Spatial-temporal data extrapolation method and system based on bidirectional dynamic coupling
By constructing a spatiotemporal map of traffic sensors and combining it with a two-way dynamic coupling strategy, the problem of insufficient single-dimensional information in spatiotemporal data extrapolation is solved, achieving efficient spatiotemporal data extrapolation in complex environments and improving the robustness and accuracy of the model.
Patent Information
- Application Number
- CN202511786551.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-03-03
AI Technical Summary
Existing spatiotemporal data extrapolation techniques struggle to effectively account for the coupling effects of time and space dimensions when dealing with data gaps, resulting in insufficient robustness of models in complex environments, especially in unmonitored areas or future time periods where generalization performance deteriorates.
A spatiotemporal data extrapolation method based on bidirectional dynamic coupling is adopted. By constructing a spatiotemporal map of traffic sensors, combined with a closed subgraph masking strategy, Transformer and diffusion graph convolutional network, global feature extraction and cross-dimensional information complementarity are achieved. The distribution offset is corrected by using a reverse masking exchange strategy, and the loss function is optimized by combining an affinity judgment mechanism.
It improves the integrity and robustness of spatiotemporal data, maintains high-precision extrapolation in scenarios with sparse or offset data, overcomes the dependence of traditional methods on complete observation data, and enhances the model's adaptability in complex environments.
Smart Images
Figure CN121597975A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of spatiotemporal data extrapolation technology, specifically to a spatiotemporal data extrapolation method and system based on bidirectional dynamic coupling. Background Technology
[0002] With the rapid development of the Internet of Things (IoT) and traffic sensor technologies, spatiotemporal data is increasingly widely used in numerous fields such as industrial production, environmental monitoring, and urban management. The dynamic patterns and correlations it contains have become core support for decision analysis. This type of data typically exhibits a deep coupling between time series and spatial location, encompassing both state changes at different time points and the interactions between different spatial units, forming a high-dimensional, highly dynamic, and complex data system. Currently, the acquisition of spatiotemporal data relies on distributed monitoring equipment and mobile sensing technologies. While this can achieve a certain range of spatiotemporal coverage, data gaps are common due to limitations such as equipment deployment costs, environmental interference, and monitoring range constraints. Monitoring data for some spatial units remains blank for extended periods, or observations at specific time points are lost due to equipment malfunctions, transmission delays, or other reasons, making it difficult to construct a complete spatiotemporal dynamic map.
[0003] Spatiotemporal extrapolation is crucial for addressing data gaps, but this process faces challenges due to temporal and spatial distribution shifts. Traditional geostatistical spatial interpolation methods (such as Kriging) rely solely on geographical proximity, failing to adequately consider the influence of factors like traffic flow and road network structure, resulting in significant discrepancies between the generated data and the actual distribution. While existing deep learning-based methods, such as large spatiotemporal models (like UrbanGPT), demonstrate advantages in emission prediction, they heavily rely on the assumption of independent and identically distributed (IOD) models, leading to a significant decrease in generalization performance in unmonitored areas or future time periods. Furthermore, current research often focuses on single-dimensional dynamic simulations, failing to capture the unique spatiotemporal coupling characteristics of traffic emissions—their distribution drift is a result of the interaction between temporal and spatial dimensions, rather than independent evolution, resulting in insufficient robustness of models in complex mobile monitoring environments.
