WO2024031520A1 - 一种基于生成对抗网络的人群移动预测方法 - Google Patents

一种基于生成对抗网络的人群移动预测方法 Download PDF

Info

Publication number
WO2024031520A1
WO2024031520A1 PCT/CN2022/111718 CN2022111718W WO2024031520A1 WO 2024031520 A1 WO2024031520 A1 WO 2024031520A1 CN 2022111718 W CN2022111718 W CN 2022111718W WO 2024031520 A1 WO2024031520 A1 WO 2024031520A1
Authority
WO
WIPO (PCT)
Prior art keywords
epidemic
crowd movement
during
crowd
module
Prior art date
Application number
PCT/CN2022/111718
Other languages
English (en)
French (fr)
Inventor
李超
李可汉
陈积明
贺诗波
杨秦敏
Original Assignee
浙江大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 浙江大学 filed Critical 浙江大学
Priority to KR1020237020703A priority Critical patent/KR102615650B1/ko
Publication of WO2024031520A1 publication Critical patent/WO2024031520A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]

Definitions

  • the invention belongs to the field of crowd movement prediction, and specifically relates to a crowd movement prediction method based on a generative adversarial network.
  • the crowd movement prediction method during the epidemic can help understand how the crowd movement pattern changes during the COVID-19 epidemic, and particularly reveals how the government Released travel restrictions and related statistics (such as confirmed cases) have dominated the movement of people during the COVID-19 pandemic. And accurately predict the level of crowd movement.
  • the current model paradigm for predicting urban crowd movement based on deep learning models is mainly trained based on rich historical data. Then, given observation data for a period of time in current and historical time, the model can output deterministic prediction results.
  • this method has certain limitations.
  • the deterministic prediction model cannot generate probability estimates of crowd mobility; in addition, the generated results are relatively fixed, making it inconvenient to flexibly change some conditions for multi-dimensional simulation.
  • government decision-makers need more flexible models, such as generative models, which can learn the distribution of data and simulate a variety of potential crowd flow patterns under different policies.
  • the government can based on the situation during the epidemic Simulate crowd movement response outcomes to develop a phased reopening plan. Therefore, the applicant proposes a conditional generative adversarial network for modeling the complex dynamics between newly confirmed cases, policies, and crowd movements.
  • the purpose of the present invention is to improve and standardize the shortcomings of existing research and technology, and propose a crowd movement prediction method based on generative adversarial networks.
  • This method accurately predicts the changing trend of crowd movement by modeling the impact of different policies and epidemic conditions during the epidemic, which is more practical. It is also conducive to analyzing potential policy impacts through the results of data prediction, and expanding the application of the method. High sex.
  • a crowd movement prediction method based on generative adversarial networks including the following steps:
  • Step 1 Divide a city into H ⁇ W grids of equal area, each grid representing an area in the city;
  • Step 2 Divide the areas in step 1 and count the crowd movement levels m in different areas;
  • Step 3 Use the regional crowd movement levels counted in step 2 to obtain the crowd movement heat map M ⁇ R H ⁇ W composed of an area in the city. Each element in the matrix represents the crowd movement level in the corresponding area;
  • Step 4 Collect daily statistical data and related policies in different regions during the epidemic period, obtain the daily newly confirmed cases C as representative statistical data during the epidemic period, obtain the changes and intensity of daily policies, and record the intensity variables of these policies.
  • P
  • Step 5 based on the above analysis, for a specific city, given the historical and current crowd movement heat map ⁇ M t-1 , M t ⁇ and the corresponding new coronavirus epidemic statistics ⁇ C t-1 , C t , C t +1 ⁇ and policy ⁇ P t-1 , P t , ⁇ , predict the crowd movement level heat map ⁇ M t+1 ⁇ in the future period.
  • the crowd movement prediction model during the epidemic period is used to predict the rules of population movement during the epidemic period, and a heat map of crowd movement levels in the future period is generated.
  • the crowd movement prediction model during the epidemic period includes a generator module and a discriminator module. And domain knowledge fusion module;
  • the generator module predicts crowd movement patterns in the future based on historical crowd movement data in the past period
  • the discriminating module is used to predict the labels of the human flow map and determine whether the generated population flow heat map is consistent with the real distribution;
  • the domain knowledge fusion module is used to integrate the influence of external factors during the epidemic.
  • the generator module models the response of population flow intensity in different regions to policy changes during the epidemic by modeling the change values of population flow between different time steps.
  • the input of the generator module can be expressed as two consecutive times. Population movement level changes between segments:
