Traffic prediction method based on location-aware dynamic graph convolution circulation network

Abstract

The invention discloses a traffic prediction method based on a position-aware dynamic graph convolution circulation network, which comprises the following steps: taking road network sensors as nodes, taking connectivity between the sensors as edges, calculating the weight of the edges, modeling a traffic road network into a weighted directed graph, and taking historical traffic observation data recorded by the sensors as graph signals; dividing the historical traffic observation data into training and verification data; using the weighted directed graph and the training and verification data as the input of a position-aware dynamic graph convolutional loop network; the position-aware dynamic graph convolution circulation network comprises a position-aware dynamic graph learning module and a dynamic graph convolution circulation module; training the position sensing dynamic graph convolution circulation network to obtain an optimal prediction model; the weighted directed graph and the traffic observation data are used as input of the optimal prediction model, the traffic state in a period of time in the future is predicted, and the problem of data distribution difference in different stages can be effectively relieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic prediction, and in particular relates to a traffic prediction method based on a location-aware dynamic graph convolutional recurrent network. Background Art

[0002] The problem of spatiotemporal data mining has received extensive attention in the field of deep learning. In Intelligent Transportation Systems (ITS), traffic prediction, as a typical spatiotemporal data mining task, has attracted much attention due to its complex spatiotemporal dependencies and great practical value.

[0003] In recent years, deep learning methods such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs) have been widely used in traffic prediction. RNNs model temporal dependencies through gated memory units, and CNNs use 1D convolutions to extract temporal dependencies. Unfortunately, these methods may not be able to adequately model the spatial relationships of traffic networks. Therefore, some researchers divide traffic networks into regular grids and use CNNs to extract spatial dependencies between these grids. However, CNNs used in Euclidean space ignore the topological relationships in the road network. Currently, graph neural networks (GNNs) model traffic networks as a graph with sensors as nodes. Graph convolution operations are used to aggregate information from neighboring nodes and update the features of central nodes to learn spatial topology. With the development of GNNs, it is found that: the graph structure defined based on predefined rules is limited by availability and heavily relies on the acquisition of prior knowledge; the adaptive graph based on learnable parameters may not necessarily be applicable to the verification and testing phases, ignoring time-varying spatial features. Summary of the invention

[0004] In order to solve the problem of data distribution differences in different stages and take into account the time-varying spatiotemporal heterogeneity and periodic characteristics, the present invention proposes a Position-Aware Dynamic Graph Convolutional Recurrent Network (PADGCRN) and performs traffic prediction based on the network.

[0005] The idea of ​​the present invention is as follows: first, the location-aware dynamic graph learning module combines the spatiotemporal position embedding matrix with the graph convolutional network to extract the spatial relationship between nodes and generate a dynamic graph under discrete time slots; then, the dynamic graph convolutional loop module combines the diffusion convolution with the gated recurrent unit (GRU) to capture the synchronous spatiotemporal correlation; finally, the present invention uses the RNN-based training strategy to effectively reduce the time and memory consumption during the training process. When traffic prediction is performed based on this network, it not only has a faster speed, but also has good prediction accuracy, and the demand for hardware resources is also reduced.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A traffic prediction method based on a location-aware dynamic graph convolutional recurrent network comprises the following steps:

[0008] Step 1: Take the road network sensors as nodes, the connectivity between sensors as edges, and calculate the edge weights. The traffic network is modeled as a weighted directed graph, and the historical traffic observation data recorded by the sensors is used as the graph signal.

[0009] Step 2, dividing the historical traffic observation data into training data and verification data;

[0010] Step 3, using the weighted directed graph and the training and verification data as inputs of a location-aware dynamic graph convolutional recurrent network; the location-aware dynamic graph convolutional recurrent network includes a location-aware dynamic graph learning module and a dynamic graph convolutional recurrent module;

[0011] Step 4, training the location-aware dynamic graph convolutional recurrent network to obtain an optimal prediction model;

[0012] Step 5: Use the weighted directed graph and traffic observation data as inputs of the optimal prediction model to predict the traffic status for a period of time in the future.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] Traffic prediction is a typical spatiotemporal data mining task. Existing studies mainly reflect real-world traffic patterns through predefined static graphs and adaptive graphs. However, traffic conditions often undergo unexpected changes, so that the graph structure learned in the training phase is not always applicable to the testing phase. Therefore, the present invention proposes a new location-aware dynamic graph learning method to obtain dynamic spatial representation. The method generates a location-aware dynamic graph through a location-aware dynamic graph learning module, and combines a gated recurrent unit with a dynamic diffusion convolutional network through a dynamic graph convolutional recurrent module to model spatiotemporal correlation. Experiments on two real datasets show that the prediction performance of the present invention is the best compared with 12 baseline methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is the summary architecture of the location-aware dynamic graph convolutional recurrent network proposed in the present invention and the implementation flow chart of each module.