[0004] Existing research focuses on dynamics in a single dimension, utilizing differentiated temporal patterns or spatial distributions to simulate environmental disturbances and thus enhance invariance representation. However, the distribution drift of spatiotemporal dynamic systems, coupled with temporal and spatial dimensional offsets, means that dynamic patterns mined from limited monitoring data in a single dimension cannot fully simulate all situations outside the distribution. The simulation of environmental disturbances suffers from significant information loss, and the invariant relationships learned by the model are not based on the underlying spatiotemporal co-operational logic, ultimately resulting in insufficient robustness under unknown spatiotemporal coupling offset scenarios. This invention leverages coupling characteristics to extend dynamic pattern recognition from a single dimension to a two-dimensional spatiotemporal perspective, introducing a spatiotemporal extrapolation method based on bidirectional dynamic coupling. Through bidirectional dynamic correlation and spatiotemporal joint training, it compensates for the deficiencies of single-dimensional information, and a reverse mask correction strategy corrects the distribution offset between the masked data and the original data, which is of great significance for improving the integrity of spatiotemporal data. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention provides a spatiotemporal data extrapolation method and system based on bidirectional dynamic coupling.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a spatiotemporal data extrapolation method based on bidirectional dynamic coupling, which predicts traffic sensor data in distributed missing data or unmonitored areas through an extrapolation model, including: Construct a spatiotemporal graph of traffic sensors, where nodes represent traffic sensors and edges represent the connectivity between traffic sensors; The traffic sensor spatiotemporal graph is preprocessed using a closed subgraph masking strategy to simulate missing traffic sensor data. The masked traffic sensor spatiotemporal map is input into the spatiotemporal information extraction module to obtain global features. The global features include global time-related information extracted using Transformer and spatial event-related information learned through graph convolutional network. The extracted global features are input into the temporal extrapolation branch and the spatial extrapolation branch, respectively. The temporal extrapolation branch performs temporal extrapolation through a multi-patch layer and a frequency domain enhancement layer, and the spatial extrapolation branch performs spatial extrapolation through a diffusion graph convolutional network. The temporal extrapolation branch and the spatial extrapolation branch provide each other with additional data distribution information. The consistency between temporal and spatial extrapolation is evaluated based on the affinity judgment mechanism, and the forward loss function is adaptively calculated. The reverse loss is calculated by dynamically adjusting the distribution offset between the observed distribution and the missing road segments using a reverse masking strategy. The total loss is calculated based on the forward loss and backward loss, and the extrapolation model is trained.
[0007] In one embodiment, the traffic sensor data includes one or more of traffic flow, traffic carbon emissions, vehicle speed, and vehicle acceleration.
[0008] In one embodiment, the construction of the traffic sensor spatiotemporal graph, where nodes represent traffic sensors and edges represent the connectivity between traffic sensors, specifically includes: Based on the connectivity of traffic sensors, a spatiotemporal graph of traffic sensors is obtained by abstracting nodes of traffic sensors within a region: the road segments within the region are divided, and the spatiotemporal graph uses traffic sensors in the resulting road segment slices as nodes and traffic sensor connectivity as edges; the spatiotemporal graph of traffic sensors... ; Let be the set of edges, representing the connectivity of traffic sensors; , The number of edges, yes The adjacency matrix; This indicates the time span of traffic sensor data; road segments in Feature vector at time step , express Traffic sensor nodes eigenvectors, It is the dimension of the feature vector of the traffic sensor node.
[0009] In one embodiment, the preprocessing of the traffic sensor spatiotemporal graph based on the closed subgraph masking strategy to simulate missing traffic sensor data specifically includes: Training images are sampled from the spatiotemporal maps of traffic sensors, and selections are made from the training images containing at least... The closed subgraphs of each traffic sensor node are masked, and the unmasked portion of the training graph is called the retained graph. Then, virtual closed subgraphs are embedded into the retained graph as basic units. An extended graph structure is constructed by inserting virtual traffic sensor nodes between the virtual closed subgraphs and the retained graph. The number of inserted virtual traffic sensor nodes is controlled during the construction of the extended graph structure: the number of inserted virtual traffic sensor nodes... ,in The number of virtual traffic sensor nodes is determined by the traffic sensor data missing rate. This represents the minimum interpolation amount.
