Traffic Flow Prediction Method and Device Based on Spatiotemporal Synchronous Graph Aggregation

Abstract

The present invention provides a traffic flow prediction method and device based on spatio-temporal synchronization graph aggregation. The method includes: obtaining the first historical traffic flow of each station in the area to be predicted; inputting the first historical traffic flow of each station into a traffic flow prediction model to obtain the predicted traffic flow of each station output by the traffic flow prediction model; the traffic flow prediction model is trained based on the second historical traffic flow of each station; the traffic flow prediction model is used to extract the spatio-temporal features corresponding to each station based on the first historical traffic flow of each station, and perform traffic flow prediction based on the spatio-temporal features corresponding to each station; the spatio-temporal features are synchronously extracted by a spatio-temporal synchronization graph aggregation module in the traffic flow prediction model. The method and device provided by the present invention optimize the modeling process, and at the same time greatly improve the prediction performance of traffic flow, realizing end-to-end traffic flow prediction.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to a traffic flow prediction method and device based on spatiotemporal synchronous graph aggregation. Background Art

[0002] Traffic forecasting is a crucial component of intelligent transportation systems. Accurate, real-time traffic forecasts facilitate real-time traffic management and guidance, as well as the alleviation of traffic congestion. In recent years, the growing demand for travel and the accumulation of traffic big data have posed new challenges to traffic forecasting. Deep learning methods, with their ability to model complex nonlinear patterns, have been widely used in this area.

[0003] Existing traffic flow prediction methods usually adopt a modular and step-by-step modeling approach that combines multiple deep learning models. However, this modeling method has a complex model structure and poor prediction performance. Summary of the Invention

[0004] The present invention provides a traffic flow prediction method and device based on spatiotemporal synchronous graph aggregation, which are used to solve the defect of poor traffic flow prediction performance in the prior art.

[0005] The present invention provides a traffic flow prediction method based on spatiotemporal synchronous graph aggregation, comprising:

[0006] Obtain the first historical traffic flow of each station in the area to be predicted;

[0007] Inputting the first historical traffic flow of each station into a traffic flow prediction model to obtain the predicted traffic flow of each station output by the traffic flow prediction model;

[0008] The traffic flow prediction model is obtained based on the second historical traffic flow training of each site; the traffic flow prediction model is used to extract the spatiotemporal characteristics corresponding to each site based on the first historical traffic flow of each site, and perform traffic flow prediction based on the spatiotemporal characteristics corresponding to each site; the spatiotemporal characteristics are synchronously extracted by the spatiotemporal synchronization graph aggregation module in the traffic flow prediction model.

[0009] According to a traffic flow prediction method based on spatiotemporal synchronization graph aggregation provided by the present invention, the spatiotemporal features corresponding to each station are extracted based on the following steps:

[0010] Based on the encoding module in the spatiotemporal synchronization graph aggregation module, the first historical traffic flow of each station is encoded to obtain the traffic characteristics corresponding to each station;

[0011] Based on the feature extraction module in the spatio-temporal synchronization graph aggregation module, the traffic features corresponding to each site, the spatial adjacency matrix, and the temporal adjacency matrix are applied for feature extraction to obtain the spatio-temporal features corresponding to each site;

[0012] The spatial adjacency matrix represents the connectivity relationship between the sites, and the temporal adjacency matrix represents the connectivity relationship between the corresponding time nodes of each site at each moment.

[0013] According to a traffic flow prediction method based on spatio-temporal synchronization graph aggregation provided by the present invention, the connectivity relationship between the sites is determined based on the following steps:

[0014] Based on the distances between the sites, an initial connectivity relationship between the sites is determined;

[0015] Based on the third historical traffic flow of each site, the initial connectivity relationship between the sites is adjusted to obtain the connectivity relationship between the sites.

[0016] According to a traffic flow prediction method based on spatio-temporal synchronization graph aggregation provided by the present invention, adjusting the initial connectivity relationship between the sites based on the third historical traffic flow of each site to obtain the connectivity relationship between the sites includes:

[0017] Based on the third historical traffic flow of each site, the correlation coefficient between the sites is determined;

[0018] Based on the difference between the correlation coefficient between the sites and the corresponding initial connectivity relationship, the initial connectivity relationship between the sites is adjusted to obtain the connectivity relationship between the sites.

[0019] According to a traffic flow prediction method based on spatio-temporal synchronization graph aggregation provided by the present invention, the connectivity relationship between the time nodes is determined based on the historical nodes and adjacent nodes corresponding to each time node, the adjacent nodes are the time nodes adjacent to each time node, and the historical nodes are the time nodes before each time node.

[0020] According to a traffic flow prediction method based on spatio-temporal synchronization graph aggregation provided by the present invention, the feature extraction module includes alternately arranged feature extraction units and batch normalization layers;

[0021] The feature extraction unit includes a plurality of parallel spatio-temporal synchronization graph aggregation channels and a feature aggregation layer connected to the plurality of spatio-temporal synchronization graph aggregation channels.

[0022] The present invention also provides a traffic flow prediction device based on spatio-temporal synchronization graph aggregation, including:

[0023] An acquisition unit for acquiring the first historical traffic flow of each site in the area to be predicted;

[0024] A prediction unit for inputting the first historical traffic flow of each site into a traffic flow prediction model to obtain the predicted traffic flow of each site output by the traffic flow prediction model;

[0025] The traffic flow prediction model is trained based on the second historical traffic flow of each site; the traffic flow prediction model is used to extract the spatio-temporal features corresponding to each site based on the first historical traffic flow of each site, and perform traffic flow prediction based on the spatio-temporal features corresponding to each site; the spatio-temporal features are synchronously extracted by a spatio-temporal synchronization graph aggregation module in the traffic flow prediction model.

[0026] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the traffic flow prediction method based on spatio-temporal synchronization graph aggregation as described in any one of the above.