  • a transformer-based encoder module was introduced into the crowd movement prediction model during the epidemic to model long-distance spatiotemporal correlation, and a multi-head self-attention mechanism module was introduced to extract feature maps.
  • the features processed by the transformer are recorded as And the output features and external condition features are spliced together and sent to the decoder to output the predicted population flow results.
  • the domain knowledge fusion module takes policies and epidemic statistical data during the epidemic as conditions and integrates them with spatiotemporal population movement characteristics. Specifically, it includes: introducing a fully connected neural network to first convert different types of domain knowledge into hidden variables. Then a gated fusion network module is introduced to activate the spatiotemporal features of different regions:
  • the crowd movement prediction model during the epidemic introduces a noise vector into the working space and time.
  • the spatio-temporal feature vector is spliced on the feature dimension, and finally the crowd movement prediction model during the epidemic introduces a cross-model connection to obtain an estimate of the population mobility level for the next time step:
  • a loss function with a mask matrix can be calculated from the generator module:
  • the loss function of the discriminator module is obtained:
  • the daily policy includes one or more indicators among travel restrictions and blockade policies, economic policies, and health system policies.
  • This invention adopts a progressive learning method and can handle the rapidly changing population flow data during the epidemic;
  • This invention introduces a domain knowledge fusion module, which can well model the impact of policies and epidemic development on changes in population mobility;
  • the generator module and the discriminator module constitute a generative adversarial network to model the uncertainty in the population flow process.
  • Figure 1 is a flow chart of the present invention
  • Figure 2 is a schematic structural diagram of the crowd movement prediction model during the epidemic period of the present invention.
  • a crowd movement prediction method based on generative adversarial networks includes the following steps:
  • Step 1 Divide a city into H ⁇ W grids of equal area, each grid representing an area in the city;
  • Step 2 Divide the areas in step 1 and count the crowd movement levels m in different areas;
  • Step 3 Use the regional crowd movement levels counted in step 2 to obtain the crowd movement heat map M ⁇ R H ⁇ W composed of an area in the city. Each element in the matrix represents the crowd movement level in the corresponding area;
  • Step 4 Collect daily statistical data and related policies in different regions during the epidemic period, obtain the daily newly confirmed cases C as representative statistical data during the epidemic period, obtain the changes and intensity of daily policies, and record the intensity variables of these policies.
  • P
  • Step 5 based on the above analysis, for a specific city, given the historical and current crowd movement heat map ⁇ M t-1 , M t ⁇ and the corresponding new coronavirus epidemic statistics ⁇ C t-1 , C t , C t +1 ⁇ (A large number of articles have achieved accurate real-time prediction of new confirmed cases) and policies ⁇ P t-1 , P t , ⁇ (can be formulated by policymakers in advance), predicting the level of crowd movement in the future period.
  • the crowd movement prediction model during the epidemic period is used to predict the pattern of population movement during the epidemic period and generate a heat map of crowd movement levels in the future.
  • the crowd movement prediction model during the epidemic period includes a generator module, a discriminator module generator module and domain knowledge fusion module; the generator module predicts crowd movement patterns in the future based on historical crowd movement data in the past period; the discriminating module is used to predict the labels of human flow diagrams and determine the generated population flow Whether the heat map is consistent with the real distribution; the domain knowledge fusion module is used to integrate the influence of external factors during the epidemic.
  • the generator module and the discriminator module constitute a generative adversarial network, which is used to model the uncertainty in the population movement process.
  • the present invention proposes a model based on a deep generative network for predicting the rules of population movement during the epidemic.
  • Generative models can learn the distribution of data and simulate a variety of potential crowd movement patterns under different policies.
  • the government can formulate phased reopening plans based on the simulated crowd movement response results during the epidemic.
  • This invention proposes a conditional generative adversarial network for complex dynamic modeling between newly confirmed cases, policies and crowd flows. Different from traditional models that model long-term temporal dependencies, the model focuses on predicting crowd mobility transitions between adjacent time periods formed by potential epidemic-related statistics, policies, and the latest developments in crowd mobility. .