[0010] In one embodiment, the masked traffic sensor spatiotemporal map is input into the spatiotemporal information extraction module to obtain global features. These global features include global time-related information extracted using a Transformer and spatial event-related information learned through a graph convolutional network, specifically including: Extracting global time-related information using Transformer: ; The feature vector representing a traffic sensor node. This represents the core attention of the Transformer. This indicates the initial time information extracted; Then, a graph neural network that updates both points and edges simultaneously learns global event-related information: ; in, This refers to a filter based on an attention mechanism. This represents the feature transformation function of traffic sensor nodes. These are the initial features of the v-th traffic sensor node input to the graph neural network. Indicates the first Features of the v-th traffic sensor node in layer 1 Refers to the first The values of adjacent traffic sensor nodes in the layer, Indicates the first Overall characteristics of traffic sensor nodes in the layer This represents the boundary value of the previous layer. This represents the weighting parameter.
[0011] In one embodiment, the spatial extrapolation branch performs spatial extrapolation through a diffusion graph convolutional network, specifically including: The spatial extrapolation branch introduces a diffusion graph convolutional network, which dynamically updates the representation of traffic sensor nodes by simulating the information diffusion process. ; and These are the forward transition matrix and the backward transition matrix, respectively. and These are learning parameters. This represents a diffusion graph neural network. Let Nv represent the feature representation of the Nv-th virtual traffic sensor node in the l-th layer, where Nv is the number of virtual traffic sensor nodes in the additional virtual closed subgraph inserted in the closed subgraph masking strategy. For the first The hidden state of non-zero traffic sensor nodes in the layer. Represents a non-linear activation function. Represents a linear transformation function; Errors in spatial extrapolation ; Indicates the number of mask nodes. This represents the loss hyperparameter used to balance multiple tasks. Represents 0 to The actual value of the mask node at any given time. Indicates spatial extrapolation branches from 0 to The predicted value of the mask node at time step.
[0012] In one embodiment, the time extrapolation branch performs time extrapolation through a multi-patch layer and a frequency domain enhancement layer, specifically including: The multi-patch layer divides the time series corresponding to the global features into non-overlapping multi-scale blocks, and extracts weighted features at different time scales: ; Represents the weight matrix. Indicates the slice length. Represents global features. This represents the operation of dividing into multi-scale blocks. This represents data after multi-scale weighting; The frequency domain enhancement layer uses discrete cosine transform to extract the frequency components of the data in each spatial channel: ; Let represent the vector after the discrete cosine transform of the i-th channel. Indicates the length of the time-domain signal. This represents the channel characteristic of the l-th channel among n channels in a spatial node. Indicates the number of channels; Temporal features are extracted by combining the frequency components of low-frequency and high-frequency channels: ; Represents frequency domain characteristics, Indicates stacking. This represents the characteristics of the nth channel after the discrete cosine transform. Errors in time extrapolation , To preserve the number of traffic sensor nodes in the graph, This represents the actual observations of the traffic sensor nodes in the retained graph at time T. For time extrapolation branch pairs Predicted values of traffic sensor nodes in a time-preserved graph.
[0013] In one embodiment, the evaluation of the consistency between temporal and spatial extrapolations based on the affinity judgment mechanism to adaptively calculate the forward loss function specifically includes: The formula for judging affinity is: ; This indicates the affinity between time extrapolation and the Kriging mission; This represents the spatiotemporal extrapolation loss of the shared parameters from the previous round; This represents the shared parameters before the time extrapolation loss iteration; This represents the shared parameters after iterating through the time extrapolation loss; The unique parameter representing the spatial extrapolation task; Whether to calculate forward loss is determined by calculating the affinity between time extrapolation and the Kriging task: if the affinity is positive, forward loss is calculated; if the affinity is negative, forward loss is not calculated. .
