Traffic state prediction method, device, electronic device and storage medium

By extracting the time-space correlation and time-domain variation of traffic state data, using graph convolution and time-convolution networks, the problem of low traffic state prediction accuracy in the prior art is solved, and a higher precision traffic state prediction is achieved.

CN115631622BActive Publication Date: 2024-11-26TSINGHUA UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211153883.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-09-21
Publication Date
2024-11-26
Estimated Expiration
2042-09-21

AI Technical Summary

Technical Problem

The existing traffic state prediction method based on machine learning cannot accurately and efficiently extract nonlinear features and spatiotemporal correlation features in traffic state data, resulting in low prediction results accuracy.

Method used

By extracting the spatial and temporal correlation and extracting the time domain change of historical traffic state data, the graph convolution network and temporal convolution network in the short-range and long-range architecture search modules are used to obtain the results of hidden space feature fusion to predict future traffic states.

Benefits of technology

The accuracy of traffic state prediction results is improved, nonlinear features and spatiotemporal correlation characteristics in the data can be more reliable, and the reliability of prediction results is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115631622B_ABST
    Figure CN115631622B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of urban computing, and provides a traffic state prediction method, device, electronic device and storage medium, the method comprising: receiving first traffic state data within a historical period; performing spatiotemporal correlation extraction and temporal change extraction based on the first traffic state data, and determining second traffic state data within a future period based on the results of the spatiotemporal correlation extraction and the results of the temporal change extraction. Since the spatiotemporal correlation extraction and the temporal change extraction can be performed based on the first traffic state data, the nonlinear features and spatiotemporal correlation features hidden in the first traffic state data can be obtained, so that the output data reliability is higher, the accuracy of the traffic state prediction results is improved, and the problem of low accuracy of the prediction results of the existing traffic state prediction methods is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of urban computing technology, and in particular to a traffic state prediction method, device, electronic equipment and storage medium. Background Art

[0002] Urban computing is a process of solving challenges faced by cities (such as environmental degradation, traffic congestion, increased energy consumption, backward planning, etc.) by continuously acquiring, integrating and analyzing various heterogeneous big data in cities. It can help people understand the nature of various urban phenomena and even predict the future of cities. The application of urban computing in the field of intelligent transportation often involves the prediction of traffic status.

[0003] Traffic status refers to the operating conditions of traffic flow, which can reflect the flow status of vehicles and pedestrians on the road. Existing traffic status prediction methods based on machine learning cannot accurately and efficiently extract the nonlinear features and spatiotemporal correlation features hidden in traffic status data due to the unreasonable design of the deep learning model structure applied, resulting in low accuracy of the final prediction results and difficulty in achieving accurate traffic status prediction. Summary of the invention

[0004] The present invention provides a traffic status prediction method, device, electronic device and storage medium, which are used to solve the problem that the traffic status prediction method in the prior art has a low prediction result accuracy due to the unreasonable design of the deep learning model structure applied, and realize higher-precision traffic status prediction.

[0005] In a first aspect, the present invention provides a traffic state prediction method, the method comprising:

[0006] Receiving first traffic status data within a historical period;

[0007] Based on the first traffic status data, spatiotemporal correlation extraction and time domain change extraction are respectively performed, and based on the results of the spatiotemporal correlation extraction and the results of the time domain change extraction, second traffic status data in a future time period is determined.

[0008] According to the traffic state prediction method provided by the present invention, the spatiotemporal correlation extraction and the temporal change extraction are respectively performed based on the first traffic state data, and based on the result of the spatiotemporal correlation extraction and the result of the temporal change extraction, the second traffic state data in the future period is determined, including:

[0009] Performing latent space transformation on the first traffic state data to obtain initial latent state data;

[0010] Performing spatiotemporal correlation extraction and time domain change extraction on the initial hidden state data respectively to obtain the spatiotemporal correlation extraction result and the time domain change extraction result;

[0011] The result of the spatiotemporal correlation extraction and the result of the time domain change extraction are fused to obtain a latent space feature fusion result;

[0012] Based on the latent space feature fusion result, the second traffic state data is predicted.

[0013] According to the traffic state prediction method provided by the present invention, the temporal and spatial correlation extraction of the initial hidden state data includes:

[0014] Inputting the initial hidden state data into a short-range architecture search module, and establishing a candidate adjacency matrix set in the short-range architecture search module;

[0015] Determining a target candidate adjacency matrix corresponding to each graph convolutional layer in the short-range architecture search module from the set of candidate adjacency matrices;

[0016] Based on the target candidate adjacency matrix, performing a graph convolution operation on the initial hidden state data to obtain first intermediate hidden state data corresponding to each of the graph convolution layers;

[0017] The first intermediate hidden state data are fused to obtain the result of the spatiotemporal correlation extraction.

[0018] According to the traffic state prediction method provided by the present invention, the fusing of the first intermediate hidden state data to obtain the result of the spatiotemporal correlation extraction includes:

[0019] Performing cropping processing on each of the first intermediate hidden state data respectively to obtain a first processing result;

[0020] Performing linear mapping processing and nonlinear activation processing on each of the first processing results respectively to obtain a second processing result;

[0021] Aggregate the second processing results to obtain the spatiotemporal correlation extraction result.

[0022] According to the traffic state prediction method provided by the present invention, extracting the time domain changes of the initial hidden state data includes:

[0023] Inputting the initial hidden state data into a long-range architecture search module, and establishing a candidate operation set in the long-range architecture search module;

[0024] Determine a target candidate operation corresponding to each vertex in the long-range architecture search module from the candidate operation set;

[0025] Performing a temporal convolution operation on the initial hidden state data based on the target candidate operation to obtain second intermediate hidden state data of each vertex;

[0026] Each of the second intermediate latent state data is fused to obtain the result of the time domain change extraction.

[0027] According to the traffic status prediction method provided by the present invention, the candidate operation set includes at least one of a zero mapping operation, an identity mapping operation, and a gated time convolution operation with different convolution kernels and convolution types.