[0016] Figure 2 This is a comparison chart of the prediction curves of PADGCRN proposed in the present invention and the latest method Ada-STNet within one day.

[0017] Figure 3This is a visualization study of the location-aware dynamic graph and the predefined graph proposed in the present invention; wherein (a) is a visualization heat map of the predefined graph, (b) is a visualization heat map of the proposed location-aware dynamic graph, and (c) is a visualization of speed fluctuations of nodes whose marked areas in (a) do not show correlation but which show strong correlation in (b). DETAILED DESCRIPTION

[0018] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0019] The present invention proposes a traffic prediction method based on location-aware dynamic graph convolutional recurrent network, focusing on how to dynamically generate graph structure and adaptively update the spatial weights between nodes in prediction applications.

[0020] The complete steps of the present invention can be described as:

[0021] Step 1: Take the road network sensors as nodes, the connectivity between sensors as edges, and calculate the edge weights. The traffic network is modeled as a weighted directed graph, and the historical traffic observation data recorded by the sensors is used as the graph signal.

[0022] In this step, the traffic network is modeled as a weighted directed graph G = (V, E, A), where V = {v 1 ,v 2 ,…,v N} represents the road network sensor set, v N represents the Nth sensor, that is, the Nth node, N represents the number of sensors, that is, the number of nodes, and the set e = {e 1 ,e 2 ,…,e M} represents the connectivity between sensors, e M represents the Mth edge between nodes or sensors, M represents the total number of edges between all nodes, is the adjacency matrix of graph G, where element A ij represents the i-th node v i The weight of the edge between the jth node and the sensor is used as the graph signal. The graph signal of N nodes at the historical time step T can be expressed as Where C represents the feature dimension.

[0023] Step 2: Divide the original historical traffic observation data into training and validation data in proportion.

[0024] In this step, for training, verification and testing, the original historical traffic observation data is divided into training, verification and test data in a ratio of 6:2:2. In practical applications, the test data can use actual data, and the data format is the same.

[0025] In step 3, the weighted directed graph from step 1 and the training and verification data from step 2 are used as input to the location-aware dynamic graph convolutional recurrent network.

[0026] The location-aware dynamic graph convolutional recurrent network of the present invention is as follows Figure 1 As shown, it includes a position-aware dynamic graph learning module and a dynamic graph convolution loop module.

[0027] Among them, the location-aware dynamic graph learning module takes a weighted directed graph and historical traffic observation data (training data is used in the training phase, and verification data is used in the verification phase) as input. It is mainly used to combine the spatiotemporal position embedding matrix with the graph convolutional network to extract the spatial relationship between nodes and generate a dynamic graph under discrete time slots.

[0028] The dynamic graph convolutional loop module includes a dynamic diffusion convolutional layer and a gated recurrent unit, which is mainly used to capture synchronized temporal context information and spatiotemporal correlation.

[0029] In addition, the present invention adopts a codec structure, the decoder extracts spatiotemporal features of the image signal, obtains the hidden state of each time step, and the decoder decodes the hidden state into the future prediction output. The specific implementation methods can be described as follows:

[0030] (1) Location-aware dynamic graph learning module

[0031] The location-aware dynamic graph learning module learns the stable spatial relationship between nodes through a weighted directed graph, and uses graph convolution operations to represent the message passing process of dynamic node states. The output of the graph convolution operation is the dynamic filter tensor DF. t ∈R N×d , which is expressed as follows:

[0032] X′ t =X t ||H t-1

[0033]

[0034] Among them, X t is the historical traffic observation data at time t, ∥ represents the tensor concatenation operation, H t-1 is the hidden state at the previous moment, X′ t The input X at the current time in the sequence t and historical information t-1 Fusion is performed to capture the temporal dependencies between adjacent time slots. *G represents the graph convolution operation, K is the convolution depth, Θ, W k represents the convolution kernel parameters, A is the adjacency matrix of the weighted directed graph, and d is the dimension, which is specifically the hidden dimension here.

