Page 125 - Special Topic Session (STS) - Volume 3

STS520 Xin Zheng et al.

Research on method of population prediction by
big data from mobile phones
2
*1
1
1
Xin Zheng , Qing Shen , Mingcui Du , Guangzhi Zhang , Changcheng Kan 3
1 Beijing Municipal Bureau of Statistics, Beijing, China
2 RongxinZhilian Network Technology Co., Ltd., Beijing, China
3 Baidu Times Network Technology Co., Ltd. Beijing, China

Abstract
The signaling and APP data from mobile phones, as the most representative
type of data among the spatial big data, enable the population analysis based
on the personal behavior of mobile phone. By combining the traditional
statistical method with the AI deep learning technology, this paper uses the
time series correlation, spatial sequence correlation and deep residual model
to predict the population size, and good results have been achieved. As the
change in population size is periodic in time, the short-term prediction effect
is good, but the effect brought by periodicity should be eliminated for the
long-term prediction.

Keywords
Time series correlation; Spatial sequence correlation; Deep residual model

1. Introduction
In recent years, with the rapid development of mobile internet
technology, mobile phones have become indispensable items in people’s
daily life. The signalling data and APP usage records from mobile phones, as
the most representative type of data among the spatial big data, record the
massive and diversified crowd time and space location information at short
intervals, enable the analysis of urban spatial characteristics based on the
personal behaviour of mobile phone and are of great significance for urban
planning, transportation, public resource allocation, and business information
mining, etc. By combining the traditional statistical method with the AI deep
learning technology, this paper uses the time series correlation, spatial
sequence correlation and deep residual model to predict the population size.

2. Methodology
(1) Using the time series correlation
If the time during a day is divided into several time periods , , ⋯ ,

1
2
by 10min, 30min or 60min, the time series correlation model can be used to
predict the population size in this area. If it is assumed that the population
size at the next time period is related to that at the previous p time periods,

114 | I S I W S C 2 0 1 9