[0014] In one embodiment, the step of dynamically adjusting the distribution offset between the observed distribution and the missing road segments using a reverse masking strategy and calculating the reverse loss specifically includes: In the reverse process, a masking strategy is used to dynamically adjust invariant learning by swapping nodes between traffic sensor nodes in the training graph and virtual closed subgraphs. Training nodes are converted into masking stations, while virtual closed subgraphs are retained as actual nodes. The system infers information from the virtual closed subgraph using the expanded graph composed of the training graph and the virtual closed subgraph, and uses known information to fill in missing parts, thereby reducing errors caused by information loss. Reverse loss... ; This represents the number of virtual closed subgraphs in the retained graph. Represents 0 to The true value of the virtual closed subgraph of the time-preserved graph. Indicates spatial extrapolation branch pairs Predicted values of the virtual closed subgraph of the time-preserved graph.
[0015] In a second aspect, the present invention provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any embodiment of the first aspect.
[0016] Compared with the prior art, the present invention has the following technical advantages: This invention first constructs a graph structure containing road segment nodes and connectivity edges to transform traffic sensor data into a spatiotemporal graph sequence. It then combines multi-patch layers and frequency domain enhancement mechanisms to perform multi-scale spatiotemporal coding, extracting hierarchical dynamic features from micro to macro levels. Based on this, it uses a diffuse graph convolutional network (DGCN) to simulate the spatial propagation law of emissions, and uses a Transformer to capture long-range time dependencies, achieving deep fusion of spatiotemporal features.
[0017] The core advantage of this invention lies in the collaborative design of bidirectional dynamic coupling and correction mechanisms: In the forward process, time extrapolation uses known future points to provide temporal references for spatial completion, while spatial extrapolation uses samples from unobserved regions to enrich the time dimension modeling. By complementing cross-dimensional information, it overcomes the limitations of a single dimension and can better characterize the spatiotemporal coupling characteristics than traditional single-task models. In the reverse process, through the insertion of virtual traffic sensor nodes and mask exchange, the deviation between simulated data and real distribution is dynamically calibrated. Combined with the affinity judgment mechanism, the loss function is adaptively optimized, enabling the model to maintain robustness in scenarios with sparse data or distribution offsets, thus solving the problem of strong dependence of traditional interpolation methods on complete observation data. At the same time, by using multi-patch layer segmentation and discrete cosine transform (DCT), the fluctuation patterns and key frequency components at different time scales are accurately captured, and the important feature weights are strengthened through the channel attention mechanism. Compared with models that only rely on local temporal or spatial features, it can more comprehensively explore spatiotemporal dynamic patterns and improve extrapolation accuracy.
[0018] This framework, through hierarchical modeling of global baseline distribution and local dynamic calibration, not only preserves the physical characteristics of road network topology and emission diffusion, but also adapts to complex environmental disturbances through self-supervised learning. Its performance under different missing rate conditions is superior to existing baseline models, providing more reliable technical support for the refined analysis of traffic data. Attached Figure Description
[0019] Figure 1 This is a flowchart of the method in an embodiment of the present invention. Detailed Implementation
[0020] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.
[0021] like Figure 1 As shown, the present invention provides a spatiotemporal data extrapolation method based on bidirectional dynamic coupling, which predicts traffic sensor data in distributed missing data or unmonitored areas through an extrapolation model, including the following steps: S1, Construct a spatiotemporal graph of traffic sensors, where nodes represent traffic sensors and edges represent the connectivity between traffic sensors; S2, perform masking preprocessing on the spatiotemporal graph of the traffic sensor based on the closed subgraph masking strategy to simulate missing traffic sensor data; S3. Input the masked traffic sensor spatiotemporal map into the spatiotemporal information extraction module to obtain global features. The global features include global time-related information extracted using Transformer and spatial event-related information learned through graph convolutional network. S4, the extracted global features are input into the temporal extrapolation branch and the spatial extrapolation branch respectively. The temporal extrapolation branch performs temporal extrapolation through a multi-patch layer and a frequency domain enhancement layer. The spatial extrapolation branch performs spatial extrapolation through a diffusion graph convolutional network. The temporal extrapolation branch and the spatial extrapolation branch provide each other with additional data distribution information. S5 evaluates the consistency between temporal and spatial extrapolation based on the affinity judgment mechanism to adaptively calculate the forward loss function; S6, dynamically adjust the distribution offset between the observed distribution and the missing road segment through the reverse masking strategy, and calculate the reverse loss; S7. Calculate the total loss based on the forward loss and backward loss, and train the extrapolation model.
