【作者】 王静；

【导师】 刘锴；

【作者基本信息】 大连理工大学 ， 交通运输规划与管理， 2024， 硕士

【摘要】 城市人口持续增长,推动了交通出行需求的增加,城市化发展模式下轨道交通网络的建设扩张为更多人提供了便捷的出行方式。与此同时,城市轨道交通系统面临着更大的运营压力和管理挑战,尤其是对精细化乘客出行管理提出了更高的要求。轨道交通的个体出行特征呈现为规律性与随机性的耦合,加上乘客具有不同的社会经济属性和出行偏好,乘客出行行为的预测具有挑战性。随着大数据技术的迅速发展和智能化技术的应用,城市轨道交通系统可以收集和分析大量的乘客出行数据,这为深入研究乘客出行行为提供了新的机遇。本文以乘客个体出行为出发点,挖掘乘客多时间尺度下的出行规律并对可预测性上限与预测方法展开研究,并建立宏观客流与微观个体行为的联系,主要工作及成果如下:针对乘客个体出行规律,提出一种基于智能卡数据的高效数据挖掘方法,旨在提取轨道交通乘客的出行模式。利用带噪声基于密度的空间聚类(Density-Based Spatial Clustering of Applications with Noise,DBSCAN),综合考虑乘客的出行时间、起终点选择及其在总出行天数中的占比,基于乘客出行链数据精确捕捉每个乘客的历史出行模式并修正乘客出行序列。进一步,以数据驱动的方式,结合乘客出行次数及时空相似性,对出行模式进行细致分类,将频繁出行的乘客划分为时间规律、空间规律、时空规律和无规律四种类型,并比较了各类乘客的出行特征。针对个体出行规律可预测上限,使用修正时空熵值来刻画乘客个体的出行模式,量化个体乘客的复杂的出行规律。将乘客历史出行序列转化为随机出行过程,应用信息论理论推导出行过程的可预测上限。进一步通过数据压缩的方法计算具有序列相关性的出行过程熵率,通过熵率估计可预测上限。进一步对可预测上限值进行分布拟合,并分析不同出行属性(出行时间、出行起点、出行终点)的与不同乘客特征带来可预测上限值的差异。结果显示,出行序列熵率服从正态分布,而可预测上限服从Gumbel分布,进、出站站点可预测性上限的众数均超过0.78,进站时间的可预测性相较于站点序列明显偏低。不同出行次数的乘客的可预测上限可分为显著提升期、平稳提升区和平稳期三个阶段,当出行次数超过40次后,提供更多的出行序列数据对出行预测来说可能变得冗余。针对乘客个体的出行预测,根据个体的历史出行规律和出行时间,考虑长期、周期和短期三种时间尺度的出行规律,以小时为预测时段,构建特征向量和标签向量。使用生成对抗网络(Generative Adversarial Network,GAN)克服出行数据不均衡,生成虚拟实例加入原始数据。搭建深度神经网络(Deep Neural Network,DNN)实现不同类型乘客在不同时间段内的出行预测。结果显示,GAN的方法生成的数据能显著提高后续模型的预测精度,当不平衡率调整到1:10后,预测结果和模型运行时间处于相对较好的状态。在客流量层面,早晚高峰时段的预测准确度较高,总体准确率达到93.77%。更多还原

【Abstract】 The continuous growth of urban populations has driven an increase in transportation demand,while the expansion of urban rail transit networks under the urbanization development model has provided more convenient travel options for people.At the same time,to address traffic congestion and environmental issues,some cities have implemented vehicle restriction policies,prompting people to shift towards rail transit.Consequently,urban rail transit systems face greater operational pressure and management challenges,especially with higher demands for fine-grained passenger travel management.The individual travel characteristics of rail transit exhibit a coupling of regularity and randomness,compounded by passengers’ diverse socioeconomic attributes and travel preferences,making the prediction of passenger travel behavior challenging.With the rapid development of big data technology and the application of intelligent technology,urban rail transit systems can collect and analyze large amounts of passenger travel data,providing new opportunities for in-depth research into passenger travel behavior.This paper takes individual passenger travel as the starting point,explores travel patterns across multiple time scales,investigates the the upper bound of predictability and prediction methods,and establishes connections between macro passenger flows and micro individual behaviors.The main contributions are as follows:Urban rail transit network construction is rapidly advancing,accompanied by a surge in passengers.This trend poses greater challenges for refined passenger travel management.The individual travel characteristics of urban rail transit manifest as a combination of regularity and randomness,compounded by passengers’ diverse socioeconomic attributes and travel preferences.Consequently,predicting passenger travel behavior becomes a formidable task.With the rapid development of big data technology,research on passenger travel in urban rail transit has gained momentum,offering more possibilities for exploring passenger travel characteristics from an individual activity perspective.This paper takes individual passenger travel as the starting point,delving into passenger travel patterns and examining the upper bound of predictability and prediction methods.The main work is as follows:We propose an efficient data mining method based on smart card data to extract passenger travel patterns in urban rail transit.By integrating the DBSCAN algorithm and considering factors such as passenger travel time,origin-destination choices,and their proportion in total travel days,we accurately capture each passenger’s historical travel patterns and adjust passenger travel sequences.Furthermore,a data-driven approach is used to classify travel patterns into four types: time regularity,spatial regularity,spatiotemporal regularity,and irregularity.This classification is based on passenger travel frequency and spatiotemporal similarity.Travel characteristics of different passenger types are then compared.We further utilize corrected spatiotemporal entropy values to characterize individual passenger travel patterns,transitioning from qualitative to quantitative evaluation.By establishing the EPC equation to calculate the upper bound of interpretability for individual travel behavior sequences,a method is derived to determine the upper bound of predictability for both single-dimensional and composite-dimensional travel sequences.Our analysis reveals that travel sequence entropy follows a normal distribution,while predictability upper bounds conform to a Gumbel distribution.The mode of predictability upper bound for station entry and exit exceeds 0.78,indicating higher predictability for station entry times compared to station sequences.Predictability upper bound for passengers with different travel frequencies can be divided into three stages: significant improvement,stable improvement,and stability.Beyond40 trips,providing additional travel sequence data may lead to redundancy in travel prediction.Cconsidering individual historical travel patterns and travel times,long-term,periodic,and short-term travel patterns are examined,predicting travel behavior at hourly intervals.Feature vectors and label vectors are constructed and GAN is employed to address data imbalances,generating virtual instances to augment the original data.Using a DNN network,travel of different passenger types during various time periods is predicted.The results show that the data generated by the GAN method significantly enhance the prediction accuracy of subsequent models.Adjusting the imbalance ratio to 1:10 yields relatively favorable prediction results and model execution times.Prediction accuracy during peak periods is relatively high,with an overall accuracy rate of 93.77%.更多还原