City Carbon Emission Prediction Method and Device Based on Collaborative Spatiotemporal Convolutional Network

Abstract

The present invention discloses a method and device for urban carbon emission prediction based on a collaborative spatio-temporal convolutional network. The method includes: taking urban grids as nodes, and constructing a heterogeneous graph of multi-perspective correlations of the city from the perspectives of spatial connectivity, geographical proximity, and carbon emission time similarity between the grids; parallelly inputting urban carbon emission and traffic density data into the heterogeneous graph, and respectively using 1 spatio-temporal convolutional module to extract the spatio-temporal features of carbon emissions and the spatio-temporal features of traffic density from the heterogeneous graph with carbon emission and traffic density data; extracting the dependence of traffic density on carbon emissions from the spatio-temporal features of traffic density, and fusing the dependence of traffic density on carbon emissions with the spatio-temporal features of carbon emissions to obtain the global spatio-temporal features of carbon emissions; combining the global spatio-temporal features of carbon emissions with external factors to perform collaborative prediction of urban carbon emissions. The present invention can improve the accuracy of urban carbon emission prediction by predicting urban private-car-based carbon emissions in a collaborative manner.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and particularly relates to a method and device for predicting urban carbon emissions based on a collaborative spatio-temporal convolutional network. Background Art

[0002] At present, with the continuous increase in the number of vehicles, especially private cars, a series of problems such as traffic congestion, excessive carbon dioxide and other pollutant emissions have emerged. The ecological environment problems caused by vehicle exhaust emissions have become increasingly prominent and have become a hot topic of social concern.

[0003] Studies at home and abroad have shown that the excessive emissions of greenhouse gases such as CO2 are the fundamental cause of the greenhouse effect. According to the report released by the Intergovernmental Panel on Climate Change (IPCC), the carbon emissions in 2019 reached 59 billion tons, and the global climate has entered a state of emergency. Among them, urban transportation (mostly private cars) has always been considered the main source of greenhouse gas emissions, and the carbon emissions caused account for 90% of the urban air pollution. In other words, private cars are the largest producers of urban carbon emissions. The large amount of carbon emissions from private cars has caused serious urban environmental problems and brought great pressure to achieve the urban emission reduction goal.

[0004] In this regard, the carbon emissions caused by urban private cars are the fundamental cause of serious climate change and serious environmental problems. Therefore, fine monitoring and prediction of urban carbon emissions are of great significance for controlling urban traffic pollution and achieving carbon neutrality. In addition, the carbon emission prediction of private cars helps traffic managers and decision-makers improve the design of urban traffic infrastructure. For example, if the vehicle emissions in urban areas can be obtained in real time, early warnings of regional pollution can be issued to help the relevant government improve the design of urban traffic infrastructure. According to the predicted carbon emissions in the urban environment, traffic managers can implement strategies in a city or region, such as using low-carbon energy to replace vehicle fuel, so as to achieve positive and negative carbon emission neutralization.

[0005] Due to the complex spatio-temporal dependence of the carbon emissions generated by private cars, which shows spatial uncertainty and temporal variation non-linearity in the spatio-temporal dimension, traditional prediction models cannot accurately predict carbon emissions well. At the same time, most of the existing studies do not consider the influence of driving characteristics and other external factors, such as weather, holidays, events, accidents, etc., and the effectiveness of their prediction results cannot be guaranteed.

[0006] In recent years, the main challenge of current research problems lies in how to fully explore the spatio-temporal variation characteristics of carbon emissions between cities, capture their highly non-linear spatial correlation and dynamic variability in the time domain. In addition, the urban transportation network is a complex and huge system. Different types of traffic data record the change information of carbon emissions from different perspectives. With the rapid development of data mining technology, how to deeply and fully explore the potential patterns and dynamics of urban carbon emissions changes from the auxiliary data collected from multiple aspects for carbon emission collaborative prediction, and then improve the effectiveness of its prediction performance, these problems urgently need further research. Summary of the Invention

[0007] The present invention provides a method for predicting urban carbon emissions based on a collaborative spatio-temporal convolutional network, constructs a multi-perspective heterogeneous graph based on urban carbon emissions and traffic density, and performs collaborative prediction of urban carbon emissions based on the multi-perspective heterogeneous graph.