[0028] According to the traffic state prediction method provided by the present invention, the first traffic state data is obtained specifically through the following process:

[0029] Obtain the original traffic data of the traffic network in the historical period;

[0030] Taking each key object in the traffic network as a node, taking the connection relationship between each key object as an edge, and establishing an initial adjacency matrix based on the connection relationship between each key object, and constructing a traffic network graph based on the nodes, the edges and the initial adjacency matrix;

[0031] Determine the traffic status of each node at different times in the historical period based on the traffic network diagram and the original traffic data;

[0032] Based on the traffic status of each node at different times, first traffic status data is obtained.

[0033] In a second aspect, the present invention further provides a traffic state prediction device, the device comprising:

[0034] A receiving module, used for receiving first traffic status data within a historical period;

[0035] The processing module is used to perform spatiotemporal correlation extraction and temporal change extraction based on the first traffic status data, and determine the second traffic status data in a future time period based on the results of the spatiotemporal correlation extraction and the results of the temporal change extraction.

[0036] In a third aspect, the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a traffic status prediction method as described above is implemented.

[0037] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the traffic status prediction methods described above.

[0038] The traffic state prediction method, device, electronic device and storage medium provided by the present invention obtain second traffic state data in a future time period by processing first traffic state data in a historical time period. Since spatiotemporal correlation extraction and time domain change extraction can be performed based on the first traffic state data respectively, nonlinear features and spatiotemporal correlation features hidden in the first traffic state data can be obtained, thereby making the output second traffic state data more reliable and improving the accuracy of the traffic state prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 It is a flow chart of the traffic state prediction method provided by the present invention;

[0041] Figure 2 Schematic diagram of the overall architecture of the traffic state prediction method in an embodiment of the present invention;

[0042] Figure 3 is a schematic diagram of the overall architecture of the traffic prediction model in an embodiment of the present invention;

[0043] Figure 4 Schematic diagram of the main architecture of the short-range architecture search module in an embodiment of the present invention;

[0044] Figure 5 Schematic diagram of the construction principle of the candidate adjacency matrix in an embodiment of the present invention;

[0045] Figure 6 Schematic diagram of the main architecture of a gated temporal convolutional neural network in an embodiment of the present invention;

[0046] Figure 7 It is a schematic diagram of the main architecture of the long-range architecture search module in an embodiment of the present invention;

[0047] Figure 8 is a schematic diagram of a sample data acquisition process in an embodiment of the present invention;

[0048] Fig. 9 It is a structural schematic diagram of the traffic state prediction device provided by the present invention;

[0049] Fig.10 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Combine the following Figures 1 to 10 The traffic status prediction method, device, electronic device and storage medium provided by the embodiments of the present invention are described.

[0052] Figure 1 The traffic state prediction method provided by an embodiment of the present invention is shown, and the method includes:

[0053] Step 101: receiving first traffic status data within a historical period;

[0054] Step 102: performing spatiotemporal correlation extraction and temporal change extraction based on the first traffic status data, and determining second traffic status data in a future period based on the results of the spatiotemporal correlation extraction and the results of the temporal change extraction.

[0055] In this embodiment, the first traffic status data can represent the relationship between the changes in the traffic status of each key object in the traffic network over time within a previous period of time (i.e., a historical period). In this embodiment, the key object refers to the key information that can reflect different locations in the traffic network, which can specifically be different sections or lanes in the traffic network.

[0056] In an exemplary embodiment, the first traffic status data in the above historical period can be obtained through the following process:

[0057] Obtain the original traffic data of the traffic network in the historical period;

[0058] Each key object in the traffic network is regarded as a node, and the connection relationship between each key object is regarded as an edge. An initial adjacency matrix is ​​established based on the connection relationship between each key object. Based on the nodes, edges and the initial adjacency matrix, a traffic network graph is constructed.

[0059] Based on the traffic network diagram and original traffic data, determine the traffic status of each node at different times in the historical period;

[0060] Based on the traffic status of each node at different times, first traffic status data is obtained.

[0061] In some embodiments, after obtaining the original traffic data of the traffic network within the historical period, the original traffic data can also be preprocessed, and then based on the traffic network diagram and the preprocessed original traffic data, the traffic status of each node at different times within the historical period can be determined, which can improve the accuracy of the first traffic status data.

[0062] In this embodiment, the original traffic data of the traffic network can be pre-acquired through data acquisition devices such as sensors or cameras installed in the traffic network. The executor of the traffic status prediction method in this embodiment can specifically be a pre-constructed traffic prediction model. The traffic prediction model can be configured in the supervision terminal equipment of the traffic management personnel, such as a computer or monitoring station used for traffic supervision, and other hardware carriers. The above-mentioned original traffic data obtained can be centrally stored in the data storage layer in the hardware carrier.

[0063] In actual application, traffic state prediction can be modeled and predicted for the traffic flow at the intersection from the two dimensions of time and space. The predicted traffic flow is the traffic flow in each direction of each intersection. In this embodiment, the traffic network can be modeled as a traffic network graph Among them, V represents the node set obtained after considering each key object in the traffic network as a node, and |V|=N, N represents the number of nodes, E represents the set of edges, and A represents the adjacency matrix, which can be the initial adjacency matrix or the candidate adjacency matrix, which is used to describe the connection relationship between nodes in the traffic network graph.

[0064] According to the above traffic network diagram, the traffic status at time t can be expressed as It can be understood that the traffic status at time t can be presented by an N×d-dimensional vector, where d represents the characteristic dimension of the traffic status of each node.

[0065] In this case, the traffic state prediction problem in this embodiment can be defined as: learning a function f(·) so that the traffic state of the historical period can be accurately mapped to the traffic state of the future period, that is, mapping the first traffic state data to the second traffic state data, which can be specifically expressed as follows:

[0066]

[0067] in, represents the traffic status at time t-T+1, represents the traffic status at time t, represents the traffic status at time t+1, represents the traffic state at time t+T′, T represents the time series length corresponding to the input first traffic state data, and T′ represents the time series length corresponding to the output second traffic state data.

