13.5.1 Overview

To validate a model’s accuracy, the model is recommended to report results that are not only representative of the model but that also match results to real-world conditions. Previous model debugging indicates that reported results are representative of what is occurring in the network which instills confidence in model results. Before a model is deemed reliable, calibrate it to certify that the results match the real-world conditions it was based on. The goal of calibration is to find the set of parameter values that best mimics observed measures. The following section provides guidance on calibration thresholds and methodology. The majority of the discussion in this section is based on the FHWA 2019 Guidelines.
Calibration is the adjustment of various model parameters to improve the accuracy of the model’s results. Parameters are modified to create congestion patterns that are accurate to observed data, such as how congestion forms and dissipates over time under specific conditions. Default parameters are not considered calibrated.
Per the FHWA 2019 Guidelines, calibration is applied using data from a single representative day. Model results are analyzed with a single Random Seed run. Note that this is different from the
FHWA Traffic Analysis Toolbox, Volume III
(2004) Guidelines, where model results for calibrated models were analyzed from an average of multiple Random Seed runs. It is still recommended that a model be ran with multiple Random Seeds to check for any errors in coding and errors that might cause gridlock conditions. A few common reasons for gridlock include vehicles, pedestrians, bicycles stuck in an unsolvable conflict, or a missing detector at an actuated, signalized intersection. The model is typically debugged further before continuing the calibration process.
FHWA 2004 Guidelines may be followed for traffic analysis for TxDOT projects that do not involve FHWA approval. Average day traffic analysis may be performed for these projects. Follow FHWA 2019 Guidelines for projects that involve IAJRs or FHWA approval. The 2019 guidelines require considerably more data than the 2004 guidelines. If sufficient data is not available or practical to collect, then the use of the 2004 guidelines may be requested as an exception for projects that require FHWA approval.
During calibration, solving one issue often highlights new issues elsewhere. It is crucial to strategize calibration by splitting up the process into logical, sequential steps. A good strategy is to:
  • Divide model parameters into two categories; and
    • Parameters adjusted during calibration
    • Parameters that are to not be adjusted
  • Divide the adjustable parameters into global and local parameters
    • Global parameters are those that affect the entire network (e.g., specific attributes of the representative day covering the entire network)
    • Local parameters are those that affect the parts of the network (e.g., parameter that affects a particular link, movement, speed decision, etc.)
Calibrating a model using DTA is similar to calibration of microscopic traffic simulation models. The basic steps in calibrating a DTA model (as defined in the FHWA
Traffic Analysis Toolbox XIV: Guidebook on the Utilization of Dynamic Traffic Assignment in Modeling
) is as follows:
  • Establish calibration objectives and review the project objective to confirm that the calibration task directly supports the project objective.
  • Identify the performance measures and critical locations against which the models will be calibrated.
  • Determine the statistical methodology to be used to compare modeled results to the field data.
  • Determine the strategy for model calibration and identify parameters within the DTA models that are the focus of adjustments.
  • Assemble field data previously collected for comparison to model outputs.
  • Conduct model calibration runs following the strategy and conduct statistical checks. When statistical analysis falls within the acceptable ranges, the model is calibrated.
  • Validation: Test or compare the calibrated model with a data set not used for calibration. If the model replicates the different data set, the calibration parameters and model are validated.