A method and system for simulating spatio-temporal dynamics of regional population attractiveness under policy driving

Abstract

This invention discloses a method and system for simulating the spatiotemporal dynamics of regional population attractiveness under policy-driven conditions, belonging to the field of geographic information services. The method includes: data acquisition and preprocessing; abstracting the overall research domain into a graph structure, using a graph attention network to embed the node features that integrate regional adjacency information to obtain the optimal embedding matrix; inputting the optimal embedding matrix into a random forest model to establish a mapping relationship between embedded features and population attractiveness; dividing the driving factors into dynamic and static factors, constructing a mechanism for allocating policy objectives from policy action units to the regional scale, updating the node feature matrix, and then inputting it into the trained graph attention network-random forest model to simulate the spatiotemporal dynamic changes of regional population attractiveness under different policy scenarios. This invention can effectively capture the spatial dependencies between regions, quantitatively assess the changes in population attractiveness under policy-driven conditions, and provide scientific support for regional development planning and policy formulation.

Description

Technical Field

[0001] This invention belongs to the field of geographic information service technology, specifically relating to a method and system for dynamic simulation of regional population attraction under policy-driven conditions. Background Technology

[0002] Population is a crucial indicator for measuring regional development levels. From the perspective of population migration, changes in a region's population directly reflect its attractiveness. With the continued advancement of urbanization and the accelerating flow of factors of production between urban and rural areas, some regions are experiencing population shrinkage, which poses certain constraints on regional economic and social development. Therefore, understanding a region's development level is of great significance.

[0003] Current research on regional population attractiveness mainly focuses on three aspects: the spatiotemporal evolution of regional population attractiveness, analysis of influencing factors, and fine-scale mapping. In terms of spatiotemporal evolution, Zhu Yujia et al., based on inter-city population flow data from the 1‰ individual database of the sixth and seventh national population censuses, examined the changes in population attractiveness of Chinese urban agglomerations from 2010 to 2020 from the dimensions of intensity, scope, and spatial equilibrium. Regarding the analysis of influencing factors, Qi et al. used population migration intensity to measure rural population attractiveness, exploring the spatial pattern and driving factors of rural population attractiveness in China. In terms of fine-scale mapping, Guo et al. used the random forest algorithm to reduce the degree of population hollowing out at the township level to a 500m grid scale, achieving fine-scale mapping of regional population attractiveness. In addition, some studies have begun to focus on the impact of policies on population distribution. For example, some studies have used difference-in-differences models to assess the impact of functional zoning planning on population distribution changes, while others have applied system dynamics models to simulate the impact of the Beijing-Tianjin-Hebei coordinated development policy on population growth trends.

[0004] However, existing technologies still have the following shortcomings: Existing research mostly focuses on analyzing the influencing factors of regional population attractiveness, revealing the degree of influence of different factors on population attractiveness and proposing policy recommendations, but it mostly remains at a qualitative perspective and cannot quantitatively assess the changes in population attractiveness under different policies; Although a few studies have simulated regional population attractiveness, they are applied to fine-scale mapping, and the simulation models constructed do not consider the interactions between regions, making it difficult to capture the synergistic or siphoning effects generated on spatial networks; In addition, existing policy impact studies mostly analyze the overall effects of policies from a macro level, lacking the technical means to finely allocate policy objectives from policy action units to the regional scale, and thus simulate the spatiotemporal dynamic changes of regional population attractiveness.

[0005] Therefore, it is necessary to propose a spatiotemporal dynamic simulation method for population attraction that can comprehensively consider the spatial dependence between regions and quantitatively assess the policy-driven effects. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of existing technologies, such as the inability to quantitatively assess changes in regional population attractiveness under different policies, the neglect of spatial interactions between regions, and the lack of fine-grained allocation of policy objectives and spatiotemporal dynamic simulation methods. This invention provides a policy-driven method for simulating the spatiotemporal dynamics of regional population attractiveness. This invention comprehensively considers the spatial dependencies between regions, characterizes the correlation features and interaction mechanisms between regions, and constructs a simulation process under policy-driven scenarios based on the target settings and constraints of various indicators in policy documents, thereby quantitatively assessing the spatiotemporal changes in regional population attractiveness under the influence of policy implementation.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for spatiotemporal dynamic simulation of regional population attractiveness driven by policy, comprising the following steps:

[0008] S1. Data Acquisition and Preprocessing: Collect household registration and resident population data for each region in the overall research area, and calculate the population attractiveness of each region based on the ratio of resident population to household registration population; select driving factors to construct feature vectors for each region; perform data matching on panel statistics data, matching panel statistics data with regional vectors one by one; perform unified projection transformation and zonal statistics on geographic data.

[0009] S2. Regional Graph Structure Representation and Regional Adjacency Information Embedding: The entire research domain is abstracted as a graph structure, with each region as a node. A node feature matrix is ​​constructed based on the feature vectors of each region, and a geographical adjacency matrix is ​​constructed based on the geographical adjacency relationships between regions. A graph attention network is used to embed the node features after fusing regional adjacency information to obtain the embedding vector of each node. The embedding vector of each node is input into a multilayer perceptron, and the model parameters are trained by backpropagation based on the loss between the predicted and the true values ​​to obtain the optimal embedding matrix.