[0068] This embodiment mainly solves the above-mentioned traffic state prediction problem by constructing a traffic prediction model.

[0069] The traffic state prediction method provided in this embodiment can be intuitively used in the actual application process of the hardware carrier. Figure 2 The module architecture shown shows that the overall architecture can be divided into several important parts, such as data storage module 201, data access conversion module 202, traffic prediction model 203 and data security module 204.

[0070] Among them, the data storage module 201 is mainly used to store the original traffic data acquired historically and in real time. The above-mentioned original traffic data can cover various original data that can affect or characterize the traffic status, such as vehicle flow, weather data, accident information, large events, POI (Point Of Interest), pedestrian flow, sensor information, road section information, intersection information and lane information.

[0071] In order to improve the robustness and stability of the traffic prediction model 203 for low-quality data, for example, when the original traffic data is relatively sparse or seriously missing, to prevent the prediction accuracy of the traffic prediction model 203 from dropping sharply, and to ensure that the trained traffic prediction model 203 can meet normal needs, reasonable data preprocessing is also an important step before the execution of the traffic state prediction method.

[0072] To this end, this embodiment also provides a data access conversion module 202, which is mainly used to pre-process the original traffic data. The data pre-processing methods include but are not limited to data cleaning, data desensitization, missing completion, topology construction, and training data generation, which can ensure that the first traffic status data subsequently input into the traffic prediction model 203 is more accurate and complete.

[0073] The traffic prediction model 203 is mainly used to receive the first traffic status data, analyze and process the input first traffic status data, and then output the second traffic status data to achieve accurate prediction of the traffic status.

[0074] Since traffic data has complex correlations in the spatiotemporal dimension, it is crucial to build a good traffic prediction model to extract spatiotemporal correlation features to improve the accuracy of traffic prediction. At the same time, the spatial correlation between different key objects in the traffic network is multifaceted and can be determined by the topological structure of the traffic network or the similarity of time series, or the similarity of points of interest between regions. In addition, due to the mixture of nonlinear changes and periodicity, the temporal dependence of the same key object at different times is also complex.

[0075] Although there are currently prediction models with spatiotemporal extraction functions, the spatiotemporal extraction modules are artificially designed and predefined, and cannot be adaptively adjusted according to different data, resulting in the accuracy of the prediction results being difficult to guarantee. To this end, this embodiment establishes a traffic prediction model to achieve more accurate traffic status prediction based on the improvement of the internal structure of the model.

[0076] In this embodiment, the traffic prediction model 203 mainly includes a multimodal embedding representation layer, a spatiotemporal correlation modeling layer, a temporal variation modeling layer, a feature fusion layer, and a prediction result conversion and export layer;

[0077] The multimodal embedding representation layer is mainly used to convert the first traffic state data from the vector corresponding to the traffic road network diagram into a higher-dimensional latent space vector, that is, the initial latent state data;

[0078] The spatiotemporal correlation modeling layer is mainly used to extract the spatiotemporal correlation of the initial hidden state data and obtain the latent space vector after the spatiotemporal correlation feature extraction, that is, the result of the spatiotemporal correlation extraction;

[0079] The time domain change modeling layer is mainly used to extract the time domain changes of the initial hidden state data, and obtain the latent space vector after the time domain change feature extraction, that is, the result of the time domain change extraction;

[0080] The feature fusion layer is mainly used to fuse the results of spatiotemporal correlation extraction and the results of temporal change extraction to obtain the latent space vector after feature fusion, that is, the latent space feature fusion result;

[0081] The prediction result conversion and export layer is mainly used to predict and output the second traffic state data based on the latent space feature fusion results.

[0082] Through the processing of each layer structure in the above traffic prediction model 203, the prediction function of the second traffic status data based on the first traffic status data can be realized.

[0083] The data security module 204 is mainly used to ensure the security of data transmission in the entire process of traffic status prediction, prevent data leakage, and resist network attacks. In actual application, the data security module 204 can run through this architecture, thereby ensuring the security of data processing and data transmission in the data storage module 201, the data access conversion module 202 and the traffic prediction model 203.

[0084] In an exemplary embodiment, performing spatiotemporal correlation extraction and temporal change extraction based on the first traffic state data, and determining the second traffic state data in the future period based on the results of the spatiotemporal correlation extraction and the results of the temporal change extraction, may specifically include:

[0085] Performing latent space transformation on the first traffic state data to obtain initial latent state data;

[0086] Performing spatiotemporal correlation extraction and time domain change extraction on the initial hidden state data respectively, and obtaining the spatiotemporal correlation extraction results and the time domain change extraction results;

[0087] The results of spatiotemporal correlation extraction and time domain change extraction are fused to obtain the latent space feature fusion result;

[0088] Based on the latent space feature fusion result, the second traffic state data is predicted.

[0089] The implementation process of obtaining the second traffic status data in the future time period based on the first traffic status data provided in this embodiment can be completed in the traffic prediction model 203, and therefore corresponds to the data transmission process between the various layer structures in the traffic prediction model 203 in the above embodiment.

[0090] Figure 3 The main architecture of the traffic prediction model 203 in this embodiment is shown. First, a full connection layer (FC) can be used, that is, Figure 3 In FC1, the first traffic state data is transformed into a latent space by linear mapping. The entire traffic prediction model can extract spatiotemporal correlation features through a spatiotemporal synchronous graph convolutional network. In each graph convolution layer, a short-range architecture search module (i.e. Figure 3 SASM) and the long-range architecture search module (i.e. Figure 3 LASM) is used to realize the extraction of spatiotemporal correlation features within a short period of time and the extraction of temporal change features within a long period of time.

[0091] Afterwards, taking the first graph convolution layer as an example, the data output by SASM is fused with the data output by LASM by summing, and the fused data can be applied to the following two aspects: on the one hand, the fused data is linearly mapped by the fully connected layer FC2, and is summed and fused with the data calculated by the residual connection and then input into the next graph convolution layer; on the other hand, the fused data is linearly mapped by the fully connected layer FC3, and then, together with the output data of the remaining graph convolution layers, it passes through the activation function ReLU and the multilayer perceptron (MLP) in sequence to output the final prediction result, i.e., the second traffic status data.