[0027] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the traffic flow prediction method based on spatio-temporal synchronization graph aggregation as described in any one of the above.

[0028] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the traffic flow prediction method based on spatio-temporal synchronization graph aggregation as described in any one of the above.

[0029] The traffic flow prediction method and device based on spatio-temporal synchronization graph aggregation provided by the present invention realize the synchronous extraction of spatio-temporal features between traffic flow data of each site through a spatio-temporal synchronization graph aggregation module in the traffic flow prediction model, retain the mutual coupling between time correlation and space correlation, and only need to use graph convolution, optimizing the modeling process. On this basis, traffic flow prediction is performed based on the extracted spatio-temporal features, so as to greatly improve the prediction performance of traffic flow and realize end-to-end traffic flow prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0031] Figure 1It is one of the flow schematic diagrams of the traffic flow prediction method based on spatio-temporal synchronous graph aggregation provided by the present invention;

[0032] Figure 2 It is the schematic diagram of the construction of the time graph structure provided by the present invention;

[0033] Figure 3 It is the schematic diagram of the structure of the traffic flow prediction model provided by the present invention;

[0034] Figure 4 It is the schematic diagram of the internal structure of the feature extraction unit provided by the present invention;

[0035] Figure 5 It is the schematic diagram of the internal structure of the spatio-temporal synchronous graph aggregation unit provided by the present invention;

[0036] Figure 6 It is the second flow schematic diagram of the traffic flow prediction method based on spatio-temporal synchronous graph aggregation provided by the present invention;

[0037] Figure 7 It is the schematic diagram of the structure of the traffic flow prediction device based on spatio-temporal synchronous graph aggregation provided by the present invention;

[0038] Figure 8 It is the schematic diagram of the structure of the electronic device provided by the present invention. Detailed implementation manners

[0039] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] As one of the most important methods, graph neural networks bring new opportunities and challenges to the accurate prediction of traffic flow and become the focus of attention. In the existing traffic flow prediction methods based on graph neural networks, in order to model the spatio-temporal correlation in traffic data, a modular and step-by-step modeling method combining multiple deep learning models is usually adopted. However, this modeling method has a complex model structure, and extracts the spatio-temporal correlation step by step, which will destroy the mutually coupled mode of spatio-temporal correlation and ignore some important information, resulting in poor prediction performance.

[0041] In view of this, the present invention provides a traffic flow prediction method based on spatio-temporal synchronous graph aggregation. Figure 1 It is one of the flow schematic diagrams of the traffic flow prediction method based on spatio-temporal synchronous graph aggregation provided by the present invention. As Figure 1 shown, the method includes:

[0042] Step 110: Obtain the first historical traffic flow of each station in the area to be predicted.

[0043] Here, the area to be predicted can be the entire city or each sub - area. Each station in the area to be predicted can be each road section or location point, or the location where the sensor is set. The embodiments of the present invention do not make specific limitations on this. After determining each station in the area to be predicted, the traffic flow data of each station in the first historical period can be obtained, and then pre - processing such as outlier processing and standardization is performed on the traffic flow data to obtain the first historical traffic flow of each station.

[0044] The first historical traffic flow can be obtained in seconds, or in minutes or hours, and can be specifically determined according to actual needs.

[0045] Step 120: Input the first historical traffic flow of each station into the traffic flow prediction model to obtain the predicted traffic flow of each station output by the traffic flow prediction model;

[0046] The traffic flow prediction model is trained based on the second historical traffic flow of each station; the traffic flow prediction model is used to extract the spatio - temporal features corresponding to each station based on the first historical traffic flow of each station, and perform traffic flow prediction based on the spatio - temporal features corresponding to each station; the spatio - temporal features are synchronously extracted by the spatio - temporal synchronization graph aggregation module in the traffic flow prediction model.

[0047] Specifically, after obtaining the first historical traffic flow of each station, the first historical traffic flow of each station can be input into the traffic flow prediction model, and the traffic flow prediction model performs traffic flow prediction according to the first historical traffic flow of each station to obtain the predicted traffic flow of each station at a future time, so as to realize the traffic flow prediction of all stations from end to end.

[0048] Considering that the model structure of the existing technology's modeling method is complex, and the spatio-temporal correlation is extracted step by step, which destroys the mutually coupled mode of spatio-temporal correlation and results in poor prediction performance. To address this problem, an embodiment of the present invention constructs a Spatial-temporal Graph Synchronous Aggregation (STGSA) module in the traffic flow prediction model. For the first historical traffic flow of each site as input, the spatio-temporal synchronous graph aggregation module realizes the synchronous extraction of spatio-temporal features through graph convolution, thereby retaining the mutual coupling between temporal correlation and spatial correlation, which is beneficial to subsequent traffic flow prediction. Moreover, only graph convolution is required, without other deep learning modules such as RNN (Recurrent Neural Network), optimizing the model construction.

[0049] On this basis, the traffic flow prediction model can perform traffic flow prediction according to the extracted spatio-temporal features corresponding to each site, thereby obtaining the predicted traffic flow of each site at future times. Here, the future time can be one or more future moments. It can be understood that when predicting the predicted traffic flow of multiple future moments, multiple output layers corresponding to different future moments can be set in the traffic flow prediction model.

[0050] It should be noted that when performing step 120, it is possible to input only the first historical traffic flow of each site, or other data such as vehicle speed that is beneficial to traffic flow prediction. The embodiment of the present invention does not make specific limitations on this.

[0051] In addition, before this, it is also necessary to pre-train a traffic flow prediction model. The specific training method can be: First, collect the traffic flow of each site during the second historical period as the second historical traffic flow; then, divide the second historical traffic flow into training samples and corresponding sample labels. Specifically, the traffic flow with an earlier time can be used as the training sample, and the traffic flow with a later time can be used as the corresponding sample label; finally, use the training samples and corresponding sample labels to train the initial model, thereby obtaining the traffic flow prediction model.

