Method for measuring carbon emission space spillover effect at plot scale, terminal and medium

Abstract

The application provides a land plot scale carbon emission space spillover effect measuring method, a terminal and a medium, the method constructs a land plot graph structure containing spatial adjacent edges and functional correlation edges, trains a land plot carbon emission change prediction model based on edge attribute enhancement based on the graph structure, obtains output parameters of the model after training, including adjacent effect prediction values of each land plot and attention weights of each edge; based on the attention weights, obtain the carbon emission spillover intensity corresponding to each edge; identify the land plot that has changed in the graph structure, perform division of the update type group for each changed land plot; according to the update type group, based on the carbon emission spillover intensity corresponding to each edge and the adjacent effect prediction value, extract the carbon emission spillover effect measuring result corresponding to each update type. The application is suitable for massive irregular land plots, and effectively improves the accuracy of the carbon emission space spillover effect measuring result.

Description

Technical Field

[0001] This invention relates to the field of carbon emission spatial analysis technology, specifically to a method, terminal, and medium for measuring the spatial spillover effect of carbon emissions at the plot scale. Background Technology

[0002] Urban land renewal activities not only alter the carbon emissions of the renewed site itself, but may also change the carbon emissions of surrounding sites through industrial linkages, factor flows, transportation connections, and technology diffusion. This effect, which extends beyond the boundaries of the renewed site and manifests as an additional increase or decrease in carbon emissions for surrounding sites, constitutes a spatial spillover effect of carbon emissions.

[0003] Currently, spatial autocorrelation analysis is commonly used to analyze the spatial spillover effects of carbon emissions. For example, global or local Moran indices are used to describe whether carbon emission variables exhibit spatial clustering; or spatial weight matrices constructed based on adjacency relationships or centroid distances are typically employed, with global spatial autoregressive coefficients used to characterize the average spillover intensity. These methods provide some analytical effectiveness for a small number of relatively regular spatial units, such as administrative regions. However, at the plot scale, the number of analytical objects typically reaches hundreds of thousands. Therefore, spatial econometric models require repeated matrix inversions related to the Nth-order spatial weight matrix during the solution process, resulting in computational complexity and workload that are difficult to bear with conventional computing resources. This makes it difficult to directly transfer analytical methods based on administrative regions to the plot scale, and thus, it is impossible to obtain results for measuring the spatial spillover effects of carbon emissions at the plot scale.

[0004] Furthermore, activities such as changes in industrial land use type, conversion of industrial to commercial and service, conversion of industrial to residential, and commercial transformation rely on different connections such as industrial chains, commuting, transportation, or local environment. This results in differences in the transmission medium and action path of carbon emission spillover effects of a single plot to surrounding plots based on different plot renewal activities. However, existing models often describe the interaction between all plots with a single global parameter. That is, different source plots, different target plots, and different connecting edges are assigned the same or homogeneous spillover patterns, making it difficult to obtain accurate carbon emission spillover effects corresponding to the renewal type. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the present invention aims to solve the following interrelated technical problems in the determination of spatial overflow at the plot scale: the infeasibility of calculation and the problem of graph structure distortion caused by a large number of irregular plots; the problem that global homogeneous parameters cannot characterize edge-level heterogeneous overflows, etc.

[0006] To achieve the above and other related objectives, the present invention first provides a method for measuring the spatial spillover effect of carbon emissions at the land parcel scale, comprising:

[0007] Using each plot as a node, a plot graph structure is constructed, containing spatial adjacency edges and functional association edges. The graph structure includes initial feature vectors corresponding to each node and edge attribute vectors corresponding to each edge. Based on the graph structure, a plot carbon emission change prediction model based on edge attribute enhancement is trained to obtain the model's output parameters after training. The model's output parameters include the predicted neighbor effect values ​​for each plot and the attention weights for each edge. Based on the attention weights, the carbon emission spillover intensity corresponding to each edge is obtained. Plots that have undergone changes are identified in the graph structure, and each changed plot is divided into update type groups. The update type group is a combination of changes that meets the change screening conditions. Based on the update type group, and using the carbon emission spillover intensity corresponding to each edge and the predicted neighbor effect value, the carbon emission spillover effect measurement results corresponding to each update type are extracted, including at least: the direction and intensity of the spatial spillover effect of carbon emissions, as well as the distance decay curve and confidence interval.

[0008] In an optional embodiment, the edge attribute-enhanced land parcel carbon emission change prediction model includes:

[0009] A multi-head graph attention encoder with edge attribute enhancement is used to obtain the final neighbor enhancement representation of each plot and the attention weights corresponding to each edge; a self-effect decoder is used to obtain the corresponding self-effect prediction value based on the initial feature vector of the plot; and a neighbor effect decoder is connected to the multi-head graph attention encoder to obtain the neighbor enhancement representation of each plot and obtain the corresponding neighbor effect prediction value.

[0010] In an optional embodiment, training the edge attribute-enhanced land parcel carbon emission change prediction model based on the graph structure includes:

[0011] The graph structure is input into a multi-head graph attention encoder with edge attribute enhancement to obtain the neighbor enhancement representations of each plot and the attention weights corresponding to each connecting edge. The initial feature vectors of each plot are input into a self-effect decoder, and the neighbor enhancement representations of each plot are input into a neighbor effect decoder to obtain the self-effect prediction value and the neighbor effect prediction value of each plot, respectively. The self-effect prediction value and the neighbor effect prediction value are added to obtain the total carbon emission change prediction value of each plot. The self-effect prediction value and the neighbor effect prediction value are added to obtain the total carbon emission change prediction value of each plot. Based on the total carbon emission change prediction value and the actual value of each plot, the trainable parameters in the model are updated by backpropagation using a joint loss function.

[0012] The joint loss function is:

[0013]

[0014] in, For the joint loss function;

[0015] The prediction accuracy loss function is used to constrain the predicted value of the total change in carbon emissions to approximate the true value of the total change in carbon emissions.

[0016] This is a multi-head differentiation regularization term used to constrain the similarity of the attention distributions learned by each attention head, so as to force each attention head to focus on neighborhood relationships in different dimensions.

[0017] It is a multi-head differentiated orthogonal weight.

[0018] In one optional embodiment, the initial feature vector is obtained by means of:

[0019] The process involves acquiring carbon emission information for each land parcel at a reference time point, and acquiring land parcel attribute information for each land parcel at both the reference and target time points. The carbon emission information and the land parcel attribute information are then concatenated to construct initial feature vectors for each land parcel. The land parcel attribute information includes at least land parcel graphic information, land use type information, and land parcel development information. The land parcel graphic information includes at least one of area, perimeter, compactness, fractal dimension, shape index, and aspect ratio. The land use type information includes primary land use types and secondary sub-industry types corresponding to industrial land. The land parcel development information includes at least one of plot ratio, distance to city center, and distance to the nearest main road.

[0020] In one optional embodiment, the method for constructing the spatial adjacency edge includes:

[0021] Delaunay triangulation is performed based on the centroid coordinates of each plot to obtain candidate connecting edges; for the candidate connecting edges connecting plot i and plot j, the equivalent radius is determined based on the area of ​​plot i and plot j respectively, and the dynamic distance truncation threshold is determined by combining the preset scaling factor and the minimum truncation distance; when the centroid Euclidean distance between plot i and plot j is not greater than the dynamic distance truncation threshold, a spatial adjacency edge between plot i and plot j is established.

[0022] The dynamic distance cutoff threshold is:

[0023]

[0024] in, For the plot of land and plot of land Dynamic truncation thresholds between; This is an empirical proportionality coefficient; This is the preset minimum cutoff distance.

[0025] In an optional embodiment, the method for constructing the functional association edge includes:

[0026] Based on the preset functional association type between the two land use categories, or based on the input-output association between the two industrial land parcels, the weight coefficient of the functional association relationship is determined; based on the product of the weight coefficient of the functional association relationship and the distance decay function between land parcel i and land parcel j, the functional association strength is determined; when the functional association strength is greater than the preset cutoff threshold, a functional association edge is established between land parcel i and land parcel j.

[0027] In one optional embodiment, the method for obtaining the weight coefficients of the functional association relationship includes:

[0028] If both plots are industrial land, the total consumption coefficient between the two sub-sectors is extracted based on the existing input-output table of the region and set as the weight coefficient of the functional relationship. This coefficient is used to characterize the sum of direct and indirect consumption of the final product of one sub-sector on the product of the other sub-sector. If the two plots belong to different land use categories and meet the preset functional relationship type, the weight coefficient of the functional relationship is set according to the pre-associated level of the land use function combination.

[0029] In an optional embodiment, the method for obtaining the distance decay curve and the confidence interval includes:

[0030] For each update type group, sample points are constructed using the distances between plots connected by each outgoing edge of the group as independent variables and the comprehensive attention weight of the outgoing edges as dependent variables. For the distance to be estimated, local samples close to the distance to be estimated are selected from the update type group. The fitting weights are determined based on the relative distance between the local samples and the distance to be estimated, and the weighted least squares method is used to regress the local samples to obtain a smooth estimate of the distance to be estimated. The smooth estimates of multiple distances to be estimated are connected in order of distance to obtain the distance decay curve. The Bootstrap resampling method is used to obtain the confidence interval of the distance decay curve.

[0031] In an optional embodiment, the method for obtaining the direction and intensity of the carbon emission spatial spillover effect includes:

[0032] According to the land parcel update type, the predicted neighbor effect values ​​of the target land parcels affected by the corresponding source land parcels are aggregated; the spillover direction of the corresponding land parcel update type is determined based on the sign and significance of the aggregated neighbor effect predicted values, and the relative spillover intensity is determined based on the attention weight of the connecting edges within the corresponding land parcel update type group.