  • the present invention designs a Policy-Human Mobility Interplay Network (PHMIN) model to estimate the mobility changes of the crowd.
  • the model s conditional inputs mainly come from statistical data and policies during the epidemic.
  • the model can flexibly learn fine-grained crowd movement dynamics and accurately extend to the prediction of multi-wave epidemics across cities.
  • the structure of the model is shown in Figure 1.
  • the generator module efficiently models the response of population flow intensity in different regions to policy changes during the epidemic by modeling the change value of population flow between different time steps.
  • a transformer-based encoder module is introduced here to model long-distance spatiotemporal correlation and introduce multi-head self-attention.
  • Force mechanism module extract feature map The features processed by the transformer are recorded as And the output features and external condition features are spliced together and sent to the decoder to output the predicted population flow results.
  • the domain knowledge fusion module improves the accuracy of prediction by taking policies and epidemic statistics during the epidemic as conditions and integrating them with spatiotemporal population movement characteristics. Specifically, it includes: introducing a fully connected neural network to first transform different types of domain knowledge into hidden variables Then a gated fusion network module is introduced to activate the spatiotemporal features of different regions:
  • the crowd movement prediction model during the epidemic introduces a noise vector into the working space and time.
  • the spatio-temporal feature vector is spliced on the feature dimension, and finally the crowd movement prediction model during the epidemic introduces a cross-model connection to obtain an estimate of the population mobility level for the next time step:
  • the loss function with a mask matrix can be calculated from the generator module:
  • the loss function of the discriminator module can be obtained as:
  • the daily policy includes one or more indicators among travel restrictions and blockade policies, economic policies, and health system policies.
  • This embodiment uses the above crowd movement prediction method during the epidemic to predict the crowd movement during the epidemic in three cities: Beijing, Dalian, and Shijiazhuang.
  • BJ represents Beijing
  • DL represents Dalian
  • SJZ represents Shijiazhuang.
  • the following serial numbers indicate different waves.
  • the PHMIN model achieves the best performance, which proves that the model has good generalization ability in cross-city crowd flow prediction scenarios.
  • the relatively simple HA and ARIMA models show better performance than DeepST and cGAN, which to a certain extent means that the most valuable information in the problem is related to the latest situation of crowd flow.
  • cGAN is slightly better than DeepST because cGAN is able to model unknown factors of crowd flow in a generative manner.
  • DeepST gives deterministic predictions based on rich historical observations, which is not suitable in epidemic scenarios due to the lack of historical data.
  • the similarity of urban geographical distribution may affect the performance of model predictions across cities. It is found from the experimental results that models trained based on DL1 usually lead to worse performance. Taking the results on DL1-BJ1, DL1-SJZ1 and DL1-SJZ2 as an example, their results are not much different from traditional ARIMA (MAE is more Good, MAPE is worse). In fact, the urban area of Dalian is surrounded by the sea and its geographical distribution is not very regular, while Beijing and Shijiazhuang have urban areas with a very regular, checkerboard-like layout. It can be inferred that the gap in spatial distribution may make it more difficult to generalize the model across cities.
  • the PHMIN model still performs better compared to the baseline model.
  • the overall performance predicted by the training set with different intensities is worse than that of the model obtained from similar intensities in Table 6.1. This shows that compared with different urban layouts, the intensity of the epidemic is a more critical indicator that affects the generalization ability of the model.
  • Models trained based on Dalian city data still perform poorly on other cities (e.g., DL1-BJ2 and DL2-SJZ2), which again illustrates the importance of similar geographical distribution of cities.
  • the dual differences in geographical distribution and epidemic intensity cause the proposed model to have worse prediction results than the baseline method ARIMA in the DL2-SJZ2 scenario.
  • the PHMIN model is better than the baseline model in all experiments.
  • the three groups of experiments SJZ1-SJZ2, DL2-DL3, and BJ1-BJ2 achieved better results compared with the results in Tables 6.1 and 6.2, which means The negative impact of disparities within the same city may be smaller than the impact of epidemic intensity.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于生成对抗网络的人群移动预测方法，以新冠肺炎(COVID-19)疫情期间城市人群流动预测为研究对象，研究多模态数据的时空特征融合的问题。由于复杂的社会背景、政策和疫情状况等复杂多模态数据的影响，使得预测疫情期间人群移动的模式变得十分困难。基于生成对抗网络的人群移动预测方法分析了国内三个城市的人群移动数据，发现尽管不同的城市之间有着很大的差距，但是它们在新冠肺炎疫情期间的人群移动模式呈现出很高的相似性。基于生成对抗网络的人群移动预测方法设计了基于条件生成对抗网络的预测模型，模型中融合建模了多模态数据对于人群移动时空特征的影响。基于条件生成对抗网络的预测模型还可以帮助政府更好地评估不同政策对人群移动性的潜在影响，优化政策制定。

