TL;DR
This paper introduces the Synthetic Sweden Mobility (SySMo) model, which generates a detailed synthetic population of Sweden, including over 10 million agents, their households, and daily activity-travel patterns. This comprehensive dataset addresses the lack of such granular data for Sweden, enabling more realistic agent-based simulations in areas like transportation and urban planning.
What Problem Does This Paper Address?
Creating realistic synthetic populations is crucial for agent-based modeling (ABM) in various fields, including transportation, land use, economics, and epidemiology. These models simulate complex systems by representing individual agents and their interactions. However, building accurate synthetic populations requires detailed, granular data that reflects real-world demographics and behaviors. While many countries have developed such datasets, Sweden has historically lacked a comprehensive synthetic population covering its entire populace. This gap hinders research and policy-making that relies on detailed simulations of Swedish society. This paper aims to fill this gap by presenting the SySMo model and its resulting datasets.
Methods
The SySMo model employs a multi-step methodology combining state-of-the-art techniques:
- Population Synthesis: This core process involves creating synthetic individuals and households. It starts with basic attributes (age, gender, civil status, residential zone) derived from official statistics. Households are then formed based on these attributes and household type statistics. Finally, advanced attributes like employment status, income class, and car ownership are assigned using a generative model that integrates:
- Machine Learning (ML): Specifically, a Neural Network Classifier (NNC) is used to predict probability distributions for categorical attributes.
- Iterative Proportional Fitting (IPF): This technique adjusts the predicted distributions to match known aggregated statistics, ensuring consistency with official data.
- Probabilistic Sampling: Random sampling from the fitted distributions generates the final attribute assignments for each agent.
- Activity Generation: This component creates daily activity schedules for each agent. It involves:
- Activity Type Assignment: Using a generative probabilistic model similar to the population synthesis step, agents are assigned activity types (home, work, school, other).
- Activity Duration Deduction: Broad duration classes are determined, followed by estimating hourly durations using an NNC, constrained by total daily travel time.
- Activity Sequence Assignment: Agents are assigned activity sequences by finding the most similar individuals in a travel survey based on activity durations.
- Schedule Generation: The start and end times for each activity are modeled, with the day assumed to start and end at 3 AM.
- Location and Mode Assignment: This final step places activities in specific locations and determines travel modes.
- Household Placement: Households are spatially located within residential buildings (detached houses, apartments).
- Non-Home Activity Assignment: Destination zones and travel modes (car, public transport, bike, walk) are assigned using OD probability matrices derived from transport models and a generative probabilistic model. Locations are then assigned to specific buildings.
Datasets
The SySMo model was built using several key input datasets:
- Statistics Sweden (SCB) Data:
- Official statistics on socio-demographic information at municipal and Demographic Statistical Areas (DeSO) levels.
- Population density data at square kilometer grids.
- Swedish National Travel Survey: Used for employment, studenthood, personal income, car ownership, and activity-travel patterns. (Conducted 2011-2016, ~40,000 participants aged 6-84).
- Origin-Destination (OD) Matrices: From Trafikverket's Sampers model, used for assigning activity zones and determining travel modes.
- Building Data: From Lantmäteriet, providing building locations, footprint areas, and usage types.
The paper introduces the Synthetic Sweden Mobility (SySMo) dataset itself, which is available via Mendeley Data (doi:10.17632/9n29p7rmn5). This dataset comprises three tables: Person, Household, and Activity-travel patterns for the year 2018.
Key Insights and Ideas
- Bridging the Data Gap: The creation of a comprehensive synthetic population for Sweden is a significant contribution, enabling detailed agent-based modeling for a country that previously lacked such a resource.
- Realistic Mobility Simulation: The methodology allows for generating simulations that realistically reflect heterogeneous human mobility behavior without compromising personal privacy, a critical aspect for sensitive data.
- Methodological Integration: The paper showcases a powerful combination of machine learning, iterative proportional fitting, and probabilistic sampling for synthesizing complex populations and behaviors. This integrated approach is key to achieving statistical representativeness and behavioral realism.
- Reproducibility and Transparency: By making the data and methodology public, the authors enhance research integrity and transparency, allowing for the evaluation and validation of research outcomes.
Key Results
- The SySMo model generates a synthetic population of over 10 million Swedish individuals (agents).
- The synthetic population's attribute distributions (gender, age) show good agreement with official statistics, with percentage differences typically within -0.5% to 0.5% for gender and -1% to 1% for age across most DeSO regions.
- Activity duration distributions (work, school, home, other) show good similarity to the travel survey, with Jensen-Shannon (JS) distances between 0.05 and 0.13.
- Travel distance distributions between home and work activities, when compared to the Sampers model, show reasonable approximation across various travel modes, indicating the validity of the generated activity-travel patterns.
Limitations and Future Work
- Travel Time Estimation: The paper acknowledges that travel durations are estimated based on time gaps between activities, which is not an accurate estimation of actual travel time. Simulating with an actual road network and traffic simulation platforms (like MATSim) is suggested for more precise estimation.
- Data Usage Caution: Due to the limitations, particularly in travel time estimation, the data should be used with caution and in conjunction with other data sources and analytical methods. Limitations should be clearly communicated to users.
- Predicting Future Behavior: While synthetic populations provide insights into current demographic and statistical trends, they may not accurately predict individual behavior concerning future technology adoption or behavioral changes.
The authors suggest that the methodology can be applied to similar data in other regions, indicating a potential for broader application and future development of synthetic population models globally.
Summary updated