Multi-train modeling and simulation plays a vital role in railway electrification during operation and planning phase. Study of peak power demand and energy consumed by each traction substation needs to be deter- mine...Multi-train modeling and simulation plays a vital role in railway electrification during operation and planning phase. Study of peak power demand and energy consumed by each traction substation needs to be deter- mined to verify that electrical energy flowing in its railway power feeding system is appropriate or not. Gauss-Seidel, conventional Newton-Raphson, and current injection methods are well-known and widely accepted as a tool for electrical power network solver in DC railway power supply study. In this paper, a simplified Newton-Raphson method has been proposed. The proposed method employs a set of current-balance equations at each electrical node instead of the conventional power-balance equation used in the conventional Newton-Raphson method. This concept can remarkably reduce execution time and computing complexity for multi-train simulation. To evaluate its use, Sukhumvit line of Bangkok transit system （BTS） of Thai- land with 21.6-km line length and 22 passenger stopping stations is set as a test system. The multi-train simulation integrated with the proposed power network solver is developed to simulate 1-h operation service of selected 5-min headway. From the obtained results, the proposed method is more efficient with approximately 18 % faster than the conventional Newton-Raphson method and just over 6 % faster than the current injection method.展开更多
Purpose–This paper aims to propose a train timetable rescheduling(TTR)approach from the perspective of multi-train tracking optimization based on the mutual spatiotemporal information in the high-speed railway signal...Purpose–This paper aims to propose a train timetable rescheduling(TTR)approach from the perspective of multi-train tracking optimization based on the mutual spatiotemporal information in the high-speed railway signaling system.Design/methodology/approach–Firstly,a single-train trajectory optimization(STTO)model is constructed based on train dynamics and operating conditions.The train kinematics parameters,including acceleration,speed and time at each position,are calculated to predict the arrival times in the train timetable.A STTO algorithm is developed to optimize a single-train time-efficient driving strategy.Then,a TTR approach based on multi-train tracking optimization(TTR-MTTO)is proposed with mutual information.The constraints of temporary speed restriction(TSR)and end of authority are decoupled to calculate the tracking trajectory of the backward tracking train.The multi-train trajectories at each position are optimized to generate a timeefficient train timetable.Findings–The numerical experiment is performed on the Beijing-Tianjin high-speed railway line and CR400AF.The STTO algorithm predicts the train’s planned arrival time to calculate the total train delay(TTD).As for the TSR scenario,the proposed TTR-MTTO can reduce TTD by 60.60%compared with the traditional TTR approach with dispatchers’experience.Moreover,TTR-MTTO can optimize a time-efficient train timetable to help dispatchers reschedule trains more reasonably.Originality/value–With the cooperative relationship and mutual information between train rescheduling and control,the proposed TTR-MTTO approach can automatically generate a time-efficient train timetable to reduce the total train delay and the work intensity of dispatchers.展开更多