[0022] In order to overcome the problems of insufficient single-dimensional information and environmental simulation bias in the existing technology of spatiotemporal extrapolation, this invention proposes a joint training framework based on bidirectional dynamic correlation.
[0023] A joint training framework based on bidirectional dynamic association includes: abstracting traffic sensors within a range into a connected graph based on their connectivity, where traffic sensors are nodes and their connectivity is represented as edges. The graph data is represented as follows: .in, This indicates the connectivity between traffic sensors. , Let be the number of edges. yes The adjacency matrix. It is the dimension of the feature vector of the traffic sensor node, which in this invention corresponds to the number of equidistant slices for each road segment. This indicates the time span of traffic sensor data. The feature vector of the road segment at a given time is represented as: express Traffic sensor nodes The feature vectors. To visualize the actual missing values, this invention uses a binary mask. To represent the observed values of traffic sensor node data, where Represents traffic sensor node data Missing Represents traffic sensor node data It was observed.
[0024] Graph neural networks process data through masking, completion, inverse masking, and spatiotemporal extrapolation. The overall workflow is as follows: 1) Masking: Masking the original data in the training graph to simulate missing data. 2) Sample Completion: Completing the data using sample graphs, filling in the missing parts with known information. 3) Inverse Masking Correction: Correcting the completed data using an inverse masking strategy to reduce errors caused by information loss. 4) Spatiotemporal Extrapolation: Finally, predicting data in unmonitored areas using spatiotemporal extrapolation techniques, and making comprehensive inferences based on spatiotemporal evolution patterns.
[0025] This invention proposes a novel joint training framework based on bidirectional dynamic correlation. After masking preprocessing, contaminated data is input into a spatiotemporal information extraction module composed of a Transformer and ENGCN (Edge and Node co-optimize Graph Convolutional Network) for global correction. The extracted information is then fed into the temporal and spatial extrapolation branches of the forward processing, where these two branches mutually provide additional data distribution information. Finally, a reverse process corrects for spatiotemporal offsets in the missing distribution.
[0026] The key implementation of this framework relies on a closed subgraph masking strategy. First, it selects subgraphs containing at least... The closed subgraphs of each traffic sensor node are masked, and then the virtual closed subgraphs are embedded into the training graph as basic units. An expanded graph structure is constructed by selecting bridging nodes between the closed subgraphs and the retained graph. Node insertion control: Adjusting the number of inserted virtual traffic sensor nodes. ,in Corresponding to the number of virtual traffic sensor nodes determined by the data missing rate, This represents the minimum interpolation amount. This strategy balances validation performance with adaptive missing rate, ensuring that the training graph simulates real road networks and data distributions, and enabling the model to generalize across spatial regions and missing patterns.
[0027] First, extract global time-related information using a transformer. Then, a graph neural network that updates both points and edges simultaneously learns the global event relevance. ,in, This refers to a filter based on an attention mechanism. Represents the node feature transformation function. It is the initial node feature vector. Indicates the first The feature of the v-th node in the layer, Refers to the first The values of adjacent nodes in the layer, This represents the node characteristics of the previous layer. This represents the boundary value of the previous layer. This represents the weight parameters. To ensure that features of different dimensions retain their original feature information after multiple convolutional layers, , .here, Represents a linear mapping. Represents the weight parameters. This represents the decay parameter of the weight matrix of the l-th layer. The feature vector representing a traffic sensor node. Represents the identity matrix. The update equation for the edge feature matrix is: , Represents the edge between the i-th traffic sensor node and the j-th traffic sensor node. The value at layer l, This represents the edge weight update function that takes into account the features of the i-th and j-th traffic sensor nodes in the (l-1)-th layer. This represents the characteristics of the i-th traffic sensor node in the (l-1)-th layer. This represents the characteristics of the j-th traffic sensor node in the (l-1)-th layer.