[0008] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0009] A method for predicting urban carbon emissions based on a collaborative spatio-temporal convolutional network, comprising:

[0010] Taking urban grids as nodes, and constructing a heterogeneous graph of multi-perspective correlation of the city from the perspectives of spatial connectivity, geographical proximity and carbon emission time similarity between grids;

[0011] Inputting urban carbon emission and traffic density data into the heterogeneous graph respectively, and each using 1 spatio-temporal convolutional module to extract the spatio-temporal characteristics of carbon emissions and the spatio-temporal characteristics of traffic density;

[0012] Extracting the dependence of traffic density on carbon emissions from the spatio-temporal characteristics of traffic density, and fusing the dependence of traffic density on carbon emissions with the spatio-temporal characteristics of carbon emissions to obtain the global spatio-temporal characteristics of carbon emissions;

[0013] Combining the global spatio-temporal characteristics of carbon emissions with current external factors to perform collaborative prediction of urban carbon emissions.

[0014] Further, the urban map is divided into several grids by using the discrete grid technology. Specifically: traversing the starting point positions of all private car trajectories, taking the outermost points in the up, down, left and right directions as the boundaries of the four directions of the city, and given the grid size s, the urban map is evenly divided into N grids.

[0015] Further, the topological structure graph constructed from the perspective of spatial connectivity between grids is denoted as G c =(V, E c ); where V is the node set, E c is the graph G cThe weight set between intermediate nodes, where the weights are characterized by the normalized values of the number of trajectories;

[0016] The topological structure diagram constructed from the perspective of the geographical proximity between grids is denoted as G d =(V, E d ), where E d is the weight set between the nodes in graph G d The relevance represented by the elements of the weight set is determined from the perspective of the actual physical distance, that is:

[0017]

[0018] In the formula, A d (i, j) is the weight of the connecting edge between grid nodes i and j in graph G d ; Haversine(i c , j c ) is the Haversine distance function between grid nodes i and j based on the regional centroid coordinates i c , j c ; d is the average value of the Haversine distances between all connected nodes in graph G d based on the regional centroid coordinates;

[0019] The topological structure diagram constructed from the perspective of the carbon emission time similarity between grids is denoted as G t =(V, E t ), where E t is the weight set of the connecting edges in graph G t The carbon emission time similarity is characterized by calculating the distances between the regions represented by different grid nodes using the dynamic time warping function:

[0020]

[0021] In the formula, A t (i, j) is the carbon emission time similarity between grid nodes i and j, x i , x j are respectively the sequences of the carbon emission time changes of grid nodes i and j, DTW(x i , x j ) is the value of the dynamic time warping function between x i , x j , and ε is the distance decay parameter.

[0022] Furthermore, the spatio-temporal convolution module includes temporal gated convolution, spatial convolution, and temporal attention mechanism.

[0023] Furthermore, the temporal gated convolution utilizes dilated causal convolution and the gated linear unit (GLU) to capture the respective correlations in the time dimension from carbon emission and traffic density data. Specifically:

[0024] 1) For a convolution kernel of size K, the input is The process of dilated causal convolution is expressed as:

[0025]

[0026] In the formula, represents carbon emissions, u = a represents traffic density; f(t) represents the filter of the t-th order dilated causal convolution, t is the input time series information, d is the dilation rate of the dilated causal convolution; t - d×s represents the positioning of the current historical moment information

[0027] 2) The gated linear unit filters the information passed by the output of each layer of dilated causal convolution and is expressed as:

[0028]