[0092] The residual connection calculation result of the i-th layer in the traffic prediction model 203 will be used as the input data of the i+1-th layer, which is expressed as follows:

[0093]

[0094] in, represents the input data of the i+1th layer, represents the fully connected layer of the residual connection part of the i-th layer, that is, the fully connected layer FC2, and Respectively represent the output data of SASM and LASM in the i-th layer, Represents the input data of the i-th layer.

[0095] In order to alleviate the gradient disappearance, this embodiment uses parameterized jump connections for each layer in the traffic prediction model 203 to obtain the final prediction result X o , which is expressed as follows:

[0096]

[0097]

[0098] Among them, X o represents the final prediction result, i.e., the second traffic state data, L represents the number of layers of the traffic prediction model, represents the skip connection result of the i-th layer output, represents the fully connected layer of the i-th layer skip connection part, that is, the fully connected layer FC3, MLP represents the multilayer perceptron, and + represents the addition operation.

[0099] In an exemplary embodiment, performing spatiotemporal correlation extraction on the initial hidden state data may specifically include:

[0100] Inputting the initial hidden state data into the short-range architecture search module, and establishing a candidate adjacency matrix set in the short-range architecture search module;

[0101] Determine the target candidate adjacency matrix corresponding to each graph convolutional layer in the short-range architecture search module from the set of candidate adjacency matrices;

[0102] Based on the target candidate adjacency matrix, perform graph convolution operation on the initial hidden state data to obtain the first intermediate hidden state data corresponding to each graph convolution layer;

[0103] The first intermediate hidden state data are fused to obtain the result of spatiotemporal correlation extraction.

[0104] In this embodiment, the functions to be implemented by the spatiotemporal correlation modeling layer can be specifically implemented by establishing a short-range architecture search module (SASM). Figure 4: shows the main architecture of the short-range architecture search module in this embodiment. The short-range architecture search module can be mainly presented by a directed acyclic graph (DAG), which consists of L vertices. Figure 4 The architecture of the directed acyclic graph is shown in the figure using four vertices 0, 1, 2, and 3 as examples. The number of L vertices in the directed acyclic graph essentially represents L graph convolutional layers. The first intermediate hidden state data of each layer can be represented as X (i) , where i = 1, 2, ···, L. The initial hidden state data is represented as X 0 In this embodiment, the first intermediate hidden state data of each layer is calculated based on the first intermediate hidden state data output by each layer before the current layer, that is, it is calculated based on all predecessors, which can be specifically expressed as:

[0105]

[0106] Among them, X (j) is the first intermediate hidden state data of the jth layer, Indicates that the first intermediate hidden state data of the i-th layer is mixed with the first intermediate hidden state data of the previous layer, that is, the first intermediate hidden state data of each layer before the j-th layer is mixed, X (i) is the first intermediate hidden state data of the i-th layer before the j-th layer.

[0107] In order to maximize the use of complex and diverse correlations between nodes (which can be specifically represented by an adjacency matrix), this embodiment designs the hybrid operation of the short-range architecture search module as follows:

[0108]

[0109] in, represents a mixed operation on the first intermediate hidden state data X, where X represents the first intermediate hidden state data of the input; A and A′ both represent candidate adjacency matrices, is a set of predefined candidate adjacency matrices, and candidate adjacency matrices A and A′ belong to the above set; and They all represent learnable architectural parameters and can be used as weighted weights corresponding to the candidate operations AX associated with the node pair (i, j).

[0110] In this embodiment, both the initial adjacency matrix and the candidate adjacency matrix can be obtained by calculating the spatial distance or time series similarity distance between different nodes in the traffic network graph, and calculating the weights between different node pairs through a threshold Gaussian kernel function. The calculation formula is as follows:

[0111]

[0112] Among them, W ij Represents node v i With node v j The edge weight between them, dist(v i ,v j ) represents node v i With node v j The spatial distance between them or the time series similarity distance based on the DTW (Dynamic Time Warping) algorithm, σ represents the standard deviation of the spatial distance or the time series similarity distance, and κ represents the preset distance threshold.

[0113] The initial adjacency matrix or candidate adjacency matrix mentioned in this embodiment can be understood as the weights W between the above different node pairs. ij It's equal.

[0114] In an exemplary embodiment, each first intermediate latent state data is fused to obtain a result of spatiotemporal correlation extraction, which may specifically include:

[0115] Performing cropping processing on each first intermediate hidden state data respectively to obtain a first processing result;

[0116] Performing linear mapping processing and nonlinear activation processing on each first processing result respectively to obtain a second processing result;

[0117] Aggregation processing is performed on each second processing result to obtain a result of spatiotemporal correlation extraction.

[0118] In the message passing process of graph convolution, the spatiotemporal information of nodes is propagated along a given graph structure, so it is meaningful to aggregate the first intermediate hidden state data of different propagation ranges to obtain comprehensive information of message passing. Nonlinear activation, linear mapping, and layer aggregation are necessary components of graph convolution. To this end, this embodiment sets a Final Activation & Fusion Module (FAFM).

[0119] First, all the first intermediate hidden state data X in the search space (i) , where i = 1, 2, ···, L uses a pruning operation, that is, for all first intermediate hidden state data X (i) Perform trimming to simplify the results of message passing.

[0120] Afterwards, this embodiment uses multiple gated linear units (GLUs) to perform linear mapping and nonlinear activation. The above operations can be expressed as:

[0121]

[0122] in, Represents the first intermediate hidden state data X (i) The result after cropping, linear mapping and nonlinear activation is the second processing result. Crop is the cropping operation, GLU i is applied to the first intermediate hidden state data X (i) The GLU module.