[0035] The location-aware dynamic graph learning module fully considers the spatiotemporal dependencies of different scales, encodes weekly and daily features, and obtains the weekly embedding vector and daily embedding vector The dimension d is specifically the embedding dimension here. Week index w t ∈[0,7) represents a day of the week, with a day index of d t ∈[0,288) represents a certain time of the day. Concatenate all the embedding vectors of time T to generate the period embedding matrix E W ∈R T×d and the daily cycle embedding matrix E D ∈R T×d In addition, spatial heterogeneity is modeled by assigning an additional embedding vector to each node, resulting in a spatial position embedding matrix E S ∈R N×d Then, the period embedding and spatial position embedding matrices are fused to obtain the spatiotemporal position embedding matrix E STP , expressed as:

[0036] E STP =E S ⊙E W ⊙E D

[0037] Where ⊙ represents the Hadamard product.

[0038] Position-aware dynamic graph learning module combined with dynamic filter tensor DF t and the spatiotemporal position embedding matrix E STP The dynamic graph under discrete time slots can be generated. Specifically, the dynamic filter tensor DF t Perform Hadamard product and nonlinear transformation with the spatiotemporal position embedding matrix to dynamically adjust the spatial correlation of nodes and generate the spatiotemporal position-aware dynamic embedding matrix DE t Finally, the dynamic graph DA is calculated based on the similarity between spatiotemporal location-aware dynamic embeddings. t , the specific implementation is:

[0039]

[0040] Among them, DA t is the position-aware dynamic graph corresponding to time t, tanh(·) is the nonlinear activation function, the activation function ReLU(·) is used for graph sparsification, and the hyperparameter δ is used to control the saturation rate of the activation function. and are two different dynamic filter tensors, and for two different spatiotemporal location-aware dynamic embedding matrices, Represents the spatiotemporal position embedding matrix at time t.

[0041] According to the above architecture, the location-aware dynamic graph learning module of the present invention first constructs an embedding matrix based on the spatiotemporal position information; then, the position embedding is combined with the dynamic filter to generate a dynamic location-aware embedding, and finally a location-aware dynamic graph is obtained.

[0042] (2) Dynamic graph convolution loop module

[0043] The dynamic diffusion convolution layer learns spatial topology and models spatial correlation by simulating the diffusion process through dynamic graphs under discrete time slots and random walks on directed weighted graphs. Specifically, the k-order bidirectional diffusion process in a directed weighted graph is defined as follows:

[0044]

[0045] in, is a learnable parameter; and represents the forward and reverse diffusion probability matrix in the k-order bidirectional diffusion process, They represent the forward and reverse diffusion probability matrices respectively, which are mainly obtained by standardizing the weighted directed graph.

[0046] Considering the dynamic changes of spatiotemporal positions, the dynamic graph under discrete time slots is used for bidirectional diffusion convolution:

[0047]

[0048] Where Z t represents the output of the diffusion convolution at time t, Θ 1 and θ 2 Represents two different convolution kernel parameters, DA t is the location-aware dynamic graph corresponding to time t, A is the adjacency matrix of the weighted directed graph, is a dynamic diffusion matrix, I represents the identity matrix, all diagonal elements are 1, and elements in other positions are 0.

[0049] right Perform row normalization as follows:

[0050]

[0051] represents the row-normalized dynamic diffusion matrix, and denote the row-normalized dynamic diffusion matrices for the forward and reverse directions, respectively, express The element with coordinates (i,j) in the .

[0052] In addition, in order to reduce the noise accumulation and error propagation caused by the stacking of graph convolutional layers, in the dynamic diffusion convolutional layer, Z is adaptively fused by using a gating mechanism t and input X t ′, the fusion process is expressed as:

[0053]

[0054] in b z is a learnable parameter, σ(·) represents the sigmoid activation function, is the output of the dynamic graph diffusion convolution module, z is the distribution of [0,1] interval to measure X t ′ and Z t The weight value of .