[0033] In a second aspect, the present invention also provides a terminal, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the carbon emission spatial spillover effect measurement method as described above.

[0034] In a third aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the carbon emission spatial spillover effect measurement method as described above.

[0035] As described above, the method, terminal, and computer storage medium for measuring the spatial spillover effect of carbon emissions at the plot scale provided in this application, by adopting an area-adaptive distance threshold and combining continuous boundary sharing and functional association continuous weighting methods, enable the constructed plot map structure to adapt to the characteristics of large differences in plot area, irregular shape, and uneven spatial density, reducing over-connection in dense areas and under-connection in sparse areas; and by introducing edge attribute information such as distance, boundary sharing, and functional association, differentiated weights are learned for each connecting edge, avoiding the use of a single global parameter to describe different update types and different transmission paths; and by constructing a network model containing a multi-head graph attention encoder with edge attribute enhancement and a dual decoder with information isolation, it is possible to separate the plot's own changes from the prediction of neighborhood spillover increments even in the absence of independent ground values ​​for self-effects and neighbor effects. This not only greatly reduces the computational complexity and difficulty at the plot scale level, but also effectively provides the objectivity and accuracy of measuring the carbon emission effect between plots. Attached Figure Description

[0036] Figure 1 This is a schematic flowchart of one embodiment of the carbon emission spatial spillover effect measurement method of the present invention;

[0037] Figure 2 This is a schematic diagram illustrating the construction process of the land parcel map structure in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of the method for obtaining the initial feature vectors of various land parcels in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the structure of the edge attribute enhanced land parcel carbon emission change prediction model in an embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram illustrating the screening principle of effective target node plots in an embodiment of the present invention;

[0041] Figure 6 This is a schematic diagram of the distance decay curve fitting for the carbon emission spillover effect in an embodiment of the present invention.

[0042] Figure 7This is a schematic flowchart of another embodiment of the carbon emission spatial spillover effect measurement method of the present invention;

[0043] Figure 8 This is a structural block diagram of the terminal according to an embodiment of the present invention. Detailed Implementation

[0044] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0045] To address the challenges of large variations in area and shape among massive land parcels, where the computational burden of existing spatial econometric methods increases dramatically with the number of parcels, and where the spillover effects of different parcel update sources are averaged by global parameters, as well as the complex interplay between parcel evolution and neighbor spillovers, this application employs a "mapping-learning-separation-extraction" overall execution logic. This involves converting real-world parcels and their spatial relationships and land use connections into a sparsely storable structural graph. A multi-head graph attention encoder with edge attribute enhancement learns differentiated information transmission weights for each "change source-impact target" connection from observational data. Then, using a dual decoder with isolated information inputs, carbon emission changes at the target time point are decomposed into self-effects caused by the change itself and neighbor effects formed by neighbor transmission. Finally, the carbon emission spillover effect measurement results corresponding to the update type are extracted from the attention weights and neighbor effect outputs stored in the training model. Based on this, the spatial spillover process of carbon emissions, which cannot be directly observed in existing technologies, is transformed into calculable and comparable parcel-level quantitative results, providing an efficient and intuitive data foundation for subsequent low-carbon land spatial planning and analysis.

[0046] The carbon emission spillover effect measurement results include at least the direction and intensity of carbon emission spillover and the distance decay curve.

[0047] In this application, spatial adjacency edges and functional association edges are two different transmission channels used to describe carbon emission spillover effects; the spatial adjacency edges are used to characterize the short-distance carbon emission transmission relationship between geographically adjacent plots, including carbon emission spillover effects caused by material exchange or personnel movement; for example, when a plot undergoes commercial transformation, the carbon emission characteristics corresponding to the traffic, consumption and other activities of people in neighboring plots (such as residential and office plots) are changed accordingly.

[0048] The functional association edge is used to characterize the carbon emission transmission relationship between land parcels that are not necessarily directly adjacent in geographical space but are functionally related, including carbon emission spillover effects caused by changes in land use function, exchanges of production activities, etc. For example, when a textile industry land changes its industrial use, it causes changes in downstream apparel and clothing industry land in the surrounding area, thereby affecting the overall carbon emissions of the region. Therefore, even if two land parcels are not directly adjacent, an emission reduction spillover effect is still formed.

[0049] Please see Figure 1 This diagram illustrates a flowchart of an embodiment of the method for measuring the spatial spillover effect of carbon emissions at the land parcel scale provided in this application; as shown below. Figure 1 As shown, the method includes the following steps:

[0050] S100, based on the initial feature vectors corresponding to each parcel, constructs a parcel graph structure containing spatial adjacency edges and functional association edges;

[0051] In this application, the construction method of the plot map structure is as follows: Figure 2 As shown, it includes:

[0052] Set each plot within the work area as a node in the graph structure. To construct the node set of the workspace Extract the initial feature vectors of each land parcel. The initial feature vector is information used to predict the carbon emission changes of a land parcel based on its own characteristics.

[0053] Each node The corresponding initial feature vector Stack the features to obtain the initial feature matrix of the working area. ;in, This represents the total number of plots of land within the work area. For the initial feature vector The corresponding feature dimension; in the initial feature matrix In the middle, the first Row vectors represent land parcels The corresponding initial feature vector .

[0054] The relationships between different plots within the work area are represented as edges in a graph structure. Specifically, based on the spatial adjacency or functional association between plots, spatial adjacency edges or functional association edges between plots are constructed respectively, and the degree of spatial or functional association between the two plots connected by the edge is extracted accordingly. The extracted degree of association is set as the edge attribute vector corresponding to each edge.

[0055] Each of the aforementioned spatial adjacency edges and each of the aforementioned functional association edges is constructed into a corresponding set of spatial adjacency edges. Set of edges associated with functions Summarize the set of spatial adjacent edges. The set of edges associated with the aforementioned functions To obtain the total set of edges in the working area. ,Right now .

[0056] The edge attribute vectors corresponding to each edge are stacked and arranged to construct the edge attribute matrix of the working area. , ;in, This represents the total number of edges in the total edge set. Let be the feature dimension of the edge attribute vector corresponding to a single edge in the total edge set; this feature dimension is the sum of the feature dimensions corresponding to spatially adjacent edges and the feature dimensions corresponding to functionally associated edges. In the edge attribute matrix... In the middle, the first The row vector is the first row vector. Strip edge edge attribute vector .

[0057] The nodes and their relationships obtained above are stored in a compressed row sparse matrix format to construct the node relationship matrix of the working area. Specifically, this format contains three one-dimensional arrays: a row offset array, a neighbor node array, and an edge index array, used for fast retrieval of any node. The set of neighboring nodes Among them, the set of neighboring nodes To be with this node The set of all nodes directly connected by spatial adjacency edges or functional association edges.

[0058] Comprehensive Node Set Initial feature matrix Total edge set Edge attribute matrix Node association matrix Build the graph structure of the workspace .

[0059] In some optional embodiments, the initial feature vector is obtained in the following ways: Figure 3 As shown, it includes:

[0060] S101, Obtain carbon emission information for each plot at the reference time point;

[0061] Wherein, the reference time point is the starting point of the time period to be predicted; for example, when the time period to be predicted is 2015 to 2020, 2015 is used as the reference time point and 2020 is used as the target time point; the carbon emissions per unit area of ​​each plot in 2015 are obtained. This serves as the carbon emission information for the land parcel at the reference time point.

[0062] S102, Obtain the land parcel attribute information of each parcel in the work area at the base time point and the target time point respectively;

[0063] The target time point is the predicted time of the period to be predicted; the land parcel attribute information is information that affects the carbon emission spillover effect of the land parcel, including at least: graphic information, land use type and land parcel development information.

[0064] Specifically, the graphic information includes the geometric and shape information of the land parcel; the geometric information is used to characterize the geometric features of the land parcel, including but not limited to the area of ​​the land parcel (…). ),perimeter( ) and compactness ( )wait;

[0065] The shape information is used to characterize the shape complexity of the land parcel, including but not limited to fractal dimension. Shape index Aspect ratios such as major and minor axes; for example, fractal dimension. To calculate the degree of folding at the site boundary using box counting; shape index To reflect the degree to which a plot of land deviates from a circular shape; the ratio of its major axis to its minor axis ( () is used to reflect the elongation or narrowness of the plot.

[0066] This application constructs different indicator features related to the plot graphic information to characterize the irregularity and complexity of the plot shape from different dimensions, so that the subsequent graph attention network can effectively learn the impact of the plot graphic information on the carbon emission spillover effect.

[0067] In this application, the land use type refers to the land use type information of the plot, including primary category and secondary category; wherein, primary category (major category) includes but is not limited to urban residential, commercial and service land, industrial land, roads, water surface, etc.; secondary category (minor category) is only for industrial land, used to characterize the subdivided industry type mainly carried by the industrial land plot, including but not limited to non-metallic mineral mining and dressing, heat production and supply, food manufacturing, textile industry, etc.

[0068] In this application, the land development information is used to characterize the location advantages and development intensity of the land, including the plot ratio, distance to the city center, distance to the nearest main road, etc.

[0069] S103, Based on the carbon emission information and land parcel attribute information, construct the initial feature vector corresponding to each land parcel.

[0070] After obtaining the carbon emission information and corresponding land attribute information for each plot, the two are concatenated, and the concatenated comprehensive information is used as the initial feature vector for each plot. .

[0071] In one optional embodiment, the method for obtaining the spatial adjacency edges between the land parcels includes:

[0072] Based on the centroid coordinates of each parcel, standard Delaunay triangulation is performed in the planar coordinate system of the working area to generate a triangulation network covering the entire region.

[0073] By performing adaptive distance truncation determination on the triangulation network, spatial adjacency edges between land parcels are constructed, and the corresponding edge attribute vectors are obtained.