Description

一种基于生成对抗网络的人群移动预测方法 技术领域
本发明属于人群移动预测领域，具体涉及涉及一种基于生成对抗网络的人群移动预测方法，该疫情期间人群移动预测方法能够帮助理解COVID-19疫情期间人群移动模式如何变化，特别揭示了疫情期间政府发布的限制出行政策和相关统计数据(如确诊病例)在新冠肺炎大流行期间主导了人群移动的规律。并精准的进行了人群移动水平的预测。
背景技术
理解COVID-19疫情期间人群移动模式如何变化对于控制流行病的传播非常重要。考虑到复杂的社会背景、个人行为的差异和极其有限的数据等因素的影响，疫情期间人群移动的模式似乎变的不可预测。
如果无法精准预测人群的移动规律，很难制定合理的政策，而过于严格的防疫政策会对经济产生较大的影响，太松散的防疫政策又会导致疫情难以控制，影响人民的生命健康。因此需要精准的预测来提供准确的政策制定参考。
当前基于深度学习模型解决城市人群移动预测的模型范式，主要是基于丰富历史数据进行训练，之后给定当前和历史时间上一段时间的观测数据，模型能够输出确定性预测结果。但是这种方式存在一定的局限性，首先，确定性的预测模型无法生成人群流动性的概率估计；此外，生成的结果比较固定，不方便灵活的改变一些条件进行多维度的仿真。新冠肺炎疫情期间，政府决策者的决策者需要更为灵活的模型，如生成式模型，它能够学习数据的分布并模拟不同政策下的多种潜在的人群流动规律，政府可以基于疫情期间的的模拟人群移动响应结果，制定分阶段的重新开放计划。因此，申请人提出了一个条件生成对抗网络，用于新增确诊病例、政策和人群移动之间的复杂动态建模。
发明内容
本发明目的在于对现有研究和技术存在的不足之处加以完善与规范化，提出一种基于生成对抗网络的人群移动预测方法。该方法通过建模疫情期间不同的政策和疫情状况的影响，来准确的预测人群移动的变化趋势，更具有实用价值；且有利于通过数据预测的结果对潜在政策影响进行分析，方法应用的扩展性高。
本发明的目的通过以下的技术方案实现：
一种基于生成对抗网络的人群移动预测方法，包括以下步骤：
步骤1，把一个城市划分为H×W个等面积的网格，每个网格代表城市中的一个区域；
步骤2，对步骤1中的区域划分，分别统计不同区域的人群移动水平m；
步骤3，利用步骤2中统计的区域人群移动水平得出城市中一片区域构成的人群移动热图M∈R H×W，矩阵中每个元素代表了相应区域的人群移动水平；
步骤4，收集疫情期间不同地区的每日统计数据和相关政策，得到每日新增确诊病例C作为疫情期间代表性的统计数据，获取每日政策的变化和强度，将这些政策的强度变量记为P；
步骤5，基于以上的分析对于特定的城市，给定历史和当前的人群移动热图{M t-1，M t}和相应的新冠肺炎疫情统计数据{C t-1，C t，C t+1}以及政策{P t-1，P t，}，预测未来一段时间的人群移动水平热图{M t+1}。
进一步地，所述步骤5中，通过疫情期间人群移动预测模型预测疫情期间人口流动的规律，生成未来一段时间的人群移动水平热图，该疫情期间人群移动预测模型包括生成器模块、判别器模块以及领域知识融合模块；
所述生成器模块基于过去一段时间的历史人群移动数据预测未来一段时间的人群移动规律；
所述判别起模块用于预测人类流动图的标签，判断生成的人口流动热图是否与真实分布一致；
所述领域知识融合模块用于融合疫情期间外部因素的影响。
进一步地，所述生成器模块通过建模不同时间步之间的人口流动变化值，来建模疫情期间不同区域人口流动强度对于政策变化的响应，生成器模块的输入可以表示为连续两个时间片段之间的人口移动水平变化：
ΔM t-1＝M t-M t-1。
进一步地，所述疫情期间人群移动预测模型中引入基于transformer的编码器模块，来建模长距离的时空相关性，引入多头自注意力机制模块，提取得到特征图
Figure PCTCN2022111718-appb-000001
经过transformer处理的特征记为
Figure PCTCN2022111718-appb-000002
并将输出的特征与外部条件特征拼接在一起输送到解码器输出预测的人口流动结果。
进一步地，所述领域知识融合模块将疫情期间的政策、疫情统计数据作为条件，和时空人口移动特征融合，具体包括：引入一个全连接的神经网络，将不同种类的领域知识首先转化为隐变量
Figure PCTCN2022111718-appb-000003
接着引入一个门控融合网络模块去激活不同区域的时空特征：
Figure PCTCN2022111718-appb-000004
疫情期间人群移动预测模型在工作的时空同时引入一个噪声向量
Figure PCTCN2022111718-appb-000005
和时空特征向量在特征维上拼接，最后疫情期间人群移动预测模型引入一个跨模型连接，得到对于下一个时间步的人口流动水平估计：
Figure PCTCN2022111718-appb-000006
进一步地，所述疫情期间人群移动预测模型引入一个掩码矩阵
Figure PCTCN2022111718-appb-000007
来减少缺少采样的区域带来的影响，可以从生成器模块中计算出带有掩码矩阵的损失函数：
Figure PCTCN2022111718-appb-000008
得到判别器模块的损失函数为：
Figure PCTCN2022111718-appb-000009