[0029] In the formula, Θ1 and Θ2 are a set of kernels, b1 and b2 are model biases, ⊙ is the vector product, and g and σ are both activation functions, is the carbon emission or traffic density feature filtered by the gated linear unit, which includes the correlation in the time dimension;

[0030] Furthermore, the spatial convolution captures the corresponding spatial heterogeneity features from the input data through graph convolution operations. Specifically, for each perspective, the topological structure graph G r correspondingly uses a graph convolutional network (GCN) r for feature extraction and is expressed as:

[0031]

[0032]

[0033] In the formula, the subscript r represents various perspectives of spatial connectivity, geographical proximity, and carbon emission time similarity; represents the graph G r 's adjacency tensor matrix, is the original adjacency matrix, I N is the identity matrix, is the trainable weight parameter in the l-th layer of the GCN r , is the degree matrix; are the output features of the l-th and (l + 1)-th layers respectively,

[0034] In each graph convolutional network GCNr The final output utilizes a multi-layer perceptron to fuse the features of each GCN r The output features are expressed as:

[0035]

[0036] In the formula, ° represents the tensor product, and W r represents the weight parameter of the science department; is the feature obtained by the fusion of the multi-layer perceptron, that is, the output feature of the spatial convolution, which includes the captured spatial heterogeneity features.

[0037] Furthermore, the temporal attention mechanism captures the global correlation of carbon emissions in the time dimension. Specifically:

[0038] First, calculate the normalized temporal correlation matrix:

[0039]

[0040] In the formula, V e , U1, U2, U3 are all learnable parameters in the temporal attention mechanism, and are the features output by the (l-1)-th layer of the graph convolutional network;

[0041] Then, through the formula weights are assigned in the time dimension to obtain the global correlation features of carbon emissions or traffic density in the time dimension When u = p is the spatio-temporal feature of carbon emissions, and when u = a is the spatio-temporal feature of traffic density.

[0042] Furthermore, the dependence of traffic density on carbon emissions is extracted from the spatio-temporal features of traffic density, and the dependence of traffic density on carbon emissions is fused with the spatio-temporal features of carbon emissions to obtain the global spatio-temporal features of carbon emissions. Specifically:

[0043] (1) Initialize two learnable parameter matrices E1 and E2, and obtain the adaptive adjacency matrix A through node embedding adp , based on the adaptive adjacency matrix A adp obtain the dependence of traffic density on carbon emissions, expressed as:

[0044] A adp = norm(ReLU(E1, E2 T ))

[0045]

[0046] In the formula, is the spatio-temporal feature of traffic density at the l-th layer, is the l-th layer

[0047] The dependence of the spatio-temporal characteristics of traffic density on carbon emissions indicates that are learnable parameters;

[0048] (2) Then, fuse the dependence of traffic density on carbon emissions with the spatio-temporal characteristics of carbon emissions to obtain the global spatio-temporal characteristics of carbon emissions, expressed as:

[0049]

[0050] In the formula, is the global spatio-temporal characteristic of carbon emissions obtained by fusion, is the spatio-temporal characteristic of carbon emissions output by the l-th layer, and Concat( ) is an aggregation function.

[0051] Furthermore, the current external factors include, but are not limited to, cold start status and weather conditions. First, use 1 fully connected layer to perform feature embedding on all encoded current factors to obtain 1 long feature tensor Then use another fully connected layer to make the long feature tensor match the shape of the global spatio-temporal characteristic of carbon emissions Finally, based on the matched long feature tensor and the global spatio-temporal characteristic of carbon emissions predict the carbon emissions for the next q time periods:

[0052]

[0053]

[0054] In the formula, represents the predicted value of carbon emissions for the next q time periods, is the global spatio-temporal characteristic of carbon emissions obtained by the l-th layer, is the cumulative global characteristic of carbon emissions of the L-th layer; and are weight parameters that can be updated and learned during carbon emission prediction.