[0123] The data processing operation of the GLU module in this embodiment can be expressed as:

[0124]

[0125] Among them, GLU(X) represents the gated linear function, All represent learnable parameters, ⊙ represents the Hadamard product, σ represents the sigmoid activation function, and X represents the input data, which can be any first intermediate hidden state data.

[0126] Finally, this embodiment aggregates all the second processing results through the maximum pooling operation, that is, performs aggregation processing on each second processing result, and the output result after the aggregation processing, that is, the result of the spatiotemporal correlation extraction can be expressed as:

[0127]

[0128] Among them, X out Represents the output result after aggregation processing, MaxPool represents the maximum pooling operation, represents a set vector consisting of the second processing results.

[0129] At the end of the short-range architecture search, this embodiment can select the candidate adjacency matrix with the highest weight value as the target candidate adjacency matrix according to the weight values ​​corresponding to each candidate adjacency matrix when processing the mapping relationship between node pairs in each graph convolution layer, and then replace each hybrid operation with the candidate operation corresponding to the maximum architecture parameter (i.e., weight value) to determine the final architecture of the short-range architecture search module, which can be specifically expressed as follows:

[0130]

[0131] in, represents the final determined mixing operation corresponding to the first intermediate hidden state data X, X represents the input first intermediate hidden state data; argmax represents finding the parameter with the maximum score, that is, determining the candidate adjacency matrix with the highest weight value, A represents the candidate adjacency matrix, is a set of predefined candidate adjacency matrices, and the candidate adjacency matrix A belongs to the above set; It represents a learnable architectural parameter, which can be used as the weighted weight corresponding to the candidate operation AX related to the node pair (i, j), that is, the weight value of the candidate adjacency matrix A.

[0132] In this embodiment, the number of graph nodes in the spatiotemporal synchronous graph convolutional network is kN, where N represents the number of original nodes in the traffic network graph, and k represents the candidate adjacency matrix A∈R kN×kN The number of adjacent time steps in , in this embodiment, k is 4.

[0133] In order to enhance interpretability, this embodiment designs 6 candidate adjacency matrices from multiple perspectives, and finally obtains the candidate adjacency matrix set Among them, O is the zero matrix, I is the identity matrix, A1, A2, A3, A4 are kN×kN matrices determined based on the initial adjacency matrix in the traffic network graph, such as Figure 4 As shown, OX, IX, A1X, A2X, A3X and A4X all represent candidate operations implemented based on the candidate adjacency matrix.

[0134] In this embodiment, taking k as 4 as an example, by combining the spatiotemporal correlation of nodes in the traffic network graph at time t1, time t2, time t3 and time t4, candidate adjacency matrices A1, A2, A3 and A4 can be obtained.

[0135] Figure 5 Only the temporal and spatial correlations between nodes in the traffic network graph at time t1 and time t2 are shown in Figure 2. TC represents the identity matrix to represent the self-connection relationship, TG represents the time graph and obtains the time series similarity between nodes through the DTW algorithm, and SG represents the space graph and constructs the similarity according to the spatial distance between nodes in the traffic network graph. In this embodiment, A1 and A3 are designed for the mutual similarity of adjacent time steps between nodes, and A2 and A4 are designed for the self-similarity of the approximate time steps of nodes.

[0136] Specifically, see Appendix Figure 5 , the candidate adjacency matrix A1 is based on the time graph matrix at time t1 Time diagram matrix from time t1 to time t2 Time diagram matrix from time t2 to time t1 Time diagram matrix at time t2 Time diagram matrix from time t2 to time t3 Time diagram matrix from time t3 to time t2 Time diagram matrix at time t3 Time diagram matrix from time t3 to time t4 Time diagram matrix from time t4 to time t3 And the time diagram matrix at time t4 Constructed

[0137] The candidate adjacency matrix A2 is the time graph matrix based on time t1 The identity matrix from time t1 to time t2 The identity matrix from time t2 to time t1 Time diagram matrix at time t2 The identity matrix from time t2 to time t3 The identity matrix from time t3 to time t2 Time diagram matrix at time t3 The identity matrix from t3 to t4 The identity matrix from time t4 to time t3 And the time diagram matrix at time t4 Constructed

[0138] The candidate adjacency matrix A3 is based on the spatial graph matrix at time t1 Spatial graph matrix from time t1 to time t2 The spatial graph matrix from time t2 to time t1 Spatial graph matrix at time t2 Spatial graph matrix from time t2 to time t3 Spatial graph matrix from time t3 to time t2 Spatial graph matrix at time t3 Spatial graph matrix from time t3 to time t4 Spatial graph matrix from time t4 to time t3 And the spatial graph matrix at time t4 Constructed

[0139] The candidate adjacency matrix A4 is based on the spatial graph matrix at time t1 The identity matrix from time t1 to time t2 The identity matrix from time t2 to time t1 Spatial graph matrix at time t2 The identity matrix from time t2 to time t3 The identity matrix from time t3 to time t2 Spatial graph matrix at time t3 The identity matrix from t3 to t4 The identity matrix from time t4 to time t3 And the spatial graph matrix at time t4 Constructed.

[0140] In an exemplary embodiment, extracting the time domain change of the initial hidden state data may specifically include:

[0141] Inputting the initial hidden state data into the long-range architecture search module and establishing a candidate operation set in the long-range architecture search module;

[0142] Determine a target candidate operation corresponding to each vertex in the long-range architecture search module from the candidate operation set;

[0143] Performing a temporal convolution operation on the initial hidden state data based on the target candidate operation to obtain the second intermediate hidden state data of each vertex;

[0144] The second intermediate hidden state data are fused to obtain the result of time domain change extraction.

[0145] The functions to be implemented by the time domain variation modeling layer can be specifically implemented by establishing a long-range architecture search module (LASM). In this embodiment, the long-range architecture search module is mainly implemented based on a time convolution neural network (TCN). The time convolution neural network is essentially a one-dimensional neural network, and its advantage is high feature extraction efficiency. In contrast, the gated time convolution neural network (GatedTCN) achieves stronger extraction capabilities without obvious efficiency loss, and has been widely used in sequence modeling. See Appendix Figure 6 In this embodiment, the gated temporal convolutional neural network uses two parallel one-dimensional neural networks (ie, TCN), which respectively use the tangent hyperbolic activation function Tanh and the sigmoid activation function σ.

