基于大数据分析的弓网参数反事实模型研究-《计算机技术与发展》

文章信息/Info

Title:: Research on Counterfactual Modeling of Pantograph Parameters Based on Big Data Analysis

Author(s):: PENG Bo; LU Jia-wei; TU Xin-yue; GUO Yi; LI Te; School of Information Science and Engineering,East China University of Science and Technology,Shanghai 201400,China

Keywords:: pantograph system; arcing; DBSCAN algorithm; causal inference; time-aware BPNet

摘要:: 弓网系统是城市轨道交通车辆输电的核心系统。如何改善车辆弓网系统发生的高温高频次燃弧现象是当前弓网系统的重要研究课题之一。针对大部分仿真模型的结果不可解释性的问题,该文提出了使用大数据分析的方式对车辆实际运行参数进行反事实模型的构建,以推断燃弧温度与弓网参数之间的因果关系,从而对弓网系统的运维工作提供理论支持。在数据预处理阶段,基于弓网系统架设原理,通过数据匹配的方式对缺失值进行了补充,并且采用 DBSCAN 算法对数据异常值进行处理,有效地提升了反事实预测模型的预测精度。在反事实预测模型构建阶段,提出了时间感知网络模型(Time-Aware BPNet)对弓网参数之间的因果关系进行模型拟合并比较。实验结果表明,相比于线性回归、多项式拟合模型、注意力机制的深度学习算法(Attention)和长短期记忆网络(LSTM),时间感知网络模型可以更好地对反事实数据进行预测。

Abstract:: The pantograph system is the core system for power transmission of urban rail transit vehicles. Improving the occurrence of high-temperature and high-frequency arcing phenomena in the pantograph system is one of the important research topics in the current pantograph system. Addressing the problem of the uninterpretability of most simulation models,we propose using big data analysis to construct counterfactual models of actual vehicle operating parameters to infer the causal relationship between arcing temperature and pan-tograph parameters,thereby providing theoretical support for the operation and maintenance of the pantograph system. In the data prepro-cessing stage,based on the principle of pantograph system installation,missing values were supplemented through data matching,and the DBSCAN algorithm was used to handle data outliers, effectively improving the prediction accuracy of the counterfactual prediction model. In the construction stage of the counterfactual prediction model,we propose the Time-Aware BPNet model to fit and compare the causal relationships between pantograph parameters. The experimental results indicate that compared to linear regression, polynomial fitting models,attention-based deep learning algorithms (Attention),and Long Short-Term Memory networks (LSTM),the time-aware network model can better predict counterfactual data.