# Online Trajectory Prediction for Metropolitan Scale Mobility Digital Twin

**Authors:** Zipei Fan, Xiaojie Yang, Wei Yuan, Renhe Jiang, Quanjun Chen, Xuan Song, Ryosuke Shibasaki  
**arXiv:** [2207.03575](https://arxiv.org/abs/2207.03575)  
**Date:** 2022-06-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.03575v1/

## Key points

- Data-driven smart city systems rely on understanding mobility patterns of inhabitants
- Mobility digital twin has potential in predicting and simulating human movements at metropolitan scale
- Challenges include daily variations in mobility, transportation network constraints, rational fine-grained trajectory generation, and high computational costs for fine-grained predictions
- Proposed solution is a two-stage human mobility predictor that stratifies coarse and fine-grained level predictions
- Citywide mobility trends are extracted as crowd contexts to encode daily variations in human mobility at a metropolitan level
- Long-term and long-distance movements are predicted at a coarse level based on these trends
- Probabilistic trajectory retrieval methods resolve coarse predictions to a fine-grained level by offloading heavy computations to an offline phase
- Tested using real-world mobile phone GPS data from the Kanto area in Japan with good prediction accuracy and time efficiency of approximately 2 minutes when predicting future 1-hour movements of around 220K mobile phone users on a single machine
- Approach provides higher-level analysis of mobility prediction for urban planning and management decisions
- Innovative solution for predicting human movement patterns in cities through an online trajectory prediction system that combines both coarse and fine-grained predictions while addressing various challenges faced by existing models.

## Summary

The concept of a data-driven smart city system relies heavily on understanding the mobility patterns of its inhabitants. In recent years, the development of a mobility digital twin has shown great potential in predicting and simulating human movements in near real-time at a metropolitan scale. However, practical solutions have been limited due to several challenges. Firstly, daily variations in human mobility are difficult to model and predict accurately. Secondly, transportation networks impose complex constraints on human movement. Thirdly, generating rational fine-grained human trajectories poses a challenge for existing machine learning models. Finally, making fine-grained predictions incurs high computational costs that are challenging for an online system. To address these difficulties, this paper proposes a two-stage human mobility predictor that stratifies coarse and fine-grained level predictions. In the first stage, citywide mobility trends are automatically extracted as crowd contexts to encode daily variations in human mobility at a metropolitan level. Long-term and long-distance movements are then predicted at a coarse level based on these trends. In the second stage, probabilistic trajectory retrieval methods resolve coarse predictions to a fine-grained level by offloading heavy computations to an offline phase. The proposed method was tested using real-world mobile phone GPS data from the Kanto area in Japan and achieved good prediction accuracy with time efficiency of approximately 2 minutes when predicting future 1-hour movements of around 220K mobile phone users on a single machine. This approach provides higher-level analysis of mobility prediction that can be used to support urban planning and management decisions effectively. Overall, this study presents an innovative solution for predicting human movement patterns in cities through an online trajectory prediction system that combines both coarse and fine-grained predictions while addressing various challenges faced by existing models. The results demonstrate its potential for use in modern urban intelligent systems towards building smarter cities globally.