[0055] In the dynamic graph convolutional loop module of the present invention, the linear transformation layer in the gated recurrent unit (GRU) is replaced by dynamic diffusion convolution to capture synchronous temporal context information and spatial correlation. Each gated recurrent unit is defined as follows:

[0056]

[0057] H t =u t ⊙H t-1 +(1-u t )⊙C t

[0058] Among them, H t is the hidden state output of the gated recurrent unit at time t, r t 、u t Represent the reset gate and update gate of the gated recurrent unit, C t represents the state of the memory cell at time t, Θ r , Θ u , Θ C are the convolution kernel parameters corresponding to the reset gate, update gate, and memory gate.

[0059] According to the above architecture, the dynamic graph convolution loop module of the present invention combines GRU with dynamic diffusion convolution to infer the spatiotemporal position-aware embedding matrix, enhancing the model's ability to process fine-grained spatiotemporal patterns.

[0060] (3) Codec Structure

[0061] The structure of the decoding layer of the present invention is the same as that of the encoding layer. The input of the encoding layer is X and the initial hidden state, and the output is the hidden state. The input of the decoding layer is only the output hidden state of the encoder. Each encoding layer consists of a position-aware dynamic graph learning module and a dynamic graph convolution cycle module. The encoder is composed of L identical encoding layers stacked together. Correspondingly, L identical decoding layers are stacked together to form a decoder. The encoding layer is based on the input of the historical time step T. Generate hidden states at different times; the decoder is based on the encoder output H en =[H t-T+1 ,H t-T ,…,H T ], generating the forecast output for the future time step P

[0062] Step 4: Train the location-aware dynamic graph convolutional recurrent network described in step 3 to obtain an optimal prediction model.

[0063] Specifically, in this step, the evaluation index and MAE loss function are defined, and the network parameters are updated through back propagation of the Adam optimizer until the loss of the validation set no longer decreases, and the optimal prediction model is obtained.

[0064] The model evaluation indicators include mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE), which are defined as follows:

[0065]

[0066]

[0067] Among them, X is the true value, is the predicted value. The MAE loss function The definition is as follows:

[0068]

[0069] Among them, X i,j,k represents the true value of feature D of node j at time i, is the predicted value.

[0070] In order to reduce the time and memory consumption required for training, a segmented training strategy can be further introduced. The segmented training strategy of the present invention can alleviate the problem of insensitivity to long-term time dependence caused by the forgetting selection mechanism in the memory unit. It uses an efficient RNN training method to only calculate the first i time steps of the sequence in the early training process. As the training deepens, the model gradually increases the input time step, so that the decoder can obtain better performance in a longer range. This training strategy can effectively reduce the time and memory consumption in the early stage of training.

[0071] Step 5, use the weighted directed graph of step 1 and the test data of step 2 (real traffic observation data can be used in actual application) as the input of the optimal prediction model of step 4 for testing, and predict the traffic status in the future period of time (such as one hour).

[0072] In order to clearly illustrate the superiority of the proposed method, a visualization of the predicted and true values ​​of a specific sensor in a given day is plotted as follows: Figure 2 As shown. PADGCRN and Ada-STNet have good fitting of the speed change curves in most time periods. However, the proposed method shows superior prediction performance in complex scenarios, such as peak traffic hours or time periods with large fluctuations. The blue magnified area shows that the proposed method responds to the dynamic changes and sudden fluctuations of peak traffic faster and more accurately than the baseline method. Therefore, PADGCRN can more effectively capture complex spatiotemporal dependencies and simulate the dynamic changes of real-time traffic conditions with higher accuracy.

[0073] like Figure 2 As shown in Figure 2, the spatial relationship between some nodes in the dynamic graph obtained through training and the predefined graph on the METR-LA dataset is visualized as a heat map. The heat map of the predefined graph is sparse and single except for the diagonal line. Figure 3 As shown in (a); in contrast, the dynamic graph learned through the spatiotemporal position embedding matrix contains rich and diverse information such as Figure 3 As shown in (b). Figure 3 Some nodes in the marked box in (a) do not reflect the correlation, but in fact, the speed fluctuations of the nodes are very similar in the same time period. Figure 3 As shown in (c), the proposed location-aware dynamic graph can preserve the latent spatial dependencies at discrete time steps, effectively complementing the predefined graph.