[0052] It can be understood that since the traffic flow during the second historical period needs to be divided into training samples and corresponding sample labels according to time, the second historical period should be a longer historical period than the first historical period.

[0053] The method provided by the embodiments of the present invention realizes the synchronous extraction of spatio-temporal features between the traffic flow data of each site through the spatio-temporal synchronization graph aggregation module in the traffic flow prediction model, retains the mutual coupling between temporal correlation and spatial correlation, and only needs to use graph convolution to optimize the modeling process. On this basis, traffic flow prediction is performed based on the extracted spatio-temporal features, so as to greatly improve the prediction performance of traffic flow and achieve end-to-end traffic flow prediction.

[0054] Based on the above embodiments, the spatio-temporal features corresponding to each site are extracted based on the following steps:

[0055] Based on the encoding module in the spatio-temporal synchronization graph aggregation module, the first historical traffic flow of each site is encoded to obtain the traffic flow features corresponding to each site;

[0056] Based on the feature extraction module in the spatio-temporal synchronization graph aggregation module, the traffic flow features corresponding to each site, the spatial adjacency matrix, and the temporal adjacency matrix are used for feature extraction to obtain the spatio-temporal features corresponding to each site;

[0057] The spatial adjacency matrix represents the connectivity relationship between each site, and the temporal adjacency matrix represents the connectivity relationship between each time node corresponding to each site at each moment.

[0058] Specifically, the spatio-temporal synchronization graph aggregation module may include an encoding module and a feature extraction module. After the first historical traffic flow of each site is input into the spatio-temporal synchronization graph aggregation module, the encoding module can first encode the first historical traffic flow of each site to obtain the traffic flow features corresponding to each site and output them to the feature extraction module.

[0059] Subsequently, the feature extraction module incorporates the spatial graph structure and the temporal graph structure, and realizes the synchronous extraction of spatio-temporal features through graph convolution. The specific process may be to use the traffic flow features corresponding to each site, the adjacency matrix of the spatial graph structure, that is, the spatial adjacency matrix, and the adjacency matrix of the temporal graph structure, that is, the temporal adjacency matrix, for feature extraction to obtain the spatio-temporal features corresponding to each site.

[0060] Here, the spatial adjacency matrix represents the connectivity relationship between each site in space, which can be determined only according to the distance between each site, or can be determined in combination with the traffic flow data of each site. The temporal adjacency matrix represents the connectivity relationship between each time node corresponding to each site at each moment, which can be determined only according to the interdependent relationship between each time node, or can be determined in combination with the strength of the interdependent relationship. The embodiments of the present invention do not make specific limitations on this.

[0061] The method provided by the embodiments of the present invention explicitly constructs a spatial and temporal graph structure, integrates the extraction of spatio-temporal features of traffic flow data into a spatial graph convolution process, and further improves the prediction accuracy without using other deep learning modules.

[0062] Based on any of the above embodiments, the connectivity relationship between each site is determined based on the following steps:

[0063] Based on the distances between each site, determine the initial connectivity relationship between each site;

[0064] Based on the third historical traffic flow of each site, adjust the initial connectivity relationship between each site to obtain the connectivity relationship between each site.

[0065] Specifically, first, according to the distances between each site in space, determine the initial connectivity relationship between each site in space, and construct an initial spatial graph structure to reflect the connectivity of the traffic road network. Immediately afterwards, considering that in the prior art when using a graph network for traffic road network modeling, an adjacency matrix is usually constructed based on the distances between road sensors. This method lacks sufficient prior knowledge guidance, cannot accurately reflect the interaction between actual nodes, and there is a risk of overfitting. For example, two sites are located on a two-way lane but in opposite directions, and the traffic flow sharing mode will be less. At this time, the distances between these two sites are close but the traffic flows are not related. Therefore, the connectivity relationship obtained solely based on the distances between these two sites is not accurate.

[0066] In response to this, the embodiments of the present invention introduce the third historical traffic flow of each site as prior knowledge, analyze the similarity of the traffic flow change trends of each site according to the third historical traffic flow of each site, and optimize and adjust the initial connectivity relationship between each site, so as to obtain the connectivity relationship between each site. It can be understood that the more similar the traffic flow change trends of any two sites are, the more similar the shared traffic flow patterns will be, and the stronger the correlation between these two sites will be. Based on this, correcting the initial connectivity relationship between each site can improve the accuracy of the connectivity relationship between each site in space.

[0067] Here, the third historical traffic flow can be the historical traffic flow after preprocessing such as outlier processing and standardization, or the originally collected historical traffic flow. The embodiments of the present invention do not make specific limitations on this. The first historical traffic flow and the third historical traffic flow can specifically come from the same historical traffic flow data or different historical traffic flow data. The embodiments of the present invention also do not make specific limitations on this.

[0068] Furthermore, the distances between each site can be calculated according to the longitude and latitude data of each site. The specific formula is as follows:

[0069]

[0070] in, Indicates site v i With site v j The distance between them, r is the radius of the earth, γ i、 γ j Site v i and v j latitude and longitude.

[0071] Based on the distance between each site, the initial spatial graph structure is constructed and the initial spatial adjacency matrix A is used s Indicates that A s Each element in can be defined as:

[0072]

[0073] in, Represents the node or site v in the spatial adjacency matrix i and v j is the initial connectivity relationship between them, σ is the standard deviation of the distance between sites, and ∈ is the threshold for adjusting the sparsity of the spatial adjacency matrix.

[0074] Based on any of the above embodiments, adjusting the initial connectivity relationships between the sites based on the third historical traffic flow of each site to obtain the connectivity relationships between the sites includes:

[0075] Determine the correlation coefficient between the stations based on the third historical traffic volume of each station;

[0076] Based on the difference between the correlation coefficients between the sites and the corresponding initial connectivity relationships, the initial connectivity relationships between the sites are adjusted to obtain the connectivity relationships between the sites.