[0010] S3. Regional population attractiveness prediction considering spatial relationships: The optimal embedding matrix is ​​used as the input of the random forest model to establish the mapping relationship between regional embedding features and population attractiveness; the key parameters of the random forest model are tuned, and the model performance is evaluated in combination with the validation set results to determine the random forest model with the best prediction effect.

[0011] S4. Policy Scenario Simulation: Driving factors are divided into dynamic and static factors; the target dynamic factors and their target values ​​are identified based on the policy content; the relationship between policy-affected units and regional units is constructed, and the policy objectives are allocated to the dynamic factor values ​​at the regional scale. The values ​​of the dynamic factors affected by the policy after a set time are calculated; for other dynamic factors not directly affected by the policy, their values ​​after a set time are estimated based on their historical natural growth rate; static factors remain unchanged; the updated node feature matrix and geographical adjacency matrix are input into the trained graph attention network-random forest model to simulate changes in regional population attractiveness under different policy scenarios.

[0012] Further, in step S1, the driving factors include dynamic factors and static factors. The dynamic factors are regional socio-economic attribute indicators that change over time, and the static factors are geographical environmental attribute indicators that do not change over time or change slowly. The dynamic factors include, but are not limited to, GDP, urbanization rate of permanent residents, number of industrial enterprises, forest coverage rate, and road network density. The static factors include, but are not limited to, elevation, slope, and distance from the city center.

[0013] Furthermore, in step S2, the element values ​​of the geographic adjacency matrix are determined as follows: if two regions share a boundary, the corresponding element is set to 1; otherwise, it is set to 0.

[0014] Furthermore, in step S2, the graph attention network adopts a multi-head attention mechanism, independently calculating the attention weights between nodes for each attention head, and concatenating or averaging the outputs of each attention head as the final output of the layer.

[0015] The graph attention network comprises at least two graph attention layers. The first layer employs multiple attention heads. Each attention head performs a linear transformation on the node features and calculates the attention score between node i and its neighbor node j in the following manner. :

[0016]

[0017] in, This represents the activation function. and Let be the feature vectors of nodes i and j after linear transformation, respectively. This is a learnable attention weight vector. This represents a vector concatenation operation;

[0018] The attention scores of all neighboring nodes are normalized using the softmax function to obtain the attention weights. :

[0019] (5)

[0020] in, Represents a node For nodes Attention weights Represents a node The set of neighboring nodes;

[0021] Then, the features of the neighboring nodes are weighted and summed, and an activation function is applied to update the node representation:

[0022]

[0023] in, For nodes The output characteristics, Neighboring nodes Feature representation, This represents the ELU activation function.

[0024] Further, in step S2, the multilayer perceptron includes at least one hidden layer. The embedding matrix output by the graph attention network is used as input. After nonlinear transformation by the hidden layer, the predicted value of population attractiveness of each region is output in the output layer. The mean squared error is used as the loss function. The parameters of the graph attention network and the multilayer perceptron are updated through backpropagation until the loss converges or the preset number of training rounds is reached.

[0025] The hidden layer activation function of the multilayer perceptron adopts the sigmoid function. During training, an early stopping strategy is adopted. When the validation set loss no longer decreases within a certain number of consecutive rounds, training is stopped, and the embedding matrix output by the graph attention network at this time is taken as the optimal embedding matrix.

[0026] Further, in step S3, the embedding vectors corresponding to each region in the optimal embedding matrix are used as input features of the random forest model. The random forest model consists of multiple regression trees, each of which makes independent predictions on the input features. The final prediction value of the regional population attractiveness is the arithmetic mean of the prediction results of all regression trees.

[0027] When tuning the parameters of a random forest model, the parameters to be tuned should include at least the number of decision trees, the maximum depth of the trees, the minimum number of samples required for internal node re-splits, the minimum number of samples required for leaf nodes, and the number of features randomly selected each time a node splits. The optimal parameters are selected by cross-validation or grid search based on the validation set, which minimizes the model's prediction error.

[0028] Further, in step S4, the driving factors are divided into a set of dynamic factors affected by policies, a set of other dynamic factors not directly affected by policies, and a set of static factors; for dynamic factors affected by policies, their target values ​​in each policy action unit are extracted according to policy documents, and the target values ​​are allocated to each region through a target allocation matrix.

[0029] Furthermore, the target allocation matrix is ​​constructed as follows: Let the base period node feature matrix be... ,in This represents the value of the i-th region in the base period on the q-th driving factor; assuming the number of policy action units is n, define the regional affiliation coefficient. When region i belongs to the r-th policy action unit Otherwise, it is 0; then the target allocation coefficient from policy action unit r to region i is... Calculated based on the proportion of the base period value:

[0030]

[0031] Where m is the total number of regions;

[0032] The regional-scale values ​​of the dynamic factor q affected by policy after time T and under policy scenario s Calculated in the following way:

[0033]

[0034] in, This represents the target value of the dynamic factor q in the r-th policy action unit after time T under the influence of policy s. This represents the base period value of the dynamic factor.

