Data-Driven Prediction System of Dynamic People-Flow in Large Urban Network Using Cellular Probe Data

Abstract EN

Cellular probe data, which is collected by cellular network operators, has emerged as a critical data source for human-trace inference in large-scale urban areas.

However, because cellular probe data of individual mobile phone users is temporally and spatially sparse (unlike GPS data), few studies predicted people-flow using cellular probe data in real-time.

In addition, it is hard to validate the prediction method at a large scale.

This paper proposed a data-driven method for dynamic people-flow prediction, which contains four models.

The first model is a cellular probe data preprocessing module, which removes the inaccurate and duplicated records of cellular data.

The second module is a grid-based data transformation and data integration module, which is proposed to integrate multiple data sources, including transportation network data, point-of-interest data, and people movement inferred from real-time cellular probe data.

The third module is a trip-chain based human-daily-trajectory generation module, which provides the base dataset for data-driven model validation.

The fourth module is for dynamic people-flow prediction, which is developed based on an online inferring machine-learning model (random forest).

The feasibility of dynamic people-flow prediction using real-time cellular probe data is investigated.

The experimental result shows that the proposed people-flow prediction system could provide prediction precision of 76.8% and 70% for outbound and inbound people, respectively.

This is much higher than the single-feature model, which provides prediction precision around 50%.