[0074] Specifically, iterate through each edge of the triangulation network, and perform the following steps for each edge:

[0075] 1) Obtain the two end nodes of the edge ( , The distance between the corresponding land parcels ;

[0076] The distance between the plots is the Euclidean distance between the centroids of the two plots. .

[0077] 2) Based on the area size of the plot, determine the equivalent radius of the plots corresponding to the two end nodes respectively, and combine it with the pre-set minimum cutoff distance to determine the dynamic cutoff threshold between the plots corresponding to the two end nodes.

[0078] Specifically, acquiring land parcels and plot of land The corresponding land parcel areas are respectively and Acquire land parcels according to their size. and plot of land The corresponding equivalent radius is:

[0079] , .

[0080] in, For the plot of land The equivalent radius; For the plot of land The area of ​​the plot; For the plot of land The equivalent radius; For the plot of land The area of ​​the plot.

[0081] Based on the equivalent radii corresponding to the two plots, and combined with the minimum cutoff distance, the plots are obtained. and plot of land The dynamic truncation threshold between them is:

[0082]

[0083] in, For the plot of land and plot of land Dynamic truncation thresholds between; This is an empirical proportionality coefficient; This is the preset minimum cutoff distance.

[0084] 3) Based on the distance between the two plots of land With the dynamic truncation threshold Based on the size relationship between the two plots, a spatial adjacency edge is constructed between them, and the edge attribute vector of the spatial adjacency edge is obtained.

[0085] Specifically, compare the distances between the two plots of land. With the dynamic truncation threshold The size relationship between them;

[0086] If the distance of the land parcel Then create a spatial adjacency edge between plot i and plot j. and add it to the set of spatial adjacent edges. In the middle; if the plot is far from If so, then discard that edge.

[0087] This application obtains an equivalent radius that is positively correlated with the area of ​​the plot. For larger plots, it can automatically increase the dynamic cutoff threshold between plots to ensure that larger plots can establish connections with their neighboring plots within a reasonable range. Conversely, for relatively smaller plots, it can automatically decrease the dynamic cutoff threshold between plots to avoid excessive computation due to overly dense connections between plots. This ensures that the edge density of the graph structure remains relatively balanced across plot regions at different scales.

[0088] For each spatial adjacent edge The spatial distance features and boundary sharing features between the land parcels corresponding to the two ends of the spatial adjacent edge are extracted respectively. The total feature obtained by combining the above features is taken as the spatial adjacent edge. The corresponding edge attribute vector; wherein, the spatial distance feature is used to characterize the absolute spatial distance between two plots; for example, the spatial distance feature adopts the centroid Euclidean distance between the two plots; the boundary sharing feature is used to characterize the degree of adjacency between the two plots; for example, the boundary sharing feature adopts the continuous boundary sharing degree between the two plots, which is:

[0089]

[0090] in, This is the actual length of the shared boundary between the two plots of land. , These are the perimeters of the two plots of land, respectively. The value ranges from [0,1]. The larger the value, the higher the degree of proximity between the two plots, which means that the physical barrier to the carbon emission spillover effect of the plots is smaller.

[0091] In an optional embodiment, constructing functional association edges between nodes corresponding to each land parcel based on the functional association relationships between land parcels includes:

[0092] For any set of plots , If the functional association between the land use types corresponding to the two plots meets the preset functional association type, then calculate the functional association strength between plot j and plot i. The preset functional association types include at least the following:

[0093] The first category is functional connections between different land use categories; that is, two plots of land belong to different land use categories, and the different land use categories have functional connections in terms of production, living, and other social functions. For example, industrial land and commercial land satisfy the connection of supporting services; commercial land and urban residential land satisfy the connection of commuting and consumption, etc.

[0094] The second category is the functional relationship between different sub-sectors within industrial land; that is, when two plots of land are both industrial land, but there is an input-output relationship between the sub-sectors they belong to, such as the upstream and downstream relationship between the textile industry and the apparel industry.

[0095] The functional association strength between plot j and plot i The calculation method is as follows:

[0096]

[0097] in, The weighting coefficients for functional relationships; For the plot of land With the plot of land Distance between plots of land; This is a distance decay function, used to characterize the pattern that the strength of functional association gradually weakens as the distance between plots increases; Functional association strength is used to characterize the strength of the functional association relationship.

[0098] In one specific implementation, the weighting coefficient of the functional association relationship The methods of obtaining it include:

[0099] If both plots are industrial land, the total consumption coefficient between the two sub-sectors is extracted based on the existing input-output table of the region and set as the weight coefficient of the functional relationship. This coefficient represents the sum of direct and indirect consumption of the final product produced by one sub-sector on the products of the other sub-sector, thus reflecting the transmission depth of the functional relationship between the sub-sectors. For example, taking the total consumption coefficient of the apparel industry on the textile industry as an example, in some regions this coefficient is 0.9, indicating that the production of the apparel industry is highly dependent on the upstream supply of the textile industry. Therefore, the stronger the transmission ability of the edge constructed based on this functional relationship, the better.

[0100] In some preferred embodiments, the total consumption coefficient is a normalized coefficient, that is, the original value of the total consumption coefficient is mapped to the [0,1] interval, so as to convert the weight coefficients of the functional relationship between different sub-sectors to the same dimension for comparison.

[0101] If the two plots belong to different land use categories and meet the preset functional association type, then the weight coefficient of the functional association relationship is set according to the association level of the land use function combination. .

[0102] The correlation level is preset based on the fundamental dependency relationship between land use functions and socio-economic factors, and is divided into three levels: strong, medium, and weak. The weight coefficients corresponding to each level are preset constants. For example:

[0103] Strong correlation: The weighting coefficient is set to 1.0, which is suitable for functional combinations with strong relationships of people flow, logistics or service supply; for example, the correlation between commercial and service land and commuting consumption in urban residential areas.

[0104] Intermediate correlation: The weighting coefficient is set to 0.6, which is suitable for functional combinations with indirect supporting service relationships; for example, the supply chain correlation between logistics and warehousing land and industrial land.

[0105] Weak correlation: The weighting coefficient is set to 0.3, which is suitable for functional combinations with weak functional connections but still potential transmission effects. For example, the correlation between commercial and service land and supporting services for industrial land.

[0106] In one specific embodiment, the distance decay function adopts a negative exponential form, as follows:

[0107]

[0108] In the formula, The distance between the centroids of plot i and plot j; The attenuation coefficient is used to control the rate at which the functional association intensity decreases as the distance between plots increases.

[0109] The functional association strength is compared with a preset cutoff threshold. The magnitude relationship between them; when the functional association strength is greater than this cutoff threshold, i.e. Then, a functional association edge is created between the two plots. And add it to the set of functional associated edges. In, and the functional correlation strength The attribute information of this edge is stored in the edge attribute matrix. When the functional association strength is not greater than this cutoff threshold, i.e. If so, the current edge is discarded. For example, a truncation threshold... for .

[0110] S200, Using the graph structure, the land parcel carbon emission change prediction model based on edge attribute enhancement is trained so as to obtain the output parameters of the model after training;

[0111] The land parcel carbon emission change prediction model is an end-to-end model structure, the main body of which is as follows: Figure 4 As shown, it includes: an edge-enhanced GAT multi-head graph attention encoder and a dual decoder; the dual decoder includes a neighbor effect decoder and a self effect decoder;

[0112] Based on the graph structure, this application utilizes an edge-attribute-enhanced multi-head graph attention encoder to learn the information transmission weights of connecting edges between different land parcels. Then, using self-effect decoders and neighbor-effect decoders with mutually isolated information, the carbon emission changes of each land parcel are converted into self-effect predictions based on the changes within the land parcel itself and neighbor-effect predictions based on changes in surrounding land parcels. The self-effect predictions and neighbor-effect predictions are added to obtain the total carbon emission change prediction for each land parcel. Based on the total carbon emission change predictions and actual values ​​for each land parcel, the trainable parameters in the model are updated using backpropagation with a joint loss function to train the land parcel carbon emission change prediction model. After training, intermediate calculation results of the model are obtained, including the attention weights of each edge and the neighbor-effect predictions for each node, which serve as the model's output parameters and participate in subsequent processing.

[0113] In this application, the edge attribute-enhanced multi-head graph attention encoder is as follows: Figure 4 As shown, it includes a first graph attention layer and a second graph attention layer connected in sequence;

[0114] The first graph attention layer receives an initial feature matrix and an edge attribute matrix; performs a trainable linear transformation on the initial feature vector of each node in the initial feature matrix and the edge attribute vector of each edge in the edge attribute matrix; for each node (as the target node), it calculates the original attention scores of each neighbor node (as the source node) and performs neighborhood normalization to obtain normalized attention weights; it aggregates the features of neighbor nodes based on the attention weights to obtain the first node representation vector corresponding to each node, which is the first neighbor enhanced representation of each node; and it outputs the first node feature matrix based on the first node representation vector corresponding to each node.

[0115] Specifically, the input feature matrix of the first graph attention layer The initial characteristic matrix in the graph structure ,Right now:

[0116] ;

[0117] In the formula, This is the input feature matrix for the attention layer of the first graph; This is the initial characteristic matrix.

[0118] The first node feature matrix is ​​output after processing by the attention layer of the first graph. ,in , Let be the feature dimension of the feature matrix of the first node; for example, The value is 128, meaning the first graph attention layer outputs a 128-dimensional first node feature matrix.

[0119] The second graph attention layer is used to receive the first node feature matrix and the edge attribute matrix; perform linear transformation and edge feature embedding on the first node representation vector of each node in the first node feature matrix and the edge attribute vector of each edge in the edge attribute matrix; for each node (as the target node), calculate the original attention score of each neighbor node (as the source node) and perform neighborhood normalization to obtain normalized attention weights; aggregate the neighbor node features based on the attention weights to obtain the second node representation vector corresponding to each node, that is, the final neighbor enhanced representation of each node; and output the second node feature matrix based on the second node representation vector corresponding to each node.