[0055] An urban carbon emission prediction device based on a collaborative spatio-temporal convolutional network includes a memory and a processor. When the computer program stored in the memory is executed by the processor, the processor implements the urban carbon emission prediction method based on the collaborative spatio-temporal convolutional network described in any one of the above technical solutions.

[0056] Beneficial effects

[0057] The present invention constructs a multi-perspective spatio-temporal heterogeneous graph through a graph structure to model the spatial heterogeneity of urban carbon emissions, and then correlates the main characteristics of carbon emissions and the auxiliary characteristics of traffic density with respect to carbon emissions as multi-variable inputs, and predicts the carbon emissions of private cars in the city in a collaborative method, which can improve the accuracy of urban carbon emission prediction. Description of the Drawings

[0058] Figure 1 It is the overall framework diagram of the method described in the embodiments of the present application. Detailed Embodiments

[0059] The following makes a detailed description of the embodiments of the present invention. Based on the technical solution of the present invention, detailed implementation manners and specific operation processes are given, and the technical solution of the present invention is further explained and illustrated.

[0060] The following describes the embodiments of the present disclosure in detail with reference to the drawings.

[0061] The following illustrates the embodiments of the present disclosure through specific specific examples. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments. The present disclosure can also be implemented or applied through other different specific implementation manners, 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 disclosure. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts belong to the scope of protection of the present disclosure.

[0062] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be obvious that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is illustrative only. Based on the present disclosure, those skilled in the art should understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, this device can be implemented and this method can be practiced using other structures and / or functions in addition to one or more of the aspects described herein.

[0063] It should also be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present disclosure schematically. The diagrams only show the components related to the present disclosure, rather than being drawn according to the number, shape, and size of the components in actual implementation. The types, quantities, and ratios of the components in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0064] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0065] An embodiment of the present disclosure provides a method for predicting urban carbon emissions based on a collaborative spatio-temporal convolutional network. The method can be applied to the process of predicting urban carbon emissions in intelligent transportation or environmental protection scenarios.

[0066] See Figure 1 , which is a schematic flowchart of a method for predicting urban carbon emissions based on a collaborative spatio-temporal convolutional network provided by an embodiment of the present disclosure. As Figure 1 shown, the method mainly includes the following steps:

[0067] Step 1: Taking urban grids as nodes, construct a heterogeneous graph of multi-perspective correlation of the city from the perspectives of spatial connectivity, geographical proximity, and carbon emission time similarity between grids.

[0068] First, use the discrete grid technology to divide the city into grids: For the starting and ending positions recorded in each private car trajectory, use the grid technology to divide the urban geographical space where the longitude and latitude positions of all trajectory starting points are located into discrete grids. Specifically, traverse all the starting point positions of the recorded trajectories, use the outermost points in the up, down, left, and right directions as the boundaries of the four directions of the city, and given the grid size s, the city map is evenly divided into N grids. Among them, each grid is regarded as a small urban area.

[0069] Then, construct a heterogeneous graph structure of multi-perspective correlation of the city from three attributes: spatial connectivity, geographical proximity, and time similarity.

[0070] If a certain trajectory starts at grid node i / j in the city and stops at grid node j / i, it is considered that there is connectivity between grid nodes i and j, and an edge can be created between these two nodes. In this way, a topological structure diagram is constructed from the perspective of spatial connectivity between grids, denoted as G c =(V, E c ), where V is the set of nodes, and E c is the set of weights between nodes in graph G c . The weight is characterized by the normalized value of the number of trajectories.

