Traffic data prediction method and system based on global spatio-temporal recurrent neural network

Abstract

The application provides a traffic data prediction method and system based on a global space-time recurrent neural network, which comprises the following steps: collecting historical traffic data to obtain an original topology graph, a graph signal sequence of each node and a geographical space feature vector; inputting the graph signal sequence and the geographical space feature vector into a global space-time recurrent neural network model to obtain a time-state mode similarity matrix of each node; training the global space-time recurrent neural network model by calculating a loss; re-encoding the geographical space feature vector of each node by the global space-time recurrent neural network model in the training stage to generate a geographical space feature embedding vector; clustering the nodes and reconstructing the original topology graph based on the clustering result to obtain a new topology graph; and inputting the new topology graph and the graph signal sequence based on the new topology graph into a prediction model to obtain an output of a target time period in the future. The application can improve the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of spatiotemporal data prediction technology, and in particular to a traffic data prediction method and system based on a global spatiotemporal recurrent neural network. Background Technology

[0002] In spatiotemporal data analysis, it is necessary to consider not only the local spatial relationships caused by the direct adjacency between spatial locations, but also the global spatial correlations between geographically distant locations to improve the accuracy of prediction methods. For example, locations with similar geographical features exhibit similar variable values. Taking traffic flow data from two different functional areas in a city—office and residential areas—as an example, people typically travel from residential areas to office areas during the morning rush hour and return during the evening rush hour. Therefore, for locations within the same office area, their morning rush hour traffic inflow is similar. However, for residential areas, due to different regional characteristics, the inflow trend will be completely different from that of the office area.

[0003] Previous studies defined the neighbors of nodes in a graph based on spatial distance, neglecting the fact that nodes geographically distant may also share similar spatiotemporal patterns. Li et al. and He et al. proposed defining adjacency relationships between nodes based on temporal pattern similarity, constructing a spatial association graph. Due to the complex dynamic changes and time drift phenomena in spatiotemporal data, traditional static distance metrics such as Euclidean distance are generally unsuitable for calculating the similarity between two time series. Therefore, the aforementioned studies employed the Dynamic Time Warping (DTW) algorithm to calculate the temporal pattern similarity between two time series. However, given time series data recorded from multiple geographical locations, directly calculating node similarity using DTW distance only considers temporal pattern similarity, ignoring the influence of spatial factors. Furthermore, time series may contain dirty data, i.e., erroneous / missing data recorded due to sensor malfunctions, etc. In such cases, considering only DTW distance leads to distortion of temporal pattern similarity.

[0004] Spatial features are "static," meaning they do not change or change very slowly over time. Furthermore, there are many types of spatial features, such as POI features, terrain features, and road connectivity, making it difficult to manually determine which type is more important in the prediction task; this requires expert knowledge in the field. In addition, the same spatial location may exhibit different temporal patterns in the morning, noon, and evening. The similarity of temporal patterns also changes over time due to the influence of human activity intensity and weather conditions in different locations. Therefore, it is necessary to consider the temporal pattern similarity in spatiotemporal data across different time periods. Summary of the Invention

[0005] To address the issue that most existing prediction methods use distance-based topological graph structures to model spatial dependencies, considering only local spatial correlations and neglecting the relationships between distant nodes, this invention aims to provide a traffic data prediction method and system based on a global spatial temporal recurrent neural network. This method utilizes an improved spatiotemporal GRU and spatial attention mechanism to establish a global spatial temporal recurrent neural network (GSTRNN), comprehensively considering node temporal pattern similarity and spatial feature similarity to model global spatial correlations and improve prediction accuracy.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] On the one hand, a traffic data prediction method based on a global spatiotemporal recurrent neural network is provided, which includes the following steps:

[0008] Historical traffic data is collected to obtain the original topology map, the graph signal sequence of each node, and the geospatial feature vector; the node refers to the location of the monitoring sensor.

[0009] The global spatiotemporal recurrent neural network model is constructed by inputting the graph signal sequence and the geospatial feature vector, and the temporal pattern similarity matrix of each node is predicted.

[0010] The loss is calculated using the predicted temporal pattern similarity matrix and the calculated temporal pattern similarity matrix, and the global spatiotemporal recurrent neural network model is trained through gradient backpropagation.

[0011] The global spatiotemporal recurrent neural network model re-encodes the geospatial feature vectors of each node during the training phase to generate geospatial feature embedding vectors for each node.

[0012] The nodes are clustered based on their geospatial feature embedding vectors, and the original topology map is reconstructed based on the clustering results to obtain a new topology map.

[0013] The new topology map and the graph signal sequence based on the new topology map are used as inputs to the prediction model to obtain the output for the future target time period.

[0014] Optionally, the geospatial features include: latitude and longitude, POI features, terrain features, surrounding buildings and facilities, and road connectivity.

[0015] Optionally, for the graph signal sequence and geospatial feature vector input to the global spatiotemporal recurrent neural network model, two fully connected layers are used to process them respectively; the geospatial feature vector is re-encoded by the fully connected layer to obtain a geospatial feature embedding vector, which is then concatenated with the processed graph signal sequence to obtain a complete node feature vector, which is input into the spatiotemporal GRU; the spatiotemporal GRU extracts the feature information of each node to generate a node state vector, and then uses a Gaussian kernel function to calculate the temporal pattern similarity between node pairs.

[0016] Optionally, for each node, the importance of different neighboring nodes is calculated using a spatial attention mechanism to generate a region information vector. After obtaining the region information vector, the spatiotemporal GRU is improved to obtain an improved spatiotemporal GRU that can aggregate the node region information vectors. The improved spatiotemporal GRU is used to generate a node state vector. Then, the temporal pattern similarity between node pairs is calculated using a Gaussian kernel function.