[0120] That is, the first node feature matrix output by the first graph attention layer. As input, the processed matrix outputs the feature matrix of the second node. ,in , Let be the feature dimension of the feature matrix of the second node; for example, The value is 64, meaning the second graph attention layer outputs a 64-dimensional second node feature matrix.

[0121] Let the feature matrix of the second node Let be the final neighbor enhancement representation matrix of the multi-head graph attention encoder, i.e.:

[0122] .

[0123] The final neighbor enhancement representation matrix contains the final neighbor enhancement representations corresponding to each parcel.

[0124] More specifically, the feature dimension of the first node's feature matrix is ​​128, and four independent attention heads are set in the first graph's attention layer, so the feature dimension output by each attention head is 32, that is:

[0125]

[0126] For each attention head, it has an independent node feature transformation matrix, edge attribute transformation matrix, and attention vector; the attention vector is used to perform an inner product with the combined vector formed by concatenating the target node transformation features, source node transformation features, and edge embedding vector, so as to map the high-dimensional concatenated information into a scalar original attention score, thereby quantifying the original overflow intensity of the source node to the target node.

[0127] The target node transformation feature is a high-dimensional feature representation obtained by projecting the input feature vector of the target node onto the node feature transformation matrix of the current attention head. It represents the attribute features of the target node in the feature subspace that the current attention head focuses on, and is used to calculate the attention score together with the features of the source node and the edge to quantify the sensitivity of the target node to external influences.

[0128] The source node transformation feature is a high-dimensional feature representation obtained by projecting the input feature vector of the source node onto the node feature transformation matrix of the current attention head. It represents the attribute features of the source node in the feature subspace that the current attention head focuses on. It is used to calculate the attention score together with the features of the target node and the edges to quantify the ability of the source node to transmit carbon emission spillover effects.

[0129] The edge embedding vector is a high-dimensional feature representation obtained by projecting the edge attribute vector of the edge onto the edge attribute transformation matrix of the current attention head. It represents the relational attribute features of the connecting edge in the feature subspace of the current attention head. It is used to introduce non-node information such as spatial distance, boundary sharing degree and functional association strength into the calculation of attention score, so that the attention weight obtained after normalization can reflect the influence of edge-level heterogeneity transmission conditions on spillover intensity.

[0130] If the initial feature dimension of a node is D and the feature dimension of an edge attribute is 4, then the dimension of the node feature transformation matrix corresponding to a single attention head is D×32, and the dimension of the edge attribute transformation matrix is ​​4×32.

[0131] The input of the second image attention layer is connected to the output of the first image attention layer, and the second image attention layer also has 4 attention heads. Therefore, the feature dimension output by each attention head is 16 dimensions, i.e.:

[0132]

[0133] It should be noted that the settings for the number of graph attention layers, the number of attention heads, and the model parameters in each graph attention layer in the multi-head graph attention encoder described in this application are all preferred parameters based on the application scenario of hundreds of thousands of urban land parcels. For those skilled in the art, the above settings should not be construed as a limitation on the features of the method in this application. That is, the multi-head graph attention encoder can be adapted according to the number of land parcels, average area, expected spillover range, etc. For example, the number of graph attention layers can be any value from 1 to 4, and the number of attention heads can be selected from 2 to 8. Specifically, when the average area of ​​the land parcels is small and the transmission chain of carbon emission spillover is long, the number of network layers can be increased; when the area of ​​the land parcels is large and the carbon emission spillover is mainly concentrated in directly adjacent land parcels, the number of network layers can be reduced accordingly; and when the model needs to pay more attention to a certain dimension of spatial distance or land use function correlation, the number of attention heads corresponding to that dimension can be increased accordingly.

[0134] After propagation through the two graph attention layers in the multi-head graph attention encoder, the final neighbor enhancement representation corresponding to each parcel includes not only the one-hop neighbor information of the adjacent parcels but also the two-hop neighbor information transmitted through the adjacent parcels.

[0135] The first-hop neighbor information is the attribute information contained in the first node representation vector output after the source node is directly connected to the target node through spatial adjacency edges or functional association edges, and aggregated by the first graph attention layer. This information reflects the carbon emission-related attributes of the land parcel directly associated with the target node and its direct impact on the target node. The second-hop neighbor information is the attribute information that the source node transmits to the target node after aggregating the information of its own neighbor nodes (i.e., the second-hop neighbors of the target node) through the first graph attention layer and the second graph attention layer. This information reflects the carbon emission spillover effect indirectly transmitted to the target node by non-directly connected land parcels through intermediate land parcels.

[0136] In this application, each attention head in the first graph attention layer uses independent trainable parameters during model training, and the process executed by each attention head during training is exactly the same; in an optional embodiment, for a single attention head During model training, it performs the following:

[0137] 1) Perform linear transformations on each initial feature vector in the initial feature matrix and each edge attribute vector in the edge attribute matrix to obtain transformed feature vectors and edge embedding vectors;

[0138] For a single node Corresponding input feature vector That is, the input feature matrix The first in The row feature vector is transformed using the node feature transformation matrix in the current attention head. The input feature vector Mapped to a 32-dimensional transformed eigenvector ,for:

[0139]

[0140] Based on the current attention head For the input feature matrix Each node in the process performs the above operations to obtain the transformed feature vector of each node under the current attention head.

[0141] For a single edge (Connecting nodes) and nodes (edge), utilizing the current attention head The edge attribute transformation matrix in , on this side The corresponding edge attribute vector The linear mapping is a 32-dimensional edge embedding vector. ,for:

[0142]

[0143] in, For attention head The edge embedding vector; For attention head The edge attribute transformation matrix satisfies ; For the edge The corresponding edge attribute vector, i.e., the edge attribute matrix The first in row feature vectors, satisfying ; represents the feature dimension of the edge attribute vector.

[0144] Compared to the multi-head graph attention encoders used in existing technologies, where the attention score is usually determined only by node features, this application adds edge attribute information based on spillover propagation to the existing attention mechanism, so that the attention weights calculated subsequently can effectively reflect the impact of edge attribute information on carbon emission spillover weights.

[0145] 2) Based on the transformed feature vector and the edge embedding vector, the attention weight corresponding to each edge is obtained using the attention vector in the current attention head;

[0146] Single node Set the node as the target node; based on the node association matrix in the graph structure, search for all nodes pointing to the target node. edge The corresponding source node As the target node Neighboring nodes; construct the target node The corresponding set of neighbor nodes .

[0147] Extract the set of neighbor nodes using the attention vector in the current attention head. The original attention scores of each neighbor node relative to the target node; after obtaining the original attention scores of all neighbor nodes, for each neighbor node... The original attention score is subjected to Softmax normalization to obtain the neighbor node. Relative to the target node attention weights ,for:

[0148]

[0149] in, express The exponential function; Representing neighboring nodes Relative to the target node The original attention score; Represents the set of neighboring nodes The sum of the original attention scores of all neighboring nodes; Neighboring nodes Relative to the target node The attention weights are used to represent the neighboring nodes' performance under the current attention head's judgment. To the target node The relative strength of the transmitted information; the larger the value, the stronger the neighboring node. For the target node The more significant the carbon emission spillover effect, the better.

[0150] In one specific embodiment, the method for obtaining the original attention score of a single neighbor node relative to the target node includes:

[0151] The target node Corresponding transformed feature vector (32-dimensional) Neighbor nodes Corresponding transformed feature vector (32-dimensional), and the edge embedding vector corresponding to the current edge. (32-dimensional) sequentially concatenated to obtain a combined vector, which is ;in, This indicates a concatenation operation, and the combined vector has a feature dimension of 96.

[0152] Attention vector using the current attention head By using this attention vector The transpose of the vector is multiplied by the combined vector to obtain a scalar value; the LeakyReLU activation function is then used to process this scalar value to obtain neighboring nodes. Relative to the target node raw attention score ,for:

[0153]

[0154] in, Let be the attention vector of the current attention head, which satisfies That is, the feature dimension of the attention vector is the same as the feature dimension of the combined vector, both being 96 dimensions;

[0155] This is the activation function, which represents that when the function input is greater than 0, the function output is equal to the input; when the function input is less than or equal to 0, the function output is equal to the input multiplied by a preset negative slope coefficient; for example, the negative slope coefficient is 0.2.

[0156] 3) Based on the attention weights, the transformed feature vectors of each neighboring node corresponding to the target node are combined to obtain the output feature vector corresponding to the target node;

[0157] Specifically, for a single target node The attention weights of its corresponding neighbor nodes Set as weighting coefficients; based on weighting coefficients The transformed feature vectors of each neighboring node of the target node are weighted and summed, and the ELU activation function is applied to the weighted sum to obtain the target node. The feature vector output under the current attention head is:

[0158]

[0159] In the formula, For the target node Attention The output feature vector satisfies That is, the dimension of the output feature vector under the current attention head is the same as the dimension of the transformed feature vector of the node, which is 32 dimensions.

[0160] The ELU activation function satisfies the following conditions: when the input is greater than 0, the function output is the same as the input; when the input is less than or equal to 0, the output is different. ,in This parameter controls the degree of negative value saturation and is used to adjust the output amplitude of the activation function in the negative input region; for example, its value is 1.

[0161] For each attention head in the attention layer of the first image, the above model training process is performed respectively, so that each attention head has its own independent set of trainable parameters, including: node feature transformation matrix. Edge attribute transformation matrix and attention vector The trainable parameter set is not shared among the attention heads.