[0077] Specifically, the correlation coefficient between each station can be determined based on the third historical traffic flow of each station, and then the difference between the correlation coefficient between each station and the corresponding initial connectivity relationship can be calculated. According to the absolute value of the corresponding difference, the initial connectivity relationship between each station can be optimized and adjusted to obtain a more accurate connectivity relationship between each station.

[0078] Furthermore, taking the Spearman correlation coefficient as an example, the Spearman correlation coefficient between each station can be calculated based on the third historical traffic flow of each station to obtain the correlation coefficient matrix A cor , A cor Each element in can be defined as:

[0079]

[0080] Among them, represents the correlation coefficient between sites v i and v j ; and and are respectively the rank sum and the average value of the ranks of the third historical traffic flow data of sites v i and v j .

[0081] Then, calculate the difference between the corresponding elements in A cor and A s . If the absolute value of the difference exceeds the threshold ε, the corresponding initial connectivity relationship can be adjusted to the correlation coefficient and used as the final connectivity relationship; considering that the initial connectivity relationship calculated based on distance can reflect the connectivity of the traffic road network and is more in line with intuitive cognition, therefore, if the absolute value of the difference does not exceed the threshold ε, no adjustment is required, and the corresponding initial connectivity relationship can be directly used as the connectivity relationship. Finally, the connectivity relationships between each site can be obtained and represented by the spatial adjacency matrix A hs represent.

[0082] The method provided by the embodiments of the present invention, by calculating the correlation coefficients between each site according to the actual historical traffic flow data and combining the physical distances between each site, comprehensively models the traffic road network to obtain a more accurate spatial graph structure and the corresponding spatial adjacency matrix, further improving the prediction accuracy.

[0083] Based on any of the above embodiments, the connectivity relationships between each time node are determined based on the historical nodes and adjacent nodes corresponding to each time node. The adjacent nodes are the time nodes adjacent to each time node, and the historical nodes are the time nodes before each time node.

[0084] Specifically, Figure 2 is a schematic diagram of the construction of the time graph structure provided by the present invention. As Figure 2 shown, by expanding each spatial node, that is, each site, in the time dimension, the time nodes corresponding to each site at each moment can be obtained. Considering that each time node is related to its adjacent time nodes, that is, adjacent nodes, and all the time nodes before it, that is, historical nodes, at the same time, that is, it will be affected by the historical nodes and adjacent nodes corresponding to it, and the two respectively represent local and long-term time dependence relationships. In this regard, the embodiments of the present invention represent this local and long-term time dependence relationship in the form of binary (0 / 1), thereby obtaining the original time adjacency matrix to represent the connectivity relationships of each spatial node in the time dimension. The specific formula is as follows:

[0085]

[0086] Among them, is the time node of site v i , where m and n are different moments, and 0 ≤ m ≤ T, 0 ≤ n ≤ T. 1 indicates that the corresponding two time nodes are connected, and 0 indicates that the corresponding two time nodes are not connected.

[0087] Taking Figure 2 in as an example, and are its neighbor nodes, and are its historical nodes, The connectivity relationship between and is 1, and the connectivity relationship with itself is also 1, while the connectivity relationship with is 0.

[0088] Furthermore, considering that in the time dimension, the strength of the connectivity relationship between different time nodes is different. In this regard, in the embodiment of the present invention, after obtaining the original time adjacency matrix , it is optimized by adaptive weighting, and thus the optimized time adjacency matrix A t ∈R T×T can be obtained. Specifically:

[0089]

[0090] Among them, w t ∈R T×T is a trainable weight and can be randomly initialized; ⊙ represents the Hadamard product, that is, the corresponding elements of the matrix are multiplied.

[0091] It should be noted that in the time dimension, heuristically modeling explicit local and global time dependencies as a time graph structure can achieve synchronous extraction of local and global spatio-temporal features, and set corresponding weights for the connectivity relationships between time nodes in the time graph structure, so that the model can adaptively learn the differences in the strength of the connectivity relationships between time nodes, and finally obtain a more accurate time adjacency matrix, further improving the prediction accuracy.

[0092] Based on any of the above embodiments, the feature extraction module includes alternately arranged feature extraction units and batch normalization layers;

[0093] The feature extraction unit includes multiple parallel spatio-temporal synchronous graph aggregation channels and a feature aggregation layer connected to the multiple spatio-temporal synchronous graph aggregation channels.

[0094] Specifically, the feature extraction module is a deep structure composed of alternately arranged feature extraction units and Batch Normalization layers, that is, a Batch Normalization layer is cascaded after each feature extraction unit, mainly used for extracting spatio-temporal features of the input data.

[0095] Among them, the feature extraction unit may include multiple parallel spatio-temporal synchronization graph aggregation channels and a feature aggregation layer connected to the multiple spatio-temporal synchronization graph aggregation channels, and the outputs of each channel can finally be aggregated through the feature aggregation layer. Further, the feature aggregation layer may be a pooling layer. To deepen the depth of the network and extract more extensive depth features, each spatio-temporal synchronization graph aggregation channel may include multiple spatio-temporal synchronization graph aggregation units.

[0096] It should be noted that the structures of each spatio-temporal synchronization graph aggregation channel are exactly the same, the weights of each channel are randomly initialized and parameter sharing is not performed, and the final objective function is the same, so that it is beneficial for each channel to converge towards the same target from different initial positions in the parameter space for the same input, obtaining different output results, and then aggregating the important information of different channels through the feature aggregation layer, thereby enhancing the generalization of the model.