[0035] Furthermore, in step S4, for other dynamic factors not directly affected by the policy, their historical natural growth rate is obtained by fitting the statistical values ​​of the factor in the years prior to the base period, and the value after time T is calculated using the following formula:

[0036]

[0037] in, Let be the historical natural growth rate of the i-th region on the q-th dynamic factor.

[0038] The policy scenarios include at least a natural development scenario and at least one policy intervention scenario. Under the natural development scenario, all dynamic factors change according to their respective historical natural growth rates, while static factors remain unchanged. Under the policy intervention scenario, at least one dynamic factor affected by the policy is adjusted according to the policy objective, while the remaining dynamic factors still change according to their historical natural growth rates, and static factors remain unchanged.

[0039] This invention also provides a policy-driven spatiotemporal dynamic simulation system for regional population attraction, used to implement the above-mentioned method, comprising:

[0040] The data acquisition and preprocessing module is used to collect household registration population data and resident population data for each region in the overall research area, and calculate the population attractiveness of each region based on the ratio of resident population to household registration population; select driving factors to construct feature vectors for each region; perform data matching on panel statistical data, matching panel statistical data with regional vectors one by one; and perform unified projection transformation and zonal statistics on geographic data.

[0041] The graph structure construction and adjacency information embedding module is used to abstract the total research domain into a graph structure, with each region as a node. A node feature matrix is ​​constructed based on the feature vectors of each region, and a geographic adjacency matrix is ​​constructed based on the geographic adjacency relationships between regions. A graph attention network is used to embed the node features after fusing regional adjacency information to obtain the embedding vector of each node. The embedding vector of each node is input into a multilayer perceptron, and the model parameters are trained by backpropagation based on the loss between the predicted value and the true value to obtain the optimal embedding matrix.

[0042] The population attractiveness prediction module is used to take the optimal embedding matrix as input to the random forest model, establish the mapping relationship between regional embedding features and population attractiveness, fine-tune the key parameters of the random forest model, evaluate the model performance in combination with the validation set results, and determine the random forest model with the best prediction effect.

[0043] The policy scenario simulation module is used to classify driving factors into dynamic and static factors; identify the target dynamic factors and their target values ​​affected by the policy based on the policy content; construct the correlation between policy-affected units and regional units, allocate policy targets to the dynamic factor values ​​at the regional scale, and calculate the values ​​of the dynamic factors affected by the policy after a set time; for other dynamic factors not directly affected by the policy, the values ​​after a set time are estimated based on their historical natural growth rate; static factors remain unchanged; the updated node feature matrix and geographical adjacency matrix are input into the trained graph attention network-random forest model to simulate the changes in regional population attractiveness under different policy scenarios.

[0044] The output module is used to output the spatiotemporal dynamic simulation results of regional population attraction under different policy scenarios.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] (i) Improve simulation accuracy. This invention abstracts the total research domain into a graph structure and uses a graph attention network to adaptively learn the spatial dependencies between regions. This effectively captures the cooperative or siphoning effects generated on the spatial network, significantly improving the accuracy of regional population attraction simulation compared to traditional methods and single machine learning methods.

[0047] (ii) Quantitative Policy Evaluation Capability. This invention establishes a quantitative allocation mechanism for policy objectives from the policy action unit to the regional scale by constructing a classification system of dynamic and static factors. This enables quantitative evaluation of the spatiotemporal dynamic changes in regional population attractiveness under different policy-driven scenarios, overcoming the shortcomings of existing research that remains at the level of qualitative analysis.

[0048] (III) Fine-scale simulation. This invention supports fine-scale simulation of regional population attractiveness at the township level and above, which can provide refined scientific support for regional development planning and rural revitalization policy formulation. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the implementation of the policy-driven spatiotemporal dynamic simulation method for regional population attraction provided in this embodiment of the invention.

[0050] Figure 2 This is a comparison chart of the model performance of the method of the present invention (GAT-RF) and existing methods (RF, OLS) in the embodiments of the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0052] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0053] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0054] like Figure 1 As shown, this invention provides a spatiotemporal dynamic simulation method for regional population attraction driven by policy, and its specific implementation steps are as follows.

[0055] S1. Data Acquisition and Preprocessing: Collect registered population data and resident population data for each region in the overall research area, and calculate the population attractiveness of each region based on the ratio of resident population to registered population; select driving factors to construct feature vectors for each region; perform data matching on panel statistical data, matching panel statistical data with regional vectors one by one; and perform unified projection transformation and zoning statistics on geographic data.

[0056] Specifically, the registered population and resident population of each region in the research area are collected to determine the region's population attractiveness (Formula 1). At the same time, relevant studies are referenced to select driving factors to construct the region's feature vector. In terms of data preprocessing, for panel statistics: (1) Data matching: Determine the unique identifier field and use the connection tool to match the panel statistics with the regional vector one by one. For geographic data: (1) Unified projection: Transform all data to a unified projection coordinate system; (2) Regional statistics: Calculate the statistical values ​​(total value, average value, etc.) of each driving factor under each regional vector unit.