[0017] Optionally, the calculated temporal pattern similarity matrix refers to the temporal pattern similarity matrix calculated using the FastDTW algorithm, specifically including:

[0018] Given the time series data recorded by any two nodes in the node set of the topology graph at a given historical time step and a future time step, calculate the DTW distance between the two time series data to obtain the sequence distance matrix; and calculate the temporal pattern similarity between the two nodes based on the sequence distance matrix.

[0019] Optionally, based on the geospatial feature embedding vectors of each node, the KMeans clustering algorithm is used to cluster the nodes, grouping nodes with similar geospatial features into the same category;

[0020] For each node, select λ nodes that are in the same category and have the highest temporal pattern similarity to it, establish edges, and add them to the original topology graph to obtain a new topology graph.

[0021] Optionally, the prediction model employs a denoised spatiotemporal graph attention network;

[0022] Given a graph signal sequence and its corresponding new topology graph as input, a denoising spatiotemporal graph attention network is trained to learn a mapping function to predict the graph signal sequence for a future target time period.

[0023] For a new given input, the trained denoised spatiotemporal graph attention network is used to predict the output.

[0024] On the other hand, a traffic data prediction system based on a global spatiotemporal recurrent neural network is provided, the system comprising:

[0025] Traffic data acquisition module, used to collect traffic data;

[0026] The prediction module is used to predict the collected traffic data according to the traffic data prediction method described above.

[0027] The monitoring app is used to provide data query, display, online update and modification, making it convenient for managers to monitor in real time.

[0028] Optionally, the traffic data acquisition module includes: traffic monitoring instruments, serial port server, management workstation, interface server, information center storage server, and platform core switch;

[0029] The traffic monitoring instrument is connected to the management workstation via the serial port server. The management workstation is connected to the interface server. The interface server is connected to the information center storage server via a one-way isolation gateway. The information center storage server is connected to the platform core switch via optical fiber. The platform core switch is connected to multiple cloud computing nodes and web servers.

[0030] The monitoring app includes: a client, a server, and a system management backend;

[0031] The client is used for user registration and login, online query, modification, and logout; the server is used for registration and login verification, as well as data transmission, addition, modification, and deletion functions; the system management backend is used for database management.

[0032] On the other hand, an electronic device is provided, the electronic device comprising:

[0033] processor;

[0034] A memory storing computer-readable instructions, which, when loaded and executed by the processor, implement the steps of the prediction method described above.

[0035] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement the steps of the prediction method described above.

[0036] The beneficial effects of the technical solution provided by this invention include at least the following:

[0037] This invention proposes a traffic data prediction method based on a global spatiotemporal recurrent neural network. It predicts the temporal pattern similarity between nodes within a target time period by establishing a global spatiotemporal recurrent neural network and encoding geospatial feature information to dynamically determine the impact of multiple types of geospatial information on node similarity in the traffic spatiotemporal data prediction task. Simultaneously considering the temporal pattern similarity and spatial feature similarity of spatial nodes in the traffic spatiotemporal data, topological relationships are added to the distance-based topology map, thereby obtaining global spatial correlation. The method is applied to a traffic dataset, and experimental results verify its effectiveness.

[0038] This invention also constructs a traffic data prediction system, which consists of three interconnected parts: a traffic data acquisition module, a prediction module, and a monitoring APP, forming a complete system that facilitates real-time monitoring and provides a useful reference for research in the field of traffic data prediction. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of a traffic data prediction method based on a global spatiotemporal recurrent neural network provided in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the training stage of the global spatiotemporal recurrent neural network provided in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of the application stage of the global spatiotemporal recurrent neural network provided in an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of the ST-GRU network provided in an embodiment of the present invention;

[0044] Figure 5 This is a schematic diagram of the data prediction process provided in an embodiment of the present invention;

[0045] Figure 6 This is a schematic diagram of the traffic data acquisition module provided in an embodiment of the present invention;

[0046] Figure 7 This is a prediction curve of taxi outflow data recorded in two different regions of the TaxiBJ dataset before the DSTGAT model provided in this embodiment of the invention integrates GSTRNN.

[0047] Figure 8 This is the prediction curve of taxi outflow data recorded in two different regions of the TaxiBJ dataset after the DSTGAT model is integrated with GSTRNN provided in this embodiment of the invention.

[0048] Figure 9 This is a schematic diagram illustrating the impact of the number of clusters k on prediction, provided in an embodiment of the present invention.

[0049] Figure 10 This is a temporal pattern similarity matrix heatmap obtained by calculating the DTW distance from graph signal sequence data for two different target time periods, as provided in this embodiment of the invention.

[0050] Figure 11 This is a schematic diagram illustrating the impact of the value of λ on prediction provided in an embodiment of the present invention;

[0051] Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0053] This invention provides a traffic data prediction method based on a global spatiotemporal recurrent neural network, such as... Figure 1 As shown, the method includes the following steps:

[0054] S1. Collect historical traffic data to obtain the original topology map, the graph signal sequence of each node, and the geospatial feature vector; the node refers to the location of the monitoring sensor.

[0055] Specifically, for each node in the original topology map, the corresponding graph signal sequence can be obtained along the time axis. The geospatial features include: latitude and longitude, POI features, terrain features, surrounding buildings and facilities, road connectivity, etc.

[0056] S2. Input the graph signal sequence and the geospatial feature vector into the constructed global spatiotemporal recurrent neural network model to predict the temporal pattern similarity matrix of each node.