[0162] After execution, for each target node Each feature vector corresponding to the output of a given attention head is obtained as follows: ;in, For the target node The feature vector output under the first attention head; And so on, without further explanation.

[0163] 4) Based on the feature vectors output by the target node under each attention head, obtain the first node feature vector output by the target node in the first graph attention layer. ;

[0164] Specifically, in obtaining the target node After processing the feature vectors output by each attention head, the feature vectors corresponding to each attention head are concatenated in order to obtain the first image attention layer's view of the target node. Complete first node feature vector ,for:

[0165]

[0166] in, The first node's feature vector has a feature dimension of 128, i.e. ;

[0167] Within the work area Each node performs the above concatenation operation to obtain the result from... The first node feature matrix is ​​formed by stacking the first node feature vectors row by row. .

[0168] In this application, the processing method performed by the second graph attention layer is exactly the same as that of the first graph attention layer, only the feature dimension is different;

[0169] Specifically, the second graph attention layer uses the first node feature matrix output by the first graph attention layer. As input.

[0170] For a single attention head in the attention layer of the second graph ( ), using node feature transformation matrix The 128-dimensional first node feature vector is mapped to a 16-dimensional transformed feature vector. ; and, using the edge attribute transformation matrix ,Will dimensional edge attribute vector Mapped to 16-dimensional edge embedding vectors .

[0171] It should be noted that during the processing of the second graph attention layer, the edge attribute vector Similar to the edge attribute vectors in the attention layer of the first graph, they are both edge attribute matrices. The eigenvectors in the model.

[0172] For the nodes at both ends of a single edge, set the first node as the target node. Set the second node as the source node. ; to target node The corresponding transformed eigenvector (16-dimensional), source node The corresponding transformed eigenvector (16-dimensional), and the edge embedding vector corresponding to the current edge. (16-dimensional) concatenations in sequence yield a combined vector (48-dimensional), which is ;in, This indicates a concatenation operation; the attention vector in the current attention head has a dimension of 48, i.e. .

[0173] Perform the same operations as in the first graph attention layer on each attention head in the second graph attention layer to obtain each target node. The feature vectors output by each attention head are concatenated in order to obtain the second graph attention layer's view of the target node. Complete second node feature vector ,for:

[0174]

[0175] Within the work area Each node performs the above concatenation operation to obtain... The second node feature vector ; will be obtained The second node feature vectors are stacked row-wise to obtain the second node feature matrix. .

[0176] Let the second node feature matrix be output by the attention layer of the second graph. The final neighbor enhancement representation matrix is:

[0177]

[0178] In the final neighbor enhancement representation matrix, the first Okay, for the plot of land The corresponding final neighbor-enhanced representation is:

[0179]

[0180] Furthermore, during the forward propagation process, this application performs a process for each edge... The second attention weights are calculated under each attention head of the second graph attention layer. ( These are saved separately for use in subsequent steps; wherein, the second attention weight is a scalar value, that is, for each edge Save the scalar value corresponding to each attention point.

[0181] Optionally, the scalar values ​​are stored using a key-value pair mapping method, that is, the key is used as the edge index and the value is a list composed of each second attention weight.

[0182] In this application, the self-effect decoder includes a first multilayer perceptron (hereinafter referred to as "first perceptron"), which is used to obtain the carbon emission change trend of the land parcel under the absence of external influence based on the land parcel's own information (excluding the attribute information and edge attribute information of neighboring land parcels).

[0183] Specifically, the self-effect decoder is as follows: Figure 4 As shown, its input is a land parcel. initial feature vector Its output is land parcels. The self-effect predicted value is:

[0184]

[0185] In the formula, For the plot of land The initial feature vector; For the plot of land The self-effect predicted value; It is the first sensor.

[0186] In one specific implementation, the first sensor includes an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence;

[0187] The input layer of the first perceptron has the following number of neurons: With the initial feature vector The feature dimensions are the same, and no activation function is applied to this layer, which is used to process the initial feature vector of the input. It is passed directly to the next level.

[0188] The first hidden layer of the first perceptron is a fully connected layer with 64 neurons, and it uses the ReLU activation function as follows:

[0189]

[0190] In the formula, This is the first weight matrix of the first perceptron; is the first bias vector of the first perceptron, and all of them are trainable parameters; For activation functions; This is the feature vector output by the first hidden layer in the first perceptron; This is the initial feature vector.

[0191] The second hidden layer is a fully connected layer with 32 neurons, and it uses the ReLU activation function as follows:

[0192]

[0193] In the formula, This is the second weight matrix of the first perceptron. is the second bias vector of the first perceptron, and all of them are trainable parameters; For activation functions; This is the feature vector output by the second hidden layer in the first perceptron.

[0194] The output layer of the first perceptron is a fully connected layer with 1 neuron, and it uses a linear activation function:

[0195]

[0196] In the formula, This is the third weight matrix of the first perceptron; This is the third bias vector of the first perceptron; For the plot of land The self-effect prediction value is used to characterize the carbon emission change predicted based solely on the attribute information of the land parcel itself, without introducing neighboring land parcel information and edge attribute information in the model input layer.

[0197] In this application, the neighbor effect decoder is set independently of the self effect decoder, and it includes a second multilayer perceptron (hereinafter referred to as the "second perceptron") for use based on the land parcel. Final Neighbor Enhancement Representation Acquired the land parcel The corresponding predicted neighbor effect value is:

[0198]

[0199] In the formula, For the plot of land The predicted value of the neighbor effect; For the plot of land The final neighbor-enhanced representation vector; It is the second sensor.

[0200] In an optional embodiment, the second perceptron has the same network structure as the first perceptron, and the neighbor effect decoder is as follows: Figure 4 As shown, it similarly includes an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence;

[0201] The second perceptron's input layer has a neuron count L and a final neighbor-enhanced representation vector. The feature dimensions are the same, and the input data is land parcels. The corresponding final neighbor augmentation representation vector That is, the graph attention encoder outputs the final neighbor enhancement representation matrix. The first in Row vectors.

[0202] The first hidden layer of the second perceptron is:

[0203]

[0204] In the formula, This is the first weight matrix of the second perceptron; is the first bias vector of the second perceptron, and all of them are trainable parameters; For activation functions; This is the feature vector output by the first hidden layer in the second perceptron.

[0205] The second hidden layer of the second multilayer perceptron is a fully connected layer, which is:

[0206]

[0207] In the formula, This is the second weight matrix of the second perceptron. is the second bias vector of the second perceptron, and all of them are trainable parameters; For activation functions; This is the feature vector output by the second hidden layer in the second perceptron.

[0208] This application sets up a second hidden layer and uses the second hidden layer to perform a nonlinear transformation on the final neighbor enhancement representation vector output by the first hidden layer, so as to extract the high-order interaction relationship between the independent contributions of each neighbor node and enhance the decoder's ability to fit complex carbon emission spillover patterns.

[0209] The output layer of the second perceptron is a fully connected layer with 1 neuron, and it uses a linear activation function:

[0210]

[0211] In the formula, This is the third weight matrix of the second perceptron. This is the third bias vector of the second perceptron. For the plot of land The neighbor effect prediction value, which is used to characterize the plot. The change in carbon emissions received through spatial and functional linkages due to land use renewal activities in surrounding areas.

[0212] like Figure 4 As shown, the land plot carbon emission change prediction model will target the land plot. The self-effect prediction value and the neighbor effect prediction value are superimposed to obtain the target plot. The projected total change in carbon emissions at the target time point is:

[0213]

[0214] in, The predicted total change in carbon emissions for target site i; Let i be the self-effect prediction value of the target plot i; Let be the predicted value of the neighbor effect for target plot i.

[0215] Since it is impossible to accurately collect carbon emission changes caused solely by the site itself, as well as carbon emission changes caused by neighboring sites, this application uses the true value of total carbon emission changes in model training. The ground truth for model training, i.e., before model training, is the true value of the total change in carbon emissions of plot i at the target time point, obtained based on the carbon emissions of plot i at the baseline and target time points, and used as the training ground truth for model training, as follows:

[0216]

[0217] in, Let i be the carbon emissions of plot i at the baseline time point. Let i be the carbon emissions of plot i at the target time point; This represents the true value of the total change in carbon emissions of plot i at the target time point, which is also the training true value.

[0218] Based on this, this application constructs a joint loss function based on the self-effect and the neighbor effect of carbon emissions, which automatically separates the constraints on the self-effect prediction value and the neighbor effect prediction value during the model training process.

[0219] In one specific embodiment, the joint loss function is:

[0220]

[0221] in, The joint loss function is used to characterize the overall optimization objective that needs to be minimized during model training. It is composed of the prediction accuracy loss and the multi-head differential regularization term through a weighted sum.

[0222] The prediction accuracy loss function is used to characterize the deviation between the model's predicted value and the actual value of the total change in carbon emissions in each plot, so as to force the sum of the outputs of the self-effect decoder and the neighbor effect decoder to numerically approximate the actual change in carbon emissions.

[0223] This is a multi-head differentiation regularization term used to characterize the similarity of the attention distributions learned by each attention head. When the attention distributions learned by different attention heads tend to be similar, this term applies a penalty, forcing each attention head to focus on neighborhood relationships in different dimensions, avoiding the multi-head effect from degenerating into a single-head effect, thereby enhancing the model's ability to capture multiple spillover propagation paths in parallel;

[0224] The weight is a multi-head differentiated orthogonal weight, and its value ranges from 0.01 to 0.04; in a preferred embodiment, the weight is set to 0.01.

[0225] In one specific implementation, the prediction accuracy loss function Calculated using the mean square error form, it is:

[0226]

[0227] In the formula, Let be the true value of the total change in carbon emissions for plot i at the target time point. Let be the predicted total change in carbon emissions for plot i at the target time point.

