Methods for Calibrating and Validating Stochastic Microsimulation Traffic Models

Abstract

The purpose of this research was to propose a multistage framework for the calibration and validation of the traffic simulation models and present results of a calibration and validation experience using CORSIM model for a network of urban streets. The study proposed a series of logical, sequential steps for the calibration and validation of micro-simulation traffic models. The test bed used for the study is an important network of traffic signals in the city of Chicago, Illinois. The internal network consisting of twelve nodes at the core of the network served as the main focus of the calibration and validation experience for this study. Base data was collected using video and manual counts for extended AM and PM peak periods.

Two methods for determining the number of model repetitions were proposed: a) use of statistical formula based on desired confidence interval and degree of confidence, and b) model-based sensitivity test which examines the number of outlier runs and the variability (distribution) in the model output from running sets of 25, 50, and 100 model runs. The study showed that both methods compliment each other in arriving at the required number of model repetitions.

Automation processes using the REXX code was used for extracting the required model outputs and perform analysis of repetitive and multiple model runs during the calibration and validation processes. The calibration strategy adopted for the test network consisted of four distinct stages: a) error checking, b) calibration of input parameters for capacity and demand (throughput comparisons), c) model tuning (link attributes), and d) demand adjustment. The study showed that the concept of split links in modeling long term blockages by curb side parked vehicles proved to be more useful as compared to the NETSIM record types for long term events and parking activity. The study also showed the use of 'In' and 'Out' throughput volumes as an efficient and effective tool in calibration of micro-simulation models for urban street networks.

Outputs from the calibrated and 'tuned' model for 100 replications were used in the model validation for test network. The research demonstrated the use of the mean stop time per vehicle and its modified form — the mean stop time per stopped vehicle as effective and efficient measures for use in validation of micro-simulation models. Between the mean stop time per vehicle and the mean stop time per stopped vehicle, the later proved to be more useful in the validation process primarily because it eliminates the difference between the model and the real-world which is purely on the basis of difference in the values of percent stops counts.

For the test network, nine alternative scenarios were used for the model validation criteria in terms of the level of significance and the proportion of links using two-sample t-test. The study showed that the answer to the question whether the model is valid is dependant upon the satisfaction of the pre-defined criteria. The answer changes as the pre-defined validation criteria change.

The key contribution of this research is the development of a multistage methodological framework for calibration and validation of micro-simulation traffic models. The methodology is quick to set up and implement on traffic networks and can be used beneficially by future analysts and researchers.