Optimization of headway with smart card data mining in urban public transportation planning

In public transport, smart card charging is becoming more common. The registration of payment and boarding with smart card is kept in large databases. These data related to smart card transactions have the potential to be used in planning activities such as understanding consumer behavior patterns, measuring line service reliability and performance, estimating the origin - destination matrix, predicting future demand, determining optimum travel frequencies, and reordering line routes. By applying various data mining methods and techniques to smart card data, consumer behaviors can be understood and some suggestions can be presented to the behavior of consumer groups. With the help of the smart card bus service data, it is possible to determine the extent to which buses behave in accordance with the schedule and reliability measurements can be made. In most smart card pricing systems, the card is shown only during boarding, and the card is not displayed again during alighting. Hence, only data on boarding stops are usually available in databases. In this case, an estimate of alighting stops from boarding stops are necessary to an origin - destination matrix and predict future demand forecasts. In the literature, it is seen that travel chain algorithm is widely used. With this method, alighting stops can be estimated for most of the trips. If the boarding and landing stops are certain, the data required for the creation of the start - arrival matrices and the demand forecast are ready. A variety of methods can be used for future forecasting. The aim of this study is; to expose the hidden potential of smart card data in public transport planning and to show the planners how to use this data. It is aimed to show how useful data can be obtained from these large data heaps by using data mining methods and how new data can be used for planning. In this study, travel chaining algorithm and Diker's random assignment estimation method are used to estimate descent stops from boarding stops. Other algorithms used: k-means, x-means clustering algorithms, origin-destination matrix computation algorithm, and feed forward back-propagational multilayer perceptron type artificial neural network for future prediction, recommended simple algorithm and genetic algorithm with target programming formulation for optimum headway. In the first part of the study, use of smart cards in public transportation and planning, the literature review of knowledge discovery, performance measurement and planning activities is presented through various data mining methods from smart card data. In the second part, literature review of the application areas of data mining, data mining process, about the methods used are presented. In the third chapter, the findings of the literature research related to the algorithms used in the study are summarized. In the third chapter, the findings of the literature research related to the algorithms used in the study are summarized. This chapter also provides information on the simple approach for determining the optimum frequency of travel. In this chapter, information on the simple approach for determining the optimum headway is presented. Many findings were obtained in the study. First of all, the algorithms that identified the alighting stops from the boarding stops was run. An average of 58.68% of the alighting stops were estimated by the trip chaining algorithm. 37,63% of the remaining alighting stops were determined by random assignment estimation method and 3,69% by assignment to final stop as alighting method. Origin - destinatiton matrices were calculated for 10 bus lines with a Konak connection and maximum demand levels were calculated for each time period at the time of data scope. From this demand data, a separate ANN was run with data on weekends and an ANN for the weekday. In addition to ANN, moving averages, weighted moving averages, exponential smoothing and regression models have been developed and the performance estimations have been tested. The most accurate predictions were obtained with ANN. In order to determine the optimum frequency, the mathematical models based on the target program were solved by both simple and genetic algorithms and the results were compared. The genetic algorithm and the algorithm we proposed gave optimum results.