[0097] Based on any of the above embodiments, Figure 3 is a schematic structural diagram of the traffic flow prediction model provided by the present invention. As Figure 3 shown, the schematic structural diagram of the traffic flow prediction model may include an encoding module, a feature extraction module, a Permute layer, and a many-to-many output module. Among them, the feature extraction module includes alternately arranged feature extraction units (i.e., Figure 3 the STGSAS in Figure 3 ) and Batch Normalization layers, the Permute layer is used to perform dimension adjustment operations, and the output module includes multiple fully connected layers (i.e.,

[0098] Figure 4 the FCs in Figure 4 ) and a concat layer. Figure 4 Figure 4 is a schematic internal structure diagram of the feature extraction unit provided by the present invention. As

[0099] shown, the feature extraction unit may include three parallel spatio-temporal synchronization graph aggregation channels and a feature aggregation layer (i.e., Figure 4 the Max Pooling in , each spatio-temporal synchronization graph aggregation channel may include two spatio-temporal synchronization graph aggregation units (i.e., Figure 4 the STGSA Block in ). The first historical traffic flows X = [X t-T+1 , X t-T+2 , …, X tInput into the traffic flow prediction model, the encoding module encodes the input data through a fully connected network, maps it to a higher dimension, and obtains the encoded input data X e , specifically:

[0100] X e = leakyRelu(XW e + b e )

[0101] Among them, X ∈ R N×T×F , N, T, and F respectively represent the number of stations, the length of the time series, and the feature dimensions (such as traffic flow, speed, etc.); W e , b e are the corresponding trainable weights and biases, and leakyRelu is the activation function.

[0102] The encoding module embeds the encoded input data into time encoding and spatial encoding to obtain the traffic feature X es , specifically:

[0103]

[0104] Among them, E s ∈ R N×1×F , E t ∈ R 1×N×F are respectively the trainable explicit position encoding and time encoding, that is, the trainable time weights and spatial weights; represents the matrix addition of automatic dimension expansion.

[0105] Subsequently, the feature extraction module obtains the traffic features output by the encoding module and incorporates the spatial graph structure (i.e., the Spatial Graph in Figure 3 ) and the temporal graph structure (i.e., the Temporal Graph in Figure 3 ) in each feature extraction unit.

[0106] The spatio-temporal graph synchronous aggregation unit in the feature extraction unit can integrate the extraction of time and space features into a synchronous convolution process. In one embodiment, its specific calculation formula can be as follows:

[0107] h (l) = Γ(A hs h (l-1) A t W1 + b1)

[0108] Among them, h (l) , h (l-1) respectively represent the output features of the l-th and l-1-th layers; A hs and A tThey are respectively the adjacency matrix representations corresponding to the constructed spatial graph structure and temporal graph structure; Γ is the GLU activation function, and W1 and b1 are respectively the trainable weights and biases.

[0109] Figure 5 It is a schematic diagram of the internal structure of the spatio-temporal synchronous graph aggregation unit provided by the present invention. As Figure 5 shown, in another embodiment, further combining the layer connection mechanism, the calculation formula of the spatio-temporal graph synchronous aggregation unit can also be as follows:

[0110] h * = concat(Γ(A hs h (l-1) A t W1 + b1), h (l-1) )

[0111] h (l) = σ(h * W2 + b2) ⊙ (h * W3 + b3)

[0112] Among them, h * represents the combined result of the output of the l-th layer of Graph AGG and the output of the l-1-th layer. concat represents the combination operation, specifically the splicing of two feature vectors in the feature dimension; σ represents the Sigmoid activation function; W2, W3 and b2, b3 are respectively the trainable weights and biases.

[0113] As Figure 4 shown, the output results of three channels will be obtained accordingly, and finally aggregated through the max-pooling operation of the feature aggregation layer. The formula is as follows:

[0114]

[0115] Among them, F (l) is the output of the l-th layer of the multi-stream parallel feature extraction unit; are respectively the output features of three parallel channels; the Maxpooling operation represents taking the maximum value of the corresponding position features.

[0116] As Figure 3 shown, each feature extraction unit STGSAS will output an aggregated feature accordingly and input it into the Batch Normalization layer. Finally, spatio-temporal features are obtained and used as the output of the feature extraction module. After the feature dimension is adjusted through the Permute layer, it is input into the many-to-many output module for traffic flow prediction at multiple future moments to obtain the final multi-step prediction result. The formula is as follows:

[0117]

[0118] Among them, represents the prediction result for the i-th future moment; h (n) is the spatio-temporal feature, and are the trainable weights and biases respectively; leakyRelu is the activation function, and its formula is defined as:

[0119]

[0120] Among them, α i is the hyperparameter.

[0121] Based on any of the above embodiments, Figure 6 is the second flowchart of the traffic flow prediction method based on spatio-temporal synchronous graph aggregation provided by the present invention. As Figure 6 shown, the specific process of this method is as follows:

[0122] Step S10: Collect the historical traffic flow data and latitude and longitude data of each station in the area to be predicted. The historical traffic flow data here can include the first historical traffic flow, the second historical traffic flow, and the third historical traffic flow.

[0123] Step S20: Calculate the distances between stations based on the latitude and longitude data of each station, and construct a spatial graph structure based on the distances between stations:

[0124] Step S201: Calculate the distances between stations based on the latitude and longitude data of each station;

[0125] Step S202: Based on the distances between stations, construct an initial spatial graph structure, and represent it with the initial spatial adjacency matrix A s to represent.

[0126] Step S30: Calculate the Spearman correlation coefficient of each station based on the collected third historical traffic flow of each station, and optimize the spatial graph structure:

[0127] Step S301: Calculate the Spearman correlation coefficient between stations based on the third historical traffic flow of each station to obtain the correlation coefficient matrix A cor .

