基于卷积神经网络的城市人群时空分布预测模型——以南京为例

Spatiotemporal Distribution Prediction Model of Urban Population Based on Convolutional Neural Network: A Case Study of Nanjing

人群时空分布模型的建构方法是城市规划学科的长期研究议题,通信技术的发展为通过建构人群时空分布模型研究城市复杂系统提供了重要的数据来源和技术支撑。本文以500 m×500 m的矩形栅格为空间分析单元,利用手机信令数据建构24小时人群密度分布特征数据集;集成32个建成环境因子形成指标库,建构建成环境特征数据集。基于卷积神经网络建构南京市人群时空分布预测模型,以城市建成环境特征为自变量、人群多时段密度分布情况为因变量,学习两者间的非线性关系,实现了对南京市人群分布情况的模拟。基于平均百分比误差值的空间分布,该模型呈现较高的拟合度。结果发现,模拟结果误差较小的空间分析单元位于南京市中心区域和远郊未集中建设区域,误差较大的空间分析单元位于城市拓展区和正在建设中的地区等人群活动不稳定的区域。

The construction method of population spatiotemporal distribution model is a longterm research topic of urban planning. The development of information and communication technology provides important data source and technical support for the construction of population spatiotemporal distribution model to study urban complex systems. In this paper,500m × 500m rectangular grid is used as the research unit to construct the 24h population density distribution characteristic data set by using the mobile phone signaling data. It integrates 32 built-up environmental factors to form an index database, and constructs the built-up environmental characteristic data set. Based on the convolution neural network, the temporal and spatial distribution prediction model of population in Nanjing is constructed. Taking the characteristics of urban built environment as the independent variable and the multi period density distribution of population as the dependent variable, the nonlinear relationship between them is learned, and the simulation of population distribution in Nanjing is realized. Based on the spatial distribution of MAPE value, the model shows a high degree of fitting. Among them, the spatial units with small error in the simulation results are located in the central area of Nanjing and the non-centralized construction area in the outer suburbs, and the spatial units with large error are located in the areas with unstable crowd activities such as urban expansion area and construction area.