[0071] Carbon emissions can affect the carbon emissions in their spatially adjacent regions. For example, future carbon emissions downstream are affected by the emission diffusion in nearby regions. Therefore, adjacent neighbors are more relevant than distant neighbors. To capture the correlation between spatial locations, in this embodiment, a topological structure diagram is constructed from the perspective of the geographical proximity between grids, denoted as G d =(V, E d ), where E d is the weight set of the connecting edges in the graph G d . The correlation represented by the elements of the weight set is determined from the perspective of the actual physical distance, that is:

[0072]

[0073] In the formula, A d (i, j) is the weight of the connecting edge between the grid nodes i and j in the graph G d . Haversine(i c , j c ) is the Haversine distance function between the grid nodes i and j based on the regional centroid coordinates i c , j c . d is the average value of the Haversine distances between all connecting nodes in the graph G d based on the regional centroid coordinates. The centroid coordinates of each grid node are obtained by matching the carbon emission data of all trajectories to each of the already divided regions through longitude and latitude information, and then iteratively aggregating the data within each region to obtain the centroid coordinates of the region. K is set to 1 and aggregated into 1 category.

[0074] Two grid node regions may be geographically far apart, but the temporal variations of their carbon emissions are similar. Therefore, in this embodiment, this temporal similarity is modeled between grid node regions. Define the carbon emission change of region i as x i . Considering a time interval of 1 hour, the time window length for 1 day is 24. Compared with the previous moment, if the carbon emission change in this region is set to 1 in the time window with the subscript of this moment, and 0 if there is no change. Then, the dynamic time warping function is used to calculate the distance between different regions, denoted as DTW(x i , x j ). Then, a semantic graph G t =(V, E t ) is constructed to capture the temporal similarity between regions:

[0075]

[0076] In the formula, A t (i, j) is the temporal similarity of carbon emissions between the grid nodes i and j, and x i,x j They are the sequences of the time variations of carbon emissions at grid nodes i and j respectively. DTW(x i ,x j ) is the value of the dynamic time warping function between x i ,x j . ε is the distance decay parameter.

[0077] Step 2: Input the urban carbon emissions and traffic density data into the heterogeneous graph in parallel, and each uses 1 spatio-temporal convolution module to extract the spatio-temporal features of carbon emissions and the spatio-temporal features of traffic density from the heterogeneous graph with carbon emissions and traffic density data respectively.

[0078] Specifically, the spatio-temporal convolution module includes temporal gated convolution, spatial convolution, and temporal attention mechanism.

[0079] Among them, the process of learning temporal dependencies through temporal gated convolution is expressed as:

[0080] The carbon emissions in a region are highly correlated with historical observations. Based on this, dilated causal convolution and gated linear unit (GLU) can be used to capture the correlations in the time dimension. The filter of the dilated causal convolution at the t-th order is defined as f, and the convolution process using it is expressed as:

[0081]

[0082] In the formula, d is the dilation rate of the dilated causal convolution. As the depth of the dilated causal convolution increases, it can receive a wider receptive field. In addition, the information transmitted through each layer will be filtered by the gated linear unit:

[0083]

[0084] In the formula, Θ1 and Θ2 are a set of kernels, b1 and b2 are model biases, ⊙ is the vector product, and g and σ are both activation functions.

[0085] After passing through the temporal gated convolution, it is passed into the spatial convolution. The spatial convolution is based on the constructed multi-view graph and captures spatial heterogeneity through convolution operations. The implementation steps are as follows:

[0086]

[0087]

[0088] In the formula, the subscript r represents various perspectives of spatial connectivity, geographical proximity, and carbon emission time similarity; represents the adjacency tensor matrix of graph G r , is the original adjacency matrix, and I N is the identity matrix. is the GCN r The trainable weight parameters in the l-th layer, is the degree matrix; are the output features of the l-th and l+1-th layers respectively,

[0089] In the formula, is the adjacency tensor matrix, I N is an identity matrix, are the trainable weight parameters in each layer, is the degree matrix, and each topological structure graph G r (i.e., G c , G d and G t ) is input into the corresponding graph convolutional network GCN r (i.e., GCN c , GCN d and GCN t ), and the output feature matrix H r (i.e., H c , H d and H t ). In addition, a multi-layer perceptron is used to fuse the feature matrices output by each layer, which is expressed as:

[0090]

[0091] In the formula, ° represents the tensor product, and then the softmax operation is performed to normalize the feature matrix.