预训练世界模型是提升强化学习样本效率的关键技术,但现有方法因视频数据缺乏显式动作标注,难以捕捉状态转移的因果机制。对此,提出多模态大模型辅助的视频动作生成预训练框架(MLM-generated Action-based Pre-training from videos for...预训练世界模型是提升强化学习样本效率的关键技术,但现有方法因视频数据缺乏显式动作标注,难以捕捉状态转移的因果机制。对此,提出多模态大模型辅助的视频动作生成预训练框架(MLM-generated Action-based Pre-training from videos for world models,MAPO),通过整合视觉语言模型的语义理解能力与动力学建模需求,突破传统预训练范式在动作语义缺失方面的局限性。具体地,MAPO在预训练阶段利用多模态大模型(QWEN2_5-VL-7B)解析视频帧序列,生成细粒度语义动作描述,构建具有因果解释性的动作-状态关联;设计上下文量化编码机制,解耦场景静态特征与动态控制因素,增强跨模态表征能力。在微调阶段,通过双网络协同架构实现预训练动力学特征与真实环境动作的端到端对齐。实验表明,MAPO在DeepMind Control Suite和Meta-World的8项任务中的平均回报较最优基线获得稳定提升,尤其在长时程任务中展现出卓越的性能。该研究为跨模态世界模型训练提供了新范式,揭示了语义动作生成在因果推理中的关键作用。展开更多
针对城轨交通单一储能系统(Energy Storage System,ESS)功率能量特性难以匹配多车异步工况的瓶颈问题,提出一种基于动态阻抗匹配与多车协同的超级电容-钛酸锂电池混合储能系统(Hybrid Energy Storage System of Supercapacitor and Lith...针对城轨交通单一储能系统(Energy Storage System,ESS)功率能量特性难以匹配多车异步工况的瓶颈问题,提出一种基于动态阻抗匹配与多车协同的超级电容-钛酸锂电池混合储能系统(Hybrid Energy Storage System of Supercapacitor and Lithium Titanium Oxide Battery,SC-LTO HESS)。通过构建多车耦合直流牵引网络动态阻抗电路模型,设计自适应双闭环控制结构,结合某市轨道交通线,对不同储能系统多车运行时进行对比与分析,并在MATLAB/Simulink仿真系统中验证了其有效性。实验结果表明,该系统显著提升了能量利用效率、牵引网稳压率和储能系统节能率。展开更多
现有基于预训练语言模型(PLM)的作文自动评分(AES)方法偏向于直接使用从PLM提取的全局语义特征表示作文的质量,却忽略了作文质量与更细粒度特征关联关系的问题。聚焦于中文AES研究,从多种文本角度分析和评估作文质量,提出利用图神经网络...现有基于预训练语言模型(PLM)的作文自动评分(AES)方法偏向于直接使用从PLM提取的全局语义特征表示作文的质量,却忽略了作文质量与更细粒度特征关联关系的问题。聚焦于中文AES研究,从多种文本角度分析和评估作文质量,提出利用图神经网络(GNN)对作文的多尺度特征进行联合学习的中文AES方法。首先,利用GNN分别获取作文在句子级别和段落级别的篇章特征;然后,将这些篇章特征与作文的全局语义特征进行联合特征学习,实现对作文更精准的评分;最后,构建一个中文AES数据集,为中文AES研究提供数据基础。在所构建的数据集上的实验结果表明,所提方法在6个作文主题上的平均二次加权Kappa(QWK)系数相较于R2-BERT(Bidirectional Encoder Representations from Transformers model with Regression and Ranking)提升了1.1个百分点,验证了在AES任务中进行多尺度特征联合学习的有效性。同时,消融实验结果进一步表明了不同尺度的作文特征对评分效果的贡献。为了证明小模型在特定任务场景下的优越性,与当前流行的通用大语言模型GPT-3.5-turbo和DeepSeek-V3进行了对比。结果表明,使用所提方法的BERT(Bidirectional Encoder Representations from Transformers)模型在6个作文主题上的平均QWK比GPT-3.5-turbo和DeepSeek-V3分别高出了65.8和45.3个百分点,验证了大语言模型(LLMs)在面向领域的篇章级作文评分任务中,因缺乏大规模有监督微调数据而表现不佳的观点。展开更多
文摘Multi-train modeling and simulation plays a vital role in railway electrification during operation and planning phase. Study of peak power demand and energy consumed by each traction substation needs to be deter- mined to verify that electrical energy flowing in its railway power feeding system is appropriate or not. Gauss-Seidel, conventional Newton-Raphson, and current injection methods are well-known and widely accepted as a tool for electrical power network solver in DC railway power supply study. In this paper, a simplified Newton-Raphson method has been proposed. The proposed method employs a set of current-balance equations at each electrical node instead of the conventional power-balance equation used in the conventional Newton-Raphson method. This concept can remarkably reduce execution time and computing complexity for multi-train simulation. To evaluate its use, Sukhumvit line of Bangkok transit system （BTS） of Thai- land with 21.6-km line length and 22 passenger stopping stations is set as a test system. The multi-train simulation integrated with the proposed power network solver is developed to simulate 1-h operation service of selected 5-min headway. From the obtained results, the proposed method is more efficient with approximately 18 % faster than the conventional Newton-Raphson method and just over 6 % faster than the current injection method.
基金This research was jointly supported by the National Natural Science Foundation of China[Grant 62203468]the Young Elite Scientist Sponsorship Program by China Association for Science and Technology(CAST)[Grant 2022QNRC001]+1 种基金the Technological Research and Development Program of China Railway Corporation Limited[Grant K2021X001]by the Foundation of China Academy of Railway Sciences Corporation Limited[Grant 2021YJ043].On behalf all authors,the corresponding author states that there is no conflict of interest.