[0057] As an optional embodiment of the present invention, the graph signal sequence and geospatial feature vector input to the global spatiotemporal recurrent neural network model are processed by two fully connected layers respectively; the geospatial feature vector is re-encoded by the fully connected layer to obtain a geospatial feature embedding vector, which is then concatenated with the processed graph signal sequence to obtain a complete node feature vector, which is input to the spatiotemporal GRU; the spatiotemporal GRU extracts the feature information of each node to generate a node state vector, and then uses a Gaussian kernel function to calculate the temporal pattern similarity between node pairs.

[0058] Furthermore, for each node, the importance of different neighboring nodes is calculated using a spatial attention mechanism to generate a region information vector. After obtaining the region information vector, the spatiotemporal GRU is improved to obtain an improved spatiotemporal GRU that can aggregate the node region information vectors. The improved spatiotemporal GRU is used to generate node state vectors. Then, the temporal pattern similarity between node pairs is calculated using a Gaussian kernel function.

[0059] S3. Calculate the loss using the predicted temporal pattern similarity matrix and the calculated temporal pattern similarity matrix, and train the global spatiotemporal recurrent neural network model through gradient backpropagation.

[0060] As an optional embodiment of the present invention, the temporal pattern similarity matrix is ​​calculated using the FastDTW algorithm, specifically including:

[0061] Given the time series data recorded by any two nodes in the node set of the topology graph at a given historical time step and a future time step, calculate the DTW distance between the two time series data to obtain the sequence distance matrix; and calculate the temporal pattern similarity between the two nodes based on the sequence distance matrix.

[0062] S4. During the training phase, the global spatiotemporal recurrent neural network model re-encodes the geospatial feature vectors of each node to generate geospatial feature embedding vectors for each node.

[0063] This step is completed during the training phase of the global spatiotemporal recurrent neural network model and can be performed simultaneously with steps S2 and S3. The geospatial feature vector is re-encoded using the fully connected layer of the global spatiotemporal recurrent neural network model to obtain the geospatial feature embedding vector.

[0064] S5. Cluster the nodes according to the geospatial feature embedding vectors of each node, and reconstruct the original topology map based on the clustering results to obtain a new topology map.

[0065] As an optional implementation of the present invention, the KMeans clustering algorithm is used to cluster the nodes according to the geospatial feature embedding vector of each node, and nodes with similar geospatial features are classified into the same category.

[0066] For each node, select λ nodes that are in the same category and have the highest temporal pattern similarity to it, establish edges, and add them to the original topology graph to obtain a new topology graph.

[0067] S6. Using the new topology map and the graph signal sequence based on the new topology map as input to the prediction model, the output for the future target time period is obtained.

[0068] As an optional embodiment of the present invention, the prediction model employs a denoised spatiotemporal graph attention network; given a graph signal sequence and its corresponding new topology graph as input, the denoised spatiotemporal graph attention network is trained to learn a mapping function to predict the graph signal sequence for the future target time period; for a newly given input, the trained denoised spatiotemporal graph attention network is used to predict the output.

[0069] This invention predicts the temporal pattern similarity between nodes in a target time period by establishing a global spatiotemporal recurrent neural network and encoding geospatial feature information to dynamically determine the impact of multiple types of geospatial information on node similarity in traffic spatiotemporal data prediction tasks. Simultaneously considering the temporal pattern similarity and spatial feature similarity of various spatial locations in the traffic spatiotemporal data, topological relationships are added to the distance-based topology map, thereby obtaining global spatial correlation. Compared with existing technologies, this improves prediction accuracy.

[0070] Specifically, the traffic data prediction problem based on a global spatiotemporal recurrent neural network is defined as follows:

[0071] The problem of traffic data prediction based on graph neural network methods is defined as follows: given historical data T... h Traffic data map signal sequence at each time point Given a graph G as input, the goal is to learn a function f to predict the next step T. p Traffic data target value at each time point

[0072] However, the input to the function f modeled by the predictive model is an information-poor graph. in, Let be the set of nodes in the graph. ε represents the number of nodes, corresponding to the N sensor nodes (locations of traffic monitoring instruments) in the traffic data; ε is the set of edges, representing the connectivity between nodes, the existence of edges between nodes depends on the spatial distance between them; A∈R N×N Let G be the adjacency matrix of graph G.

[0073] To address the aforementioned problems, this invention proposes a method that integrates temporal pattern similarity and geospatial feature similarity among nodes to consider the global spatial correlation between nodes, thereby adding new node associations ε to information-scarce graph G.sim The specific definition is as follows:

[0074] Given a graph signal sequence Geospatial feature vectors of each node and As input, by learning a mapping function Obtain new spatially rich graphs in, Represents node n i Spatial eigenvectors, F s This represents the corresponding vector dimension.

[0075]

[0076] Based on the above problem definition, the original traffic data prediction problem definition is modified to: given a traffic data map signal sequence and its corresponding new graph G new As input, learn a mapping function. Predicting the next T p Traffic data target value at each time step in Represents node n i From t+1 to t+T P The value of the target variable within the time period.

[0077]

[0078] The spatial correlation modeling based on Global Spatiotemporal Recurrent Neural Network (GSTRNN) proposed in this invention mainly includes two parts: node temporal pattern similarity prediction and spatial node clustering.

[0079] In node temporal pattern similarity prediction, historical T is used. h The graph signal sequence at each time point and the geospatial feature vectors of each node Predicting the future T p The temporal pattern similarity of nodes at a given time point (i.e., the future target time period) is calculated; then, spatial node clustering is used to partition the nodes, grouping nodes with similar geospatial features into the same cluster. Under the combined constraints of temporal pattern similarity and spatial feature similarity, the numerical changes of the node target variable will show similar trends, thus obtaining the global spatial correlation of the nodes. This allows for the reconstruction of the topology structure and improvement of the accuracy of traffic data prediction. The above process is divided into two stages: a training stage and an application stage. Figure 2 and Figure 3 As shown.