[0057] (1)

[0058] in, Indicates the first The population attractiveness of a region Indicates the first The permanent resident population of each region Indicates the first The registered population of each region.

[0059] S2. Regional Graph Structure Representation and Regional Adjacency Information Embedding: The entire research domain is abstracted as a graph structure, with each region as a node. A node feature matrix is ​​constructed based on the feature vectors of each region, and a geographical adjacency matrix is ​​constructed based on the geographical adjacency relationships between regions. A graph attention network is used to embed the node features after fusing regional adjacency information to obtain the embedding vector of each node. The embedding vector of each node is input into a multilayer perceptron, and the model parameters are trained by backpropagation based on the loss between the predicted and the true values ​​to obtain the optimal embedding matrix.

[0060] Specifically, this invention abstracts the entire research domain into a graph structure, treating each region as a node, and constructs a node feature matrix in the graph structure based on the feature vectors of each region after data preprocessing. Adjacency relationships are determined through a geographical adjacency matrix. This representation is used to capture the spatial relationships between regions within the study area. Regarding the embedding of region adjacency information, this invention employs Graph Attention Networks (GAT) to embed the fused region adjacency information. The input is a node feature matrix. and geographic adjacency matrix The specific process is as follows: (1) Through linear transformation (2) Map the original features of the nodes initially; (3) Calculate the attention score based on the node features. (3) Use softmax to normalize the neighbor nodes to obtain the attention weights. (4) Perform a weighted summation of the features of neighboring nodes and apply the ELU activation function to update the node representation. (5) Input the embedding vector of each node into a multi-layer perceptron (MLP), map it to its corresponding target value, calculate the loss using mean squared error (MSE) based on the predicted value and the true value, backpropagate, update the gradient to train the parameters of the model, and finally output the best embedding vector containing the geographical environment attributes of itself and its neighbors.

[0061] S3. Regional population attractiveness prediction considering spatial relationships: The optimal embedding matrix is ​​used as the input of the random forest model to establish the mapping relationship between regional embedding features and population attractiveness; the key parameters of the random forest model are tuned, and the model performance is evaluated in combination with the validation set results to determine the random forest model with the best prediction effect.

[0062] Specifically, the optimal embedding vector obtained in step S2 is used as the input to a Random Forest (RF) model to establish a mapping relationship between regional embedding features and population attractiveness. Furthermore, by optimizing key parameters of the random forest (such as the number of decision trees and the maximum tree depth), and combining the validation set results, the model performance (such as mean squared error, root mean square error, and coefficient of determination) is evaluated. Finally, the random forest model with the best predictive performance is determined for use in the next step of policy scenario simulation.

[0063] S4. Policy Scenario Simulation: Driving factors are divided into dynamic and static factors; the target dynamic factors and their target values ​​are identified based on the policy content; the relationship between policy-affected units and regional units is constructed, and the policy objectives are allocated to the dynamic factor values ​​at the regional scale. The values ​​of the dynamic factors affected by the policy after a set time are calculated; for other dynamic factors not directly affected by the policy, their values ​​after a set time are estimated based on their historical natural growth rate; static factors remain unchanged; the updated node feature matrix and geographical adjacency matrix are input into the trained graph attention network-random forest model to simulate changes in regional population attractiveness under different policy scenarios.

[0064] This step specifically includes: (1) dividing the driving factors into dynamic factors (such as GDP, number of enterprises, etc.) and static factors (such as elevation, slope, etc.); (2) identifying the target dynamic factors of the policy based on the policy content. and target values; (3) Construct the correlation between policy action units and regional units, allocate policy objectives to dynamic factor values ​​at the regional scale, and calculate the time. The values ​​of dynamic factors subsequently affected by policies are used; for other dynamic factors not directly affected by policies, the time frame is estimated based on their historical natural growth rate. The static factor remains unchanged after the value is taken; (4) The updated node feature matrix under different scenarios Geographic adjacency matrix Input the data into the trained GAT-RF model to simulate changes in regional population attractiveness under different scenarios.

[0065] Among the above steps, the regional map structure representation and embedding of regional adjacency information, prediction of regional population attractiveness, and simulation of policy scenarios are the key steps in this method, which will be discussed in detail below.

[0066] 1. Region graph structure representation and embedding of region adjacency information

[0067] First, the entire research domain is modeled as a graph structure consisting of nodes and edges. Each node represents a feature of a region, while edges describe the relationships between regions. The features of each node are represented by region attribute vectors, which reflect the driving factor information of each region. The attribute information of all regions is represented as a feature matrix. ,in It is a regional feature dimension. It refers to the number of regions.

[0068] Next, construct the geographic adjacency matrix. , of which elements Indicates the region and region The geographical adjacency relationship between regions is set to 1 if the regions are directly adjacent (e.g., sharing a boundary), and 0 otherwise. The calculation formula is as follows:

[0069] (2)

[0070] Subsequently, GAT was used to embed region adjacency information. GAT allows the model to adaptively assign weights to different connections based on task requirements and the graph's topology. This flexibility enables the model to focus its attention on key adjacency relationships, thereby effectively improving prediction accuracy.