[0128] Step S302: Based on the correlation coefficient matrix A cor obtained in step S301, optimize the spatial graph structure obtained in step S20. The specific steps are as follows:

[0129] Calculate the difference between the corresponding elements in A cor and A s , and represent its absolute value as Ae ;

[0130] Judge A e For the relative magnitudes of each element and the threshold ε, store the indices of the elements with values greater than ε in the set id s , and store the indices of the elements with values less than or equal to ε in the set id cor ;

[0131] Set the elements at the corresponding index positions in the initial spatial adjacency matrix A s in the set id s to zero to obtain A'. s ;

[0132] Set the elements at the corresponding index positions in the correlation coefficient matrix A cor in the set id cor to zero to obtain A'. cor ;

[0133] The optimized spatial adjacency matrix is the result of adding the corresponding elements of A' s and A' cor , denoted as A hs .

[0134] Step S40: Construct a heuristic temporal graph structure based on explicit local and long-term dependencies:

[0135] Step S401: As Figure 2 shown, expand each spatial node in the time dimension. In the time dimension, each time node is related to both adjacent time nodes and all historical time nodes simultaneously, which respectively represent local and long-term temporal dependencies. Represent this local and long-term relationship in the form of binary (0 / 1) to obtain the original temporal adjacency matrix to characterize the connectivity relationship of each spatial node in the time dimension.

[0136] Step S402: After obtaining the original temporal adjacency matrix , optimize it through adaptive weighting to obtain the optimized temporal adjacency matrix A t ∈R T×T .

[0137] Step S50: Construct a spatio-temporal graph synchronous aggregation model based on the constructed spatial graph structure and temporal graph structure:

[0138] Step S501: Define a simplified version of the graph convolution process in the node domain and incorporate the temporal graph, integrating the extraction of temporal and spatial features into a synchronous convolution process, which is called the spatio-temporal graph synchronous aggregation unit STGSA Block.

[0139] Step S502: Based on the STGSA Block constructed in step S501, construct a feature extraction unit with multi-stream parallelism, as Figure 4 shown. The same input data is input into three parallel spatio-temporal synchronization graph aggregation channels. Each spatio-temporal synchronization graph aggregation channel is composed of multiple stacked STGSA Blocks, and there is no information exchange between different channels. The outputs of each channel are finally aggregated through a max pooling operation.

[0140] Step S503: Based on the feature extraction unit with multi-stream parallelism constructed in step S502, construct a final traffic flow prediction model, as Figure 3 shown. The model mainly includes two parts: a multi-layer stacked feature extraction module and a many-to-many output module. The feature extraction module is a deep structure composed of multiple feature extraction units with multi-stream parallelism and Batch Normalization layers stacked on each other, mainly used for extracting spatio-temporal features of the input data; the output module is composed of multiple fully connected layers, used to obtain multi-step prediction results.

[0141] Step S60: Preprocess the collected historical traffic flow data:

[0142] Step S601: Data aggregation: Aggregate the collected historical traffic flow data at a specified time interval Δt, that is, starting from time t0, the traffic flow within the range of t0 to t0 + Δt is counted as a data point in turn, and thus a traffic flow sequence with time as the index and Δt as the time interval is obtained;

[0143] Step S602: Outlier handling: Statistically analyze the outliers in the traffic flow sequence obtained in step S201, and fill them with the average value of the previous d data points corresponding to the outlier data points;

[0144] Step S603: Sample construction: Set the sliding window length to 2*T, and segment the historical traffic flow data after outlier handling in the form of a sliding window. For example, the interval between two adjacent sliding windows can be set to 1, and thus a sample data set with a sample length of 2*T can be obtained. Further, the sample data set is divided into a training set and a test set at a ratio of τ;

[0145] Step S604: Standardization processing: Adopt the mean-variance standardization method, calculate the mean and variance σ through the training set, and the training set and the test set can be standardized based on and σ. Further, the standardized samples are segmented into input samples and corresponding sample labels at a sample length of T. The input samples and corresponding sample labels are applied to train the initial model, and thus a trained traffic flow prediction model is obtained.

[0146] In addition, for the first historical traffic flow of each site input in practical applications, a similar method as described above can be used for preprocessing. The difference is that the length of the sliding window can be T. Correspondingly, there is no need to perform segmentation after the normalization process.

[0147] Step S70: Input the preprocessed first historical traffic flow into the spatio-temporal synchronous aggregation module in the traffic flow prediction model to extract spatio-temporal features:

[0148] Step S701: Encode the input data X = [X t-T+1 , X t-T+2 , …, X t through a fully connected network layer to map it to a higher dimension, obtaining the encoded input data X e .

[0149] Step S702: Embed the encoded input data with time encoding and space encoding to obtain the traffic flow feature X es .

[0150] Step S703: The feature extraction module extracts features based on the traffic flow feature X est , the spatial adjacency matrix, and the temporal adjacency matrix to obtain spatio-temporal features.

[0151] Step S80: Based on the extracted spatio-temporal features, output the predicted traffic flow of all sites in the future through the output module:

[0152] Input the spatio-temporal features extracted by the feature extraction module into different output layers in the output module to obtain the predicted traffic flow at different future times. Finally, the predicted results at multiple future times can be obtained.

[0153] The method provided in the embodiment of the present invention provides a large-scale road network traffic flow prediction method based on a spatio-temporal graph synchronous aggregation model, which obtains the historical traffic flow data and longitude and latitude position data of each site in the area to be predicted; performs preprocessing on the collected historical traffic flow data, such as aggregation, outlier processing, and normalization; optimizes the connectivity relationship between nodes in the spatial graph structure. At the same time, in the time dimension, heuristically model the explicit local and global time dependencies as a time graph structure; model the extraction of spatio-temporal features between traffic flow data as a spatio-temporal graph convolution process, which significantly improves the prediction accuracy without using other deep learning modules.

[0154] Next, a traffic flow prediction device based on spatio-temporal synchronous graph aggregation provided by the present invention will be described. The traffic flow prediction device based on spatio-temporal synchronous graph aggregation described below can be correspondingly referred to the traffic flow prediction method based on spatio-temporal synchronous graph aggregation described above.