[0092] After capturing the correlation between spatial positions, a temporal attention mechanism is used to capture the global correlation of carbon emissions in the time dimension. The execution process is as follows:

[0093] Define E as the normalized time correlation matrix, and its calculation process is expressed as:

[0094]

[0095] In the formula, V e , U1, U2, U3 are learnable parameters. The value of the element Ei,j in E semantically represents the dependence strength between time i and j. The input outputs the corresponding correlation matrix through the temporal attention mechanism, and then weights are assigned in the time dimension through the formula to adjust the input information.

[0096] In addition, as an auxiliary feature for collaborative prediction of carbon emissions, the execution process of extracting the spatio-temporal features of traffic density is as follows:

[0097] The traffic density and carbon emissions have the same graph structure. In this part, a spatio-temporal convolution module similar to that of carbon emissions is designed, which also includes temporal gated convolution, spatial convolution, and temporal attention mechanism. Correspondingly, the output feature matrix is also similar to that of carbon emissions, expressed as

[0098] Step 3: Extract the dependence of traffic density on carbon emissions from the spatio-temporal features of traffic density, and fuse the dependence of traffic density on carbon emissions with the spatio-temporal features of carbon emissions to obtain the global spatio-temporal features of carbon emissions.

[0099] This step processes the internal correlation of traffic density on carbon emissions by proposing an interaction layer. By capturing this internal dependence, the hidden states of them in the spatio-temporal dimension are further established. The execution process is as follows:

[0100] To simulate the mutual dependence between traffic density as auxiliary features and carbon emissions as main features, two matrices E1 and E2 of learnable parameters are randomly initialized. Through the node embedding method, an adaptive adjacency matrix A adp = norm(ReLU(E1, E2 T )) is proposed to adaptively deduce the implicit spatial relationship between variables. Specifically, through the convolution operation, the most relevant auxiliary hidden states are propagated to the required main nodes on the adaptive adjacency matrix, expressed as:

[0101]

[0102]

[0103] In the formula, represents the influence of the spatio-temporal features of traffic density on carbon emissions in the l-th layer, is the spatio-temporal feature of traffic density in the l-th layer, are learnable parameters. At the same time, to aggregate this unordered information (the influence of main features and auxiliary features), Concat(·) is selected as the aggregation function, and the spatio-temporal feature matrix of carbon emissions is output for each layer

[0104] Step 4: Combine the global spatio-temporal features of carbon emissions with the current external factors to perform collaborative prediction of urban carbon emissions.

[0105] In this step, a collaborative fusion strategy is designed to combine the global spatio-temporal features of carbon emissions integrated in the interaction layer with some other external factors. Among them, the implementation steps are as follows:

[0106] In addition to spatio-temporal dependence, carbon emissions are also affected by driving behaviors (such as cold start status) and other external factors (such as weather conditions). Therefore, this embodiment identifies some global factors and embeds them into the final prediction of collaborative fusion to ensure the effectiveness of the prediction results. Specifically, first, the cold start status is encoded and represented as E CS In addition, external factors such as weather conditions also provide additional information for carbon emission prediction. Therefore, this embodiment also encodes weather attributes such as sunny, rainy, foggy, and windy, denoted as E WC .

[0107] Subsequently, two fully connected layers are stacked to perform feature embedding of global factors. First, in the first layer, the current external factors are fused into a long feature tensor, expressed as:

[0108]

[0109] where is the concatenation operation between vectors.

[0110] Subsequently, the second fully connected layer is used to ensure that the output tensor matches the shape of the spatio-temporal feature matrix of carbon emissions for collaborative prediction of global factors and global spatio-temporal features of carbon emissions:

[0111]

[0112] where represents the predicted value of carbon emissions for the next q time periods; is the global spatio-temporal feature of carbon emissions obtained from the l-th layer, is the accumulated global feature of carbon emissions for L layers; and are weight parameters that can be updated and learned during carbon emission prediction.

