地理学报 ›› 2019, Vol. 74 ›› Issue (2): 203-221.doi: 10.11821/dlxb201902001

所属专题: 人口与城市研究

• 人口与城市研究 •    下一篇

中国人口省际流动重力模型的参数标定与误差估算

赵梓渝1(),魏冶2(),杨冉2,王士君2,朱宇3   

  1. 1. 宁波大学公共管理系,宁波 315211
    2. 东北师范大学地理科学学院,长春 130024
    3. 福建师范大学地理研究所,福州 350007
  • 收稿日期:2017-07-12 出版日期:2019-02-25 发布日期:2019-01-29
  • 基金资助:
    国家自然科学基金项目(41401172, 41630749)

Gravity model coefficient calibration and error estimation: Based on Chinese interprovincial population flow

ZHAO Ziyu1(),WEI Ye2(),YANG Ran2,WANG Shijun2,ZHU Yu3   

  1. 1. Department of Public Administration, Ningbo University, Ningbo 315211, Zhejiang, China
    2. School of Geographical Sciences, Northeast Normal University, Changchun 130024, China
    3. Institute of Geography, Fujian Normal University, Fuzhou 350007, China
  • Received:2017-07-12 Online:2019-02-25 Published:2019-01-29
  • Supported by:
    National Natural Science Foundation of China, No.41401172, No.41630749

摘要:

空间交互模型被广泛应用于地理要素关系强度的模拟,然而目前大量研究或建立在模型参数标定理想化、模式化的假设条件下,或是在暗箱中完成,由此导致模拟结果与实际的偏差却被严重低估。基于2015年中国春运期间人口省际流动的城市间O-D数据,在逐日、分市的研究精度下,实证推算人口流动重力模型变量的回归系数,探究模型代理变量影响效应的空间异质性,并评估重力模型在人口流动模拟上的误差。结果显示:① 重力模型参数标定的复杂性体现在交互对象代理变量影响程度的非对称性,和变量回归系数的空间异质性随研究精度加深显著加剧两个方面,因此模型参数标定的模式化将导致估算结果空间差异的趋势收敛;② 2015年春运期间中国人口省际流动距离衰减系数为1.970,在地级行政单元视角下,人口流出地距离衰减系数值域为0.712(驻马店)~7.699(乌鲁木齐),人口流入地系数值域为0.792(三亚)~8.223(乌鲁木齐);③ 应用重力模型模拟人口流动结果与实测流(百度迁徙数据)存在显著误差。就加权绝对平均误差而言,拟合总误差为85.54%,其中空间相互作用效应造成了86.09%的实测流与模拟流的最大误差,相对流出力、相对吸引力分别造成57.73%、49.34%的模型误差。因此,空间交互效应仍然是当前最难以模式化的因素。

关键词: 重力模型, 回归系数, 距离衰减系数, 误差估算, 人口流动, 中国

Abstract:

Simulations based on spatial interaction models have been widely applied to understand the strength of relationships between geographical elements, but many issues remain unclear and deviations between actual and simulated results have often been seriously underestimated. A high-precision Baidu migration process combined with mass relationships is applied in this study and enables the generation of regression coefficients of gravity model based on programmed large-scale regression simulations. A series of accuracy assessments are then developed for 2015 empirical projection daily regression coefficients that can be applied to Chinese spring interprovincial mobile gravity model variables as well as spatiotemporal research that utilizes regression coefficients within a heterogeneity research model. This approach also enables the error within the gravity model to be assessed in terms of floating population simulations. The results of this analysis lead to a number of clear conclusions, including the fact that parameter calibration complexity for the Chinese population mobility gravity model is reflected in the degree of influence asymmetry within spatial object interaction variables, and that the spatial heterogeneity of the variable regression coefficient increases in two distinct fashions. The first of these increases has to do with the overall influence of specific variables, including the fact that differences between proxies tend to be higher than inflow-outflow characteristics. In contrast, the second set of increases is related to economic levels, industrial scales, the proportion of the tertiary industry, and public service facilities. In this latter case, two-way population flow exerts a more profound influence on results and thus the scope of possible explanations for phenomena is more extensive. The regression coefficient for the existence of positive and negative proxy variables therefore relates to differences in spatial heterogeneity, including at the city level, and also assumes that floating population gravity model regression coefficients ignore spatiotemporal changes in the heterogeneity coefficient. This leads to spatial differences in estimated results and thus convergence trends, but further enables the identification of anisotropic interactions in extension space. The second main conclusion of this research is that the national scale population flow distance attenuation coefficient was 1.970 during the spring of 2015, while at the level of prefectural administrative units and given population outflow, the range encapsulated by this coefficient fell between 0.712 (Zhumadian) and 7.699 (Urumqi). Data also reveal a population inflow coefficient of 0.792 for this year that ranged as high as 8.223 in both Sanya and Urumqi. Population flow simulation results using the gravity model and including Baidu migration measured flow data were also subject to significant error. Third, the results of this analysis reveal a total fitting error of 85.54% in weighted absolute mean; the spatial interaction effect within this is responsible for a maximum error of 86.09% in actual and simulated flows, while relative outflow force and attractiveness encompass 57.73% and 49.34% of model error, respectively. These results show that the spatial interaction effect remains most difficult to model in terms of current factors.

Key words: gravity model, regression coefficient, distance attenuation coefficient, error estimation, population flow, China