最终合并生成器和判别器的损失函数得到：
Figure PCTCN2022111718-appb-000010
通过训练模型达到生成器和判别器损失函数的的鞍点，表明模型训练完成。
进一步地，所述每日政策包括出行限制和封锁政策、经济政策、卫生系统政策中的一种或多种指标。
与现有技术相比，本发明的有益效果为：
1)本发明采用了渐进式的学习方式，能够处理快速变化的疫情期间人口流动数据；
2)本发明引入了领域知识融合模块，能够很好的建模政策和疫情发展状况对人口流动变化的影响；
3)本发明中生成器模块和判别器模块构成了生成对抗网络，来建模人口流动过程中的不确定性。
附图说明
图1为本发明流程图；
图2为本发明中疫情期间人群移动预测模型结构示意图。
具体实施方式
在本发明的描述中，需要理解的是，术语“一端”、“另一端”、“外侧”、“上”、“内侧”、“水平”、“同轴”、“中央”、“端部”、“长度”、“外端”等指示的方位或位置关系为基于附图所示的方位或位置关系，仅是为了便于描述本发明和简化描述，而不是指示或暗示所指的装置或元件必须具有特定的方 位、以特定的方位构造和操作，因此不能理解为对本发明的限制。
下面结合附图对本发明作进一步说明。
请参阅图1，一种基于生成对抗网络的人群移动预测方法，该方法包括以下步骤：
步骤1，把一个城市划分为H×W个等面积的网格，每个网格代表城市中的一个区域；
步骤2，对步骤1中的区域划分，分别统计不同区域的人群移动水平m；
步骤3，利用步骤2中统计的区域人群移动水平得出城市中一片区域构成的人群移动热图M∈R H×W，矩阵中每个元素代表了相应区域的人群移动水平；
步骤4，收集疫情期间不同地区的每日统计数据和相关政策，得到每日新增确诊病例C作为疫情期间代表性的统计数据，获取每日政策的变化和强度，将这些政策的强度变量记为P；
步骤5，基于以上的分析对于特定的城市，给定历史和当前的人群移动热图{M t-1，M t}和相应的新冠肺炎疫情统计数据{C t-1，C t，C t+1}(大量文章已经实现了对新增确诊病例的准确实时预测)以及政策{P t-1，P t，}(可以由政策制定者提前制定)，预测未来一段时间的人群移动水平热图{M t+1}。
请参阅图2，所述步骤5中，通过疫情期间人群移动预测模型预测疫情期间人口流动的规律，生成未来一段时间的人群移动水平热图，该疫情期间人群移动预测模型包括生成器模块、判别器模块以及领域知识融合模块；所述生成器模块基于过去一段时间的历史人群移动数据预测未来一段时间的人群移动规律；所述判别起模块用于预测人类流动图的标签，判断生成的人口流动热图是否与真实分布一致；所述领域知识融合模块用于融合疫情期间外部因素的影响。
其中，生成器模块和判别器模块构成生成对抗网络，用来建模人口流动过程中的不确定性。
结合疫情期间政策和新增确诊人数等统计情况进行人群移动估计的方法。本发明提出了一种基于深度生成网络的模型用于预测疫情期间人口流动的规律。生成式模型能够学习数据的分布并模拟不同政策下的多种潜在的人群流动规律，政府可以基于疫情期间的的模拟人群移动响应结果，制定分阶段的重新开放计划。本发明提出了一个条件生成对抗网络，用于新增确诊病例、政策和人群流动之间的复杂动态建模。模型不同于传统的模型对长期的时序依赖关系进行建模，而是专注于预测由潜在的疫情相关统计数据、政策和人群流动性的最新动态形成的相邻时间段之间的人群流动性转变。具体来说，本发明设计了一个Policy-Human Mobility Interplay Network(PHMIN)模型来估计人群的移动性变化。在这里，模型的条件输入主要来自于疫情期间的统计数据和政策。模型可以灵活地学习细粒度的人群移动动态，并准确地扩展到跨城市的多波疫情的预测，模型的结构如图1所示。
进一步地，所述生成器模块通过建模不同时间步之间的人口流动变化值，来高效的地建模疫情期间不同区域人口流动强度对于政策变化的响应，生成器模块的输入可以表示为连续两个时间片段之间的人口移动水平变化：ΔM t-1＝M t-M t-1。
进一步地，所述疫情期间人群移动预测模型中另一方面，考虑到多尺度的时空人口移动特征，在这里引入基于transformer的编码器模块，来建模长距离的时空相关性，引入多头自注意力机制模块，提取得到特征图
Figure PCTCN2022111718-appb-000011
经过transformer处理的特征记为
Figure PCTCN2022111718-appb-000012
并将输出的特征与外部条件特征拼接在一起输送到解码器输出预测的人口流动结果。
进一步地，所述领域知识融合模块通过将疫情期间的政策、疫情统计数据作为条件，和时空人口移动特征融合，来提高预测的精度。具体包括：引入一个全 连接的神经网络，将不同种类的领域知识首先转化为隐变量
Figure PCTCN2022111718-appb-000013
接着引入一个门控融合网络模块去激活不同区域的时空特征：
Figure PCTCN2022111718-appb-000014
同时考虑到人口流动预测过程中的不确定性，疫情期间人群移动预测模型在工作的时空同时引入一个噪声向量
Figure PCTCN2022111718-appb-000015
和时空特征向量在特征维上拼接，最后疫情期间人群移动预测模型引入一个跨模型连接，得到对于下一个时间步的人口流动水平估计：
