A Deep Learning Framework for Large-Scale Traffic Flow Digital Twin Development with MATSim

Abstract

We present a novel framework that bridges the gap between deep learning-based OD traffic flow models and realistic, large-scale simulation environments for the development and validation of digital twins. Leveraging the PyTorch deep learning library, we implement a gradient-based OD estimation model and evaluate its performance on the Utah transportation network. Our simulation encompasses 3,275 nodes and 3,785 edges, modeling over 3 million agents using the MATSim platform. All code and datasets associated with this work are publicly available at: https://directlab.github.io/MatsimAI/.

Figures

Software Framework

Overview of the digital twin construction pipeline for large-scale traffic network simulations. The left branch shows the preprocessing of real-world geographic data, beginning with network extraction from OpenStreetMaps via JOSM, followed by data cleaning, conversion to MATSim format, and parsing into tensor representations. These are used to derive a graph structure Ĝ(V,E), which is clustered using K-means to form the clustered graph G*(C,E) and time-dependent subgraphs Gₜ(V,E), integrating UDOT traffic flows. The center illustrates the application of a traffic assignment matrix algorithm to infer origin-destination (OD) matrices. The right branch shows the iterative OD estimation process, where weights W are optimized using the ADAM optimizer to generate MATSim plans.

ADAM Optimization Results

The mean average absolute difference between the predicted flows and the actual flows as recorded by the UDOT sensors.

ADAM Optimization Results (Interactive)

Simulation Results (Interactive)

@inproceedings{peterson2024digitaltwins, author = {Isaac Peterson and Christopher Allred and Chandler Justice and Mario Harper}, title = {A Deep Learning Framework for Large-Scale Traffic Flow Digital Twin Development with MATSim}, booktitle = {IEEE Conference}, year = {2024}, }