[0080] (1) During the GSTRNN model training phase, the DTW distance generation temporal pattern similarity matrix O∈R is first calculated for the graph signal sequence data of historical and target time periods. N×N Then, the historical graph signal sequence data X... in The geospatial feature vectors S of each node, constructed from geographical features such as POIs, are input into the GSTRNN model to predict the temporal pattern similarity matrix for the target time period. Using the predicted matrix The loss (Loss) is calculated using the matrix O obtained from the DTW distance calculation, and the GSTRNN model parameters are updated through gradient backpropagation.

[0081] Specifically, the fast-DTW algorithm is used to calculate the temporal pattern similarity of graph signal sequences recorded by different nodes in the topological graph G, and the similarity between the obtained node pairs is used as the temporal pattern similarity of the nodes.

[0082] Specifically, given history T h A time step and the future T p Total time steps T a The set of nodes in the topological graph G at each time step. In the context of any two spatial nodes n d and n l The recorded time series data containing F variables and calculate and The DTW distance yields the sequence distance matrix. set up express and If the matching degree is:

[0083]

[0084] Here, min(·) represents taking the minimum value among multiple values. As shown in Algorithm 1, the DTW distance between the corresponding graph signal sequences of two nodes is calculated. The reciprocal of the value is used as the temporal pattern similarity between the two nodes.

[0085]

[0086] (2) In the GSTRNN application stage, for the model obtained in the above training stage, the historical image signal sequence X is used. in The geospatial feature vectors S of each node are input into the model to obtain the temporal pattern similarity matrix. Furthermore, during the training phase, the model re-encodes the geospatial feature input S of each node to obtain the geospatial feature embedding vectors of each node. Clustering is performed to classify nodes into categories. Finally, for each node, a connection is established between the λ nodes with the highest temporal pattern similarity from the set of nodes in the same cluster category. This connection serves as the result of GSTRNN modeling spatial correlation and is input into the spatiotemporal data prediction model.

[0087] The GSTRNN model contains two types of input: (a) graphical signal sequences For each node in Figure G, the corresponding multivariate time series can be obtained along the time axis. The time series data recorded by monitoring stations / sensors in traffic data, as a representation of the spatiotemporal characteristics of the region, implicitly contains a large amount of "dynamic" feature information. The correlation between nodes changes dynamically over time; therefore, when modeling spatial correlation, it is necessary to simultaneously consider the "dynamic" feature information of each node that changes over time.

[0088] (b) Geospatial feature vectors of each node The data focuses on the geospatial features surrounding each sensor, such as latitude and longitude, POI characteristics, terrain features, surrounding buildings and facilities, and road connectivity. Excluding special cases like human intervention, these features typically have a "static" meaning, meaning they do not change over time, or change so slowly that they can be approximated as constant. Since the evolution of spatiotemporal data is influenced by its own geospatial features, geospatial feature information needs to be considered in spatiotemporal data analysis tasks.

[0089] In this embodiment of the invention, the aforementioned "static" geospatial features, which are typically generated manually using one-hot encoding methods, contain a large amount of redundant information. Therefore, the geospatial feature vectors of each node in the topology map are re-encoded using a fully connected layer, enabling the model to adaptively consider the influence weights of various features on the prediction task and generate compact node geospatial feature embedding vectors.

[0090] Specifically, for two different inputs X in And S, using two fully connected layers to process them respectively. The "static" geospatial features of each node generated from spatial features are embedded into a vector. With the "dynamic" characteristics of nodes Concatenate them to form a complete node feature vector. Input spatiotemporal GRU network, where Node n in graph G i The sequence of eigenvectors.

[0091] S emb =W S S+b S (4)

[0092] Xe =W in X in +b in (5)

[0093]

[0094] in, and These correspond to the learnable parameters in the two fully connected layers, This indicates a splicing operation.

[0095] To address the temporal pattern similarity of nodes, a spatiotemporal GRU is used to extract the feature information of each node to generate a node state vector. Then, a Gaussian kernel function is used on the state vector to calculate the similarity between nodes (locations).

[0096] In node temporal pattern similarity prediction, the simplest method is to assign each node obtained above from 1 to T. h Node feature vector sequence at time step The node state vectors are generated by sequentially inputting them into the spatiotemporal GRU. However, the above simple method only considers the characteristics of the nodes themselves and ignores their regional information.

[0097] Therefore, for each node, we consider extracting the feature information of its neighboring geospatial regions at each time step to obtain a regional information vector. Then, we aggregate the information from the regional information vector as part of the node's features and incorporate it into the node's state. Specifically, given node n... i By using a spatial attention mechanism to calculate the importance of different neighboring nodes, a regional information vector is generated.

[0098]

[0099] As shown in Algorithm 2, based on the set of neighbor nodes Generate node n i Regional information vector The aggregation operation AGG(·) is defined by the following formula:

[0100]

[0101] in These are learnable parameters. β z Indicates compared to n i Other neighboring nodes, neighboring node z to node n i Importance. β z It is calculated using the following formula:

[0102]

[0103] Where relu(·) represents the ReLU activation function, These are learnable parameters.