[0155] Based on any of the above embodiments, the present invention provides a traffic flow prediction device based on spatio-temporal synchronous graph aggregation. Figure 7 It is a schematic structural diagram of the traffic flow prediction device based on spatio-temporal synchronous graph aggregation provided by the present invention. As Figure 7 shown, the device includes:

[0156] An acquisition unit 710, configured to acquire the first historical traffic flow of each station in the area to be predicted;

[0157] A prediction unit 720, configured to input the first historical traffic flow of each station into a traffic flow prediction model, and obtain the predicted traffic flow of each station output by the traffic flow prediction model;

[0158] The traffic flow prediction model is trained based on the second historical traffic flow of each station; the traffic flow prediction model is used to extract the spatio-temporal features corresponding to each station based on the first historical traffic flow of each station, and perform traffic flow prediction based on the spatio-temporal features corresponding to each station; the spatio-temporal features are synchronously extracted by a spatio-temporal synchronous graph aggregation module in the traffic flow prediction model.

[0159] The device provided by the embodiment of the present invention realizes the synchronous extraction of spatio-temporal features between the traffic flow data of each station through the spatio-temporal synchronous graph aggregation module in the traffic flow prediction model, retains the mutual coupling between time correlation and space correlation, and only needs to use graph convolution, optimizing the modeling process. On this basis, traffic flow prediction is performed based on the extracted spatio-temporal features, so as to greatly improve the prediction performance of traffic flow and realize end-to-end traffic flow prediction.

[0160] Based on any of the above embodiments, the spatio-temporal features corresponding to each station are extracted based on the following steps:

[0161] Based on the encoding module in the spatio-temporal synchronous graph aggregation module, encode the first historical traffic flow of each station to obtain the flow features corresponding to each station;

[0162] Based on the feature extraction module in the spatio-temporal synchronous graph aggregation module, apply the flow features corresponding to each station, the spatial adjacency matrix, and the temporal adjacency matrix to perform feature extraction to obtain the spatio-temporal features corresponding to each station;

[0163] The spatial adjacency matrix represents the connectivity relationship between each station, and the temporal adjacency matrix represents the connectivity relationship between each time node corresponding to each station at each moment.

[0164] Based on any of the above embodiments, the connectivity relationship between each station is determined based on the following steps:

[0165] Determine the initial connection relationship between each site based on the distance between each site;

[0166] Based on the third historical traffic flow of each site, adjust the initial connection relationship between each site to obtain the connection relationship between each site.

[0167] Based on any of the above embodiments, based on the third historical traffic flow of each site, adjust the initial connection relationship between each site to obtain the connection relationship between each site, including:

[0168] Based on the third historical traffic flow of each site, determine the correlation coefficient between each site;

[0169] Based on the difference between the correlation coefficient between each site and the corresponding initial connection relationship, adjust the initial connection relationship between each site to obtain the connection relationship between each site.

[0170] Based on any of the above embodiments, the connection relationship between each time node is determined based on the historical node and the adjacent node corresponding to each time node. The adjacent node is the time node adjacent to each time node, and the historical node is the time node before each time node.

[0171] Based on any of the above embodiments, the feature extraction module includes alternately arranged feature extraction units and batch normalization layers;

[0172] The feature extraction unit includes a plurality of parallel spatio-temporal synchronous graph aggregation channels and a feature aggregation layer connected to the plurality of spatio-temporal synchronous graph aggregation channels.

[0173] Figure 8 Illustrates a schematic diagram of the physical structure of an electronic device, such as Figure 8As shown in the figure, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840. Among them, the processor 810, the communications interface 820, and the memory 830 complete communication with each other through the communication bus 840. The processor 810 can call the logical instructions in the memory 830 to execute a traffic flow prediction method based on spatio-temporal synchronization graph aggregation. The method includes: obtaining the first historical traffic flow of each station in the area to be predicted; inputting the first historical traffic flow of each station into a traffic flow prediction model to obtain the predicted traffic flow of each station output by the traffic flow prediction model; the traffic flow prediction model is trained based on the second historical traffic flow of each station; the traffic flow prediction model is used to extract the spatio-temporal features corresponding to each station based on the first historical traffic flow of each station, and perform traffic flow prediction based on the spatio-temporal features corresponding to each station; the spatio-temporal features are synchronously extracted by a spatio-temporal synchronization graph aggregation module in the traffic flow prediction model.

[0174] In addition, when the logical instructions in the above-mentioned memory 830 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a 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: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, etc., which can store program codes.

[0175] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the traffic flow prediction method based on spatio-temporal synchronization graph aggregation provided by the above-mentioned various methods. The method includes: obtaining the first historical traffic flow of each station in the area to be predicted; inputting the first historical traffic flow of each station into a traffic flow prediction model to obtain the predicted traffic flow of each station output by the traffic flow prediction model; the traffic flow prediction model is trained based on the second historical traffic flow of each station; the traffic flow prediction model is used to extract the spatio-temporal features corresponding to each station based on the first historical traffic flow of each station, and perform traffic flow prediction based on the spatio-temporal features corresponding to each station; the spatio-temporal features are synchronously extracted by a spatio-temporal synchronization graph aggregation module in the traffic flow prediction model.

[0176] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is implemented to execute the traffic flow prediction method based on spatio-temporal synchronization graph aggregation provided by the above-mentioned various methods. The method includes: obtaining the first historical traffic flow of each station in the area to be predicted; inputting the first historical traffic flow of each station into a traffic flow prediction model to obtain the predicted traffic flow of each station output by the traffic flow prediction model; the traffic flow prediction model is trained based on the second historical traffic flow of each station; the traffic flow prediction model is used to extract the spatio-temporal features corresponding to each station based on the first historical traffic flow of each station, and perform traffic flow prediction based on the spatio-temporal features corresponding to each station; the spatio-temporal features are synchronously extracted by a spatio-temporal synchronization graph aggregation module in the traffic flow prediction model.