[0228] By constructing the prediction accuracy loss function described above This makes the model's predicted total change in carbon emissions... Numerically close to the true value This allows for constraints on the predicted total change in carbon emissions.

[0229] In an optional embodiment, the multi-head differential regularization term is:

[0230]

[0231] in, In the second layer of the graph attention network, the first... One's attention is on the side Normalized attention weights on; In the second layer of the graph attention network, the first... Attention is focused on the same side. Normalized attention weights on the graph; M is the total number of edges involved in the computation in the second-layer graph attention network, that is, the total number of directed connections involved in message passing in the second-layer graph attention network; It is the set of all directed connections in the second-layer graph attention network.

[0232] The mechanism of the multi-head differentiation regularization term is as follows: it imposes constraints by comparing the consistency of the attention weight distribution of any two attention heads on each directed connection edge. Specifically, if the weight distribution assigned to each edge by the two attention heads tends to be the same, the larger the regularization term, the more consistent the attention focus of the two attention heads on each edge, resulting in attention head degradation. Conversely, if the weight distribution of the two attention heads on each edge differs significantly, the smaller the regularization term, the more differentiated the attention focus of the two attention heads has become. By minimizing this regularization term, the weight distribution of different attention heads is forced to differentiate from each other, thereby guiding each attention head to focus on different spillover transmission factors. For example, the first attention head focuses on spatially close neighbors, the second attention head focuses on neighbors with strong land use function associations, and other attention heads capture interaction patterns not covered by the first two. Thus, the model can extract information from multiple spillover transmission paths in parallel, enhancing its ability to represent differentiated spillover effects.

[0233] Compared to existing technologies, this application creatively introduces a self-effect decoder and a neighbor-effect decoder into the model. The self-effect decoder takes only the initial feature vector of the land parcel itself as input. This vector is directly taken from the initial node feature matrix constructed in step S100, without any graph aggregation operations, thus completely isolating it from neighboring land parcels at the information level. Based on this input structure, it is ensured that the self-effect decoder can only reason based on the land parcel's own geometric attributes, land use type, development intensity, carbon emission information, etc., making it structurally insensitive to the existence and changes of neighboring land parcels. Therefore, during training, the self-effect decoder can only output the carbon emission change determined by the evolution of the land parcel's own attributes.

[0234] The neighbor effect decoder takes only the final neighbor-enhanced representation vector of the land parcels, aggregated by the multi-head graph attention encoder, as input. This representation vector is obtained through two layers of message passing, highly integrating information from one-hop and two-hop neighbors, with the fusion weight of each neighbor's information differentiated by the edge attribute-enhanced attention mechanism. Based on this input structure, the neighbor effect decoder tends to capture the carbon emission change contributed by the carbon emission spillover effect of surrounding neighboring land parcels.

[0235] Although the two decoders have similar structures, their parameters are independent of each other to match the different input sources of the two decoders, and the information representation also differs significantly.

[0236] The sum of the outputs constitutes the predicted total change in carbon emissions for the land parcel. Constrained by the loss in prediction accuracy, the predicted total change must approximate the true value (i.e., the true observed value). Therefore, during training, the model assigns variations explainable by its own attributes to the self-effect decoder and variations explainable only by neighbor information to the neighbor effect decoder. This structural division of labor replaces the dependence on independent ground values ​​for self-effects and neighbor effects, allowing the decomposition process to be completed automatically without additional supervisory labels.

[0237] All land parcels within the work area are divided using a spatially continuous partitioning method based on a preset proportional relationship to obtain training, validation, and test sets.

[0238] In one optional embodiment, the specific methods for obtaining the training set, validation set, and test set include:

[0239] First, the entire work area is divided according to a regular grid or administrative boundaries to create several spatially contiguous blocks. Using these blocks as units, the corresponding land parcels are randomly assigned to three sets—training set, validation set, and test set—according to a preset ratio. This ensures that the number of parcels in the training set, validation set, and test set conforms to a preset ratio. For example, the preset ratio is 7:1.5:1.5, meaning that the land parcels are divided according to a 7:1.5:1.5 ratio to construct the training set, validation set, and test set respectively.

[0240] Optionally, this application employs the Adam optimizer to update model parameters during the training process using the training set; and after every few training rounds, it calculates the prediction accuracy loss using the validation set. If, after several consecutive rounds, the prediction accuracy loss does not reach a new historical low, training is terminated, and the model parameters are restored to the parameters at which the validation loss was optimal. This application, by setting the aforementioned early stopping mechanism, can prevent overfitting.

[0241] After the model training is completed, all trainable parameters in the model are fixed so that they are no longer updated, thus obtaining the current trained model.

[0242] Then, using the complete graph structure data Perform a forward propagation, extracting the following during the forward propagation process:

[0243] 1) The attention weights calculated for each edge by all attention heads in the attention layer;

[0244] 2) Neighbor effect decoder outputs neighbor effect prediction values ​​for each plot.

[0245] The attention weights and neighbor effect predictions extracted above are used as the output parameters of the model.

[0246] S300, Based on the output parameters of the model, extract the carbon emission spillover intensity corresponding to each edge;

[0247] Specifically, for each edge The attention weights are calculated for each attention head in the second graph attention layer. Take the average to obtain the comprehensive attention weight for each edge. The intensity of carbon emission spillover corresponding to each edge is defined as follows:

[0248] in, For the second attention layer One attention point ( The attention weights corresponding to ) range from [0,1] and are used to characterize the edge attention mechanism. The transmission intensity of carbon emission spillover effects between the two connected plots.

[0249] S400, based on the distribution of outgoing edges, extract each candidate source node from the graph structure; identify the changed land parcels from the candidate source nodes, and perform update type grouping on each changed land parcel;

[0250] The term "outgoing edge" refers to the existence of at least one message passing edge from a single node to other nodes, meaning that the plot corresponding to the node has a connection relationship in the graph structure to pass information to other plots; the term "candidate source node" refers to a node with outgoing edges.

[0251] To efficiently obtain the correlation between the regeneration activities of various land parcels and the carbon emission spillover effects, it is necessary to obtain the regeneration type of the land parcels corresponding to each candidate source node (hereinafter referred to as "candidate source node land parcels") at the target time point, including:

[0252] From the initial feature vector of the land parcel corresponding to each candidate source node, extract the land use type of the land parcel at the reference time and the land use type at the target time. By comparing whether the two are consistent, identify whether the land use type of the land parcel has changed. Based on the preset attribute change filtering conditions, extract the updated land parcels that meet the filtering conditions, and divide the filtered updated land parcels and their corresponding outgoing edges into different update type groups according to the update type.

[0253] Specifically, the attribute change filtering conditions include:

[0254] Screening criterion 1: The major land use category has changed between the baseline time point and the target time point, such as changing from "industrial" to "commercial and service".

[0255] Screening criterion 2: The major land use category remains "industrial", but the minor land use category has changed between the baseline time point and the target time point, such as changing from "textile industry" to "computer communication and electronics".

[0256] After acquiring the updated land parcel, extract the corresponding update type for the updated land parcel. The update type includes the change in the major land use type before and after the land parcel update. For example, when the land use type of the land parcel changes from "industrial" to "commercial and service", the corresponding update type is "industrial to commercial and service". When the major land use type of the land parcel remains "industrial land", but the minor land use type changes from "textile industry" to "computer, communication and electronics", the corresponding update type is "textile industry to computer, communication and electronics".

[0257] The outgoing edges corresponding to the updated land parcels are assigned to the corresponding update types to construct corresponding update type groups. For example, the outgoing edges of the updated land parcels that will be changed from "industrial" to "commercial" are assigned to the "industrial to commercial" group.

[0258] S500, Based on the update type group, and based on the carbon emission spillover intensity corresponding to each edge and the neighbor effect prediction value, extract the carbon emission spillover effect measurement results corresponding to each update type;

[0259] The carbon emission spillover effect measurement results include at least: the direction and intensity of the carbon emission spillover effect, as well as the distance decay curve and confidence interval of the carbon emission spillover effect.

[0260] Yes, for each partitioned update type group, it collects the "distance-intensity" data pairs corresponding to each outgoing edge in that update type group. ;in, This represents the distance between the plots connected by the outgoing edge. The comprehensive attention weight corresponding to the outgoing edge is used to characterize the carbon emission spillover intensity corresponding to the outgoing edge.

[0261] Curve fitting is performed on the distance-intensity data pairs corresponding to each outgoing edge in the update type group to obtain the distance decay curve corresponding to that update type group; for example, the distance decay curve is as follows: Figure 6 As shown, the vertical axis represents the carbon emission spillover intensity, and the horizontal axis represents the distance between plots; each data pair is fitted, and the fitted curve is set as the distance decay curve.

[0262] Specifically, assuming the update type group is Having in this update type group Outgoing edges; for update type groups For each outgoing edge in the array, obtain its corresponding "distance-intensity" data pair, which is: the distance between the plots connected by the outgoing edge. and comprehensive attention weight ;

[0263] Based on the distance of the land parcel and comprehensive attention weight Each outgoing edge is used to construct a sample point to obtain the update type group. The corresponding set of sample points is:

[0264]

[0265] in, For the set of sample points; In order to be with the first Sample points corresponding to each outgoing edge; For the first The distance between the plots corresponding to the outgoing edges; For the first The overall attention weight corresponding to each outgoing edge.

[0266] Since the overall attention weight is usually affected by factors such as land parcel attributes and functional association strength, different outgoing edges with the same or similar distances may have significantly different overall attention weights. Therefore, when constructing the decay fitting curve, it is not possible to directly connect the discretely distributed sample points in sequence.

[0267] In a specific embodiment of this application, a local weighted regression method is required to smoothly fit each sample point in the same update type group, including: for a certain distance to be estimated First, select the sample points from all sample points in the update type group that are at a distance from the current location. Recent Each sample point is used as a local neighborhood.

