POI-based double-layer graph convolution network traffic prediction method

Abstract

The application discloses a POI-based double-layer graph convolution network traffic prediction method, and obtains a traffic flow prediction value by processing historical traffic flow data, and comprises the following steps: constructing a model based on POI and flow node functions; obtaining a function similarity matrix between nodes based on the constructed model; clustering nodes with similar functions to construct a regional network; respectively extracting spatial features on the nodes and the regional network; and fusing the spatial features extracted on the nodes and the regional network to output a prediction result of traffic flow. The POI-based double-layer graph convolution network traffic prediction method disclosed by the application has high prediction accuracy.

Description

Technical Field

[0001] This invention relates to a traffic prediction method based on a two-layer graph convolutional network using Points of Interest (POI), belonging to the field of traffic technology. Background Technology

[0002] Traffic forecasting is a typical task in multivariate time series forecasting and has been widely used in modern intelligent transportation systems, such as predicting future traffic congestion and planning travel routes in advance. Traffic forecasting mainly focuses on predicting future traffic characteristics (such as flow rate and speed) on road networks.

[0003] In traffic volume forecasting, Points of Interest (POIs) are used to describe the inherent geographical connections between nodes. POIs typically refer to geographical objects that people have specific interests in, such as restaurants, medical services, shopping malls, residences, and businesses. The distribution of POIs reflects the function of traffic nodes and is closely related to traffic flow.

[0004] Mainstream methods using Points of Interest (POIs) to assist traffic flow prediction typically utilize POIs as feature embeddings to calculate static functional similarity matrices. For example, in the paper Lv, M., Hong, Z., Chen, L., Chen, T., Zhu, T., and Ji, S. Temporal multi-graph convolutional network for traffic flow prediction. IEEE Transactions on Intelligent Transportation Systems 22.6:3337-3348, 2020., TF-IDF equations are used to calculate functional relationships between nodes; in the paper Zhang, M., Li, T., Li, Y., and Hui, P. Multi-view joint graph representation learning for urban region embedding. Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, 2021., TF-IDF equations are also used to obtain POI models with inherent regional attributes; and in the paper Zheng, C., Fan, X., Wen, C., Chen, L., Wang, C., and In Li, J. DeepSTD: Mining spatio-temporal disturbances of multiple context factors for citywide traffic flow prediction. IEEE Transactions on Intelligent Transportation Systems 21.9:3744-3755, 2019, it was proposed to use the Inherent Influence Factor (IIF) to calculate the inherent impact of POIs on traffic flow. IIF means the potential function of each area, which consists of three parts: the frequency of POI distribution, the density of POI distribution, and the imbalance of POI distribution.

[0005] However, traditional methods neglect the dynamic impact of Points of Interest (POIs) on inherent traffic patterns. For example, business and residential areas may experience heavy traffic during weekday peak hours, while shopping areas may be busier in the evening. Furthermore, cities are naturally divided into regions due to the functional similarity of nodes, reflecting their inherent geographical relationships. For instance, factories cluster together to centrally manage pollution, and retail businesses cluster to form commercial districts. Traffic nodes within the same region exhibit similar traffic flow patterns; therefore, mining potential traffic patterns within the region can aid in node-level traffic flow prediction.

[0006] Therefore, it is necessary to explore traffic forecasting methods more deeply in order to improve forecast accuracy. Summary of the Invention

[0007] To overcome the above problems, the inventors conducted in-depth research and designed a traffic prediction method based on a two-layer graph convolutional network using Point of Interest (POI). This method processes historical traffic flow data to obtain predicted traffic flow values, and includes the following steps:

[0008] S1. Construct a model for the functionality of traffic nodes based on POI;

[0009] S2. Obtain the functional similarity matrix between nodes based on the constructed model;

[0010] S3. Cluster nodes with similar functions to construct a regional network;

[0011] S4. Extract spatial features from the node and regional networks respectively;

[0012] S5. Fuse the spatial features extracted from nodes and regional networks to output the prediction results of traffic flow.