[0071] In the GAT model, the input includes the feature matrix of the nodes. and geographic adjacency matrix First, a linear transformation is used to map the node features, resulting in a preliminary feature representation for each node:

[0072] (3)

[0073] in, It is a node Feature representation, It is a node The input feature vector, It is the weight matrix that needs to be learned. This defines the dimension of the output features. Subsequently, node values ​​are computed using an attention mechanism. Its neighboring nodes Correlation score between The score is calculated using the following formula:

[0074] (4)

[0075] in Indicates the node and nodes Embedded vector concatenation, It is the attention weight vector that needs to be learned. It is an activation function used to introduce nonlinearity.

[0076] To ensure the normalization of attention scores, a softmax function is used for each node. The neighboring nodes are weighted and normalized to calculate the relationship with the neighboring nodes. attention weights :

[0077] (5)

[0078] in Represents a node The neighborhood set, Represents a node For nodes The importance weights. Next, based on the attention weights... For nodes The neighboring node features are weighted and summed to update the node. The representation of:

[0079] (6)

[0080] in It is a node Output features This is the activation function; the activation function chosen in this invention is the ELU function. The final result is the region adjacency information embedding matrix. .

[0081] Generate output matrix Next, using an MLP, the embedding vector of each node is mapped to its corresponding target value. Then, based on the predicted values... and the true value Calculate the loss and backpropagate, update the gradient to train the model's parameters, and obtain the optimal embedding matrix. .

[0082] The mathematical expression of MLP is as follows:

[0083] (7)

[0084] in It is the first The output of the hidden layer, It is the first The weight matrix of the layer, It is the number of hidden units in this layer. It is the first Layer bias, It is an activation function; the activation function selected in this invention is the sigmoid function, and the feature representation of the input layer is as follows: At the final level, the MLP outputs regional population attractiveness. :

[0085] (8)

[0086] in , It is the bias of the output layer. Indicates the first Predicted population attractiveness values ​​for each region. To optimize model parameters. We use mean squared error (MSE) as the loss function, which is defined as follows:

[0087] (9)

[0088] The model is for the first The predicted value of population attractiveness for each region, It is the first The true value of a region's population attractiveness. This is the total number of regions.

[0089] 2. Regional population attractiveness prediction considering spatial relationships

[0090] The optimal embedding matrix obtained in step S2 As input to the random forest model, where the th The embedding vectors corresponding to each region are represented as follows:

[0091] (10)

[0092] Constructing the training sample set:

[0093] (11)

[0094] in, This represents the true value of the region's population attractiveness. The number of samples in the region. For the embedded vector dimension.

[0095] Let random forest be composed of If the regression trees are composed of 1, then the 1st regression tree is... The predicted population attractiveness values ​​for each region are:

[0096] (12)

[0097] in, Indicates the first The output of the regression tree, Indicates the first Population attractiveness projections for each region.

[0098] Let the parameter combination of the random forest model be:

[0099] (13)

[0100] in, For the number of decision trees, For the maximum tree depth, To minimize the number of split samples, The minimum number of leaf node samples. The maximum number of features randomly selected during each partition.

[0101] Based on the validation set results, select the optimal parameter combination that minimizes the model error:

[0102] (14)

[0103] in, Indicates parameter combination The main evaluation metrics (such as mean squared error, root mean squared error, coefficient of determination, etc.) on the validation set. This indicates that extreme values ​​are selected based on the attributes of the evaluation indicators. When the main evaluation indicator is an error-related indicator, the minimum value is taken, and when the main evaluation indicator is a goodness-of-fit indicator, the maximum value is taken.

[0104] Ultimately, the best-performing random forest model was determined and used as the simulation model for policy scenario simulation.

[0105] 3. Policy Scenario Simulation

[0106] Assume the total research domain has a total of Each region has [number] areas, and each area has [number] areas. The driving factors, and the base period node feature matrix are denoted as:

[0107] (15)

[0108] in, Indicates the base period The region in the first The values ​​that can be taken on each driving factor.

[0109] The driving factors are divided into a set of dynamic factors influenced by policies. Other dynamic factors not directly affected by policies and static factor set .

[0110] For dynamic factors affected by policies Let the policy objective vector be:

[0111] (16)

[0112] in, Indicates the number of policy-implementing units.

[0113] Define the region affiliation coefficient as:

[0114] (17)

[0115] The target allocation matrix from policy scope units to regional units is as follows:

[0116] (18)

[0117] Among them, the target allocation coefficient Calculated based on the proportion of the base period value:

[0118] (19)

[0119] Then time The values ​​of the dynamic factors affected by subsequent policies at the regional scale are as follows:

[0120] (20)

[0121] in, express Dynamic factors in In policy Under the influence of time The value after that, This represents the base period value before the simulation begins.

[0122] For other dynamic factors that are not directly affected by policies The time was estimated based on its historical natural growth rate. The value after:

[0123] (twenty one)

[0124] in, Indicates the first The region in the first Historical natural growth rate on dynamic factors.