[0177] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0178] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment 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 essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0179] 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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A traffic flow prediction method based on spatio-temporal synchronous graph aggregation, characterized in that, Including: Obtain the first historical traffic flow of each station in the area to be predicted; Input the first historical traffic flow of each station into a traffic flow prediction model to obtain the predicted traffic flow of each station output by the traffic flow prediction model; The traffic flow prediction model is trained based on the second historical traffic flow of each station; the traffic flow prediction model is used to extract the spatio-temporal features corresponding to each station based on the first historical traffic flow of each station, and perform traffic flow prediction based on the spatio-temporal features corresponding to each station; the spatio-temporal features are synchronously extracted by a spatio-temporal synchronous graph aggregation module in the traffic flow prediction model, and the spatio-temporal synchronous graph aggregation module realizes the synchronous extraction of spatio-temporal features through graph convolution; Among them, the spatio-temporal synchronous graph aggregation module includes a feature extraction module, and the feature extraction module includes alternately arranged feature extraction units and batch normalization layers; each feature extraction unit includes a plurality of parallel spatio-temporal synchronous graph aggregation channels and a feature aggregation layer connected to the plurality of spatio-temporal synchronous graph aggregation channels. Each spatio-temporal synchronous graph aggregation channel has the same structure, the weights of each spatio-temporal synchronous graph aggregation channel are randomly initialized and parameter sharing is not performed, and the final objective function is the same; The spatio-temporal features corresponding to each station are extracted based on the following steps: Based on the encoding module in the spatio-temporal synchronous graph aggregation module, encode the first historical traffic flow of each station to obtain the traffic flow features corresponding to each station; Based on the feature extraction module in the spatio-temporal synchronous graph aggregation module, apply the traffic flow features corresponding to each station, the spatial adjacency matrix, and the temporal adjacency matrix for feature extraction to obtain the spatio-temporal features corresponding to each station; The spatial adjacency matrix represents the connectivity relationship between each station, the temporal adjacency matrix represents the connectivity relationship between each time node corresponding to each station at each moment, and the spatio-temporal graph synchronous aggregation unit in the feature extraction unit realizes the integration of the extraction of time and space features into a synchronous convolution process.

2. The traffic flow prediction method based on spatio-temporal synchronous graph aggregation according to claim 1, wherein The connectivity relationship between each station is determined based on the following steps: Based on the distance between each station, determine the initial connectivity relationship between each station; Based on the third historical traffic flow of each station, adjust the initial connectivity relationship between each station to obtain the connectivity relationship between each station.

3. The traffic flow prediction method based on spatio-temporal synchronization graph aggregation according to claim 2, wherein The step of adjusting the initial connectivity relationship between each station based on the third historical traffic flow of each station to obtain the connectivity relationship between each station includes: Based on the third historical traffic flow of each station, determine the correlation coefficient between each station; Based on the difference between the correlation coefficient between each station and the corresponding initial connectivity relationship, adjust the initial connectivity relationship between each station to obtain the connectivity relationship between each station.

4. The traffic flow prediction method based on spatio-temporal synchronous graph aggregation according to claim 1, characterized in that, The connectivity relationship between each time node is determined based on the historical nodes and adjacent nodes corresponding to each time node. The adjacent nodes are the time nodes adjacent to each time node, and the historical nodes are the time nodes before each time node.

5. A traffic flow prediction device based on spatio-temporal synchronous graph aggregation, characterized in that Including: An acquisition unit for acquiring the first historical traffic flow of each station in the area to be predicted; A prediction unit for inputting the first historical traffic flow of each station into a traffic flow prediction model to obtain the predicted traffic flow of each station output by the traffic flow prediction model; The traffic flow prediction model is trained based on the second historical traffic flow of each station; the traffic flow prediction model is used to extract the spatio-temporal features corresponding to each station based on the first historical traffic flow of each station, and perform traffic flow prediction based on the spatio-temporal features corresponding to each station; the spatio-temporal features are synchronously extracted by a spatio-temporal synchronization graph aggregation module in the traffic flow prediction model, and the spatio-temporal synchronization graph aggregation module realizes the synchronous extraction of spatio-temporal features through graph convolution; Wherein, the spatio-temporal synchronization graph aggregation module includes a feature extraction module, and the feature extraction module includes alternately arranged feature extraction units and batch normalization layers; each feature extraction unit includes a plurality of parallel spatio-temporal synchronization graph aggregation channels and a feature aggregation layer connected to the plurality of spatio-temporal synchronization graph aggregation channels. Each spatio-temporal synchronization graph aggregation channel has the same structure, the weights of each spatio-temporal synchronization graph aggregation channel are randomly initialized and do not share parameters, and the final objective function is the same; The spatio-temporal features corresponding to each station are extracted based on the following steps: Based on the encoding module in the spatio-temporal synchronization graph aggregation module, encode the first historical traffic flow of each station to obtain the traffic flow features corresponding to each station; Based on the feature extraction module in the spatio-temporal synchronization graph aggregation module, apply the traffic flow features corresponding to each station, the spatial adjacency matrix, and the temporal adjacency matrix for feature extraction to obtain the spatio-temporal features corresponding to each station; The spatial adjacency matrix represents the connectivity relationship between each station, the temporal adjacency matrix represents the connectivity relationship between each time node corresponding to each station at each moment, and the spatio-temporal graph synchronization aggregation unit in the feature extraction unit realizes the integration of the extraction of time and space features into a synchronous convolution process.

6. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the traffic flow prediction method based on spatio-temporal synchronization graph aggregation according to any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the traffic flow prediction method based on spatio-temporal synchronization graph aggregation according to any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the traffic flow prediction method based on spatio-temporal synchronization graph aggregation according to any one of claims 1 to 4.

Images