[0013] In a preferred embodiment, in S1, the IIF method is used for functional modeling, and the model is represented as follows:

[0014]

[0015]

[0016]

[0017] IIF(n,k)=Fre(n,k)×Den(n,k)×IBD(k)

[0018] Where Fre(n,k) represents the frequency of the k-th POI category of the n-th node, Den(n,k) represents the density of the k-th POI category of the n-th node, IBD(k) represents the imbalance of the k-th POI category, IIF(n,k) represents the function of the k-th POI category of the n-th node, and POI(n,k) represents the number of the k-th POI category of the n-th node.

[0019] In a preferred embodiment, in S2, the functions of different POI categories of each node are merged to obtain a node function matrix, and a cosine similarity function is used to obtain a function similarity matrix between different nodes.

[0020] In a preferred embodiment, in S3, nodes are clustered using a spectral clustering method, with the functional similarity matrix used as the adjacency matrix, and the Laplacian matrix obtained as the input for spectral clustering.

[0021] In a preferred embodiment, step S4, extracting spatial features at a node, includes the following sub-steps:

[0022] S411. Temporal features are extracted using temporal convolution.

[0023] S412. Use POI-based dynamic graph convolution to obtain the spatial correlation features between nodes over time.

[0024] S413. Employ the attention mechanism to obtain the temporal correlation of spatial correlation features between nodes, thereby obtaining the spatial features of nodes.

[0025] In a preferred embodiment, in S411, the temporal convolution extraction process is represented as follows:

[0026] (α1,α2)=split(W*X (l) )

[0027]

[0028] Where `split` represents the splitting function, dividing the result into two parts, α1 and α2; W represents the learnable parameters of the convolution filter; * represents the dilated convolution operation; and X... (l) This represents the input to the l-th temporal convolution. represents the output of the l-th temporal convolution, represents the extracted temporal features, ⊙ represents the element-wise multiplication operator, and σ is the sigmoid activation function.

[0029] In a preferred embodiment, step S412 includes the following sub-steps:

[0030] Using trainable tensor H cThe temporal relationships between POIs are described to obtain the dynamic function similarity tensor P describing the nodes;

[0031] Using the extracted temporal features as input, the spatial correlation features between nodes over time are obtained based on the dynamic function similarity tensor between nodes.

[0032] In a preferred embodiment, step S4, extracting spatial features on the regional network, includes the following sub-steps:

[0033] S421. Temporal features are extracted using temporal convolution.

[0034] S422. Spatial attention map convolution is used to extract spatial correlation features between regions;

[0035] S423. Use the temporal attention mechanism to obtain the temporal correlation features of the spatial correlation features between regions, and obtain the spatial features of the regional network.

[0036] In a preferred embodiment, in S422, the spatial attention map convolution is represented as:

[0037]

[0038] in, This represents the output of the spatial attention map convolution at layer l of the regional network. This represents the forward adjacency matrix of a regional network. Let represent the backward adjacency matrix of the regional network, and k represent the k-th power of the matrix.

[0039] In a preferred embodiment, in S5, the fusion is performed via a dynamic transition matrix Bel based on an attention mechanism. d Implementation, represented as:

[0040] Bel d =σ(E d )*Bel

[0041] E d =V e σ((X) T U1)U2((X R U3) T )+b e )

[0042] E d =E d -mean(E d dim=0)

[0043] Among them, V e b eU1, U2, and U3 are learnable parameters, and X represents the historical traffic flow data of all nodes. R This represents the historical traffic flow data for all regions, σ represents the sigmoid activation function, mean represents the averaging function of the tensor along the dim dimension, and Bel is the relationship matrix between different nodes and different regions.

[0044] The beneficial effects of this invention include:

[0045] (1) A regional traffic network is constructed using a functional similarity matrix, which is helpful for node-level traffic prediction;

[0046] (2) Use dynamic graph convolutional networks and spatiotemporal attention to capture dynamic functional similarity between nodes;

[0047] (3) It has high prediction accuracy and can obtain interpretable dynamic functional similarity relationships between nodes. Attached Figure Description

[0048] Figure 1 The diagram illustrates a flow chart of a POI-based two-layer graph convolutional network traffic prediction method according to a preferred embodiment of the present invention. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present invention will become clearer and more apparent.