[0028] Specifically, in order to simulate the propagation dynamics of traffic emissions in irregular road networks, the spatial extrapolation module introduces a diffusion graph convolutional network (DGCN), which dynamically updates node representations by simulating the information diffusion process. . .here, and These are the forward transition matrix and the backward transition matrix, respectively. and These are the learning parameters. The error in this process is... .
[0029] Specifically, the time extrapolation module performs time extrapolation through a multi-patch layer and a frequency domain enhancement layer. The multi-patch layer divides the time series into non-overlapping multi-scale blocks and extracts weighted features from different time scales: The frequency domain enhancement layer uses discrete cosine transform to extract the frequency components of the data in each channel: The final result extracts temporal features by concatenating the frequency components of the low-frequency and high-frequency channels. The error in this process is... In temporal extrapolation, the values of known points in the future can provide additional time series data for spatial extrapolation, while information about existing unobserved points in spatial extrapolation can provide additional spatial samples for temporal extrapolation. , Therefore, the second part, spatial extrapolation loss and temporal extrapolation loss, can be expressed as: , .
[0030] During the forward pass, the affinity judgment mechanism improves generalization ability by evaluating the consistency between temporal and spatial extrapolations. This ensures consistent predictions across scales and strengthens invariant modeling in complex environments. The formula for affinity judgment is: The impact of parameters from the spatiotemporal information extraction component after time extrapolation is evaluated on spatial extrapolation, thereby determining whether the directions of time extrapolation and spatial extrapolation are consistent. This represents the shared parameters before the time extrapolation loss iteration. This represents the shared parameters after iterative time extrapolation loss. This represents the unique parameter for the spatial extrapolation task. For temporal extrapolation, whether to compute the loss is determined by calculating its affinity to the kriging task: if the affinity is positive (i.e., it has some influence), then the loss is computed. From the perspective of distribution offset, this means using values consistent with the temporal and spatial extrapolation distributions to correct for spatial extrapolation, thus preserving the temporal extrapolation loss during weight iterations. Otherwise, no iterations are performed, and the process is as follows: .
[0031] In the reverse process, a masking strategy is used to dynamically adjust invariant learning by swapping nodes between training nodes and virtual stations. During this process, training nodes are transformed into masked stations, while virtual stations are retained as actual nodes. It infers information about the virtual stations from the sample image and uses known information to fill in missing parts, thereby reducing errors caused by information loss. This can be expressed as: The reverse loss in this process is: The total loss consists of two parts: forward loss and backward loss. The backward loss embodies the core of the framework: it dynamically adjusts the spatiotemporal offset and missing road segments of the observation distribution, while the forward loss evaluates the consistency between temporal extrapolation and spatial extrapolation through similarity judgment. .
[0032] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.
[0033] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0034] In one embodiment, a computer system is provided, which may be a server. The computer system includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data used in the methods described above. The network interface communicates with external terminals via a network connection. The computer program is executed by the processor to implement the methods described above.
[0035] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0036] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.
