Mid term calibration

MT Model Calibration

Calibration Framework

For predicting close to reality, MT models are calibrated further with network performance data (i.e. Screen line traffic counts and travel times). The framework follows an approach in which Within-day and Supply loop is run few times before a Pre-day component is made part of the loop. However, Preday component is ran externally with manual changes in the parameters to match the mode share and time-of-day choices mainly in terms of closely looking at simulated screen line counts and observed screen line counts based on overall aggregation and aggregating w.r.t. time of day. This is done because the whole MT loop that involves running preday, within-day and supply is not automated and also there is no exhaustive Pathsets (all-node to all-node) available for entire Singapore. In current settings Preday runs separately, and every run of Preday requires generation of path set for new OD’s, generated due to random allocation of nodes, as a result of having destination choice model especially for shopping and other activities within the Preday. The path set table currently stored paths between several O-Ds, however, the generation process only requires paths for new O-Ds. There was some memory issue experienced to store this huge amount of data, however, recently this has been enhanced and hopefully the Pathset generation can now be run for all-node to all-node. Once this task is finished with automation of whole MT model, the calibration framework can integrate all components of MT running in one loop. So gradient computation (based on plus and minus perturbation) may involve Preday parameters as well, which is not the case in current setup.

Data Used for Calibration

Two types of data is used for calibration in current framework. a. Screen-line counts (these are classified counts for every half-hour) b. Travel times (these are zones to zones travel time for every half-hour) The input data provided to run calibration script is stored in vishnu’s and shahita’s hpc account where Calibration script can be triggered.

Calibration Algorithm

SPSA and WSPSA algorithms are being currently used in the calibration script with OLS based objective function for both screen line count and travel time comparison. However, different weights are provided to each parameter for each part of the objective function. Parameters related to route choice are given more weight for screen line counts function and parameters related to capacity and speed density function are given more weightage for travel times. More details of it can be observed by looking at the python calibration script.

For more details on the SPSA/WSPSA methodology watch the following video by Carlos Avezedo: https://www.youtube.com/watch?v=WAWSIJsyBME

Calibration Parameters

Parameters for the Calibration are as follows: Within-day and supply parameters: These are speed-density function parameters based on 7-link category types. Capacity of the segments based on seven-link category and capacity of all segments that are approaching intersections. Route choice parameters for private route choice and public transport. Preday Parameters: These are not directly part of the calibration framework, however, based on the aggregate results of screen line counts and travel times, we have made runs of Preday several times especially changing the mode choice model parameters and time-of-day choice model parameters. In earlier simulations, huge congestion was noticed and therefore day pattern level parameters are changed to generate reasonable number of activities.

Current Results

The current calibration results are stored in Vishnu’s PC and can be found in vishnu’s and Shahita’s HPC account. However, brief summary of the results can be seen in attached posters. Remember, the objective function is OLS based, however, results are reported by computing Root Mean Square Normalized error (RMSN). The las reported RMSN for Calibration is around 0.52, giving a weightage of 0.8 to screen line measurements and 0.2 to travel time measurements. The same weights are used in OLS based objective function.

Calibration Framework Enhancement

We are working on enhancing the framework in many ways:

1. All MT models in one loop requires automation and all-node to all-node pathsets
2. Taxi roaming model need to be implemented
3. Use of Ez-link data for calibrating number of boarding in PT
4. Bus fleet management need to be incorporated, currently Bus supply model only follow dispatching of buses from their terminal stop with time-of-day based given headway.
5. If O-D matrix is available from LTA for 2012, we can have a two-step calibration. Within-day and supply can be calibrated first based on this O-D. Thereafter, Preday model can be calibrated for producing the similar O-D. This gives more control in Calibration, and better results can be achieved in short amount of simulation runs.

---