[0050] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0051] This invention provides a traffic prediction method based on a two-layer graph convolutional network using Points of Interest (POI). This method processes historical traffic flow data to obtain predicted traffic flow values, such as... Figure 1 As shown, it includes the following steps:

[0052] S1. Construct a model for the functionality of traffic nodes based on POI;

[0053] S2. Obtain the functional similarity matrix between nodes based on the constructed model;

[0054] S3. Cluster nodes with similar functions to construct a regional network;

[0055] S4. Extract spatial features from the node and regional networks respectively;

[0056] S5. Fuse the spatial features extracted from nodes and regional networks to output the prediction results of traffic flow.

[0057] In this invention, in S1, the constructed model is a directed graph model. Traffic flow prediction is a time series modeling problem, which uses historical traffic characteristics (e.g., flow rate, speed) to predict future traffic characteristics. Generally, when predicting traffic flow, the traffic network is modeled as a directed graph G = (V, E, A), where V ∈ R. N Let A be the set of road nodes, E be the set of edges, and A ∈ R. N×N Let N represent the adjacency matrix of graph G, where N is the number of traffic nodes.

[0058] According to a preferred embodiment of the present invention, in S1, the IIF method is used for functional modeling. The specific process of the IIF (Inherent Influence Factor) method can be found in the literature Zheng, C., Fan, X., Wen, C., Chen, L., Wang, C., and Li, J. DeepSTD: Mining spatio-temporal disturbances of multiple context factors for citywide traffic flow prediction. IEEE Transactions on Intelligent Transportation Systems 21.9:3744-3755, 2019. It will not be elaborated upon in this invention. IIF means the potential function of each region, which consists of three parts: the frequency of POI distribution, the density of POI distribution, and the imbalance of POI distribution.

[0059] In this invention, based on the IFF method, Shannon entropy is used to measure the degree of imbalance of POI, and the obtained model is expressed as follows:

[0060]

[0061]

[0062]

[0063] IIF(n,k)=Fre(n,k)×Den(n,k)×IBD(k)

[0064] Where Fre(n,k) represents the frequency of the k-th POI category of the n-th node, Den(n,k) represents the density of the k-th POI category of the n-th node, IBD(k) represents the imbalance of the k-th POI category, IIF(n,k) represents the function of the k-th POI category of the n-th node, and POI(n,k) represents the number of the k-th POI category of the n-th node.

[0065] Shannon entropy is used to measure the imbalance of POIs, which allows POIs with fewer distributions to be given higher weights.

[0066] According to the present invention, in S2, the functions of different POI categories of each node are merged to obtain a node function matrix.

[0067] In this invention, the functional similarity matrix can be obtained using various methods, such as cosine similarity, JS divergence, and Euclidean distance. Preferably, the functional similarity matrix between different nodes is obtained using the cosine similarity function, which has better performance.

[0068] The merger is represented as follows:

[0069] IIF(n)=[IIF(n,1),IIF(n,2),…,IIF(n,k)]∈R K

[0070] Where IIF(n) represents the function of the nth node, and K represents the total number of POI categories.

[0071] The functional matrix IIF of different nodes is represented as follows:

[0072] IIF=[IIF(1),IIF(2),…,IIF(n)]∈R N×K

[0073] The functional similarity matrix Sim(i,j) is represented as:

[0074] Sim(i,j)=cosine(IIF(i),IIF(j))

[0075] Where i and j represent different nodes, and cosine represents the cosine similarity function.

[0076] The distribution of Points of Interest (POIs) naturally influences traffic flow. Traffic nodes with similar POI distributions exhibit similar traffic patterns. To explore these potential traffic patterns, according to the present invention, in S3, all nodes are clustered into N. R Each region contains nodes with similar functions. By clustering, traffic areas can be obtained and regional traffic maps can be constructed.