[0125] For static factors , remains unchanged, that is:

[0126] (twenty two)

[0127] Finally, the updated node feature matrices under different scenarios are obtained, and they are compared with the geographical adjacency matrix. Inputting the data into the trained GAT-RF model yields simulation results of regional population attractiveness under different scenarios.

[0128] (twenty three).

[0129] This embodiment uses a coastal province in my country as the overall research domain, selecting the township level as the regional scale, and conducting experiments in its 922 townships (excluding subdistricts). Township-level registered and resident populations from the Sixth National Population Census (2010) and the Seventh National Population Census (2020) were collected to calculate township population attractiveness. Eight independent variables were selected as driving factors: elevation, slope, GDP, urbanization rate of resident population, number of industrial enterprises, forest coverage rate, road network density, and distance from the city center for the corresponding years.

[0130] First, data preprocessing was used to obtain the population attractiveness and driving factors of each township. Then, each township was treated as a node, and a node feature matrix and geographical adjacency matrix were constructed. The GAT module was configured as a two-layer graph attention network. The first layer contained 8 attention heads, each with an output dimension of 8, and the second layer had an output dimension of 32. The activation function was ELU, Dropout was 0.3, the learning rate was 0.005, and the weight decay coefficient was 0.00005. The network was trained for 300 epochs using the Adam optimizer, with 30 early stopping epochs. The 32-dimensional spatial embedding features from the GAT output were concatenated with the original 8 driving factors and input into a random forest regression model. The random forest parameters were set as follows: 500 decision trees, maximum tree depth 12, minimum number of split samples 4, minimum number of leaf node samples 2, maximum number of features 6, and random seed 42. The training and validation sets were divided into 80% and 20% datasets based on 2010 data, respectively, and the test set was based on 2020 data, thus constructing a township population attractiveness simulation model. Two scenarios were then set up to simulate the population attraction of townships in 2030: (1) Natural development scenario: dynamic factors (GDP, urbanization rate of permanent residents, number of industrial enterprises, forest coverage rate, road network density) increased according to the historical growth rate, while static factors (elevation, slope, distance from the city center) remained unchanged; (2) Revitalization of old revolutionary base areas and former Soviet areas scenario: with reference to the policy, GDP and urbanization rate of permanent residents were used as policy factors. According to the policy objectives, the GDP of old revolutionary base areas and former Soviet areas was set to grow at an average annual rate of 6.5%, and the urbanization rate of permanent residents was set to grow at an average annual rate of 0.64%. These were then allocated to each township. The remaining dynamic factors developed according to the historical growth rate, while the static factors remained unchanged; (3) Ecological protection scenario: with reference to the policy, forest coverage rate was used as a policy factor. According to the requirement that the forest coverage rate reach 67.5% by 2030, the townships were spatially allocated. The remaining dynamic factors developed according to the historical growth rate, while the static factors remained unchanged.

[0131] The results show that: (1) the model of the present invention outperforms traditional methods and machine learning methods (results are shown in Figure 1). Figure 2As shown), it can capture spatial correlations well and be used to simulate the population attractiveness of townships. (2) In terms of policy scenario simulation, the average population attractiveness of townships, the Gini coefficient of township population attractiveness and spatial distribution are selected to evaluate the changes in population attractiveness under different scenarios. Among them, under the natural development scenario, the average population attractiveness increased by 0.084 and the Gini coefficient increased by 0.06. Spatially, the population attractiveness of townships in coastal areas increased, while the population attractiveness of townships in mountainous areas increased slightly or remained unchanged. Under the revitalization scenario of old revolutionary base areas and Soviet areas, the average value increased by 0.087 and the Gini coefficient decreased by 0.07. Spatially, the gap in population attractiveness between coastal areas and mountainous areas narrowed, and both increased overall. Under the ecological protection scenario, the average value increased by 0.079 and the Gini coefficient increased by 0.02. Spatially, the gap between mountainous areas and coastal areas remained basically unchanged.

[0132] This embodiment also provides a policy-driven spatiotemporal dynamic simulation system for regional population attraction, used to implement the above-mentioned method, including:

[0133] The data acquisition and preprocessing module is used to collect household registration population data and resident population data for each region in the overall research area, and calculate the population attractiveness of each region based on the ratio of resident population to household registration population; select driving factors to construct feature vectors for each region; perform data matching on panel statistical data, matching panel statistical data with regional vectors one by one; and perform unified projection transformation and zonal statistics on geographic data.

[0134] The graph structure construction and adjacency information embedding module is used to abstract the total research domain into a graph structure, with each region as a node. A node feature matrix is ​​constructed based on the feature vectors of each region, and a geographic adjacency matrix is ​​constructed based on the geographic adjacency relationships between regions. A graph attention network is used to embed the node features after fusing regional adjacency information to obtain the embedding vector of each node. The embedding vector of each node is input into a multilayer perceptron, and the model parameters are trained by backpropagation based on the loss between the predicted value and the true value to obtain the optimal embedding matrix.

