350 likes | 467 Views
San Francisco DTA Model: Working Model Calibration Part 1: Process Greg Erhardt Dan Tischler Neema Nassir. DTA Peer Review Panel Meeting July 25 th , 2012. Agenda. 9:00 Background 9:30 Technical Overview – Part 1 Development Process and Code Base/Network Development 10:15 Break
E N D
San Francisco DTA Model: Working Model CalibrationPart 1: ProcessGreg ErhardtDan TischlerNeema Nassir SAN FRANCISCO COUNTY TRANSPORTATION AUTHORITY DTA Peer Review Panel Meeting July 25th, 2012
SAN FRANCISCO COUNTY TRANSPORTATION AUTHORITY Agenda • 9:00 Background • 9:30 Technical Overview – Part 1 • Development Process and Code Base/Network Development • 10:15 Break • 10:30 Technical Overview – Part 2 • Calibration and Integration Strategies • 12:00 Working Lunch / Discussion • 2:00 Panel Caucus (closed) • 3:30 Panel report • 5:30 Adjourn
Outline • Model Overview • Calibration Approach • Speed Flow Parameters • Presented by Dan Tischler & Neema Nassir • Model Calibration Runs • Current Model Parameters • Key Findings
Model Overview Natural breakpoint at San Bruno Mountain Park 976 TAZs 22 external stations 1,115 signals 3,726 stop controlled intersections
Model Overview PM Peak Model from 4:30-6:30 pm 1 hour warm-up time 3 hour network clearing time 270,000 internal trips 180,000 IX , XI or XX trips Dynameq Software Platform
Calibration Approach • Ensure quality inputs • Measure anything that can be measured • Evaluate the results qualitatively • Evaluate the results quantitatively • Make defensible adjustments
Ensure Quality Inputs Identify and investigate failed signal imports Spot check stop-control—some issues with direction of 2-way stops Automate as much as possible
Measure Anything that can be Measured Measure speed flow parameters Change perceived cost instead of measured speed and capacity Avoid arbitrary demand changes
Evaluate Qualitatively Example of extreme congestion
Evaluate Quantitatively Relative gap, RMSE, GEH, R-Squared Scatter plots, maps Tables by: area type, facility type, speed, turn type, time period, etc. Corridor plots Speeds
Make Defensible Adjustments Evaluate results and investigate worse offenders Hypothesize problems and propose changes
Base Case – July 6 Test • Change(s): • Results: • RMSE: Links = 133 (58%), Movements = 64 (80%) • GEH: Links = 7.17, Movements = 4.59 • Overall Vol/Count Ratio: Links = 0.6527, Movements = 0.7145 • This test includes intrazonal trips (assigned to the nearest centroid) and ambiguous two-way stop signs re-assigned as all-way stops • At this stage, there were still network and signal issues that have since been dealt with
Test 1 – Speed-Flow Curve Changes • Change(s): Free-flow speed, response time factor, effective length factor • Results: • RMSE: Links = 132 (57%), Movements = 64 (80%) • GEH: Links = 7.04, Movements = 4.56 • Overall Vol/Count Ratio: Links = 0.6467, Movements = 0.7051 • Increasing RTF and decreasing speeds caused gridlock in the CBD • Without bus-only lanes, these changes have more impact • With bus-only lanes included, capacities are too low and CBD is full of gridlock
Test 2 – Removing Bus-only Lanes • Change(s): Bus-only lanes no longer specified as bus-only • Results: • RMSE: Links = 135 (59%), Movements = 64 (80%) • GEH: Links = 7.32, Movements = 4.59 • Overall Vol/Count Ratio: Links = 0.6459, Movements = 0.7085 • Got rid of gridlock in CBD • People are allowed to use these lanes for right turns – how can we model that? • Need to add them back in some way while still allowing for limited use – next test.
Test 3 – Increasing Demand • Change(s): Increasing internal demand by 30% • Results: • RMSE: Links = 155 (68%), Movements = 72 (90%) • GEH: Links = 8.18, Movements = 4.86 • Overall Vol/Count Ratio: Links = 0.6316, Movements = 0.7526 • Significant gridlock all over the network • Previously about 30% low on counts, but more demand overloads the network • Need to fix flow patterns and speeds, not demand
Test 4 – Penalizing Locals & Collectors DTA Volumes Static Volumes
Test 4 – Penalizing Locals & Collectors DTA Volumes Static Volumes
Test 4 – Penalizing Locals & Collectors • Change(s): Local and collector links had penalty of 1*FFTime added to generalized cost • Results: • RMSE: Links = 122 (53%), Movements = 61 (76%) • GEH: Links = 6.85, Movements = 4.47 • Overall Vol/Count Ratio: Links = 0.8074, Movements = 0.855 • Arterial Plus flows are still much lower than expected – looking at speed-flow curves • Important to test this again with transit-only lanes added back in some way
Response Time Factors * Response times corresponding to RTF equal to 1.1 and 1.2 are respectively 1.375 and 1.5 seconds
Assignment Specification • These values define the period of the simulation: • Start of demand: 15:30 • End of demand: 18:30 • End of simulation period: 21:30 • Transit lines simulation: Yes • Re-optimization: No • Re-optimization iteration(s): 0
Demand Specification • Demand and generalized cost for cars: • class: Car_NoToll • matrix: car_notoll • paths: 20 • intervals: 12 • types (%): Car=100, • generalized cost: movement expression + link expression • movement expression: ptime+(left_turn_pc*left_turn)+ (right_turn_pc*right_turn) • link expression: fac_type_pen*(3600*length/fspeed) • Demand and generalized cost for trucks: • class: Truck_NoToll • matrix: truck_notoll • paths: 20 • intervals: 12 • types (%): Truck=100, • generalized cost: movement expression + link expression movement expression: ptime+(left_turn_pc*left_turn)+(right_turn_pc*right_turn) • link expression: fac_type_pen*(3600*length/fspeed)
Control Plans and Results Specifications • Signals are applied during this period: • excelSignalsToDynameq: 15:30 - 18:30 • These settings specify the time steps used by Dynameq. The purpose of these settings is just for analysis of the DTA results and doesn’t have any bearing on the results themselves. • Simulation results: 15:30:00 - 21:30:00 -- 00:05:00 • Lane queue animation: 15:30:00 - 21:30:00 -- 00:05:00 • Transit results: 15:30:00 - 21:30:00 -- 00:05:00
Advanced Specifications • These values are settings for the DTA method used by Dynameq. • Traffic generator: Conditional • Random seed: 1 • Travel times averaged over: 450 s • Path pruning: 0.001 • MSA reset: 3 • Dynamic path search: No • MSA method: Flow Balancing • Effective length factor: 1.00 • Response time factor: 1.00
Key Findings • Model is sensitive to changes, and can easily regress into gridlock. • Bus-only lanes matter. • Penalizing locals and collectors helps. • Increasing internal demand 10% helps. Increasing demand 30% causes gridlock. • Most runs show less congestion than we would anticipate.