[0077] Furthermore, nodes are clustered using spectral clustering, with the functional similarity matrix used as the adjacency matrix, and their Laplace matrices obtained as input for spectral clustering. Spectral clustering is used to obtain the functional region division of the transportation network, capturing hierarchical geographical relationships. Spectral clustering is a commonly used method based on spectral graph theory, and will not be elaborated upon in this invention.

[0078] After clustering, nodes with similar functions will belong to the same region; for example, all residential areas will belong to the same region. This is based on historical traffic data of length T1. It can generate regional historical data of length T1. Where D R It is the feature number of the region, and the flow data of all nodes at time t is represented as: Where N is the number of traffic nodes. Let D represent the signal from node i at time t, where D is the number of input features.

[0079] In this invention, the average flow of all nodes in the region is used as the flow characteristic of the region.

[0080] Cities are naturally divided into regions due to the functional similarity of nodes, reflecting their inherent geographical relationships. Traditional methods ignore the impact of region division on traffic flow. In this invention, by constructing a regional network and extracting spatial features from both nodes and the regional network, the constructed regional network facilitates node-level prediction. In this invention, nodes and the regional network are used as inputs to obtain the spatial features E of the nodes. d Using a regional network as input, the spatial features E of the region are obtained on the regional network. R .

[0081] Specifically, in S4, extracting spatial features at nodes includes the following sub-steps:

[0082] S411. Temporal features are extracted using temporal convolution.

[0083] S412. POI-based dynamic graph convolution is used to obtain the spatial correlation features between nodes over time, which are used to characterize the dynamic functional similarity between nodes.

[0084] S413. Use an attention mechanism to obtain the temporal correlation of spatial correlation features between nodes.

[0085] In a preferred embodiment, in S411, dilated causal convolution is used to expand the receptive field of the temporal convolution.

[0086] In a preferred embodiment, in S411, the temporal convolution extraction process is represented as follows:

[0087] (α1,α2)=split(W*X (l) )

[0088]

[0089] Where `split` represents the splitting function, dividing the result into two parts, α1 and α2, representing the node network data and the region network data respectively; W is the learnable parameter of the convolution filter; * is the dilated convolution operation; X (l) This represents the input to the l-th temporal convolution. represents the output of the l-th temporal convolution, represents the extracted temporal features, ⊙ represents the element-wise multiplication operator, and σ is the sigmoid activation function.

[0090] In graph convolution, the adjacency matrix is ​​used to represent the spatial relationships between nodes. Traditional methods primarily use static matrices to handle spatial correlations between nodes, employing POI distributions to describe these static spatial relationships. The adjacency matrix is ​​pre-computed as input to the graph convolution. However, the functional similarity between two nodes changes over time. For example, traffic flow between industrial and residential areas peaks in the morning and evening, while traffic flow between commercial and residential areas is greater in the evening. Although trainable matrices can be used to mine dynamic relationships between nodes, considering the periodicity of traffic flow data, using the same trainable similarity matrix for each slice of the day results in a complexity of O(T′×N). 2 ), where T′ is the number of time slices in a day, resulting in a large number of training parameters and excessively high computational requirements.

[0091] In this invention, a dynamic graph convolutional structure based on POIs is designed to describe functional similarity. Since people's travel is purposeful and they move between different POIs, this invention models the dynamic similarity between POIs and uses linear combination of tensors to obtain the functional similarity relationship between nodes.

[0092] In a preferred embodiment, step S412, obtaining spatial correlation features using POI-based dynamic graph convolution, includes the following sub-steps:

[0093] Using trainable tensor H c The temporal relationships between POIs are described to obtain the dynamic function similarity tensor P describing the nodes;

[0094] Using the extracted temporal features as input, the spatial correlation features between nodes over time are obtained based on the dynamic function similarity tensor between nodes.