Figure PCTCN2022111718-appb-000016
进一步地，所述疫情期间人群移动预测模型优化模型的过程中引入一个掩码矩阵
Figure PCTCN2022111718-appb-000017
来减少缺少采样的区域带来的影响，因此可以从生成器模块中计算出带有掩码矩阵的损失函数：
Figure PCTCN2022111718-appb-000018
更进一步可以得到判别器模块的损失函数为：
Figure PCTCN2022111718-appb-000019
最终合并生成器和判别器的损失函数得到：
Figure PCTCN2022111718-appb-000020
通过训练模型达到生成器和判别器损失函数的的鞍点，表明模型训练完成。
进一步地，所述每日政策包括出行限制和封锁政策、经济政策、卫生系统政策中的一种或多种指标。
实施例
本实施例采用上述疫情期间人群移动预测方法对北京、大连、石家庄三座城市的疫情期间人群移动情况进行预测，其中BJ代表北京，DL代表大连，SJZ代表石家庄，后面的序号表明不同波次的疫情期间人口移动数据。考虑到上述三个城 市共经历了七波新冠肺炎疫情，本实施例根据实际的情况把疫情划分为两种强度，进而得到三组实验：1)不同城市间相似强度下的实验；2)不同城市不同强度的实验；3)同一个城市相似或者不同强度场景下的实验的结果进行分层分析。这些实验的设计旨在回答一个基本问题：模型是否可以在不同的环境中有效的泛化，用于预测下一波疫情到来时政策对人群流动的影响。下文将详细地讨论这个问题。
表6.1：跨城市相似疫情强度预测实验
Figure PCTCN2022111718-appb-000021
不同城市相似强度：根据表6.1中显示的实验结果，可以得到以下几个结论：
总体上来说，与基线模型相比，PHMIN模型达到了最好的性能，这证明了模型在跨城市的人群流动预测场景下具有良好泛化能力。具体来说，相对简单的HA和ARIMA模型表现出比DeepST和cGAN更好的性能，这在一定程度上意味着问题中最有价值的信息与人群流动的最新情况相关。此外，cGAN略优于DeepST，因为cGAN能够以生成方式对人群流动的未知因素进行建模。另一方面，DeepST给出的是基于丰富历史观察的确定性预测，在疫情场景中由于缺少历史数据不太适合。
给定相同的测试城市作为预测目标，发现基于不同城市训练集得到模型会有 不同的表现。在DL1、SJZ1和SJZ2上进行测试时，基于BJ1训练的模型明显显示出比其他城市更好的结果。这可以归因于北京的数据质量更高，有着更多的可用样本和较高的人群密度。
城市地理分布相似性可能会影响模型跨城市预测的效果。从实验结果中发现，基于DL1训练的模型通常会导致更差的性能，以DL1-BJ1、DL1-SJZ1和DL1-SJZ2上的结果为例，它们的结果与传统的ARIMA相差不大(MAE更好，MAPE更差)。事实上，大连市区四面环海，地理上的分布也不是很规则，而北京和石家庄则有着非常规则的、类似棋盘式的布局的市区。据此可以推测出空间分布的差距可能会使模型在城市间的泛化变得更加困难。
表6.2：跨城市不同强度预测实验
Figure PCTCN2022111718-appb-000022
不同城市不同强度：为了研究模型在不同强度疫情之间的泛化表现，进一步设计并进行了表6.2中所示的实验，实验的结果说明了：
在大多数情况下，与基线模型相比，PHMIN模型仍然表现更好。然而，对于同一个测试城市，不同强度的训练集预测的整体性能比从表6.1中相似强度得到的模型要差一些。这说明与不同的城市布局相比，疫情的强度是影响模型泛化能 力更关键的指标。
基于大连市的数据训练的模型在其他城市上仍然表现不佳(例如，DL1-BJ2和DL2-SJZ2)，这再次说明了城市地理分布相似的重要性。此外，地理分布和疫情强度的双重差异导致提出的模型在DL2-SJZ2场景下的预测结果比基线方法ARIMA更差。
表6.3：同城市相似或者不同强度预测实验
Figure PCTCN2022111718-appb-000023
同城相似或者不同强度：直观地来说，在同一个城市进行实验，模型应该更容易泛化取得更好的结果。但是表6.3中的实验数据揭示了一些不同的结果：
首先，PHMIN模型在所有的实验中都优于基线模型，此外，SJZ1-SJZ2、DL2-DL3、BJ1-BJ2三组实验与表6.1和6.2中的结果相比达到了更好的效果，这意味着同一个城市内的差距带来的负面影响可能还要小于疫情强度的作用。
但是在DL1-DL2和DL1-DL3两组实验中结果出现了异常，其结果分别比BJ2-DL2和BJ2-DL3还要差。这可以归因于BJ2、DL2和DL3，三个数据集尽管在不同的城市，但是经历的疫情强度相似。比较DL2-DL3和DL1-DL3之间的结果，可以推断出提出的模型在同一城市内的相似疫情强度的场景进行人群流动预测时，可以表现出更好的泛化性能。
最后应说明的是：以上各实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述各实施例对本发明进行了详细的说明，本领域的普通技术人员 应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分或者全部技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的范围。