[0037] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A spatiotemporal data extrapolation method based on bidirectional dynamic coupling, which predicts traffic sensor data in distributed missing data or unmonitored areas through an extrapolation model, characterized in that, include: Construct a spatiotemporal graph of traffic sensors, where nodes represent traffic sensors and edges represent the connectivity between traffic sensors; The traffic sensor spatiotemporal graph is preprocessed using a closed subgraph masking strategy to simulate missing traffic sensor data. The masked traffic sensor spatiotemporal map is input into the spatiotemporal information extraction module to obtain global features. The global features include global time-related information extracted using Transformer and spatial event-related information learned through graph convolutional network. The extracted global features are input into the temporal extrapolation branch and the spatial extrapolation branch, respectively. The temporal extrapolation branch performs temporal extrapolation through a multi-patch layer and a frequency domain enhancement layer, and the spatial extrapolation branch performs spatial extrapolation through a diffusion graph convolutional network. The temporal extrapolation branch and the spatial extrapolation branch provide each other with additional data distribution information. The consistency between temporal and spatial extrapolation is evaluated based on the affinity judgment mechanism to adaptively calculate the forward loss; The reverse loss is calculated by dynamically adjusting the distribution offset between the observed distribution and the missing road segments using a reverse masking strategy. The total loss is calculated based on the forward loss and backward loss, and the extrapolation model is trained.
2. The spatiotemporal data extrapolation method based on bidirectional dynamic coupling according to claim 1, characterized in that, The traffic sensor data includes one or more of the following: traffic flow, traffic carbon emissions, vehicle speed, and vehicle acceleration.
3. The spatiotemporal data extrapolation method based on bidirectional dynamic coupling according to claim 1, wherein constructing a spatiotemporal graph of traffic sensors, where nodes represent traffic sensors and edges represent the connectivity between traffic sensors, specifically includes: Based on the connectivity of traffic sensors, a spatiotemporal graph of traffic sensors is obtained by abstracting nodes of traffic sensors within a region: the road segments within the region are divided, and the spatiotemporal graph uses traffic sensors in the resulting road segment slices as nodes and traffic sensor connectivity as edges; the spatiotemporal graph of traffic sensors... ; Let be the set of edges, representing the connectivity of traffic sensors; , The number of edges, yes The adjacency matrix; This indicates the time span of traffic sensor data; The section is Feature vector at time step , express Traffic sensor nodes eigenvectors, It is the dimension of the feature vector of the traffic sensor node.
4. The spatiotemporal data extrapolation method based on bidirectional dynamic coupling according to claim 1, characterized in that, The preprocessing of the traffic sensor spatiotemporal graph based on the closed subgraph masking strategy to simulate missing traffic sensor data specifically includes: Training images are sampled from the spatiotemporal maps of traffic sensors, and selections are made from the training images containing at least... The closed subgraphs of each traffic sensor node are masked, and the unmasked portion of the training graph is called the retained graph. Then, virtual closed subgraphs are embedded into the retained graph as basic units. An extended graph structure is constructed by inserting virtual traffic sensor nodes between the virtual closed subgraphs and the retained graph. The number of inserted virtual traffic sensor nodes is controlled during the construction of the extended graph structure: the number of inserted virtual traffic sensor nodes... ,in The number of virtual traffic sensor nodes is determined by the traffic sensor data missing rate. This represents the minimum interpolation amount.
5. The spatiotemporal data extrapolation method based on bidirectional dynamic coupling according to claim 1, characterized in that, The masked traffic sensor spatiotemporal map is input into the spatiotemporal information extraction module to obtain global features. These global features include global time-related information extracted using a Transformer and spatial event-related information learned through a graph convolutional network. Specifically, they include: Extracting global time-related information using Transformer: ； The feature vector representing a traffic sensor node. This represents the core attention of the Transformer. This indicates the initial time information extracted; Then, a graph neural network that updates both points and edges simultaneously learns global event-related information: ； in, This refers to a filter based on an attention mechanism. This represents the feature transformation function of traffic sensor nodes. These are the initial features of the v-th traffic sensor node input to the graph neural network. Indicates the first Features of the v-th traffic sensor node in layer 1 Refers to the first The values of adjacent traffic sensor nodes in the layer, Indicates the first Overall characteristics of traffic sensor nodes in the layer This represents the boundary value of the previous layer. This represents the weighting parameter.
