Abstract
This research presents a significant advancement in urban planning technology by integrating three traditionally siloed domains—land use, travel demand, and traffic assignment—into a single, high-performance modular pipeline. By leveraging GPU acceleration and open-source tools like UrbanSim and ActivitySim, the system can model millions of individual agents across an entire metropolitan region with unprecedented detail and speed, capturing the complex feedback loops between real estate markets and traffic congestion.
Introduction
In regional planning, the "gold standard" is to understand how land use (where people live and work) and transportation (how they move) interact over decades. However, achieving this at a "disaggregate" level—modeling individual people, households, and vehicles rather than broad averages—is computationally grueling. Historically, models have been data-hungry, slow, and often failed to represent non-driving modes due to low spatial resolution. This framework aims to break these barriers by creating a modular, open-source architecture capable of simulating metropolitan-scale dynamics at the level of individual city blocks.
Integrated Modeling Framework
The proposed architecture is organized into two primary stages: a data initialization pipeline and a modular microsimulation pipeline. The system is designed to be flexible, allowing researchers to swap specific model steps through a templating system.
The core microsimulation pipeline consists of three primary components:
- UrbanSim (Land Use): A long-term model predicting household and firm locations over a 30-year horizon using discrete choice models to simulate real estate market-clearing processes.
- ActivitySim (Travel Demand): An agent-based modeling (ABM) platform that predicts the daily itineraries of synthetic agents, including destinations, departure times, and transport modes.
- Traffic Assignment & Microsimulation: The system integrates standard Static User Equilibrium with a novel GPU-parallelized traffic microsimulator. By moving vehicle behavior simulation (lane changing, car following) to the GPU, the system can process millions of agents simultaneously.
Implementation and Data Strategy
To overcome the performance limitations of Python, the framework utilizes C-implemented math libraries like NumPy for vectorized calculations. A key innovation is the use of a "traffic atlas" to discretize roads into memory-efficient layouts for GPU processing.
The framework relies on a mix of public and proprietary data sources:
- OpenStreetMap (OSM): Processed via the OSMnx library to generate simplified, realistic road networks.
- Synthetic Populations: Generated from Census and MPO data, representing millions of individuals (e.g., 6.9 million for the SF Bay Area).
- Real Estate Data: County assessor records and commercial sources provide the basis for price and rent modeling.
By moving away from arbitrary "traffic analysis zones" toward network-based accessibility (using tools like Pandana), the model achieves the spatial resolution necessary to analyze walking and biking infrastructure.
Performance and Validation
The system demonstrates remarkable computational efficiency. A single UrbanSim iteration for 2.6 million households takes approximately 11 minutes, while a full 30-year scenario can be completed in 5.5 hours. ActivitySim, while more intensive, processes the same population in roughly 16 hours.
Validation against industry-standard tools like SUMO showed that the GPU-based traffic microsimulator produced results within a 6% margin of error for occupancy values, despite operating at a significantly larger metropolitan scale. Furthermore, the model successfully captures the feedback loop where congested travel times eventually influence relocation decisions in the land-use model.
Conclusion and Future Work
While the current framework provides a robust foundation for regional planning, future iterations will focus on optimizing bottlenecks in travel demand modeling and improving the integration of workplace choice and vehicle ownership. The modular design is specifically intended to eventually incorporate emerging technologies such as Transportation Network Companies (TNCs) and Autonomous Vehicles (AVs), ensuring the tool remains relevant as urban mobility continues to evolve.
Summary updated