[0135] The population attractiveness prediction module is used to take the optimal embedding matrix as input to the random forest model, establish the mapping relationship between regional embedding features and population attractiveness, fine-tune the key parameters of the random forest model, evaluate the model performance in combination with the validation set results, and determine the random forest model with the best prediction effect.

[0136] The policy scenario simulation module is used to classify driving factors into dynamic and static factors; identify the target dynamic factors and their target values ​​affected by the policy based on the policy content; construct the correlation between policy-affected units and regional units, allocate policy targets to the dynamic factor values ​​at the regional scale, and calculate the values ​​of the dynamic factors affected by the policy after a set time; for other dynamic factors not directly affected by the policy, the values ​​after a set time are estimated based on their historical natural growth rate; static factors remain unchanged; the updated node feature matrix and geographical adjacency matrix are input into the trained graph attention network-random forest model to simulate the changes in regional population attractiveness under different policy scenarios.

[0137] The output module is used to output the spatiotemporal dynamic simulation results of regional population attraction under different policy scenarios.

[0138] This embodiment also provides a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the above-described method.

[0139] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0140] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0141] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0142] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0143] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A policy-driven method for spatiotemporal dynamic simulation of regional population attractiveness, characterized in that, Includes the following steps: S1. Data Collection and Preprocessing: Collect household registration population data and resident population data for each region in the research area, and calculate the population attractiveness of each region based on the ratio of resident population to household registration population. Select driving factors to construct feature vectors for each region; perform data matching on panel statistics, matching panel statistics with regional vectors one by one; perform unified projection transformation and zonal statistics on geographic data; S2. Regional Graph Structure Representation and Embedding of Regional Adjacency Information: The entire research domain is abstracted into a graph structure, with each region as a node. A node feature matrix is ​​constructed based on the feature vectors of each region, and a geographical adjacency matrix is ​​constructed based on the geographical adjacency relationships between regions. A graph attention network is used to embed the node features after fusing the adjacency information of the region to obtain the embedding vector of each node; the embedding vector of each node is input into a multilayer perceptron, and the model parameters are trained by backpropagation based on the loss between the predicted value and the true value to obtain the optimal embedding matrix. S3. Regional population attractiveness prediction considering spatial relationships: The optimal embedding matrix is ​​used as the input of the random forest model to establish the mapping relationship between regional embedding features and population attractiveness; the key parameters of the random forest model are tuned, and the model performance is evaluated in combination with the validation set results to determine the random forest model with the best prediction effect. S4. Policy Scenario Simulation: Driving factors are divided into dynamic and static factors; the target dynamic factors and their target values ​​are identified based on the policy content; the relationship between policy-affected units and regional units is constructed, and the policy objectives are allocated to the dynamic factor values ​​at the regional scale. The values ​​of the dynamic factors affected by the policy after a set time are calculated; for other dynamic factors not directly affected by the policy, their values ​​after a set time are estimated based on their historical natural growth rate; static factors remain unchanged; the updated node feature matrix and geographical adjacency matrix are input into the trained graph attention network-random forest model to simulate changes in regional population attractiveness under different policy scenarios.

2. The spatiotemporal dynamic simulation method for regional population attraction driven by policy according to claim 1, characterized in that, In step S1, the driving factors include dynamic factors and static factors. The dynamic factors are regional socio-economic attribute indicators that change over time, and the static factors are geographical environmental attribute indicators that do not change over time or change slowly. The dynamic factors include GDP, urbanization rate of permanent residents, number of industrial enterprises, forest coverage rate, and road network density. The static factors include elevation, slope, and distance from the city center.

3. The spatiotemporal dynamic simulation method for regional population attraction driven by policy according to claim 1, characterized in that, In step S2, the element values ​​of the geographic adjacency matrix are determined as follows: if two regions share a boundary, the corresponding element is set to 1; otherwise, it is set to 0.

4. The spatiotemporal dynamic simulation method for regional population attraction driven by policy according to claim 1, characterized in that, In step S2, the graph attention network adopts a multi-head attention mechanism, independently calculating the attention weights between nodes for each attention head, and concatenating or averaging the outputs of each attention head as the final output of the layer. The graph attention network comprises at least two graph attention layers. The first layer employs multiple attention heads. Each attention head performs a linear transformation on the node features and calculates the attention score between node i and its neighbor node j in the following manner. : in, This represents the activation function. and Let be the feature vectors of nodes i and j after linear transformation, respectively. This is a learnable attention weight vector. This represents a vector concatenation operation; The attention scores of all neighboring nodes are normalized using the softmax function to obtain the attention weights. : (5) in, Represents a node For nodes Attention weights Represents a node The set of neighboring nodes; Then, the features of the neighboring nodes are weighted and summed, and an activation function is applied to update the node representation: in, For nodes The output characteristics, Neighboring nodes Feature representation, This represents the ELU activation function.