[0146] In addition, this embodiment also introduces one-dimensional dilated convolution to enhance the receptive field of TCN. However, due to the oversimplification of convolution kernels and convolution types, traditional TCN has difficulty in processing long sequences and capturing long-range time patterns in spatiotemporal modeling. To this end, this embodiment attempts to improve traditional TCN from the following two aspects: on the one hand, various convolution types such as standard convolution, separation convolution, and dilated convolution need to be considered; on the other hand, different convolution kernel sizes should be used in combination to capture time dependencies of different ranges (especially long ranges).

[0147] Based on the above analysis, this embodiment designs a long-range architecture search module. Figure 7 The main structure of the long-range architecture search module in this embodiment is shown. The search space of the long-range architecture search module is a space composed of N v The input data in the search space, that is, the initial hidden state data, is represented as X 0 , the second intermediate hidden state data of the i-th vertex is represented as Y (i) ,i=1,2,…N v, the output of the long-range architecture search module can be obtained by summing (i.e. Figure 7 ADD operation in the middle) can be expressed as:

[0148]

[0149] Among them, Y out represents the output result of the long-range architecture search module, that is, the result of the time domain change extraction; o and o′ both represent candidate operations, and different candidate operations form a set and and Both represent learnable architecture parameters, Specifically, it can be understood as the weighted weight corresponding to the candidate operation O related to the vertex pair (i, j), Y (i) Represents the second intermediate hidden state data of the i-th vertex.

[0150] In this embodiment, the candidate operation set may specifically include at least one of a zero mapping operation, an identity mapping operation, and a gated temporal convolution operation with different convolution kernels and convolution types. Figure 7 Figure 2 shows gated temporal convolution operations including a zero mapping operation, an identity mapping operation, and two gated temporal convolution operations, Gated TCN1 and Gated TCN2, such as standard convolution with a 1×3 convolution kernel, dilated convolution with a 1×5 convolution kernel, or separated convolution with a 1×7 convolution kernel.

[0151] At the end of the long-range architecture search, this embodiment can also select the candidate operation with the highest weight value as the target candidate operation according to the weight value corresponding to each candidate operation when processing the mapping relationship between each vertex pair, and then replace each hybrid operation with the target candidate operation corresponding to the maximum architecture parameter to obtain the final architecture of the long-range architecture search module, which can be specifically expressed as follows:

[0152]

[0153] Among them, (i,j) represents the target candidate operation for the vertex pair (i, j), argmax represents finding the parameter with the maximum score, that is, determining the candidate operation with the highest weight value. represents the weight corresponding to the candidate operation o associated with the vertex pair (i, j).

[0154] In the traffic prediction model provided in this embodiment, the architecture parameters include weight values ​​of different candidate adjacency matrices, such as It also includes the weights of candidate operations, such as those in the long-range architecture search module.

[0155] The network parameters represent all learnable parameters in the model except the architecture parameters, such as the parameters in the fully connected layer, the multilayer perceptron, and the gated temporal convolution in the candidate operation. When training the traffic prediction model, the present embodiment first uses the training set and the validation set to alternately update the network parameters and the architecture parameters until the stopping criteria are met. Then, the obtained model architecture is further trained with the training set until the stopping criteria are met, thereby obtaining a trained traffic prediction model.

[0156] Finally, the trained traffic prediction model is tested through the test set. In this embodiment, mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are selected as evaluation indicators. The calculation methods of the above three evaluation indicators are as follows:

[0157]

[0158] Among them, represents the true value y and the predicted value The mean absolute error between |Ω| represents the total number of samples, y i represents the i-th true value, represents the i-th predicted value.

[0159]

[0160] in, Represents the true value y and the predicted value The root mean square error between |Ω| represents the total number of samples, y i represents the i-th true value, represents the i-th predicted value.

[0161]

[0162] in, Represents the true value x and the predicted value The mean absolute percentage error between |Ω| represents the total number of samples, y i represents the i-th true value, represents the i-th predicted value.

[0163] See attached Figure 8 When acquiring sample data, this embodiment uses a sliding time window method to generate sample data, and divides the data into a training set, a validation set, and a test set in a certain proportion.

[0164] In the actual application process, the traffic prediction model in this embodiment is trained by the Adam optimizer, the batch size is 32, the learning rate is 0.001, and the loss function adopts the above MAE. This embodiment adopts the early stopping method to train the model to improve the generalization ability of the model and reduce the risk of overfitting.

[0165] This embodiment tests the traffic prediction model based on four public traffic data sets PEMS03, PEMS04, PEMS07 and PEMS08, and compares it with the existing traffic prediction models. The existing traffic prediction models are established based on FC-LSTM, DCRNN, STGCN, ASTGCN(r), Gradh WaveNet, STSGCN, STFGNN, and AutoSTG, respectively. The traffic prediction model in this embodiment can be defined as being established based on AutoSTS, which refers to an improved traffic prediction model established based on the above-mentioned short-range architecture search module and long-range architecture search module. The evaluation indicators selected in this embodiment include mean absolute error MAE, root mean square error RMSE, and mean absolute percentage error MAPE. The test results of the traffic prediction model provided in this embodiment and the existing eight traffic prediction models are shown in Table 1:

[0166] Table 1 Comparison data of model test results

[0167]

[0168]

[0169] Based on Table 1 above, it can be seen that the traffic prediction model provided in this embodiment, compared with the other eight existing traffic prediction models, the results of testing on four public traffic data sets show that the three evaluation indicators of mean absolute error MAE, root mean square error RMSE and mean absolute percentage error MAPE are significantly lower than those of the other eight existing traffic prediction models. It can be seen that the prediction performance of the traffic prediction model provided in this embodiment is greatly improved, and the prediction accuracy is higher.