[0113] In this embodiment, by training the above model, the mean absolute error (MAE) loss between the predicted carbon emissions and the true values is minimized, expressed as:

[0114]

[0115] Specifically, the experimental data is divided into 60% as the training set, 20% as the test set, and 20% as the validation set; the grid size s is set to 2500 meters; the node embeddings are randomly initialized using a uniform distribution; the Adam algorithm is used as the optimizer for the model; the initial value of the learning rate is set to 0.001, and the model automatically decays 90% of the initial value every 100 iterations; the batch size for model training is 64; an early stopping strategy is set for the experiment. When the loss of the test set is equal to the loss of the training set, the model training reaches the optimal state, that is, the training is stopped to prevent the model from overfitting.

[0116] The above embodiments are the preferred embodiments of the present application. Those of ordinary skill in the art can also make various transformations or improvements on this basis. Without departing from the general concept of the present application, these transformations or improvements should all fall within the scope of protection required by the present application.

Claims

1. A method for predicting urban carbon emissions based on a collaborative spatio-temporal convolutional network, characterized in that, Including: Construct a heterogeneous graph of multi-perspective correlation of a city with the urban grid as nodes and from the perspectives of spatial connectivity, geographical proximity, and carbon emission time similarity between grids respectively. Input the urban carbon emission and traffic density data into the heterogeneous graph respectively, and use 1 spatio-temporal convolution module each to extract the spatio-temporal features of carbon emissions and the spatio-temporal features of traffic density. Extract the dependence of traffic density on carbon emissions from the spatio-temporal features of traffic density, and fuse the dependence of traffic density on carbon emissions with the spatio-temporal features of carbon emissions to obtain the global spatio-temporal features of carbon emissions. Combine the global spatio-temporal features of carbon emissions with the current external factors to conduct collaborative prediction of urban carbon emissions. Among them, the topological structure diagram constructed from the perspective of the spatial connectivity between grids is represented as G c = (V, E c ); where V is the node set, and E c is the weight set between the nodes in the graph G c , and the weight is characterized by the normalized value of the number of trajectories; The topological structure diagram constructed from the perspective of the geographical proximity between grids is denoted as G d = (V, E d ), where E d is the weight set between the nodes in the graph G d The relevance represented by the elements of the weight set is determined from the perspective of the actual physical distance, that is: ； In the formula, is the weight of the connecting edge between d the grid nodes in graph G; is the distance function between the grid nodes based on the regional centroid coordinates ; d is the average value of the distances between all connected nodes in graph G d based on the regional centroid coordinates; is the average distance. The topological structure diagram constructed from the perspective of the similarity of carbon emission time between grids is denoted as G t = (V, E t ), where E t is the weight set of the connecting edges in graph G t . The dynamic time warping function is used to calculate the distance between the regions represented by different grid nodes to characterize the similarity of carbon emission time: ； In the formula, is the carbon emission time similarity between grid nodes, are the sequences of the carbon emission time changes of grid nodes respectively, is the dynamic time warping function value between is the distance decay parameter; The current external factors include, but are not limited to, cold start state and weather conditions. First, a fully connected layer is used to perform feature embedding on all encoded current factors to obtain a long feature tensor , and then another fully connected layer is used to match the shape of the long feature tensor with the global spatio-temporal features of carbon emissions . Finally, based on the matched long feature tensor and the global spatio-temporal features of carbon emissions , the carbon emissions for the future time periods are predicted: ； ； In the formula, represents the predicted carbon emission values for future time periods, is the global spatio-temporal feature of carbon emissions obtained from the th layer, is the accumulated global feature of carbon emissions for layer L; and are weight parameters that can be updated and learned during carbon emission prediction.

2. The urban carbon emission prediction method according to claim 1, wherein Use the discrete grid technology to divide the urban map into several grids. Specifically: traverse the starting point positions of all private car trajectories, use the outermost points in the up, down, left, and right directions as the boundaries of the four directions of the city, and given the grid size s, divide the urban map into N grids on average.