[0074] Table 1 shows the performance comparison of PADGCRN and 12 baseline methods at different prediction times under the same experimental conditions. The best results are in bold and the second best results are underlined. Compared with other baselines, PADGCRN has the best overall prediction accuracy on two real traffic speed datasets. GMAN uses the self-attention mechanism to achieve a global receptive field and has an advantage in capturing long-term spatiotemporal dependencies. However, it performs poorly in extracting local temporal dependencies and has difficulty handling situations with too many nodes.

[0075] Table 1 Experimental comparison results on two real traffic datasets

[0076]

[0077] Therefore, the present invention adopts a traffic prediction method based on a location-aware dynamic graph convolutional recurrent network, which combines the spatiotemporal position embedding matrix with the graph convolutional network to capture the spatial relationship between nodes and generate a discrete time dynamic graph; then, the diffusion convolution is combined with the GRU to capture the synchronous spatiotemporal correlation, and the RNN-based training strategy is used to effectively reduce time and memory consumption. Experiments on two traffic speed data sets show that the method outperforms most of the comparison methods in the prediction interval, proving its superiority.

Claims

1. A traffic prediction method based on location-aware dynamic graph convolutional recurrent network, characterized in that: The steps include: Step 1: Take the road network sensors as nodes, the connectivity between sensors as edges, and calculate the edge weights. The traffic network is modeled as a weighted directed graph, and the historical traffic observation data recorded by the sensors is used as the graph signal. Step 2, dividing the historical traffic observation data into training data and verification data; Step 3, using the weighted directed graph and the training and verification data as inputs of a location-aware dynamic graph convolutional recurrent network; the location-aware dynamic graph convolutional recurrent network includes a location-aware dynamic graph learning module and a dynamic graph convolutional recurrent module; Step 4, training the location-aware dynamic graph convolutional recurrent network to obtain an optimal prediction model; Step 5: Use the weighted directed graph and traffic observation data as inputs of the optimal prediction model to predict the traffic status for a period of time in the future.

2. The traffic prediction method based on location-aware dynamic graph convolutional recurrent network according to claim 1 is characterized in that: In step 1, the weighted directed graph is represented as: G=(V,E,A) where V = {v1,v2,…,v N } represents the road network sensor set, v N represents the Nth sensor, that is, the Nth node, N represents the number of sensors, that is, the number of nodes, E = {e1, e2, …, e M } represents the connectivity between sensors, e M represents the Mth edge between nodes or sensors, M represents the total number of edges between all nodes, is the adjacency matrix of graph G, where element A ij represents the i-th node v i The weight of the edge between the jth node and the jth node, the graph signal of N nodes at the historical time step T is expressed as Where C represents the feature dimension.

3. The traffic prediction method based on location-aware dynamic graph convolutional recurrent network according to claim 1 is characterized in that: The location-aware dynamic graph learning module takes a weighted directed graph and historical traffic observation data as input, combines the spatiotemporal position embedding matrix with a graph convolutional network, extracts the spatial relationship between nodes, and generates a dynamic graph under discrete time slots; The dynamic graph convolutional loop module includes a dynamic diffusion convolutional layer and a gated recurrent unit to capture synchronized temporal context information and spatiotemporal correlation.