[0170] Since the traffic prediction model provided in this embodiment is an end-to-end deep learning model, after the traffic prediction model is trained, the application process of the traffic prediction model only needs to input the adjacency matrix and historical time series (i.e., the first traffic status data) of different key objects in the traffic network (such as different road sections or sensors) into the traffic prediction model. The output of the traffic prediction model is the predicted value of the traffic conditions for a period of time in the future (i.e., the future time period) (i.e., the second traffic status data).

[0171] In actual application, the length of the input historical time series and the length of the output prediction time series can be set in advance. For example, the time granularity can be set to five minutes, the input historical series length can be set to 1 hour, and the length of the prediction time training can also be 1 hour, that is, predicting the traffic status in the next 1 hour. In this case, the number of time steps of the input and output time series is 60 / 5=12 steps. This embodiment is only an exemplary explanation of the setting scheme of the preset parameters. The specific values ​​can be reasonably set according to the actual application needs, and no further details are given here.

[0172] The traffic state prediction device provided by the present invention is described below. The traffic state prediction device described below and the traffic state prediction method described above can be referred to each other.

[0173] Fig. 9 The traffic state prediction device provided by an embodiment of the present invention is shown, and the device includes:

[0174] A receiving module 901 is used to receive first traffic status data within a historical period;

[0175] The processing module 902 is used to perform spatiotemporal correlation extraction and temporal change extraction based on the first traffic status data, and determine the second traffic status data in the future period based on the results of the spatiotemporal correlation extraction and the results of the temporal change extraction.

[0176] In an exemplary embodiment, the processing module 902 may be specifically used for:

[0177] Performing latent space transformation on the first traffic state data to obtain initial latent state data;

[0178] Performing spatiotemporal correlation extraction and time domain change extraction on the initial hidden state data respectively, and obtaining the spatiotemporal correlation extraction results and the time domain change extraction results;

[0179] The results of spatiotemporal correlation extraction and time domain change extraction are fused to obtain the latent space feature fusion result;

[0180] Based on the latent space feature fusion result, the second traffic state data is predicted.

[0181] Furthermore, the processing module 902 can specifically extract the spatiotemporal correlation of the initial hidden state data in the following manner:

[0182] Inputting the initial hidden state data into the short-range architecture search module, and establishing a candidate adjacency matrix set in the short-range architecture search module;

[0183] Determine the target candidate adjacency matrix corresponding to each graph convolutional layer in the short-range architecture search module from the set of candidate adjacency matrices;

[0184] Based on the target candidate adjacency matrix, perform graph convolution operation on the initial hidden state data to obtain the first intermediate hidden state data corresponding to each graph convolution layer;

[0185] The first intermediate hidden state data are fused to obtain the result of spatiotemporal correlation extraction.

[0186] Furthermore, the processing module 902 may specifically implement the fusion of each first intermediate latent state data in the following manner to obtain the result of spatiotemporal correlation extraction:

[0187] Performing cropping processing on each first intermediate hidden state data respectively to obtain a first processing result;

[0188] Performing linear mapping processing and nonlinear activation processing on each first processing result respectively to obtain a second processing result;

[0189] Aggregate the second processing results to obtain a result of spatiotemporal correlation extraction.

[0190] Furthermore, the processing module 902 can specifically implement time domain change extraction of the initial hidden state data in the following manner:

[0191] Inputting the initial hidden state data into the long-range architecture search module and establishing a candidate operation set in the long-range architecture search module;

[0192] Determine a target candidate operation corresponding to each vertex in the long-range architecture search module from the candidate operation set;

[0193] Performing a temporal convolution operation on the initial hidden state data based on the target candidate operation to obtain the second intermediate hidden state data of each vertex;

[0194] The second intermediate hidden state data are fused to obtain the result of time domain change extraction.

[0195] In an exemplary embodiment, the candidate operation set may specifically include at least one of a zero mapping operation, an identity mapping operation, and a gated temporal convolution operation having different convolution kernels and convolution types.

[0196] In an exemplary embodiment, the first traffic status data may be obtained through the following process:

[0197] Obtain the original traffic data of the traffic network in the historical period;

[0198] Each key object in the traffic network is regarded as a node, and the connection relationship between each key object is regarded as an edge. An initial adjacency matrix is ​​established based on the connection relationship between each key object. Based on the nodes, edges and the initial adjacency matrix, a traffic network graph is constructed.

[0199] Based on the traffic network diagram and original traffic data, determine the traffic status of each node at different times in the historical period;

[0200] Based on the traffic status of each node at different times, first traffic status data is obtained.

[0201] In summary, the traffic status prediction device provided by the embodiment of the present invention receives the first traffic status data through the receiving module, and processes the first traffic status data in the historical time period through the processing module to obtain the second traffic status data in the future time period. Since the spatiotemporal correlation extraction and the time domain change extraction can be performed based on the first traffic status data respectively, the nonlinear characteristics and spatiotemporal correlation characteristics hidden in the first traffic status data can be obtained, thereby making the output second traffic status data more reliable and improving the accuracy of the traffic status prediction results.

[0202] Fig.10 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.10 As shown, the electronic device may include: a processor 1010, a communication interface 1020, a memory 1030 and a communication bus 1040, wherein the processor 1010, the communication interface 1020 and the memory 1030 communicate with each other through the communication bus 1040. The processor 1010 may call the logic instructions in the memory 1030 to execute the traffic state prediction method, which includes: receiving first traffic state data within a historical period; performing spatiotemporal correlation extraction and temporal change extraction based on the first traffic state data, and determining second traffic state data within a future period based on the results of the spatiotemporal correlation extraction and the results of the temporal change extraction.

[0203] In addition, the logic instructions in the above-mentioned memory 1030 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk.

[0204] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the traffic status prediction method provided by the above-mentioned embodiments, the method including: receiving first traffic status data within a historical period; performing spatiotemporal correlation extraction and time domain change extraction based on the first traffic status data, and determining second traffic status data within a future period based on the results of the spatiotemporal correlation extraction and the results of the time domain change extraction.