5. The spatiotemporal dynamic simulation method for regional population attraction driven by policy according to claim 1, characterized in that, In step S2, the multilayer perceptron includes at least one hidden layer. The embedding matrix output by the graph attention network is used as input. After nonlinear transformation by the hidden layer, the predicted value of population attractiveness of each region is output in the output layer. The mean squared error is used as the loss function. The parameters of the graph attention network and the multilayer perceptron are updated through backpropagation until the loss converges or the preset number of training rounds is reached. The hidden layer activation function of the multilayer perceptron adopts the sigmoid function. During training, an early stopping strategy is adopted. When the validation set loss no longer decreases within a certain number of consecutive rounds, training is stopped, and the embedding matrix output by the graph attention network at this time is taken as the optimal embedding matrix.

6. The spatiotemporal dynamic simulation method for regional population attraction driven by policy according to claim 1, characterized in that, In step S3, the embedding vectors corresponding to each region in the optimal embedding matrix are used as input features of the random forest model. The random forest model consists of multiple regression trees, each of which makes independent predictions for the input features. The final prediction value of the regional population attractiveness is the arithmetic mean of the prediction results of all regression trees. When tuning the parameters of a random forest model, the parameters to be tuned should include at least the number of decision trees, the maximum depth of the trees, the minimum number of samples required for internal node re-splits, the minimum number of samples required for leaf nodes, and the number of features randomly selected each time a node splits. The optimal parameters are selected by cross-validation or grid search based on the validation set, which minimizes the model's prediction error.

7. The spatiotemporal dynamic simulation method for regional population attraction driven by policy according to claim 1, characterized in that, In step S4, the driving factors are divided into a set of dynamic factors affected by policies, a set of other dynamic factors not directly affected by policies, and a set of static factors. For dynamic factors affected by policies, their target values ​​in each policy action unit are extracted according to policy documents, and the target values ​​are allocated to each region through a target allocation matrix.

8. The spatiotemporal dynamic simulation method for regional population attraction driven by policy, as described in claim 7, is characterized in that... The target allocation matrix is ​​constructed as follows: Let the base period node feature matrix be... ,in This represents the value of the i-th region in the base period on the q-th driving factor; assuming the number of policy action units is n, define the regional affiliation coefficient. When region i belongs to the r-th policy action unit Otherwise, it is 0; then the target allocation coefficient from policy action unit r to region i is... Calculated based on the proportion of the base period value: Where m is the total number of regions; The regional-scale values ​​of the dynamic factor q affected by policy after time T and under policy scenario s Calculated in the following way: in, This represents the target value of the dynamic factor q in the r-th policy action unit after time T under the influence of policy s. This represents the base period value of the dynamic factor.

9. The spatiotemporal dynamic simulation method for regional population attraction driven by policy according to claim 1, characterized in that, In step S4, for other dynamic factors not directly affected by the policy, their historical natural growth rate is obtained by fitting the statistical values ​​of the factor in the years prior to the base period, and the value after time T is calculated using the following formula: in, Let be the historical natural growth rate of the i-th region on the q-th dynamic factor. The policy scenarios include at least a natural development scenario and at least one policy intervention scenario. Under the natural development scenario, all dynamic factors change according to their respective historical natural growth rates, while static factors remain unchanged. Under the policy intervention scenario, at least one dynamic factor affected by the policy is adjusted according to the policy objective, while the remaining dynamic factors still change according to their historical natural growth rates, and static factors remain unchanged.

10. A policy-driven spatiotemporal dynamic simulation system for regional population attraction, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to collect household registration population data and resident population data for each region in the overall research area, and calculate the population attractiveness of each region based on the ratio of resident population to household registration population. Select driving factors to construct feature vectors for each region; perform data matching on panel statistics, matching panel statistics with regional vectors one by one; perform unified projection transformation and zonal statistics on geographic data; The graph structure construction and adjacency information embedding module is used to abstract the total research domain into a graph structure, with each region as a node. The node feature matrix is ​​constructed based on the feature vectors of each region, and the geographic adjacency matrix is ​​constructed based on the geographic adjacency relationship between regions. A graph attention network is used to embed the node features after fusing the adjacency information of the region to obtain the embedding vector of each node; the embedding vector of each node is input into a multilayer perceptron, and the model parameters are trained by backpropagation based on the loss between the predicted value and the true value to obtain the optimal embedding matrix. The population attractiveness prediction module is used to take the optimal embedding matrix as input to the random forest model and establish a mapping relationship between regional embedding features and population attractiveness. The key parameters of the random forest model are tuned, and the model performance is evaluated by combining the validation set results to determine the random forest model with the best prediction effect. The policy scenario simulation module is used to classify driving factors into dynamic and static factors; identify the target dynamic factors and their target values ​​affected by the policy based on the policy content; construct the correlation between policy-affected units and regional units, allocate policy targets to the dynamic factor values ​​at the regional scale, and calculate the values ​​of the dynamic factors affected by the policy after a set time; for other dynamic factors not directly affected by the policy, the values ​​after a set time are estimated based on their historical natural growth rate; static factors remain unchanged; the updated node feature matrix and geographical adjacency matrix are input into the trained graph attention network-random forest model to simulate the changes in regional population attractiveness under different policy scenarios. The output module is used to output the spatiotemporal dynamic simulation results of regional population attraction under different policy scenarios.

Images