[0095] Specifically, the trainable tensor H c ∈R K×K×T′ Where T′ is the number of time slices in a day, and after modeling, information from each time slice can be extracted to obtain the relationship matrix H between POIs. c(t) ∈R K×KWhere t∈{1,2,…,T′}; using two static matrices H s =H d =IIF∈R N×K Let H represent the distribution of POIs at the start and end points respectively. c(t) H s H d Generate tensor P∈R N ×N×T′ To describe the dynamic function similarity between nodes, it is represented as:

[0096]

[0097] P i,j,t =softmax(P i,j,t dim=0)

[0098] In this context, the subscripts i, j, q, r, and t represent different elements in each dimension of the matrix, and dim = 0 indicates that the softmax operation is performed in the 0th dimension of the input tensor.

[0099] According to the present invention, the extracted time features As input, the spatial correlation features between nodes over time are obtained based on the dynamic function similarity tensor, and are represented as follows:

[0100]

[0101] in, This represents the spatial correlation features of the output of the l-th layer of the graph convolution. The k-th power represents the forward transition matrix of the transportation network. The k-th power represents the backward transition matrix of the transportation network. This represents the k-th power of the dynamic function similarity tensor P in the time dimension t. It is a learnable matrix.

[0102] In S413, an attention mechanism is used to obtain the temporal correlation of spatial correlation features between nodes. This attention mechanism is expressed as follows:

[0103] E = V e σ((X) T U1)U2((XU3) T )+b e )

[0104]

[0105]

[0106] V e , These are learnable parameters. The temporal correlation features representing the spatial correlation between nodes, i.e., the node spatial features, are represented by E, which represents the learnable attention matrix. The subscripts i and j represent different elements in the two dimensions of the matrix, and C represents the temporal correlation features representing the spatial correlation between nodes. l It is the number of features of the input data of the l-th layer of the network. It is the length of the time dimension of the l-th layer of the network.

[0107] According to the present invention, by using an attention mechanism, the output features of the previous layer are used as the input features of the next layer, and after multiple convolutional operations, the spatial features E of the nodes can be obtained. d .

[0108] According to the present invention, in S4, spatial features are extracted on the regional network using spatial attention map convolution, and the process is related to the spatial features E of the nodes. d The difference is that the input has been changed from nodes and regional networks to regional networks.

[0109] Specifically, it includes the following sub-steps:

[0110] S421. Temporal features are extracted using temporal convolution.

[0111] S422. Spatial attention map convolution is used to extract spatial correlation features between regions;

[0112] S423. Use the temporal attention mechanism to obtain the temporal correlation features of the spatial correlation features between regions, and obtain the spatial features of the regional network.

[0113] The S421 process is the same as the S411 process, except that the input to the temporal convolution is a region network, and the regional temporal features output from the l-th layer of the network are represented as follows:

[0114] In S422, the spatial attention map convolution is represented as:

[0115]

[0116] in, This represents the output of the spatial attention map convolution at layer l of the regional network. This represents the forward adjacency matrix of a regional network. Sim represents the backward adjacency matrix of the regional network, where k represents the k-th power of the matrix. R This represents the functional similarity matrix between regional networks.

[0117] Similarly, Sim r The method for obtaining it is the same as the method for obtaining Sim.

[0118] According to the present invention, the process of S423 is the same as that of S413, except that the input is the spatial correlation characteristics between nodes in the process that change over time. Replace with spatial correlation characteristics between regions over time. This allows us to obtain the spatial features of the regional network output from the l-th layer.

[0119] Similarly, by using the attention mechanism, the output features of the previous layer are used as the input features of the next layer. After multiple convolutional operations, the spatial features E of the regional network can be obtained. r .