6. The spatiotemporal data extrapolation method based on bidirectional dynamic coupling according to claim 4, characterized in that, The spatial extrapolation branch performs spatial extrapolation through a diffusion graph convolutional network, specifically including: The spatial extrapolation branch introduces a diffusion graph convolutional network, which dynamically updates the representation of traffic sensor nodes by simulating the information diffusion process. ； and These are the forward transition matrix and the backward transition matrix, respectively. and These are learning parameters. This represents a diffusion graph neural network. Let Nv represent the feature representation of the Nv-th virtual traffic sensor node in the l-th layer, where Nv is the number of virtual traffic sensor nodes in the additional virtual closed subgraph inserted in the closed subgraph masking strategy. For the first The hidden state of non-zero traffic sensor nodes in the layer. Represents a non-linear activation function. Represents a linear transformation function; Errors in spatial extrapolation ; Indicates the number of mask nodes. This represents the loss hyperparameter used to balance multiple tasks. Represents 0 to The actual value of the mask node at any given time. Indicates spatial extrapolation branches from 0 to The predicted value of the mask node at time step.
7. The spatiotemporal data extrapolation method based on bidirectional dynamic coupling according to claim 6, characterized in that, The time extrapolation branch performs time extrapolation through multiple patch layers and frequency domain enhancement layers, specifically including: The multi-patch layer divides the time series corresponding to the global features into non-overlapping multi-scale blocks, and extracts weighted features at different time scales: ； Represents the weight matrix. Indicates the slice length. Represents global features. This represents the operation of dividing into multi-scale blocks. This represents data after multi-scale weighting; The frequency domain enhancement layer uses discrete cosine transform to extract the frequency components of the data in each spatial channel: ； Let represent the vector after the discrete cosine transform of the i-th channel. Indicates the length of the time-domain signal. This represents the channel characteristic of the l-th channel among n channels in a spatial node. Indicates the number of channels; Temporal features are extracted by combining the frequency components of low-frequency and high-frequency channels: ； Represents frequency domain characteristics, Indicates stacking. This represents the characteristics of the nth channel after the discrete cosine transform. Errors in time extrapolation , To preserve the number of traffic sensor nodes in the graph, This represents the actual observations of the traffic sensor nodes in the retained graph at time T. For time extrapolation branch pairs Predicted values of traffic sensor nodes in a time-preserved graph.
8. The spatiotemporal data extrapolation method based on bidirectional dynamic coupling according to claim 7, characterized in that, The consistency assessment between temporal and spatial extrapolation based on the affinity judgment mechanism, and the adaptive calculation of the forward loss function, specifically includes: The formula for judging affinity is: ; This indicates the affinity between time extrapolation and the Kriging mission; This represents the spatiotemporal extrapolation loss of the shared parameters from the previous round; This represents the shared parameters before the time extrapolation loss iteration; This represents the shared parameters after iterating through the time extrapolation loss; The unique parameter representing the spatial extrapolation task; Whether to calculate forward loss is determined by calculating the affinity between time extrapolation and the Kriging task: if the affinity is positive, forward loss is calculated; if the affinity is negative, forward loss is not calculated. .
9. The spatiotemporal data extrapolation method based on bidirectional dynamic coupling according to claim 1, characterized in that, The process of dynamically adjusting the distribution offset between the observed distribution and the missing road segments using a reverse masking strategy, and calculating the reverse loss, specifically includes: In the reverse process, a masking strategy is used to dynamically adjust invariant learning by swapping nodes between traffic sensor nodes in the training graph and virtual closed subgraphs. Training nodes are converted into masking stations, while virtual closed subgraphs are retained as actual nodes. The system infers information from the virtual closed subgraph using the expanded graph composed of the training graph and the virtual closed subgraph, and uses known information to fill in missing parts, thereby reducing errors caused by information loss. Reverse loss... ; This represents the number of virtual closed subgraphs in the retained graph. Represents 0 to The true value of the virtual closed subgraph of the time-preserved graph. Indicates spatial extrapolation branch pairs Predicted values of the virtual closed subgraph of the time-preserved graph.
10. A computer system comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.