[0104] After calculating the region information vector, the original spatiotemporal GRU can be improved using the region information vector to obtain the spatiotemporal GRU (ST-GRU) computation node state vector that can aggregate node region information. Figure 4 This is a schematic diagram of an ST-GRU network. Given the node n obtained after fusing geospatial feature information... i Feature vector sequence Represents node n at time t i The feature vectors and the sequence of region information vectors For each time step t, the state vector is calculated iteratively using the following formula. in

[0105]

[0106]

[0107]

[0108]

[0109]

[0110] in, and For learnable parameters, The symbol represents a vector concatenation operation, and ⊙ represents the Hadamard product; These are the update gate, reset gate, and region gate, respectively. Equations (9)-(11) represent the calculation processes of the update gate, reset gate, and region gate, respectively. The above gate mechanism controls the transmission of information by considering that the state vector of a node at the current moment is affected by the node's current information. Previous time state vector and current area information The influence of this process generates the node state vector corresponding to the current time step.

[0111] In the above formula, the result values ​​of all gates are within the range [0,1]. In formula (12) This indicates how much information from the previous time step's state vector is retained when generating the state vector; This indicates how much regional information is retained. Formula (13) represents the calculation process of the new state vector. This means that some information from the previous time step is discarded when updating the state vector; This indicates that the node and region information at the current time step are introduced when updating the state vector.

[0112] Finally, for each node state vector obtained from the above calculation process, the Gaussian kernel function is used to calculate the similarity between node pairs as the output, where ε is the standard deviation.

[0113]

[0114] The output of GSTRNN is a similarity matrix. matrix The value in the i-th row and j-th column Represents node n in the topological graph G i and node n j Temporal pattern similarity.

[0115] Next, the spatial nodes are clustered. The geospatial feature vectors of each node are then analyzed. As part of the dataset, geospatial features differ in nature from sensor-recorded variable data. Sensor-recorded variable data is inherently numerical, while geospatial features largely require manual conversion into numerical form. Geospatial features encompass multiple categories, such as latitude and longitude, POI features, terrain features, surrounding buildings and facilities, and road connectivity. In the early stages of spatiotemporal data analysis, each category of feature needs to be manually encoded into numerical form using methods such as one-hot encoding. Then, the encoded results of multiple categories are concatenated to obtain the nodes in the aforementioned S. geospatial feature vector In spatiotemporal data analysis tasks, the importance of different categories of geospatial features contained in the node geospatial feature vectors is different. Therefore, it is necessary to re-encode the node geospatial feature vectors in the model.

[0116] In this invention, the GSTRNN model automatically learns the importance of different types of geospatial features during the training phase, thereby obtaining the geospatial feature embedding vectors for each node.

[0117] For the geospatial feature embedding vectors S generated by each node in the GSTRNN model emb The nodes are clustered based on the KMeans clustering algorithm, with specific steps as shown in Algorithm 3. The clustering method groups nodes with similar geospatial features into the same category (cluster). Under the dual constraints of temporal pattern similarity and spatial feature similarity, the global spatial correlation of nodes is modeled.

[0118]

[0119] The node similarity matrix obtained through the GSTRNN model is a dense matrix, where each pair of nodes has a numerical value representing the temporal pattern similarity between them. Although spatial node clustering only considers the temporal pattern similarity between nodes in the same category, it may still contain a large number of nodes. To focus on strong correlations and reduce computational complexity, it is necessary to identify the closely related nodes for each node and then construct a sparse matrix by removing low-weight edges from the dense matrix. To this end, for each node in graph G, the λ nodes with the highest temporal pattern similarity in the same category from the clustering results are selected as the set of far-distance spatially related nodes for that node. That is, for each node, edges are established with the λ nodes in the same category with the highest temporal pattern similarity, and these edges are added to the original topological graph G to obtain a new topological graph Gi. new .

[0120] by Figure 5 Taking traffic data from six monitoring stations as an example, Figure 5 In Figure (a), the topological graph G is constructed based on the distance between monitoring stations. For ease of understanding, the node numbers are represented by numbers on the nodes; the connections between nodes are the edges ε contained in G, and the existence of an edge between nodes indicates that there is a spatial correlation between them. Figure 5 In (b), a new topological graph structure G is obtained by adding new spatial relationships through GSTRNN and spatial node clustering. new In the example, the dashed lines between node 1 and node 4, node 2 and node 5, and node 3 and node 6 represent the new spatial relationships between nodes constructed through GSTRNN and spatial node clustering, i.e., the new edges ε. sim .

[0121] Algorithm 4 presents the spatiotemporal data prediction steps based on GSTRNN, using GSTRNN and spatial node clustering methods to model global spatial correlations and obtain a new topological graph G. new The spatial relationships inherent in the original topological graph G, which is rich in information but lacks rich details. For example... Figure 5 As shown in (c), a denoised spatiotemporal graph attention network is used as the prediction model to obtain a new topological graph G. new and based on G new Using graph signal sequences with topological structures as input to the prediction model improves the accuracy of spatiotemporal data prediction.

[0122]

[0123]

[0124] Accordingly, embodiments of the present invention also provide a traffic data prediction system based on a global spatiotemporal recurrent neural network, the system comprising:

[0125] Traffic data acquisition module, used to collect traffic data;

[0126] The prediction module is used to predict the collected traffic data according to the traffic data prediction method described above.

[0127] The monitoring app is used to provide data query, display, online update and modification, making it convenient for managers to monitor in real time.

[0128] The prediction module described in this embodiment can be used to execute... Figure 1 The technical solutions of the illustrated method embodiments are similar in principle and technical effect, and will not be described again here. The system constructed by this invention consists of three interconnected parts: a traffic data acquisition module, a prediction module, and a monitoring APP, forming a complete system that facilitates real-time monitoring.

