Quantify Uncertainty in Travel Forecasts Jason Chen, RSG Vince Bernardin, RSG Thomas Adler, RSG Nikhil Sikka, RSG Steven Trevino, RSG Steve Tuttle, RSG May, 2017
Acknowledgement & Disclaimer This project was funded as part of the TMIP Toolbox Disclaimer The views and opinions expressed in this presentation are those of the presenters and do not represent the official policy or position of FHWA and do not constitute an endorsement, recommendation, or specification by FHWA.
Introduction
Reasons for Quantifying Uncertainty Benefits of quantifying uncertainty Providing comprehensive results Evaluating the risk to key stakeholders Describing how changes to key assumption could affect the outcome Accounting for highly uncertain assumptions
Methods for Quantifying Uncertainty Historical / Retrospective Analytic Univariate sensitivity analyses Decomposition/Incremental Scenario testing Response surface simulation
Overview
Response Surface Simulation Steps Step 1: Quantify the key uncertainties or risks Land use Cost Travel Preferences/Behavior ETC.
Response Surface Simulation Steps Step 2: Design and Conduct Experiments Experimental Design Many Travel Model Runs
Response Surface Simulation Steps Step 3: Estimate and Apply Response Surface Model in Simulation Regression model to connect model output with uncertain inputs Monte Carlo Simulation to produce probability distributions of performance measures from model outputs
Response surface simulation steps Model Parameters Forecasting Model Future Conditions Response Surface Analysis Model review Historical Data Synthesized Model Probability Distribution of Model Parameters Probability Distribution of Future Conditions Monte Carlo Simulation Forecasting Distribution
Case Study 1 Toledo, Ohio Trip based model Cube Around 600K population
Case Study 2 Chattanooga, Tennessee Activity based model TransCAD Around 450K population
Step 1: Quantifying the Uncertainties
Step 1: Quantify the Key Uncertainties Sources of Uncertainty Generational Modal Preferences Telecommuting Parking cost Transit fare Fuel costs Land Use More …
Probability Distribution for Land Use Three mutually exclusive areas: Urban core Boom city Halo area Distinct scenarios: Default 2045 from agency High growth Low growth Medium growth Boom city
Probability Distribution for Land Use
Probability Distribution for Telecommuting Census/ACS/SIPP Annual Growth 2045 Telecommuting 𝑃𝑐𝑡 2045 = 𝑃𝑐𝑡 2010 ∗ 1+ 𝐺 10−11 ∗ 1+ 𝐺 11−12 ∗ 1+ 𝐺 12−13 ∗…(1+ 𝐺 44−45 )
Probability Distribution for Fuel Cost 1994 to 2014 retail gasoline 2045 distribution of fuel costs was simulated by growing the 2010 prices
Probability distribution for Transit Fare Policy driven Weak relationship with economic factors Assumed to be discrete distribution
Step 2: Design and Conduct Experiments
Computational Experiments Two Component Steps Experimental Design Run Travel Demand Forecasting Models
Experimental Design Key sources of uncertainty in two case studies Land Use Telecommuting Parking Cost Transit Fare Fuel Cost Generational Modal Preferences Influencing factors for output Urban Core Population Growth Urban Core Employment Growth Halo Area Population Growth Halo Area Employment Growth Boom City Growth Generational Modal Preferences Telecommuting Parking cost Transit fare Fuel cost
Experimental Design 20 Orthogonal fractional factorial experiments
Step 3: Response Surface Modeling
Response Surface Modeling Two Component Steps Response Surface Model Estimation Linear regression in SPSS Monte Carlo Simulation Used Excel add-in Performance Measures of Interest Vehicle Miles Traveled (VMT) Vehicle Hours Traveled (VHT) / Delay Auto and Truck Emissions Transit Ridership
Estimated Coefficients for VMT Model Trip Based Model (Toledo) Activity Based Model (Chattanooga) Coefficients Beta Std. Error Constant 15757601.0 184225.3 13860181.3 554104.1 Auto Cost 13092.2 5343.5 -26994.5 6312.8 Telecommute -42453.3 6765.1 -5861.3 15507.4 Transit Cost 35241.2 45100.9 38232.2 71316.1 Parking Cost 2793.4 6132.7 -612.0 7188.1 Non Auto Preference -308132.0 105704.9 -363941.9 118890.3 Urban Population Growth 21009.9 2074.9 21103.015 16712.6 Urban Employment Growth -1695.0 2271.5 -20064.1 14440.7 Halo Zones Population Growth 29493.7 13068.7 26934.5 6028.0 8587.5 2035.7 28613.6 10257.8 Boom City 929590.8 69483.9 418178.7 109600.2 Halo Population Growth Square 878.7 350.3 Adjusted R Square 0.98 0.89
Monte Carlo Simulation
Output Probability Distribution: VMT
Final Thoughts
Translating Analysis into Insights Understanding the Impact of Uncertainties A few things (growth, particularly suburban) may drive uncertainty in outcomes (VMT) Other things may have limited impact (on VMT)
Conclusions Uncertainties inherent in demand forecasts Demand models not efficient for simulating the probability distributions of demand Simple sensitivity analyses do not provide robust information Response surface models can effectively simulate risks associated with much more complex travel demand forecasting models
Jason Chen jason.chen@rsginc.com 802.359.6431 Vince Bernardin, PhD CONSULTANT jason.chen@rsginc.com 802.359.6431 Vince Bernardin, PhD DIRECTOR vince.bernardin@rsginc.com 812.200.2351