[0120] According to the present invention, in S5, the node spatial feature E d Spatial characteristics of regional networks E R The fusion is performed using a dynamic transition matrix Bel based on an attention mechanism. d Implementation, represented as:

[0121] Bel d =σ(E d )*Bel

[0122] E d =V e σ((X) T U1)U2((X R U3) T )+b e )

[0123] E d =E d -mean(E d dim=0)

[0124] Among them, V e b e U1, U2, and U3 are learnable parameters. U1∈R N , C l It is the number of features of the input data of the l-th layer of the graph convolutional network. X is the length of the time dimension of the l-th layer of the graph convolutional network; X represents the historical traffic flow data of all nodes. R This represents historical traffic flow data for all regions, where σ represents the sigmoid activation function, and mean represents the averaging function over the tensor along the dim dimension. The relationship matrix between different nodes and different regions is represented as follows:

[0125]

[0126] in,

[0127] Through attention matrix Bel d The output features F of the node can be obtained. out and the output characteristics of the region

[0128] Specifically, the output feature F of the node out Represented as:

[0129] F out =concat(F,Bel) d *F R )

[0130] Where concat represents the join function, F represents the input features of the node, and F R The input features of the region are represented.

[0131] Output characteristics of the region Represented as:

[0132]

[0133] In S5, after feature fusion, the output feature F of the node is used. out and the output characteristics of the region The input is passed through multiple fully connected layers, and the output is the traffic prediction result.

[0134] Preferably, skip connections are used to combine the outputs of fully connected layers.

[0135] In a preferred embodiment, the POI-based two-layer graph convolutional network is trained using mean absolute error (MAE).

[0136] The mean absolute error Loss is expressed as:

[0137]

[0138] Where T2 represents the prediction time length, s represents different segments within the T2 time length, N represents the number of traffic nodes, and p represents different traffic nodes. This represents the actual traffic flow at the p-th node during the time period T1+s. This represents the predicted traffic flow at the p-th node during the time period T1+s.

[0139] Example

[0140] Example 1

[0141] We used publicly available datasets (Jinan Traffic Speed ​​Dataset and Xi'an Traffic Speed ​​Dataset) to perform traffic prediction and validate the effectiveness of the prediction method.

[0142] The Jinan traffic dataset contains 561 nodes, and the Xi'an traffic speed dataset contains 792 nodes. The total number of samples in the two datasets is 52,286. Both datasets contain the average speed of road segments sampled every 10 minutes. The datasets are divided into three parts: 60% for training, 20% for validation, and 20% for testing.

[0143] During the prediction process, POIs were divided into 23 categories. The Jinan traffic dataset and the Xi'an traffic dataset contain 72,722 and 117,865 POIs, respectively.

[0144] Traffic prediction is performed using a two-layer graph convolutional network based on Points of Interest (POI), including the following steps:

[0145] S1. Construct a model for the functionality of traffic nodes based on POI;

[0146] S2. Obtain the functional similarity matrix between nodes based on the constructed model.

[0147] S3. Cluster nodes with similar functions to construct a regional network;

[0148] S4. Extract spatial features from the node and regional networks respectively;

[0149] S5. Fuse the spatial features extracted from nodes and regional networks to output the prediction results of traffic flow.

[0150] In S1, the IIF method is used for functional modeling, and the resulting model is represented as follows:

[0151]

[0152]

[0153]

[0154] IIF(n,k)=Fre(n,k)×Den(n,k)×IBD(k)

[0155] In S2, the functions of different POI categories of each node are merged to obtain the node function matrix, and the cosine similarity function is used to obtain the function similarity matrix between different nodes.

[0156] In S3, all nodes are clustered into N. R There are N regions, among which the Jinan traffic data is centrally located. R =20, Xi'an traffic data central N R=30, where nodes in each region have similar functions. Nodes are clustered using spectral clustering, with the functional similarity matrix used as the adjacency matrix and its Laplacian matrix as input. In the clustered regional network, the output features F of the nodes are... out and the output characteristics of the region Obtained through feature fusion.

[0157] In S4, extracting spatial features at nodes includes the following sub-steps:

[0158] S411. Temporal features are extracted using temporal convolution.

[0159] S412. Use POI-based dynamic graph convolution to obtain the spatial correlation features between nodes over time.

[0160] S413. Employ the attention mechanism to obtain the temporal correlation of spatial correlation features between nodes, thereby obtaining the spatial features of nodes.