[0268] Optional, The quantity is the product of the total number of sample points in the update type group and a preset bandwidth parameter; wherein, the bandwidth parameter is used to characterize the proportion of the local neighborhood size to the total number of sample points in the update type group, that is, to control the number of sample points participating in local weighted regression, and its size affects the smoothness of the fitting curve; for example, the bandwidth parameter is 0.1, that is:

[0269]

[0270] in, This represents the number of sample points within the local neighborhood. The total number of sample points within the update type group; This indicates rounding up to the nearest integer.

[0271] Then, the distance between each sample point in the local neighborhood and the distance to be estimated are calculated. The relative distance difference between them is:

[0272]

[0273]

[0274] in, Distance to be estimated The set of sample points within a local neighborhood; The maximum distance difference within the local neighborhood; For any sample point within the local neighborhood The corresponding distance to the land parcel; The distance to be estimated;

[0275] Let be the relative distance difference between sample points within a local neighborhood, and let be the distance between any sample point within the local neighborhood. Distance to be estimated The ratio between the distance difference between two samples and the maximum distance difference within their local neighborhood, i.e., for any sample point... Distance to be estimated within the local neighborhood The normalized distance between them ranges from [0,1].

[0276] Furthermore, for each sample point within the local neighborhood, based on the relative distance difference between the sample points, a cubic kernel function is used to construct the fitting weights corresponding to the sample points, as follows:

[0277]

[0278] In the formula, Distance to be estimated The fitting weights of the corresponding sample points; This represents the relative distance difference between sample points within a local neighborhood; when the distance between sample points and their corresponding plots... The closer to the distance to be estimated The larger the fitting weight, the greater the corresponding fitting weight; conversely, the smaller the fitting weight, the greater the fitting weight.

[0279] Based on the fitting weights, each sample point in the local neighborhood is weighted, and a first-order polynomial of weighted least squares is used to fit the weighted sample points to obtain the distance to be estimated. The smoothed estimate at that point.

[0280] Within the range of distances to be estimated (usually from the minimum to the maximum distance between plots), select multiple distances to be estimated at equal intervals (e.g., 500 uniformly distributed distance values). Perform the above steps for each distance to be estimated to obtain smoothed estimates at all corresponding distances. Connect these estimates in order of distance to obtain a complete empirical distance decay curve.

[0281] This application constructs an empirical distance decay curve by employing a locally weighted regression method, which presents the curve in a non-parametric form. This eliminates the need to pre-define the decay curve as a specific curve such as an exponential function, a reciprocal distance function, or a Gaussian function. The function type can be data-driven based on the attention weights of the updated type group, thereby enabling the constructed empirical distance decay curve to more closely resemble the curve shape of the sample data itself.

[0282] After obtaining the empirical distance decay curve, the Bootstrap resampling method is used to estimate the confidence interval, i.e., for the update type group. In Each edge is resampled multiple times with replacement from the original sample points, and a locally weighted regression is performed on the resampled data that is exactly the same as the original sample data to obtain the corresponding empirical distance decay curve. Repeat execution After this, the complete empirical distance decay curves are obtained as follows:

[0283]

[0284] in, To repeat the empirical distance decay curve obtained from the first execution, To repeat the first The empirical distance decay curve obtained at this time; it should be noted that, in actual execution, It can be set to 500 times, 1000 times or more; preferably, It can be used 1000 times.

[0285] For any plot of land distance Obtain the distance of all curves to this plot of land. The estimated value at that location is There are 10 different estimates, namely:

[0286]

[0287] The estimated values ​​are sorted from smallest to largest, and the 2.5% quantile is taken as the lower bound of the 95% confidence interval, and the 97.5% quantile is taken as the upper bound of the 95% confidence interval. For a single estimated value, its corresponding upper and lower bounds are connected to construct the 95% confidence interval band of the distance decay curve corresponding to the update type. This band is used to characterize the statistical robustness of the decay law. That is, the narrower the interval, the more sufficient the sample size of the update type group and the more reliable the law. The wider the interval, the fewer the samples of the update type group or the greater the variability of the law itself.

[0288] In an optional embodiment, the method for determining the direction and intensity of the carbon emission spillover effect includes:

[0289] Based on the neighbor effect prediction value of the neighboring plots corresponding to the updated plots in each update type group and the carbon emission spillover intensity corresponding to each edge, the carbon emission spillover direction and spillover intensity corresponding to each update type group are determined.

[0290] Specifically, each updated plot in the update type group is designated as a source node plot. Based on the comprehensive attention weight corresponding to each edge obtained in step S300, each target node plot corresponding to the source node plot is filtered, and only target node plots that meet the preset influence strength condition are designated as valid target node plots. All valid target node plots in the update type group are then obtained. The preset influence strength condition is as follows: for any source node plot k in the current update type group, any corresponding target node plot t is obtained; if the comprehensive attention weight of plot k to plot t is greater than or equal to the maximum value among the comprehensive attention weights contributed by all source node plots to plot t, then plot t is determined to meet the preset influence strength condition and is included in the set of valid target node plots of the update type group.

[0291] To facilitate understanding of the preset influence intensity condition, Figure 5 Taking this as an example, the implementation process of filtering effective target node plots based on the preset influence intensity conditions is explained in detail, including:

[0292] For the current source plot k1, extract each target plot it can affect, including t1, t2, and t3; for target plot t1, obtain the comprehensive attention weights of all source plots pointing to target plot t1, including: k1—0.6, k2—0.2, k3—0.1, k5—0.1; extract the maximum value (k1—0.6) among the comprehensive attention weights; compare the comprehensive attention weight of k1 to plot t1 with the maximum value. Since the comprehensive attention weight of k1 to plot t1 is the same as the maximum value, it is determined that target plot t1 meets the preset influence intensity condition (marked in blue in the figure); for target plot t2, similarly obtain the comprehensive attention weights of all source plots pointing to target plot t1, and extract the maximum value, which is k6—0.6; since the comprehensive attention weight of k1 to plot t1 (0.3) is less than the maximum value (0.6), it is determined that target plot t2 does not meet the preset influence intensity condition (marked in black in the figure). Similarly, for target plot t3, since the comprehensive attention weight of k1 to plot t3 (0.3) is less than the maximum value (0.7), it is determined that target plot t3 does not meet the preset influence intensity condition (marked in black in the figure).

[0293] This application ensures that the neighbor effect prediction value of the target node plot included in the spillover direction statistics is mainly contributed by the current source node plot by constructing the above-mentioned influence intensity conditions, thus excluding the interference of other mixed sources.

[0294] For a single target node plot j among all valid target node plots, extract the neighbor effect prediction value corresponding to target node plot j. ;

[0295] The arithmetic mean of the predicted neighbor effects for each target node plot is calculated as follows:

[0296]

[0297] in, This represents the number of target node plots within the current update type group after the above filtering.

[0298] It is the intra-group arithmetic mean of the neighbor effect prediction values ​​corresponding to each target node plot. The sign of the value can reflect the spillover direction of the source plot to its surrounding plots under this update type.

[0299] It should be noted that the predicted neighbor effect value of a single effective target node plot, which characterizes the net increment of the influence of all its external source plots on that plot, may still include residual interference components from source node plots that are not of the same update type. However, through the screening based on the preset influence intensity condition, the sample included in the statistics has been limited to the current source node plot that contributes the largest weight among all external sources of its target node plot, that is, the current source node plot is the dominant contributor to the neighbor effect change of the target node plot. On this basis, in the set of effective target node plots after screening, the interference components of source node plots that are not in the same group have been substantially weakened, and their residual effects are further canceled out when the mean is calculated by grouping by update type. Therefore, the mean within the group is... The symbol can reliably reflect the direction of the average net external impact of the source node plot of this update type on the surrounding plots in a statistical sense.

[0300] when When negative, it indicates that the carbon emissions of surrounding land parcels decrease under the current land parcel renewal type. This means the renewal of the source node land parcel has a positive spillover effect on the carbon emissions of surrounding land parcels. For example, if a source node land parcel changes from high-carbon-emission chemical plant land to low-carbon-emission computer, communication, and electronic equipment manufacturing land, the change in land use types in the surrounding area is transmitted through the industrial chain, thereby reducing the carbon emissions of the surrounding land parcels. When the value is positive, it indicates that the carbon emissions of the surrounding plots are increasing under the current plot renewal type, meaning that the carbon emission effect is a negative spillover.

[0301] Perform a one-sample two-tailed test on the above within-group arithmetic mean. The test is used to determine whether the degree of deviation from zero is statistically significant;

[0302] For example, let's assume that This means that the current update type group has no significant carbon emission spillover effect. Setting the significance level to 0.05, then the two-sided... The test statistic is:

[0303]

[0304] in, The sample standard deviation of the neighbor effect prediction values ​​corresponding to each effective target node plot;

[0305] According to the two sides The test statistic, combined with the degrees of freedom, is obtained through a query. Distribution table determines two-sided Value; based on both sides The relationship between the value and the significance threshold, combined with the zero-biasedness of the neighbor effect prediction value, determines the carbon emission spillover effect corresponding to the current update type group;

[0306] Specifically, if and If the current update type group has a significant positive carbon emission spillover effect, it means that the update of the source node plot can drive carbon emission reduction of the surrounding plots.

[0307] like and If the current update type has a significant negative carbon emission spillover effect, it is determined that the update of the source node plot leads to an increase in carbon emissions from surrounding plots.

[0308] like If the current update type has no significant statistical effect on carbon emission spillover, then there is no sufficient evidence to prove that the update of the source node plot has a promoting or inhibiting effect on the carbon emissions of the surrounding plots.