[0129] Furthermore, such as Figure 6 As shown, the traffic data acquisition module includes: traffic monitoring instruments, serial port server, management workstation, interface server, information center storage server, and platform core switch;

[0130] The traffic monitoring instrument is connected to the management workstation via the serial port server. The management workstation is connected to the interface server. The interface server is connected to the information center storage server via a one-way isolation gateway. The information center storage server is connected to the platform core switch via optical fiber. The platform core switch is connected to multiple cloud computing nodes and web servers.

[0131] The traffic data acquisition module is a support system responsible for traffic data acquisition. It is developed using C++ and incorporates multiple data communication protocols from IEC 60870-5, including 101, 102, 103, 104, Modbus, CDT, and DISA. Its modeling conforms to the interface reference model, common information model (CIM), and component interface specification (CIS) requirements of IEC 61970, meeting international standards and allowing for seamless integration with various systems as middleware. It enables the access of data from monitoring systems, integrated management systems, metering, fault analysis, and alarm push notification systems. The system supports the access of multiple devices and has the ability to parse multiple protocols.

[0132] The traffic monitoring instruments are used to collect data such as traffic flow and upload it centrally to the cloud platform. The information center storage server, i.e., the cloud data center, uses a database to store multi-dimensional spatiotemporal data and provides retrieval services. The database adopts a real-time database system design. A real-time database system is a new type of database management system software. Based on a 64-bit system, its high-speed database engine and advanced distributed cluster architecture make this invention suitable for the collection, storage, retrieval, and publication of massive real-time / historical data. It has excellent horizontal scalability and high availability, and can handle dynamic data that changes rapidly over time, improving the speed and efficiency of data retrieval.

[0133] The monitoring app mainly includes a client, a server, and a system management backend. The client is developed using the MUI front-end framework and HTML5, CSS, and JavaScript for user registration and login, online querying, modification, and logout. The server uses the ThinkJS server-side framework and a MySQL database for registration and login verification, as well as data transmission, addition, modification, and deletion. The system management backend is developed using HTML5, CSS, and JavaScript for database management.

[0134] The monitoring app is simple and convenient to use, with a clean and aesthetically pleasing interface. It offers real-time monitoring, allowing registered users to log in from anywhere via their mobile phones. The system provides automatic query and display functions, as well as user registration information management capabilities.

[0135] The effectiveness of the method of the present invention will be verified through specific embodiments below.

[0136] First, the required dataset is obtained through the traffic data acquisition module. Here, this invention uses the TaxiBJ traffic flow dataset to verify the effectiveness of the GSTRNN model. Details of this dataset are as follows:

[0137] The TaxiBJ dataset records the number of taxis entering / leaving each of 315 areas in Beijing per hour from February to June 2015, serving as the hourly inflow / outflow for that area. The data missing rate is 5.32%. The prediction task is to use the historical taxi inflow and outflow data for each area to predict the taxi outflow for the next 12 hours.

[0138] Using Beijing POI data, road network data, and GPS coordinates, geospatial feature vectors containing geospatial information were constructed for the aforementioned dataset. The POI data includes 982,829 points of interest across 668 categories in Beijing; the road network data includes information such as road classification and number of lanes for 690,242 roads in Beijing. The geospatial feature vectors for each node were obtained by statistically analyzing features such as the number of different types of points of interest and the number of roads within the corresponding area.

[0139] The proposed GSTRNN model is implemented using the PyTorch deep learning framework. In experiments, GCRN, ASTGCN, and DSTGAT are used as benchmark models to verify the effectiveness of the proposed GSTRNN in modeling spatial correlations. During the experiments, the TaxiBJ dataset is divided into training, validation, and test sets in a 7:1:2 ratio; the time ranges are T... h =24 and T p =12, the parameters of ST-GRU in the GSTRNN model are set to F respectively. g =84, F h =72, F con =72; The dimension of the node spatial feature embedding vector is set to F. se =36, the dimension of the fully connected layer parameter for processing graph signal sequence input is set to F. e =36; number of clusters k=5. Batch size is set to 16, Adam optimization algorithm is selected, initial learning rate is set to 0.0005, decay rate is 0.92 per epoch, and model is trained for 50 epochs.

[0140] The experimental results are analyzed as follows:

[0141] (1) Prediction performance comparison

[0142] On the TaxiBJ dataset, the effectiveness of the proposed GSTRNN model is verified by comparing the prediction accuracy before and after integrating the GSTRNN model with three benchmark models: GCRN, ASTGCN, and DSTGAT. For ease of representation, all models integrated with the GSTRNN model are named "Benchmark Model+". The average prediction results of each model over the next 12 hours (12 time steps) are shown in Table 1. It can be seen that the MAE and RMSE of the prediction results of the three benchmark models are improved to varying degrees before and after integrating the GSTRNN model. On the TaxiBJ dataset, the MAE improved by an average of approximately 5%, and the RMSE improved by an average of 3%. Specifically, DSTGAT+ improved the MAE by 5.25% and the RMSE by 3.41% compared to DSTGAT over the 12-hour average prediction results.

[0143] The above comparative experimental results demonstrate the effectiveness of the GSTRNN model proposed in this invention. While GCRN, ASTGCN, and DSTGAT models all use distance-based topology graph structures, GSTRNN obtains the associations between distant nodes in the topology graph through temporal pattern similarity and spatial feature similarity of nodes in the topology graph, enriching the spatial association information in the topology graph. This enables GCRN, ASTGCN, and DSTGAT to more accurately model spatial dependencies in spatiotemporal data.

[0144] Furthermore, compared to the GCRN and ASTGCN models, the DSTGAT model proposed in this invention achieves better prediction performance on the TaxiBJ taxi traffic prediction dataset, further verifying the superiority of DSTGAT.