3. The urban carbon emission prediction method according to claim 1, wherein The spatio-temporal convolution module includes a temporal gated convolution, a spatial convolution, and a temporal attention mechanism.

4. The urban carbon emission prediction method according to claim 3, characterized in that The temporal gated convolution uses dilated causal convolution and a gated linear unit (GLU) to capture the respective correlations in the time dimension from the carbon emission and traffic density data. Specifically: 1) For a convolutional kernel of size K, the input is The dilated causal convolution process is expressed as: ； In the formula, represents carbon emissions, represents traffic density; represents the filter of the t-th order dilated causal convolution, where t is the input time series information and d is the dilation rate of the dilated causal convolution; represents the positioning of the information at the current historical moment 2) The gated linear unit filters the information transmitted by the output of each layer of dilated causal convolution, expressed as: to perform the filtering, which is expressed as: ； In the formula, and are a set of nuclei, b1 and b2 are model biases, ⊙ is a vector product, and both g and σ are activation functions. is the carbon emission or traffic density feature obtained by filtering through the gated linear unit, including the correlation in the time dimension.

5. The urban carbon emission prediction method according to claim 3, characterized in that, The spatial convolution captures corresponding spatial heterogeneity features from the input data through graph convolution operations. Specifically, the topological structure diagrams of each perspective correspondingly use a graph convolutional network for feature extraction, expressed as: ； ； In the formula, the subscript represents various perspectives of spatial connectivity, geographical proximity, and similarity in carbon emission time; denotes the adjacency tensor matrix of graph ; is the original adjacency matrix, is the identity matrix, is the trainable weight parameter in the th layer, is the degree matrix; are the output features of the th layer respectively, ; In the graph convolutional network The final output of uses a multi-layer perceptron to fuse the output features of each and is expressed as: ； In the formula, represents the tensor product, represents the learnable weight parameter; is the feature obtained by multi-layer perceptron fusion, that is, the output feature of spatial convolution, which includes the captured spatial heterogeneity features.

6. The urban carbon emission prediction method according to claim 3, wherein The temporal attention mechanism captures the global correlation of carbon emissions in the time dimension. Specifically: First calculate the normalized time correlation matrix: ； wherein, are all learnable parameters in the temporal attention mechanism, is the feature output by the -th layer of the graph convolutional network; Then, through the formula weights are assigned in the time dimension to obtain the global correlation features of carbon emissions or traffic density in the time dimension ; when at this time it is the spatio-temporal feature of carbon emissions, and when at this time it is the spatio-temporal feature of traffic density.

7. The urban carbon emission prediction method according to claim 1, wherein The extraction of the dependence of traffic density on carbon emissions from the spatio-temporal features of traffic density and the fusion of the dependence of traffic density on carbon emissions with the spatio-temporal features of carbon emissions to obtain the global spatio-temporal features of carbon emissions is specifically as follows: (1) Initialize two learnable parameter matrices , and obtain an adaptive adjacency matrix through node embedding . Based on the adaptive adjacency matrix , obtain the dependence of traffic density on carbon emissions, expressed as: ； ； In the formula, is the spatio-temporal characteristic of the traffic density of the th layer, is the representation of the dependence of the spatio-temporal characteristic of the traffic density in the th layer on carbon emissions, is a learnable parameter; (2) Then fuse the dependence of traffic density on carbon emissions with the spatio-temporal features of carbon emissions to obtain the global spatio-temporal features of carbon emissions, expressed as: ； wherein, is the global spatio-temporal feature of carbon emissions obtained by fusion, is the spatio-temporal feature of carbon emissions output by the -th layer, is an aggregation function.

8. An urban carbon emission prediction device based on a collaborative spatio-temporal convolutional network, comprising a memory and a processor, wherein a computer program is stored in the memory, characterized in that, When the computer program is executed by the processor, the processor is caused to implement the method according to any one of claims 1 to 7.

Images