Development of dynamic OD matrix based TDM framework using cell phone data

Abstract

A framework of travel demand modeling is devised in this work that uses cell-phone network generated data to estimate the operational demand matrices in fine spatial and temporal resolution. Due to its high penetration rate, extensive coverage and ubiquitous use, opportunistically collected cell-phone data has a great potential to be used for passive probing of traffic. For megacities like Dhaka, this method has definite advantage over others because of lack of permanent data collection infrastructure, high market penetration rate (over 90%) of cell-phones, spatially densified network structure and less expense involved in the process. Also, from a questionnaire survey done for this thesis it is found that 80% of mobile phone users make six or more calls per day and 28% of them make more than 10 calls during their trip. Motivated by the prospect, a large dataset containing one month call data records (CDR) of nearly 1.3 billion network connections for about 5 million subscribers was collected from the largest network operator of Bangladesh (Grameen Phone). To the knowledge of the author, this is the largest billing dataset used in this purpose. Computationally efficient and simplified algorithms are designed to process and analyze this big data of 80 GB. A part of the simplification is achieved by introducing temporal resolution and filtering out infrequent callers. This simplification has reduced the computation time by 35 times. The database is processed to impute trips and synthesize time-dependent trip matrices. Derived matrices are important indicator of travel pattern or characteristics of persontrips. In order to convert these ‘pattern’ or ‘seed’ matrices to operational vehicle-trip matrices, scaling factors for different time periods are developed. Surveyed commute flow data (DHUTS, 2010) is used to derive these factors. The factors are determined separately for each OD pair due to variations in person-trip to vehicle-trip conversion factors, operator subscription, network coverage, market penetration rate and type of technology used. Also, clustering of matrices for identical time periods is done to get representative trip matrices for those periods. Therefore, devised framework is capable of incorporating more dataset and producing more reliable factors for similar time period. ‘Seed’ matrices for that corresponding time periods are factored and the factored matrices are assigned in modeling software, TransCAD, using prevalent demand modeling techniques. A validation framework is presented at the end to find out the accuracy of our methodology. Video recording made during the time period was used as ‘ground truth’ data. But due to unavailability of this data for the entire modeled period and network, some complementary data from secondary source was used. The result shows less than 40 percent deviation in network flows for 85 percent links of the network that is relatively less compared to other demand modeling practices where it is common to get greater percentage of deviation. Additional information regarding users’ behavior in the network is also extracted. The dynamic calling pattern map shows us population distribution and activity-space over time. It is found that the number of calls a user make in a day follows a diminishing distribution pattern. The logarithm of inter-call time follows a quadratic frequency distribution. Cumulative distribution function for the number of imputed trips by a user can be well approximated by a normal distribution or an exponential distribution. The relationship between tripmaking and callmaking of each user reveals that a minimum number of callmaking is required to sense certain number of trips for a specific user. Also, most of the data points (99.5%) are contained within the limit of 20 trips per day. Number of trip sensed is at the maximum for Thursdays and at the minimum for Fridays. Moreover, an absence of morning peak is found for weekends and peak-spreading of morning peak is observed for weekdays.