[0145] Table 1 Comparison of prediction performance of each model before and after integrating GSTRNN

[0146]

[0147] (2) Visualization of prediction results

[0148] Figure 7 and Figure 8 The graph shows the prediction results of taxi outflow data for two different regions in the TaxiBJ dataset before and after integrating GSTRNN into the DSTGAT model. "Ground Truth" represents the actual outflow data for these two regions; the horizontal axis represents the time axis, with an interval of 1 hour. It can be seen that DSTGAT+, combined with the GSTRNN model, achieves more accurate prediction results for different regions in the TaxiBJ dataset compared to the original DSTGAT model. In contrast, the DSTGAT model's prediction results are more conservative, which is more evident in the prediction results for taxi outflow data in region one. For the peak outflow in region one, i.e., the 600-700 range, DSTGAT's prediction results are often significantly lower, failing to accurately fit the actual value. However, by integrating GSTRNN, the resulting DSTGAT+ model more accurately fits the peak value of the actual taxi outflow.

[0149] (3) Spatial node clustering analysis

[0150] Figure 9The impact of different cluster numbers k (k=3, 4, 5, 6) on the prediction accuracy of DSTGAT+ is presented, using the average prediction results at the 12-hour prediction level. The horizontal axis represents the number of clusters, the left vertical axis represents the MAE of the prediction result, and the right vertical axis represents the RMSE of the prediction result. It can be seen that the prediction accuracy of DSTGAT+ initially increases and then decreases with the increase of the number of clusters k. The worst prediction result is achieved when the number of clusters k=3. As the number of clusters increases, the model's prediction accuracy shows an upward trend. The best performance is achieved when the number of clusters is 5. However, with further increases in the number of clusters, the prediction accuracy of DSTGAT+ begins to decrease.

[0151] Experimental results show that when the number of clusters is too small, it becomes difficult to distinguish differences between spatial locations, essentially relying solely on the temporal pattern similarity of nodes to obtain spatial correlation, leading to a decline in prediction performance. When the number of clusters is too large, the spatial correlation obtained by GSTRNN focuses more on the spatial feature similarity of nodes. If the number of nodes in the same category is too small, issues such as missing values ​​may prevent the selection of nodes with truly similar temporal patterns, thus affecting the model's prediction accuracy. In summary, setting an appropriate number of clusters allows GSTRNN to simultaneously consider both temporal pattern similarity and spatial feature similarity to obtain the spatial correlation between nodes.

[0152] (4) Temporal pattern similarity analysis

[0153] like Figure 10 The temporal pattern similarity comparison of the nodes shown is illustrated, where, Figure 10 Images (a) and (b) are heatmaps of temporal pattern similarity matrices obtained by calculating DTW distance for graph signal sequence data from two different target time periods. It can be seen that the temporal pattern similarity of nodes changes dynamically over time. Figure 10 In the middle (c), the temporal pattern similarity of the target time period 1 node obtained using GSTRNN is shown, which fits the data well. Figure 10 The results in (a) demonstrate the effectiveness of GSTRNN in calculating temporal pattern similarity. Figure 10 (d) presents a heatmap of node temporal pattern similarity obtained by calculating DTW distance using all graph signal sequence data in the training set at once. It can be seen that it is similar to... Figure 10 The differences between (a) and (b) are significant because temporal patterns in spatiotemporal data change over time. The same monitoring station may exhibit different temporal patterns in the morning, noon, and evening. Furthermore, the similarity of temporal patterns between different monitoring stations is influenced by factors such as the intensity of human activity and meteorological conditions in their respective areas, causing these patterns to change continuously over time. Therefore, it is necessary to consider the temporal pattern similarity in spatiotemporal data across different time periods. This is provided by the embodiments of the present invention.

[0154] Figure 11 The paper presents the impact of DSTGAT+ on the average prediction results of taxi outflow in various regions over 12 hours on the TaxiBJ dataset for multiple cases where λ = 2, 5, 7, and 9. The model achieves the best performance when λ = 5. This is because when the number of λ values ​​is too small, the spatial relationships contained in the topological graph structure become insufficient. Specifically, when λ = 0, it corresponds to the prediction results of the DSTGAT model without using the proposed GSTRNN model to model spatial correlations.

[0155] In summary, this invention proposes establishing a global spatiotemporal recurrent neural network to predict the temporal pattern similarity between nodes in a target time period, and encodes geospatial feature information to dynamically determine the impact of multiple types of spatial information on node similarity in spatiotemporal data prediction tasks. Simultaneously considering the temporal pattern similarity and spatial feature similarity of various spatial locations in the spatiotemporal data, topological relationships are added to the distance-based topology map, thereby obtaining global spatial correlation. The method is applied to a traffic dataset, and experimental results verify the effectiveness of the proposed method.

[0156] The invention also constructs a traffic data prediction system, which consists of three interconnected parts: a traffic data acquisition module, a prediction module, and a monitoring APP, forming a complete system that facilitates real-time monitoring and provides a useful reference for research in the field of traffic data prediction.

[0157] In an exemplary embodiment, the present invention also provides an electronic device, the electronic device comprising:

[0158] processor;

[0159] A memory storing computer-readable instructions, which, when loaded and executed by the processor, implement the steps of the prediction method described above.

[0160] Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 12 As shown, the electronic device 300 may include a processor 3001 and a memory 3002. Optionally, the electronic device 300 may also include a transceiver 3003. The processor 3001, memory 3002, and transceiver 3003 may be connected via a communication bus. The memory 3002 stores computer-readable instructions, which, when executed by the processor 3001, implement the steps of the prediction method described above.

[0161] In a specific implementation, as one example, the processor 3001 may include one or more CPUs, for example... Figure 12 CPU0 and CPU1 are shown in the diagram.