文摘Purpose–This paper aims to propose a train timetable rescheduling(TTR)approach from the perspective of multi-train tracking optimization based on the mutual spatiotemporal information in the high-speed railway signaling system.Design/methodology/approach–Firstly,a single-train trajectory optimization(STTO)model is constructed based on train dynamics and operating conditions.The train kinematics parameters,including acceleration,speed and time at each position,are calculated to predict the arrival times in the train timetable.A STTO algorithm is developed to optimize a single-train time-efficient driving strategy.Then,a TTR approach based on multi-train tracking optimization(TTR-MTTO)is proposed with mutual information.The constraints of temporary speed restriction(TSR)and end of authority are decoupled to calculate the tracking trajectory of the backward tracking train.The multi-train trajectories at each position are optimized to generate a timeefficient train timetable.Findings–The numerical experiment is performed on the Beijing-Tianjin high-speed railway line and CR400AF.The STTO algorithm predicts the train’s planned arrival time to calculate the total train delay(TTD).As for the TSR scenario,the proposed TTR-MTTO can reduce TTD by 60.60%compared with the traditional TTR approach with dispatchers’experience.Moreover,TTR-MTTO can optimize a time-efficient train timetable to help dispatchers reschedule trains more reasonably.Originality/value–With the cooperative relationship and mutual information between train rescheduling and control,the proposed TTR-MTTO approach can automatically generate a time-efficient train timetable to reduce the total train delay and the work intensity of dispatchers.
文摘预训练世界模型是提升强化学习样本效率的关键技术,但现有方法因视频数据缺乏显式动作标注,难以捕捉状态转移的因果机制。对此,提出多模态大模型辅助的视频动作生成预训练框架(MLM-generated Action-based Pre-training from videos for world models,MAPO),通过整合视觉语言模型的语义理解能力与动力学建模需求,突破传统预训练范式在动作语义缺失方面的局限性。具体地,MAPO在预训练阶段利用多模态大模型(QWEN2_5-VL-7B)解析视频帧序列,生成细粒度语义动作描述,构建具有因果解释性的动作-状态关联;设计上下文量化编码机制,解耦场景静态特征与动态控制因素,增强跨模态表征能力。在微调阶段,通过双网络协同架构实现预训练动力学特征与真实环境动作的端到端对齐。实验表明,MAPO在DeepMind Control Suite和Meta-World的8项任务中的平均回报较最优基线获得稳定提升,尤其在长时程任务中展现出卓越的性能。该研究为跨模态世界模型训练提供了新范式,揭示了语义动作生成在因果推理中的关键作用。
文摘针对城轨交通单一储能系统(Energy Storage System,ESS)功率能量特性难以匹配多车异步工况的瓶颈问题,提出一种基于动态阻抗匹配与多车协同的超级电容-钛酸锂电池混合储能系统(Hybrid Energy Storage System of Supercapacitor and Lithium Titanium Oxide Battery,SC-LTO HESS)。通过构建多车耦合直流牵引网络动态阻抗电路模型,设计自适应双闭环控制结构,结合某市轨道交通线,对不同储能系统多车运行时进行对比与分析,并在MATLAB/Simulink仿真系统中验证了其有效性。实验结果表明,该系统显著提升了能量利用效率、牵引网稳压率和储能系统节能率。
文摘现有基于预训练语言模型(PLM)的作文自动评分(AES)方法偏向于直接使用从PLM提取的全局语义特征表示作文的质量,却忽略了作文质量与更细粒度特征关联关系的问题。聚焦于中文AES研究,从多种文本角度分析和评估作文质量,提出利用图神经网络(GNN)对作文的多尺度特征进行联合学习的中文AES方法。首先,利用GNN分别获取作文在句子级别和段落级别的篇章特征;然后,将这些篇章特征与作文的全局语义特征进行联合特征学习,实现对作文更精准的评分;最后,构建一个中文AES数据集,为中文AES研究提供数据基础。在所构建的数据集上的实验结果表明,所提方法在6个作文主题上的平均二次加权Kappa(QWK)系数相较于R2-BERT(Bidirectional Encoder Representations from Transformers model with Regression and Ranking)提升了1.1个百分点,验证了在AES任务中进行多尺度特征联合学习的有效性。同时,消融实验结果进一步表明了不同尺度的作文特征对评分效果的贡献。为了证明小模型在特定任务场景下的优越性,与当前流行的通用大语言模型GPT-3.5-turbo和DeepSeek-V3进行了对比。结果表明,使用所提方法的BERT(Bidirectional Encoder Representations from Transformers)模型在6个作文主题上的平均QWK比GPT-3.5-turbo和DeepSeek-V3分别高出了65.8和45.3个百分点,验证了大语言模型(LLMs)在面向领域的篇章级作文评分任务中,因缺乏大规模有监督微调数据而表现不佳的观点。
