference degree of each charging station and the integrated travel cost. An NSGA-III framework was used to design the algorithm to solve the proposed model. The algorithm was tested using simulation experiments that were designed based on the road network and charging stations in Beijing. The final result is an experimental analysis of the weight matrices for the three different preferences of minimum energy consumption cost, minimum time cost, and minimum fee cost, which yields a difference of about 4.4% between the optimal energy consumption cost and the maximum energy cost, about 2.9% between the optimal time cost and the maximum time cost, and about 10% between the optimal fee cost and the maximum fee cost under these three different preferences, respectively. The proposed multi-objective optimization model is able to provide users with reliable charging station selection by incorporating their personalized charging preference characteristics and charge guidance schemes.
- 27 June 2023
- Published by MDPI AG in World Electric Vehicle Journal
- Vol. 14 (7) , 171
Open AccessA Charging Guidance Optimization Model for Electric Vehicle Travel by Considering Multi-Dimensional Preferences of Users
The dual-carbon strategy advocates a green, environmentally friendly, and low-carbon lifestyle. In the field of transportation, electric vehicles (EVs) have been regarded as an effective solution to reduce carbon emissions and to conserve energy. Developing a reasonable charging guidance scheme for users is a feasible way to solve problems, such as the range anxiety of EV users, and has a great application value for the promotion of EVs in the future. In practical situations, how to develop charging induction schemes for users that better meet their needs according to the type of user and their multi-dimensional preferences is the focus of this paper. To this end, this study utilized charging behavioral data to investigate the multi-dimensional charging preference of users based on the collaborative filtering algorithm. Then, a multi-objective optimization model was established based on the preference degree of each charging station and the integrated travel cost. An NSGA-III framework was used to design the algorithm to solve the proposed model. The algorithm was tested using simulation experiments that were designed based on the road network and charging stations in Beijing. The final result is an experimental analysis of the weight matrices for the three different preferences of minimum energy consumption cost, minimum time cost, and minimum fee cost, which yields a difference of about 4.4% between the optimal energy consumption cost and the maximum energy cost, about 2.9% between the optimal time cost and the maximum time cost, and about 10% between the optimal fee cost and the maximum fee cost under these three different preferences, respectively. The proposed multi-objective optimization model is able to provide users with reliable charging station selection by incorporating their personalized charging preference characteristics and charge guidance schemes.
- Open Access
A Hybrid Genetic Algorithm Based on Imitation Learning for the Airport Gate Assignment Problem
Airport gates are the main places for aircraft to receive ground services. With the increased number of flights, limited gate resources near to the terminal make the gate assignment work more complex. Traditional solution methods based on mathematical programming models and iterative algorithms are usually used to solve these static situations, lacking learning and real-time decision-making abilities. In this paper, a two-stage hybrid algorithm based on imitation learning and genetic algorithm (IL-GA) is proposed to solve the gate assignment problem. First of all, the problem is defined from a mathematical model to a Markov decision process (MDP), with the goal of maximizing the number of flights assigned to contact gates and the total gate preferences. In the first stage of the algorithm, a deep policy network is created to obtain the gate selection probability of each flight. This policy network is trained by imitating and learning the assignment trajectory data of human experts, and this process is offline. In the second stage of the algorithm, the policy network is used to generate a good initial population for the genetic algorithm to calculate the optimal solution for an online instance. The experimental results show that the genetic algorithm combined with imitation learning can greatly shorten the iterations and improve the population convergence speed. The flight rate allocated to the contact gates is 14.9% higher than the manual allocation result and 4% higher than the traditional genetic algorithm. Learning the expert assignment data also makes the allocation scheme more consistent with the preference of the airport, which is helpful for the practical application of the algorithm.
Open AccessA Hybrid Genetic Algorithm Based on Imitation Learning for the Airport Gate Assignment Problem
Airport gates are the main places for aircraft to receive ground services. With the increased number of flights, limited gate resources near to the terminal make the gate assignment work more complex. Traditional solution methods based on mathematical programming models and iterative algorithms are usually used to solve these static situations, lacking learning and real-time decision-making abilities. In this paper, a two-stage hybrid algorithm based on imitation learning and genetic algorithm (IL-GA) is proposed to solve the gate assignment problem. First of all, the problem is defined from a mathematical model to a Markov decision process (MDP), with the goal of maximizing the number of flights assigned to contact gates and the total gate preferences. In the first stage of the algorithm, a deep policy network is created to obtain the gate selection probability of each flight. This policy network is trained by imitating and learning the assignment trajectory data of human experts, and this process is offline. In the second stage of the algorithm, the policy network is used to generate a good initial population for the genetic algorithm to calculate the optimal solution for an online instance. The experimental results show that the genetic algorithm combined with imitation learning can greatly shorten the iterations and improve the population convergence speed. The flight rate allocated to the contact gates is 14.9% higher than the manual allocation result and 4% higher than the traditional genetic algorithm. Learning the expert assignment data also makes the allocation scheme more consistent with the preference of the airport, which is helpful for the practical application of the algorithm.