[0161] S412 includes the following sub-steps:

[0162] Using trainable tensor H c The temporal relationships between POIs are described to obtain the dynamic function similarity tensor P describing the nodes, which is represented as:

[0163]

[0164] P i,j,t =softmax(P i,j,t dim=0)

[0165] Using the extracted temporal features as input, the spatial correlation features between nodes over time are obtained based on the dynamic function similarity tensor between nodes, as represented as:

[0166]

[0167] Extracting spatial features from regional networks includes the following sub-steps:

[0168] S421. Temporal features are extracted using temporal convolution.

[0169] S422. Spatial correlation features between regions are extracted using spatial attention map convolution; the spatial attention map convolution is represented as:

[0170]

[0171] S423. Temporal attention mechanism is used to obtain the temporal correlation features of spatial correlation features between regions, and the spatial features of the regional network are obtained. In S5, after feature fusion, the traffic prediction results are output after passing through multiple fully connected layers.

[0172] Among them, the number of regional features D R =2, l=2.

[0173] Furthermore, during the prediction process, both the input time interval and the prediction time interval were set to 12.

[0174] Comparative Example 1

[0175] The same experiments as in Example 1 were conducted, except that historical average (HA) and existing prediction methods, including LSTM, GRU, GCRN, OGCRNN, GWNET, MTGNN, HGCN, and DeepSTD, were used for prediction.

[0176] LSTM was proposed in the paper "Hochreiter, Sepp, and Jörgn Schmidhuber. Long short-term memory. Neural computation 9.8 (1997): 1735-1780."

[0177] GRU was proposed in the paper "Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555.2014."

[0178] GCRN was proposed in the document "Seo, Y.; Defferrard, M.; Vandergheynst, P.; and Bresson, X. Structured Sequence Modeling with Graph Convolutional Recurrent Networks. In International Conference on Learning Representations (ICLR), 2017.";

[0179] OGCRNN Proceedings “Guo, K. and Hu, Y. and Qian, ZS and Sun, Y. and Gao, J. and Yin, B. AnOptimized TemporalSpatial Gated Graph Convolution Network for Traffic Forecasting.IEEE Intelligent Transportation Systems Magazine,2020b.” Source:

[0180] Proceedings of the GWNET “Wu, Z. and Pan, S. and Long, G. and Jiang, J. and Zhang, C. Graph WaveNetfor Deep Spatial-Temporal Graph Modeling.

[0181] “Gao, J., Zhang, X., Tian, ​​L., Liu, Y., Wang, J., Li, Z., and Hu, X. MTGNN: Multi-Task Graph Neural Network Based Few-Shot Learning for Diseasesimilarity Measurement.Methods 198(2022):88-95.”

[0182] HGCN published “Guo, K., Hu, Y., Sun, Y., Qian, S., Gao, J., and Yin, B. Hierarchicalgraph convolution network for traffic forecasting.

[0183] DeepSTD was proposed in the document "Zheng, C., Fan,

[0184] Experimental Example 1

[0185] The results of Example 1 and Comparative Example 1 are shown in Table 1 and Table 2. Table 1 shows the prediction results on the Jinan traffic speed dataset, and Table 2 shows the prediction results on the Xi'an traffic speed dataset. The prediction times are 30 minutes, 60 minutes, and 120 minutes, respectively.

[0186] Table 1

[0187]

[0188] Table 2

[0189]

[0190] The prediction accuracy was characterized by Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE). Table 1 shows that the results of Example 1 demonstrate the best prediction performance at 30 and 60 minutes, and its performance at 120 minutes is close to that of HGCN. Table 2 shows that the method in Example 1 only performs worse than HGCN at 30 minutes, while Example 1 performs best at 60 minutes and has a smaller MAE loss at 120 minutes. This indicates that the method in Example 1 has higher prediction accuracy compared to existing methods.

[0191] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front," and "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship in the working state of this invention, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0192] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0193] The present invention has been described above with reference to preferred embodiments; however, these embodiments are merely exemplary and illustrative. Various substitutions and modifications can be made to the present invention based on these embodiments, all of which fall within the scope of protection of the present invention.