Claims (7)

  1. 一种基于生成对抗网络的人群移动预测方法，包括以下步骤：
    步骤1，把一个城市划分为H×W个等面积的网格，每个网格代表城市中的一个区域；
    步骤2，对步骤1中的区域划分，分别统计不同区域的人群移动水平m；
    步骤3，利用步骤2中统计的区域人群移动水平得出城市中一片区域构成的人群移动热图M∈R H×W，矩阵中每个元素代表了相应区域的人群移动水平；
    步骤4，收集疫情期间不同地区的每日统计数据和相关政策，得到每日新增确诊病例C作为疫情期间代表性的统计数据，获取每日政策的变化和强度，将这些政策的强度变量记为P；
    步骤5，对于特定的城市，给定历史和当前的人群移动热图{M t-1，M t}和相应的新冠肺炎疫情统计数据{C t-1，C t，C t+1}以及政策{P t-1，P t，}，预测未来一段时间的人群移动水平热图{M t+1}。
  2. 根据权利要求1所述的一种基于生成对抗网络的人群移动预测方法，其特征在于，所述步骤5中，通过疫情期间人群移动预测模型预测疫情期间人口流动的规律，生成未来一段时间的人群移动水平热图，该疫情期间人群移动预测模型包括生成器模块、判别器模块以及领域知识融合模块；
    所述生成器模块基于过去一段时间的历史人群移动数据预测未来一段时间的人群移动规律；
    所述判别起模块用于预测人类流动图的标签，判断生成的人口流动热图是否与真实分布一致；
    所述领域知识融合模块用于融合疫情期间外部因素的影响。
  3. 根据权利要求2所述的一种基于生成对抗网络的人群移动预测方法，其特征在于，所述生成器模块通过建模不同时间步之间的人口流动变化值，来建模疫情期间不同区域人口流动强度对于政策变化的响应，生成器模块的输入可以表示为连续两个时间片段之间的人口移动水平变化：
    ΔM t-1＝M t-M t-1。
  4. 根据权利要求2所述的一种基于生成对抗网络的人群移动预测方法，其特征在于，所述疫情期间人群移动预测模型中引入基于transformer的编码器模块，来建模长距离的时空相关性，引入多头自注意力机制模块，提取得到特征图
    Figure PCTCN2022111718-appb-100001
    经过transformer处理的特征记为
    Figure PCTCN2022111718-appb-100002
    并将输出的特征与外部条件特征拼接在一起输送到解码器输出预测的人口流动结果。
  5. 根据权利要求2所述的一种基于生成对抗网络的人群移动预测方法，其特征在于，所述领域知识融合模块将疫情期间的政策、疫情统计数据作为条件，和时空人口移动特征融合，具体包括：引入一个全连接的神经网络，将不同种类的领域知识首先转化为隐变量
    Figure PCTCN2022111718-appb-100003
    接着引入一个门控融合网络模块去激活不同区域的时空特征：
    Figure PCTCN2022111718-appb-100004
    疫情期间人群移动预测模型在工作的时空同时引入一个噪声向量
    Figure PCTCN2022111718-appb-100005
    和时空特征向量在特征维上拼接，最后疫情期间人群移动预测模型引入一个跨模型连接，得到对于下一个时间步的人口流动水平估计：
    Figure PCTCN2022111718-appb-100006
  6. 根据权利要求2所述的一种基于生成对抗网络的人群移动预测方法，其特征在于，所述疫情期间人群移动预测模型引入一个掩码矩阵
    Figure PCTCN2022111718-appb-100007
    来减少缺少采样的区域带来的影响，可以从生成器模块中计算出带有掩码矩阵的损失函数：
    Figure PCTCN2022111718-appb-100008
    得到判别器模块的损失函数为：
    Figure PCTCN2022111718-appb-100009
    最终合并生成器和判别器的损失函数得到：
    Figure PCTCN2022111718-appb-100010
    通过训练模型达到生成器和判别器损失函数的的鞍点，表明模型训练完成。
  7. 根据权利要求1所述的一种基于生成对抗网络的人群移动预测方法，其特征在于，所述每日政策包括出行限制和封锁政策、经济政策、卫生系统政策中的一种或多种指标。
PCT/CN2022/111718 2022-08-09 2022-08-11 一种基于生成对抗网络的人群移动预测方法 WO2024031520A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
KR1020237020703A KR102615650B1 (ko) 2022-08-09 2022-08-11 생성적 적대 신경망 기반의 군중 이동 예측 방법

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210948304.5A CN115238589A (zh) 2022-08-09 2022-08-09 一种基于生成对抗网络的人群移动预测方法
CN202210948304.5 2022-08-09

Publications (1)

Publication Number Publication Date
WO2024031520A1 true WO2024031520A1 (zh) 2024-02-15

Family

ID=83680178

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/111718 WO2024031520A1 (zh) 2022-08-09 2022-08-11 一种基于生成对抗网络的人群移动预测方法

Country Status (2)

Country Link
CN (1) CN115238589A (zh)
WO (1) WO2024031520A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115983142B (zh) * 2023-03-21 2023-08-29 之江实验室 基于深度生成对抗式网络的区域人口演化模型构造方法
CN116258652B (zh) * 2023-05-11 2023-07-21 四川大学 基于结构注意和文本感知的文本图像修复模型及方法