[0162] In a specific implementation, as one example, the electronic device 300 may also include multiple processors, for example... Figure 12 The processors 3001 and 3004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, "processor" can refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0163] The memory 3002 is used to store the software program that executes the present invention, and is controlled by the processor 3001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0164] The transceiver 3003 is used to communicate with network devices or with terminal devices.

[0165] Optionally, the transceiver 3003 may include a receiver and a transmitter. The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0166] Optionally, the transceiver 3003 can be integrated with the processor 3001 or exist independently and be coupled to the processor 3001 through the interface circuit of the electronic device 300. This embodiment of the invention does not specifically limit this.

[0167] It should be noted that, Figure 12 The structure of the electronic device 300 shown is not intended to limit the electronic device. Actual electronic devices may include more or fewer components than shown, or combine certain components, or have different component arrangements. Furthermore, the technical effects of the electronic device 300 can be understood by referring to the technical effects of the above-described method embodiments, and will not be repeated here.

[0168] In an exemplary embodiment, the present invention also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the steps of the prediction method described above. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc.

[0169] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0170] The use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0171] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0172] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0173] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0174] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0175] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0176] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0177] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0178] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0179] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A traffic data prediction method based on a global spatiotemporal recurrent neural network, characterized in that, Includes the following steps: Historical traffic data is collected to obtain the original topology map, the graph signal sequence of each node, and the geospatial feature vector; the node refers to the location of the monitoring sensor. The geospatial features include: latitude and longitude, POI features, terrain features, surrounding buildings and facilities, and road connectivity. The global spatiotemporal recurrent neural network model is constructed by inputting the graph signal sequence and the geospatial feature vector, and the temporal pattern similarity matrix of each node is predicted. For the graph signal sequence and geospatial feature vector input to the global spatiotemporal recurrent neural network model, two fully connected layers are used to process them respectively. The geospatial feature vector is re-encoded by the fully connected layer to obtain a geospatial feature embedding vector, which is then concatenated with the processed graph signal sequence to obtain a complete node feature vector, which is input into the spatiotemporal GRU. The spatiotemporal GRU extracts the feature information of each node to generate a node state vector, and then uses a Gaussian kernel function to calculate the temporal pattern similarity between node pairs. For each node, the importance of different neighboring nodes is calculated using a spatial attention mechanism to generate a region information vector. After obtaining the region information vector, a spatiotemporal GRU that can aggregate node region information vectors is used to generate a node state vector. Then, a Gaussian kernel function is used to calculate the temporal pattern similarity between node pairs. The loss is calculated using the predicted temporal pattern similarity matrix and the calculated temporal pattern similarity matrix, and the global spatiotemporal recurrent neural network model is trained through gradient backpropagation. The global spatiotemporal recurrent neural network model re-encodes the geospatial feature vectors of each node during the training phase to generate geospatial feature embedding vectors for each node. The nodes are clustered based on their geospatial feature embedding vectors, and the original topology map is reconstructed based on the clustering results to obtain a new topology map. The new topology map and the graph signal sequence based on the new topology map are used as inputs to the prediction model to obtain the output for the future target time period.

2. The traffic data prediction method based on a global spatiotemporal recurrent neural network according to claim 1, characterized in that, The calculated temporal pattern similarity matrix refers to the temporal pattern similarity matrix calculated using the FastDTW algorithm, specifically including: Given the time series data recorded by any two nodes in the node set of the topology graph at a historical time step and a future time step, calculate the DTW distance between the two time series data to obtain the sequence distance matrix; and calculate the temporal pattern similarity between the two nodes based on the sequence distance matrix.

3. The traffic data prediction method based on a global spatiotemporal recurrent neural network according to claim 1, characterized in that, Based on the geospatial feature embedding vectors of each node, the KMeans clustering algorithm is used to cluster the nodes, and nodes with similar geospatial features are grouped into the same category. For each node, select λ nodes that are in the same category and have the highest temporal pattern similarity to it, establish edges, and add them to the original topology graph to obtain a new topology graph.

4. The traffic data prediction method based on a global spatiotemporal recurrent neural network according to claim 1, characterized in that, The prediction model employs a denoised spatiotemporal graph attention network. Given a graph signal sequence and its corresponding new topology graph as input, a denoising spatiotemporal graph attention network is trained to learn a mapping function to predict the graph signal sequence for a future target time period. For a new given input, the trained denoised spatiotemporal graph attention network is used to predict the output.

5. A traffic data prediction system based on a global spatiotemporal recurrent neural network, characterized in that, The system includes: Traffic data acquisition module, used to collect traffic data; The prediction module is used to predict the collected traffic data using the traffic data prediction method according to any one of claims 1 to 4; The monitoring app is used to provide data query, display, online update and modification, making it convenient for managers to monitor in real time.

6. The traffic data prediction system based on a global spatiotemporal recurrent neural network according to claim 5, characterized in that, The traffic data acquisition module includes: traffic monitoring instruments, serial port server, management workstation, interface server, information center storage server, and platform core switch; The traffic monitoring instrument is connected to the management workstation via the serial port server. The management workstation is connected to the interface server. The interface server is connected to the information center storage server via a one-way isolation gateway. The information center storage server is connected to the platform core switch via optical fiber. The platform core switch is connected to multiple cloud computing nodes and web servers. The monitoring app includes: a client, a server, and a system management backend; The client is used for user registration and login, online query, modification, and logout; the server is used for registration and login verification, as well as data transmission, addition, modification, and deletion functions; the system management backend is used for database management.

7. An electronic device, characterized in that, The electronic device includes: processor; A memory storing computer-readable instructions that, when loaded and executed by the processor, implement the method as described in any one of claims 1 to 4.

Images