Claims

1. A POI-based double-layer graph convolution network traffic prediction method, characterized in that, The application discloses a POI-based double-layer graph convolution network traffic prediction method. S1, constructing a model based on POI and traffic node function; S2, obtaining a function similarity matrix between nodes based on the constructed model; S3, clustering nodes with similar functions to construct a regional network; S4, extracting spatial features on the nodes and the regional network respectively; S5, fusing the spatial features extracted on the nodes and the regional network to output a prediction result of traffic flow; In S4, the spatial features extracted on the nodes include the following sub-steps: S411, extracting time features by time convolution; S412, obtaining spatial correlation features between nodes changing over time by dynamic graph convolution based on POI; S413, obtaining time correlation of spatial correlation features between nodes by an attention mechanism to obtain node spatial features. In S412, the following sub-steps are included: Using trainable tensors H c The time sequence relationship between POIs is described to obtain a dynamic function similarity tensor P between nodes. The extracted time features are taken as input, and spatial correlation features between nodes changing over time are obtained based on a dynamic function similarity tensor between nodes.

2. The POI-based double-layer graph convolution network traffic prediction method according to claim 1, wherein in S1, an IIF method is used for function modeling, and the model is represented as: IIF(n, k) = Fre(n, k) x Den(n, k) x IBD(k) wherein Fre(n, k) represents a frequency of a kth POI category of an nth node, Den(n, k) represents a density of the kth POI category of the nth node, IBD(k) represents an imbalance degree of the kth POI category, IIF(n, k) represents a function of the kth POI category of the nth node, and POI(n, k) represents a number of the kth POI category of the nth node.

3. The POI-based double-layer graph convolution network traffic prediction method according to claim 1, wherein in S2, functions of different POI categories of each node are combined to obtain a node function matrix, and a cosine similarity function is used to obtain a function similarity matrix between different nodes.

4. The POI-based double-layer graph convolution network traffic prediction method according to claim 1, wherein in S3, nodes are clustered by a spectral clustering method, and a function similarity matrix is used as an adjacency matrix to obtain a Laplacian matrix as an input of spectral clustering.

5. The POI-based double-layer graph convolution network traffic prediction method according to claim 1, wherein in S411, the time convolution extraction process is represented as:

6. The POI-based double-layer graph convolution network traffic prediction method according to claim 1, wherein in S4, the spatial features extracted on the regional network include the following sub-steps: S421, extracting time features by time convolution; S422, extracting spatial correlation features between regions by a spatial attention graph convolution; S423, obtaining time correlation of spatial correlation features between regions by a time attention mechanism to obtain regional network spatial features. (α1, α2) = split(W * X (l) ) where split denotes a split function that divides the operation result into two parts, and W is a learnable parameter of the convolution filter, * is a dilated convolution operation, X (l) denotes the input of the l-th layer temporal convolution, denotes the output of the l-th layer temporal convolution, represents the extracted temporal feature, denotes an element multiplication operator, and σ is a sigmoid activation function. ​ ​ ​ ​ ​ 7. The POI-based double-layer graph convolutional network traffic prediction method according to claim 6, characterized in that, In S422, the spatial attention graph convolution is expressed as: wherein, represents the output of the region network layer-1 spatial attention map convolution, represents the forward adjacency matrix of the region network, represents the backward adjacency matrix of the region network, and k represents the k-th power of the matrix.

8. The POI-based double-layer graph convolutional network traffic prediction method according to claim 1, characterized in that, In S5, the fusion is performed by a dynamic transfer matrix Bel based on an attention mechanism d Implementation, denoted as: Bel d = σ(E d )* Bel E d = V e σ((X) T U1)U2((X R U3) T + b e ) E d = E d - mean(E d , dim = 0) where V e , b e , U1, U2, U3 are learnable parameters, X represents the historical traffic flow data of all nodes, X R represents the historical traffic flow data of all regions, σ represents the sigmoid activation function, mean represents the average function of the tensor in the dim dimension, and Bel is the relationship matrix between different nodes and different regions.

Images