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170103172A1 (en) * 2015-10-07 2017-04-13 The Arizona Board Of Regents On Behalf Of The University Of Arizona System And Method To Geospatially And Temporally Predict A Propagation Event
CN110993119A (zh) * 2020-03-04 2020-04-10 同盾控股有限公司 基于人口迁移的疫情预测方法、装置、电子设备及介质
CN111612206A (zh) * 2020-03-30 2020-09-01 清华大学 一种基于时空图卷积神经网络的街区人流预测方法及系统
CN111723693A (zh) * 2020-06-03 2020-09-29 云南大学 一种基于小样本学习的人群计数方法
CN111798991A (zh) * 2020-07-09 2020-10-20 重庆邮电大学 基于lstm的新冠肺炎疫情群体态势预测方法
CN111916218A (zh) * 2020-08-10 2020-11-10 重庆邮电大学 一种新型冠状病毒肺炎疫情多级预警方法及系统
CN112257934A (zh) * 2020-10-26 2021-01-22 辽宁工程技术大学 一种基于时空动态神经网络的城市人流预测方法
CN112309584A (zh) * 2020-11-19 2021-02-02 天津大学 基于疫情因素预测疫情期间城市迁徙规模的模型构建方法
CN114464329A (zh) * 2021-12-31 2022-05-10 中国科学院深圳先进技术研究院 一种城市疫情时空预测方法、系统、终端以及存储介质

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170103172A1 (en) * 2015-10-07 2017-04-13 The Arizona Board Of Regents On Behalf Of The University Of Arizona System And Method To Geospatially And Temporally Predict A Propagation Event
CN110993119A (zh) * 2020-03-04 2020-04-10 同盾控股有限公司 基于人口迁移的疫情预测方法、装置、电子设备及介质
CN111612206A (zh) * 2020-03-30 2020-09-01 清华大学 一种基于时空图卷积神经网络的街区人流预测方法及系统
CN111723693A (zh) * 2020-06-03 2020-09-29 云南大学 一种基于小样本学习的人群计数方法
CN111798991A (zh) * 2020-07-09 2020-10-20 重庆邮电大学 基于lstm的新冠肺炎疫情群体态势预测方法
CN111916218A (zh) * 2020-08-10 2020-11-10 重庆邮电大学 一种新型冠状病毒肺炎疫情多级预警方法及系统
CN112257934A (zh) * 2020-10-26 2021-01-22 辽宁工程技术大学 一种基于时空动态神经网络的城市人流预测方法
CN112309584A (zh) * 2020-11-19 2021-02-02 天津大学 基于疫情因素预测疫情期间城市迁徙规模的模型构建方法
CN114464329A (zh) * 2021-12-31 2022-05-10 中国科学院深圳先进技术研究院 一种城市疫情时空预测方法、系统、终端以及存储介质

Also Published As

Publication number Publication date
CN115238589A (zh) 2022-10-25

Similar Documents

Publication Publication Date Title
WO2024031520A1 (zh) 一种基于生成对抗网络的人群移动预测方法
CN109508360B (zh) 一种基于元胞自动机的地理多元流数据时空自相关分析方法
WO2023123625A1 (zh) 一种城市疫情时空预测方法、系统、终端以及存储介质
Liu et al. Smart environment design planning for smart city based on deep learning
WO2022257190A1 (zh) 一种基于量子游走的行为轨迹序列多特征模拟方法
Xu et al. A graph spatial-temporal model for predicting population density of key areas
Zhang et al. Investigating time, strength, and duration of measures in controlling the spread of COVID-19 using a networked meta-population model
Dai et al. Spatio-temporal deep learning framework for traffic speed forecasting in IoT
Qu et al. Forecasting fine-grained urban flows via spatio-temporal contrastive self-supervision
Bao et al. Covid-gan+: Estimating human mobility responses to covid-19 through spatio-temporal generative adversarial networks with enhanced features
Gong et al. CARESim: An integrated agent-based simulation environment for crime analysis and risk evaluation (CARE)
Shi et al. Big data technology in infectious diseases modeling, simulation, and prediction after the COVID-19 outbreak
Zeng et al. Attention based multi-component spatiotemporal cross-domain neural network model for wireless cellular network traffic prediction
Wang et al. A hypergraph-based hybrid graph convolutional network for intracity human activity intensity prediction and geographic relationship interpretation
Qin Intelligent technologies and methodological transformations in the social sciences
Li et al. Policy and newly confirmed cases universally shape the human mobility during COVID-19
KR102615650B1 (ko) 생성적 적대 신경망 기반의 군중 이동 예측 방법
Gupta et al. Emergence of Blockchain Applications with the 6G-Enabled IoT-Based Smart City
Zhang et al. Multi-modal information fusion-powered regional covid-19 epidemic forecasting
Hu et al. Online footprints of workforce migration and economic implications
Yang et al. The spatiotemporal prediction method of urban population density distribution through behaviour environment interaction agent model
Ma et al. MetaSTNet: Multimodal Meta-learning for Cellular Traffic Conformal Prediction
Gong et al. Data-driven agent-based model of intra-urban activities
Lv et al. A transportation Revitalization index prediction model based on Spatial-Temporal attention mechanism
Li Intelligent System of Smart City Construction Planning: from Data Mining to Data Perception

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22954487

Country of ref document: EP

Kind code of ref document: A1