[0205] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the traffic status prediction method provided in the above-mentioned embodiments is implemented. The method includes: receiving first traffic status data within a historical period; performing spatiotemporal correlation extraction and time domain change extraction based on the first traffic status data, and determining second traffic status data within a future period based on the results of the spatiotemporal correlation extraction and the results of the time domain change extraction.

[0206] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0207] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A traffic state prediction method, characterized in that: include: Receiving first traffic status data within a historical period; Based on the first traffic status data, respectively, a spatiotemporal correlation extraction and a temporal change extraction are performed, and based on the result of the spatiotemporal correlation extraction and the result of the temporal change extraction, second traffic status data in a future period is determined; The step of performing spatiotemporal correlation extraction and temporal change extraction based on the first traffic status data, and determining second traffic status data in a future period based on the result of the spatiotemporal correlation extraction and the result of the temporal change extraction, comprises: Performing latent space transformation on the first traffic state data to obtain initial latent state data; Performing spatiotemporal correlation extraction and time domain change extraction on the initial hidden state data respectively to obtain the spatiotemporal correlation extraction result and the time domain change extraction result; The result of the spatiotemporal correlation extraction and the result of the time domain change extraction are fused to obtain a latent space feature fusion result; Based on the latent space feature fusion result, predict and obtain the second traffic state data; Extracting spatiotemporal correlation from the initial hidden state data includes: Inputting the initial hidden state data into a short-range architecture search module, and establishing a candidate adjacency matrix set in the short-range architecture search module; Determining a target candidate adjacency matrix corresponding to each graph convolutional layer in the short-range architecture search module from the set of candidate adjacency matrices; Based on the target candidate adjacency matrix, performing a graph convolution operation on the initial hidden state data to obtain first intermediate hidden state data corresponding to each of the graph convolution layers; Fusing each of the first intermediate hidden state data to obtain the result of the spatiotemporal correlation extraction; The fusing of the first intermediate hidden state data to obtain the result of the spatiotemporal correlation extraction includes: Performing cropping processing on each of the first intermediate hidden state data respectively to obtain a first processing result; Performing linear mapping processing and nonlinear activation processing on each of the first processing results respectively to obtain a second processing result; Aggregate the second processing results to obtain the spatiotemporal correlation extraction result; Extracting the time domain changes of the initial hidden state data includes: Inputting the initial hidden state data into a long-range architecture search module, and establishing a candidate operation set in the long-range architecture search module; Determine a target candidate operation corresponding to each vertex in the long-range architecture search module from the candidate operation set; Performing a temporal convolution operation on the initial hidden state data based on the target candidate operation to obtain second intermediate hidden state data of each vertex; Each of the second intermediate latent state data is fused to obtain the result of the time domain change extraction.

2. The traffic state prediction method according to claim 1, characterized in that: The candidate operation set includes at least one of a zero mapping operation, an identity mapping operation, and a gated temporal convolution operation with different convolution kernels and convolution types.

3. The traffic state prediction method according to claim 1, characterized in that: The first traffic status data is obtained specifically through the following process: Obtain the original traffic data of the traffic network in the historical period; Taking each key object in the traffic network as a node, taking the connection relationship between each key object as an edge, and establishing an initial adjacency matrix based on the connection relationship between each key object, and constructing a traffic network graph based on the nodes, the edges and the initial adjacency matrix; Determine the traffic status of each node at different times in the historical period based on the traffic network diagram and the original traffic data; The first traffic status data is obtained based on the traffic status of each node at different times.

4. A traffic status prediction device, characterized in that: include: A receiving module, used for receiving first traffic status data within a historical period; a processing module, configured to perform spatiotemporal correlation extraction and temporal change extraction based on the first traffic status data, and determine second traffic status data in a future time period based on the results of the spatiotemporal correlation extraction and the results of the temporal change extraction; The step of performing spatiotemporal correlation extraction and temporal change extraction based on the first traffic status data, and determining second traffic status data in a future period based on the result of the spatiotemporal correlation extraction and the result of the temporal change extraction, comprises: Performing latent space transformation on the first traffic state data to obtain initial latent state data; Performing spatiotemporal correlation extraction and time domain change extraction on the initial hidden state data respectively to obtain the spatiotemporal correlation extraction result and the time domain change extraction result; The result of the spatiotemporal correlation extraction and the result of the time domain change extraction are fused to obtain a latent space feature fusion result; Based on the latent space feature fusion result, predict and obtain the second traffic state data; Extracting spatiotemporal correlation from the initial hidden state data includes: Inputting the initial hidden state data into a short-range architecture search module, and establishing a candidate adjacency matrix set in the short-range architecture search module; Determining a target candidate adjacency matrix corresponding to each graph convolutional layer in the short-range architecture search module from the set of candidate adjacency matrices; Based on the target candidate adjacency matrix, performing a graph convolution operation on the initial hidden state data to obtain first intermediate hidden state data corresponding to each of the graph convolution layers; Fusing each of the first intermediate hidden state data to obtain the result of the spatiotemporal correlation extraction; The fusing of the first intermediate hidden state data to obtain the result of the spatiotemporal correlation extraction includes: Performing cropping processing on each of the first intermediate hidden state data respectively to obtain a first processing result; Performing linear mapping processing and nonlinear activation processing on each of the first processing results respectively to obtain a second processing result; Aggregate the second processing results to obtain the spatiotemporal correlation extraction result; Extracting the time domain changes of the initial hidden state data includes: Inputting the initial hidden state data into a long-range architecture search module, and establishing a candidate operation set in the long-range architecture search module; Determine a target candidate operation corresponding to each vertex in the long-range architecture search module from the candidate operation set; Performing a temporal convolution operation on the initial hidden state data based on the target candidate operation to obtain second intermediate hidden state data of each vertex; Each of the second intermediate latent state data is fused to obtain the result of the time domain change extraction.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the traffic status prediction method as described in any one of claims 1 to 3 is implemented.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the traffic status prediction method according to any one of claims 1 to 3 is implemented.