Unplanned Disruptions Characteristics Analysis And Operational Delay Prediction In Urban Railway Systems Based On Multi-source Data

The urban railway systems play a critical role in mitigating congestion in large cities,given their characteristics of low carbon emission,high capacity,high efficiency,convenient transfer,high reliability of travel time,comfort,as well as safety.With the fast-growing passenger demand,the urban railway system becomes more vulnerable once disruptions happened.Few studies were reported on unplanned disruption analysis in urban railway systems or limited in its analysis and modeling due to the lack of disruption data,which seriously impacts the service capability and safety level of the urban railway system operation.To this end,7 years’unplanned disruptions(UPDs)datasets collected from urban railway systems are used to understand disruption patterns,explore the impacts of UPDs on the level of service and passengers’travel behavior,identify influencing factors important factors impacting operation delays and affected areas,and predict operational delays in urban subway systems.First,disruption records,railway operational network characteristics and weather information are integrated to establish the suitable dataset for comprehensive analysis of UPDs in urban railway systems.Cross-tabulation and descriptive statistical analysis are used to analyze the distribution of the occurrence frequency and operational delays of UPDs from five dimensions:time,space,operation,weather,and accident type.The results show that the general distribution of operational delays has a positively skewed distribution.The delays caused by UPDs tend to be longer in non-peak periods.Disruptions that occurred on driverless trains have the largest operation delays but lowest frequency.The majority of UPDs are caused by communication equipment failure,vehicle equipment failure,and passengers alarm devices.Second,the data-driven approach is used to explore the impacts of UPDs on system performance and individual responses in urban railway systems by extracting measurements from smart card data.For the impacts of level of service,a comparative analysis is conducted to investigate changes in passenger accumulations and average travel time of operational systems or lines,as well as bus bridging performance between normal and disruption conditions.The number of passengers tap in and tap out at the same station after disruptions occurrence,the distribution of passengers’ waiting time within stations,and response behaviors are analyzed to investigate the impacts of travel behavior.The results show that UPDs have a more significant impact on operational lines,and most passengers choose to change their travel behavior after the accident.In addition,many passengers tap in and tap out at the same station after disruption occurrence.Third,the quantile regression models are developed to explore the causes of operation delays under unplanned disruptions.The significant factors include the time of day,weather condition,signal control system(moving/fixed block),line types(urban/suburban),line operation direction,disruption location(underground/ground/elevated),the number of affected stations,and disruption types(e.g.,tracing,locomotive and rolling stock,passengers,and operation).A binary logit model is developed to explore the variables contributing to the affected areas(single or multiple stations).The results show that the affected area is significantly influenced by the signal control system,line types,line operation direction,disruption location,terminal/departure station involved or not,transfer station involved or not,and disruption types.Finally,Extremely Gradient Boosting Algorithm is used to establish a high-accuracy operational delay prediction model based on UPDs datasets.The Bayesian Optimization Algorithm is employed to optimize the parameters of the prediction model.A comparative analysis is conducted to evaluate its prediction accuracy,computational efficiency,and interpretability of explanatory variables between traditional statistical models and machine learning models.The results indicate that the Extremely Gradient Boosting model outperforms other traditional machine learning models in terms of the prediction accuracy and its computational time is acceptable.Disruption types,affected areas,spatial and operational factors have a greater contribution to predicting UPDs operational delays.In addition,there are differences in the interpretability of explanatory variables between machine learning models and statistical models.The findings provide useful insights on unplanned disruptions and support the development of engineering and policy countermeasures to prevent and mitigate unplanned disruption effects on operations and services.