[0309] In this application, the method for obtaining the relative overflow intensity index includes: for the current update type group The average of the comprehensive attention weights corresponding to each inner edge is taken, and the mean of the average is set as the relative spillover intensity index, i.e.:

[0310]

[0311] in, The relative spillover intensity index is a dimensionless value in the range [0,1]. The larger the value, the greater the relative carbon emission spillover intensity, meaning that the current update type has a higher overall attention weight to neighboring plots compared to other update types.

[0312] This application, by calculating a relative spillover intensity index, enables horizontal comparisons of carbon emission spillover intensity across different land renewal types, thereby better supporting the prioritization of low-carbon land renewal models in decision-making; for example, prioritizing certain models during decision-making. Larger and The update type is advanced for negative values ​​to maximize the synergistic emission reduction effect within the region.

[0313] Furthermore, the method described in this application is as follows: Figure 7 As shown, it also includes:

[0314] S600, based on the direction and intensity of the carbon emission spillover effect, as well as the distance decay curve and confidence interval of the carbon emission spillover effect, perform subsequent low-carbon national spatial planning and analysis;

[0315] Specifically, in regional low-carbon decision-making, based on the direction and intensity of carbon emission spillover, the types of land renewal activities that should be prioritized, proceeded cautiously, or postponed are initially selected.

[0316] And, distance decay curves based on various land use renewal activity types. And a 95% confidence interval, used for reference when delineating the impact range of land renewal projects, and to support the determination of how far the impact of a certain type of land renewal activity can extend to the surrounding carbon emissions.

[0317] To address the aforementioned technical problems in the prior art, the present invention also provides a terminal for implementing the method for measuring the spatial spillover effect of carbon emissions at the plot scale as described above.

[0318] Please see Figure 8 The diagram shows the structure of the terminal; as shown below. Figure 8 As shown, the terminal 40 includes a processor 41 and a memory 42 connected to each other; the memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory, so that when the terminal is executed, it can realize all the steps in the method for measuring the spatial spillover effect of carbon emissions at the plot scale as described above.

[0319] The specific steps of the method for measuring the spatial spillover effect of carbon emissions at the plot scale, and the beneficial effects obtained by applying the terminal provided in this embodiment are the same as those in the above embodiments, and will not be repeated here.

[0320] It should be noted that the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Similarly, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0321] This invention also provides a computer-readable storage medium storing a computer program that, when called by a processor, implements the method for determining the spatial spillover effect of carbon emissions at the plot scale as described in the above embodiments.

[0322] A computer-readable storage medium can be a tangible device capable of holding and storing instructions used by an instruction execution device. Computer-readable storage media can be, for example, (but not limited to) electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, and mechanical encoding devices.

[0323] The computer-readable program described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network, to an external computer or external storage device. A network adapter card or network interface in each computing / processing device receives computer-readable program instructions from the network and forwards these instructions to the computer-readable storage medium in the respective computing / processing device.

[0324] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for measuring the spatial spillover effect of carbon emissions at the plot scale, characterized in that, include: Using each parcel as a node, a parcel graph structure is constructed, which includes spatial adjacency edges and functional association edges; the graph structure includes the initial feature vector corresponding to each node and the edge attribute vector corresponding to each edge; Based on the graph structure, a land parcel carbon emission change prediction model based on edge attribute enhancement is trained to obtain the model's output parameters after training; wherein, the model's output parameters include the neighbor effect prediction values ​​of each parcel and the attention weights of each edge; Based on the attention weights, the carbon emission spillover intensity corresponding to each edge is obtained; The changed land parcels are identified in the graph structure, and each changed land parcel is divided into update type groups; the update type group is a combination of changes that meets the change screening conditions. Based on the update type group, and based on the carbon emission spillover intensity corresponding to each edge and the neighbor effect prediction value, extract the carbon emission spillover effect measurement results corresponding to each update type, including at least: the direction and intensity of the carbon emission spatial spillover effect, as well as the distance decay curve and confidence interval.

2. The method for measuring the spatial spillover effect of carbon emissions according to claim 1, characterized in that, The land parcel carbon emission change prediction model based on edge attribute enhancement includes: A multi-head graph attention encoder with edge attribute enhancement is used to obtain the final neighbor enhancement representation of each plot and the attention weights corresponding to each edge; The self-effect decoder is used to obtain the corresponding self-effect prediction value based on the initial feature vector of the land parcel; The neighbor effect decoder, connected to the multi-head graph attention encoder, is used to obtain the neighbor enhancement representations of each plot and obtain the corresponding neighbor effect prediction values. The training of the land parcel carbon emission change prediction model based on edge attribute enhancement, based on the graph structure, includes: The graph structure is input into a multi-head graph attention encoder with edge attribute enhancement to obtain the neighbor enhancement representation of each parcel and the attention weight corresponding to each connecting edge; the initial feature vector of each parcel is input into the self-effect decoder, and the neighbor enhancement representation of each parcel is input into the neighbor effect decoder to obtain the self-effect prediction value and the neighbor effect prediction value of each parcel respectively; the self-effect prediction value and the neighbor effect prediction value are added to obtain the total carbon emission change prediction value of each parcel. The self-effect prediction value is added to the neighbor effect prediction value to obtain the total carbon emission change prediction value for each plot. Based on the total carbon emission change prediction value and the actual value for each plot, the trainable parameters in the model are updated by backpropagation using the joint loss function.

3. The method for measuring the spatial spillover effect of carbon emissions according to claim 2, characterized in that, The joint loss function is: ； in, For the joint loss function; The prediction accuracy loss function is used to constrain the predicted value of the total change in carbon emissions to approximate the true value of the total change in carbon emissions. This is a multi-head differentiation regularization term used to constrain the similarity of the attention distributions learned by each attention head, so as to force each attention head to focus on neighborhood relationships in different dimensions. It is a multi-head differentiated orthogonal weight.

4. The method for measuring the spatial spillover effect of carbon emissions according to claim 1, characterized in that, The method for obtaining the initial feature vector includes: Obtain carbon emission information for each plot at the baseline time point, and obtain plot attribute information for each plot at the baseline time point and the target time point respectively; The carbon emission information and the land parcel attribute information are combined to construct an initial feature vector corresponding to each land parcel; wherein, the land parcel attribute information includes at least land parcel graphic information, land use type information, and land parcel development information; The plot graphic information includes at least one of area, perimeter, compactness, fractal dimension, shape index, and aspect ratio; the land use type information includes primary land use type and secondary sub-industry type corresponding to industrial land; the plot development information includes at least one of plot ratio, distance to city center, and distance to nearest main road.

5. The method for measuring the spatial spillover effect of carbon emissions according to claim 1, characterized in that, The method for constructing spatial adjacent edges includes: Delaunay triangulation is performed based on the centroid coordinates of each plot to obtain candidate connecting edges; for the candidate connecting edges connecting plot i and plot j, the equivalent radius is determined based on the area of ​​plot i and plot j respectively, and the dynamic distance truncation threshold is determined by combining the preset scaling factor and the minimum truncation distance; when the centroid Euclidean distance between plot i and plot j is not greater than the dynamic distance truncation threshold, a spatial adjacency edge between plot i and plot j is established. The dynamic distance cutoff threshold is: ； in, For the plot of land and plot of land Dynamic truncation thresholds between; This is an empirical proportionality coefficient; This is the preset minimum cutoff distance.

6. The method for measuring the spatial spillover effect of carbon emissions according to claim 1, characterized in that, The method for constructing the functional association edge includes: Based on the preset functional association type between the two land use categories, or based on the input-output association between the two industrial land parcels, the weight coefficient of the functional association relationship is determined; based on the product of the weight coefficient of the functional association relationship and the distance decay function between land parcel i and land parcel j, the functional association strength is determined; when the functional association strength is greater than the preset cutoff threshold, a functional association edge is established between land parcel i and land parcel j.

7. The method for determining the spatial spillover effect of carbon emissions according to claim 6, characterized in that, The method for determining the weight coefficients of the functional association relationships includes: If both plots are industrial land, the total consumption coefficient between the two sub-sectors is extracted based on the existing input-output table of the region and set as the weight coefficient of the functional relationship. This coefficient is used to characterize the sum of direct and indirect consumption of the products of the other sub-sector by the production unit of the final product of one sub-sector. If the two plots belong to different land use categories and meet the preset functional association type, then the weight coefficient of the functional association relationship is set according to the preset association level of the land use function combination.

8. The method for measuring the spatial spillover effect of carbon emissions according to claim 1, characterized in that, The method for obtaining the distance decay curve and confidence interval includes: For each update type group, sample points are constructed using the distances between plots connected by each outgoing edge of the group as independent variables and the comprehensive attention weight of the outgoing edges as dependent variables. For the distance to be estimated, local samples that are close to the distance to be estimated are selected from the update type group. The fitting weights are determined based on the relative distance between the local samples and the distance to be estimated, and the weighted least squares method is used to regress the local samples to obtain a smooth estimate of the distance to be estimated. The smooth estimates of multiple distances to be estimated are connected in order of distance to obtain the distance decay curve. The confidence interval of the distance decay curve is obtained by using the Bootstrap resampling method.

9. The method for measuring the spatial spillover effect of carbon emissions according to claim 1, characterized in that, The method for obtaining the direction and intensity of the carbon emission spatial spillover effect includes: Summarize the predicted neighbor effect values ​​of target plots affected by the corresponding source plots according to the plot renewal type; The spillover direction of the corresponding land parcel update type is determined based on the sign and significance of the aggregated neighbor effect prediction values, and the relative spillover intensity is determined based on the attention weight of the connecting edges within the corresponding land parcel update type group.

10. A terminal, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the carbon emission spatial spillover effect measurement method according to any one of claims 1 to 9.

Images