4. The traffic prediction method based on location-aware dynamic graph convolutional recurrent network according to claim 3 is characterized in that: The location-aware dynamic graph learning module learns the stable spatial relationship between nodes through a weighted directed graph, and uses graph convolution operations to represent the message passing process of dynamic node states. The output of the graph convolution operation is a dynamic filter tensor DF. t ∈R N ×d : X t ′=X t ∥H t-1 Among them, X t is the historical traffic observation data at time t, ∥ represents the tensor concatenation operation, H t-1 is the hidden state of the previous moment, X t ′ is the input X at the current moment in the sequence t and historical information t-1 Fusion, *G represents the graph convolution operation, K is the convolution depth, Θ, W k represents the convolution kernel parameters, A is the adjacency matrix of the weighted directed graph, and d is the dimension; The spatiotemporal position embedding matrix is ​​expressed as: AND STP =And S ⊙E W ⊙E D where ⊙ represents the Hadamard product, E S ∈R N×d is the spatial position embedding matrix, which is obtained by assigning an additional embedding vector to each node to model spatial heterogeneity; E W , E D ∈R T×d They are the weekly period embedding matrix and the daily period embedding matrix, which are generated by concatenating the embedding vectors of all T times; They are the weekly embedding vector and the daily embedding vector, which are obtained by encoding the weekly features and the daily features; Combined with dynamic filter tensor DF t and the spatiotemporal position embedding matrix E STP Generate a dynamic graph under discrete time slots, expressed as: Among them, DA t is the position-aware dynamic graph corresponding to time t, tanh(·) is the nonlinear activation function, the activation function ReLU(·) is used for graph sparsification, and the hyperparameter δ is used to control the saturation rate of the activation function. and are two different dynamic filter tensors, and for two different spatiotemporal location-aware dynamic embedding matrices, Represents the spatiotemporal position embedding matrix at time t.

5. The traffic prediction method based on location-aware dynamic graph convolutional recurrent network according to claim 3 or 4 is characterized in that: The dynamic diffusion convolution layer learns spatial topology and models spatial correlation by simulating the diffusion process through the dynamic graph under discrete time slots and the random walk on the directed weighted graph, where the k-order bidirectional diffusion process of the directed weighted graph is defined as follows: in, is a learnable parameter; and Represents the forward and reverse diffusion probability matrices in the k-order bidirectional diffusion process; Use the dynamic graph under discrete time slots for bidirectional diffusive convolution: Where Z t represents the output of the diffusion convolution at time t, Θ1 and Θ2 represent two different convolution kernel parameters, DA t is the location-aware dynamic graph corresponding to time t, A is the adjacency matrix of the weighted directed graph, and I is the identity matrix; right Perform row normalization as follows: represents the row-normalized dynamic diffusion matrix, and denote the row-normalized dynamic diffusion matrices for the forward and reverse directions, respectively, express The element with coordinates (i,j) in the .

6. The traffic prediction method based on location-aware dynamic graph convolutional recurrent network according to claim 5 is characterized in that: In the dynamic diffusion convolution layer, Z is adaptively fused through the gating mechanism. t and X t ′, the fusion process is expressed as: in b z is a learnable parameter, σ(·) represents the sigmoid activation function, is the output of the dynamic graph diffusion convolution module, z is the distribution of [0,1] interval to measure X t ′ and Z t The weight value of .

7. The traffic prediction method based on location-aware dynamic graph convolutional recurrent network according to claim 5 is characterized in that: In the dynamic graph convolution cycle module, the linear transformation layer in the gated cycle unit is replaced by dynamic diffusion convolution, and each gated cycle unit is defined as follows: H t =u t ⊙H t-1 +(1-u t )⊙C t Among them, H t is the hidden state output of the gated recurrent unit at time t, r t 、u t Represent the reset gate and update gate of the gated recurrent unit, C t represents the state of the memory cell at time t, Θ r , Θ u , Θ C are the convolution kernel parameters corresponding to the reset gate, update gate, and memory gate.

8. The traffic prediction method based on location-aware dynamic graph convolutional recurrent network according to claim 5 is characterized in that: A position-aware dynamic graph learning module and a dynamic graph convolution loop module form an encoding layer. The decoding layer has the same structure as the encoding layer. L identical encoding layers are stacked to form an encoder. The encoding layer is based on the input of the historical time step T. Generate hidden states at different times; L identical decoding layers are stacked to form a decoder. The decoder is based on the encoder output H en =[H t-T+1 ,H t-T ,…,H T ], generating the forecast output for the future time step P 9. The traffic prediction method based on location-aware dynamic graph convolutional recurrent network according to claim 1 is characterized in that: In step 4, a segmented training strategy is used for training. The segmented training strategy uses an RNN training method to only calculate the first i time steps of the sequence during the early training process. As the training progresses, the model gradually increases the input time steps.

10. The traffic prediction method based on location-aware dynamic graph convolutional recurrent network according to claim 1 or 9, characterized in that: In step 4, the evaluation index and MAE loss function are defined, and the network parameters are updated by back propagation of the Adam optimizer until the loss of the validation set no longer decreases, thereby obtaining the optimal prediction model.

Images