Calibration of Vissim for Bus Rapid Transit Systems in Beijing Using GPS Data

Bus Rapid Transit systems have grown in popularity in recent years. With the rapid development of computer technologies, using microscopic simulation models to study various strategies on planning, implementation and operation of BRT systems has become a hot research area in the field of public transportation. To make the simulation models accurately replicate field traffic conditions, model calibration is crucial. This paper presents an approach for calibrating the microscopic traffic simulation model VISSIM using GPS data for application to Beijing BRT systems. The Sum of Squared Error (SSE) of the collected versus simulated vehicle speeds at the cross-sections along the test route is specified as the evaluation index. A Genetic Algorithm is adopted as the optimization tool to minimize the SSE. Taking the Beijing North-South Central Axis BRT Corridor as a case study, it shows that the proposed approach is a practical and effective method for the model calibration. Journal of Public Transportation, 2006 BRT Special Edition 240 Introduction Bus Rapid Transit (BRT) systems have grown in popularity in recent years. With the rapid development of computer technologies, using microscopic simulation models to study various strategies on planning, implementation and operation of BRT systems has become a hot research area in the field of public transportation, particularly in cases where field experiments are difficult or expensive to conduct. There are plenty of available microscopic simulation models used worldwide, such as VISSIM, CORSIM, PARAMICS, etc. In such models, there are a number of parameters describing the traffic flow characteristics, driving behavior, and traffic system operations, which have significant effects on simulation results. Although these models provide a set of default values for each parameter and users can conduct a simulation without calibrating them, the default values may not always be representative of the traffic situation under study. For example, the driving behavior of BRT vehicles on the exclusive lanes may be different from those on urban streets or freeways because BRT has some unique traffic characteristics (e.g., dispatching according to schedule, stopping at bus stops for serving paSSEngers, etc.). Even BRT systems in different countries or different cities may have different characteristics. For a simulation study of BRT systems, adequate calibration based on observed traffic conditions can result in accurate and reliable simulation results, which can help transit operators make more appropriate decisions for BRT planning and implementation. So, when using a simulation model for different geographic and traffic conditions, the most important and difficult step is the calibration and validation of the model. The calibration is the process by which the values of a simulation model input parameters are refined and adjusted so that the model accurately replicates field-measured and observed traffic conditions. The aim of this paper is to propose an approach for the automatic calibration of the driving behavior parameters of VISSIM using GPS data for application to Beijing BRT systems. A Genetic Algorithm (GA) is used for finding the best combination of VISSIM driving behavior parameters, and a particular computer simulation program named AUTOSIM is designed to run the VISSIM simulation automatically and consecutively. The validity of the proposed approach was demonstrated via a case study for the Beijing North-South Central Axis BRT Corridor. The results show that it is a practical and efficient approach for the calibration of VISSIM. Calibration of Vissim for BRT Systems in Beijing Using GPS Data 241 Review of Calibration Methodologies The problem of model calibration is very complex because of the absence of a clear analytical formulation for model users to follow. In recent years, more and more transportation researchers have realized the importance of model calibration and made great efforts to develop various methodologies to calibrate traffic simulation models. In earlier studies, manual changes were used for calibrating model parameters (Daigle et al. 1998), which was found not efficient and practical. Fellendorf and Vortisch (2001) calibrated the car following behavior of VISSIM with measurement on the level of single vehicles, i.e., data about headways, perception thresholds, and driving characteristics. However, it is difficult for model users to collect some of such data in the field. Merritt (2003) proposed a methodology for the calibration and validation of CORSIM using empirical data. He found that extensive field data need to be collected to improve accuracy of the model calibration. With the recent applications of ITS technologies and computational resources, there are more opportunities to calibrate simulation models based on optimization theories and algorithms. Ben-Akiva et al. (2004) presented a framework for the calibration of microscopic traffic simulation models using aggregate data. They adopted a systematic search approach based on Box’s Complex algorithm for calibration, which did not require calculations of derivatives of the objective function. Nevertheless, their study found that efficient algorithms are still required to perform the calibration step. Some other algorithms, such as sequential simplex algorithm (Kim 2003) and simulated annealing algorithm (Wieland 2004), also have been studied by several researches. In recent years, microscopic traffic simulation models have been widely used as an important tool for the analysis and design of transportation systems in China. However, many users conduct simulations simply with the default parameters provided by the model without calibrating them. The study on the calibration of traffic simulation models in China is also scarce. Sun and Yang (2004) proposed a procedure for microscopic simulation model calibration in China. They designed the experiment by using Latin Square algorithm and calibrated four of the driving behavior parameters of VISSIM, including waiting time before diffusion, minimum headway, observed vehicles, and average standstill distance. However, it takes much time to finish all the simulation experiments and these four parameters cannot represent the whole set of driving behavior parameters of VISSIM. Journal of Public Transportation, 2006 BRT Special Edition 242 Proposed Calibration Approach Identification of Calibration Parameters in VISSIM VISSIM is a microscopic, time-step and behavior-based simulation model developed to model urban traffic and public transit operations. It provides significant enhancements in terms of driver behavior, multi-modal transit operations, interface with planning/forecasting models, and 3-D simulation. VISSIM contains a psycho-physical car-following model for longitudinal vehicle movement and a rule-based algorithm for lateral movements. Ten calibration parameters are selected in VISSIM, including: • Waiting Time before Diffusion—It defines the maximum amount of time a vehicle can wait at the emergency stop position waiting for a gap to change lanes in order to stay on its route. When this time is reached the vehicle is taken out of the network (diffusion) and a warning message will be written to the error file denoting the time and location of the removal. • Minimum Headway (front/rear)—defines the minimum distance to the vehicle in front that must be available for a lane change in standstill condition. • Maximum Deceleration—the fastest a vehicle can slow down or stop. • -1 per Distance—used to reduce the maximum deceleration with increasing distance to the emergency stop position. • Accepted Deceleration—the value of it is smaller than maximum deceleration but bigger than minimum deceleration, and the vehicle can slow down safely without any dangerous with accepted deceleration. • Maximum Look Ahead Distance—the maximum distance that a vehicle can see forward in order to react to other vehicles either in front or to the side of it (within the same link). This value relates to human’s physical observation ability. • Average Standstill Distance—defines the average desired distance between stopped cars and also between cars and stop lines (signal heads, priority rules, etc.) • Additive Part of Desired Safety Distance—this parameter and the next one (i.e. Multiple Part of Desired Safety Distance) contained with the car following model determine the saturation flow rate for VISSIM. The saturation flow rate defines the number of vehicles that can free flow through a VISSIM model during one hour. Calibration of Vissim for BRT Systems in Beijing Using GPS Data 243 • Multiple Part of Desired Safety Distance—described above. • Distance of Standing at 50 km/h—the safety distance between two parallel cars at both the condition of stop and moving. As the parameters mentioned above directly affect the vehicle interaction and thus can cause substantial differences in simulation results, calibration of these parameters become very important. To this end, a scientific approach is needed to calibrate these parameters. Selection of an Optimization Algorithm For calibration of a traffic simulation model, the difficulty is to select the best combination of the parameters being calibrated. However, all of these parameters need to be calibrated simultaneously, and each may have a different value range, which make the calibration process very complicated and time consuming. So, to identify the best parameter set for the model, an optimization algorithm is required. A Genetic Algorithm (GA) is a search technique used in computer science to find approximate solutions to optimization and search problems. It is a particular class of evolutionary algorithms that use techniques inspired by evolutionary biology, such as inheritance, mutation, natural selection, and recombination (or crossover). It models each possible parameter set as a separate chromosome, and each chromosome is evaluated by a fitness function that represents how well it fits a given problem (Kim 2001). GA is considered robust because it performs a search from multiple points instead of starting the search at a single point. So, using the GA approach can considerably reduce the number of search steps needed and the amount of time required to complete the search when the search space is large and complex. Index of Simulation Accuracy To evaluate the quality of the simulation in the calibration, an evaluation index needs to be defined. There are various indexes that can be used, such as traffic volumes, average travel time, average travel speed, queue lengths, etc. This paper uses the Sum of Squared Error (SSE) between the vehicle speeds collected and those simulated at pre-defined cross-sections at a 20-meter interval along the test route, which is calculated by the following equation: