Spatial-temporal data prediction method based on spatial element graph convolutional recurrent neural network

Abstract

The invention relates to a spatio-temporal data prediction method based on a spatial element graph convolutional recurrent neural network, and belongs to the technical field of spatio-temporal data prediction. The method comprises the following steps: dynamic meta-graph construction: establishing a meta-node library, and generating a meta-graph after parameters are randomly initialized; calculating geographic space similarity, wherein data set nodes correspond to road sensors, and a similarity matrix is generated according to geographic latitude and longitude coordinates of the nodes; convolution space feature enhancement: extracting matrix depth space features by using a convolution module; self-adaptive feature fusion: introducing trainable parameters, and performing weighted summation on the enhanced matrix and the meta-graph to obtain a support matrix; and spatio-temporal joint prediction: modeling through an encoder-decoder composed of GCRUs, inputting node past data by the encoder, and outputting a future node prediction result by the decoder. According to the method, geographic space information and a convolutional layer are introduced into a space-time modeling unit, the perception ability of the model for spatial heterogeneity and complex relations is enhanced, and the prediction accuracy of the model is improved in combination with a dynamic fusion mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of spatiotemporal data prediction, and in particular relates to a spatiotemporal data prediction method based on a spatial meta-graph convolutional recurrent neural network. Background Art

[0002] With the rapid development of information technology, spatiotemporal data plays an increasingly important role in many fields such as traffic flow forecasting, urban service facility planning, and weather forecasting. However, spatiotemporal data has complex spatial dependencies and temporal dynamics, which makes accurate prediction a very challenging problem. Traditional spatiotemporal data prediction methods mainly rely on statistical models and machine learning algorithms, such as time series analysis and support vector machines. Although these methods can capture the laws of data to a certain extent, they are often difficult to effectively handle the complexity of spatiotemporal data, especially when faced with large-scale, high-dimensional data, their performance and accuracy will be limited.

[0003] In recent years, deep learning technology has brought new opportunities for spatiotemporal data prediction. Graph convolutional networks (GCN) and recurrent neural networks (RNN) are two important technologies. Graph convolutional networks can effectively capture spatial dependencies, while recurrent neural networks are good at processing dynamic changes in time series data. The spatiotemporal graph convolutional network (STGCN) that combines the two has achieved remarkable results in tasks such as traffic flow prediction. However, the existing spatiotemporal graph convolutional networks still have shortcomings in dealing with spatiotemporal heterogeneity and non-stationarity. For example, traffic flow data is not only affected by the road network structure, but also non-stationary due to factors such as emergencies and weather changes. To solve this problem, researchers have begun to explore more advanced spatiotemporal modeling methods. For example, MegaCRN proposed a meta-graph learning mechanism that can better adapt to the heterogeneity and non-stationarity of spatiotemporal data by introducing a meta-graph convolutional recurrent network (MegaCRN). In addition, the dynamic graph convolutional recurrent neural network (DGCRN) has also been applied in the field of spatiotemporal prediction. Under the seq2seq architecture, it uses a recurrent neural network as the framework and uses graph convolution to obtain spatial correlation, thereby achieving accurate prediction of spatiotemporal data. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a spatiotemporal data prediction method based on spatial meta-graph convolutional recurrent neural network.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions: The present invention provides a traffic flow prediction method based on a spatial meta-graph convolutional recurrent neural network, comprising the following steps: S1. Dynamic meta-graph construction: Build a meta-node library, where parameters are initialized by random variables that obey a standard normal distribution, and generate a meta-graph through the meta-node library; S2. Geospatial similarity calculation: Each node in the dataset corresponds to a sensor on the road in the real world, and a similarity matrix is ​​generated based on the geographic latitude and longitude coordinates of the node. ; The data set records the sensor's geographic location information and the time series data collected by the sensor; S3. Convolutional spatial feature enhancement: the similarity matrix The deep spatial features are extracted through the convolution module to obtain the feature-enhanced similarity matrix ; S4. Adaptive feature fusion: Introducing two trainable parameters and As a weight, the similarity matrix for feature enhancement With the first element graph , Second Element Graph Perform weighted summation to obtain two support matrices ; S5. Spatiotemporal joint prediction: The spatiotemporal dependency is modeled through an encoder-decoder architecture consisting of graph convolutional gated recurrent units (GCRUs). Both the encoder and decoder contain one layer of GCRU units, which supports the matrix Input the GCRU unit and perform graph convolution operation; the encoder inputs the past The decoder outputs the data of the future The prediction results at each time point.

[0006] Furthermore, step S1 specifically includes: Initialize a meta node library ,in Represents the number of memory items, d represents the dimension of each memory item, the hidden layer features are input into a fully connected layer to obtain the query vector, and then the query vector and the memory items in the meta-node library are used to calculate the attention, and finally the reconstructed representation of all nodes is obtained. The formula is as follows: , , , in, Indicates The hidden features of nodes, represents the learnable weights of the fully connected layer, represents the multiplication operation, represents the learnable bias of the fully connected layer, Represents the query vector, scalar Representation vector With memory item The attention score between Represents the first Transpose of rows, meta-node vector Indicates The reconstructed representation of nodes is No. OK, Indicates the number of nodes; The meta-graph is obtained by the following formula : , , , , in, denote the first embedding vector and the second embedding vector respectively, denote the transpose of the first embedding vector and the second embedding vector respectively, represents the meta-node library, g1 and g2 represent the first fully connected layer and the second fully connected layer, Represent the weights of the first fully connected layer and the second fully connected layer, Indicates the number of memory items; represents the first element graph, represents the second element graph, represents the softmax activation function, Represents the relu activation function.

[0007] Furthermore, the geographic space similarity calculation in step S2 specifically includes: The geopy library in Python is used to calculate the straight-line distance between nodes through the longitude and latitude of the nodes. , all nodes form a distance matrix D, The weight is obtained by inputting the threshold Gaussian kernel function. The mathematical expression of the threshold Gaussian kernel function is as follows: , in, Represents the weight matrix No. Row and The elements in the column, Representation Node and The distance between represents the distribution coefficient, Represents the sparse coefficients, which are used to control the matrix The distribution and sparsity of the weight matrix The zero values ​​in are assigned to 1e-9, and then the weight matrix Each row of is transformed by applying the Softmax function to obtain the similarity matrix .

[0008] Furthermore, step S3 specifically includes: The convolution module includes a convolution layer and a softmax activation function; the convolution kernel size of the convolution layer is 3×3, the number of input and output channels is 1, and the padding is set to 1; Similarity Matrix After processing by the convolution module, the feature-enhanced similarity matrix is ​​obtained , the formula is as follows: , in, Represents the operation of a convolutional layer.

[0009] Furthermore, step S4 specifically includes: Similarity matrix for feature enhancement and the first element graph Perform weighted summation to obtain the first support matrix , the formula is as follows: , in, represents the first learnable weight coefficient, Represents the first support matrix; similarly, the similarity matrix for feature enhancement and the first element graph Perform weighted summation to obtain the second support matrix , the formula is as follows: , in, represents the second learnable weighting coefficient.

[0010] Furthermore, in step S5, the collected original time series data is input into the encoder-decoder architecture composed of graph convolution gated recurrent units GCRU, and the past time step data Input into the encoder for processing to obtain the hidden state of the data at time t ; The hidden state of the data and the reconstructed representation of the node After splicing, the splicing hidden state is obtained , input the concatenated hidden state into the decoder for prediction, and get the future Prediction results of time nodes .

[0011] Furthermore, the graph convolution gated recurrent unit GCRU is specifically: Common graph convolution operations in graph convolution gated recurrent units (GCRUs) The formula is as follows: , in, represents the input of the graph convolution operation, represents the graph convolution operation, Represents the parameters of the graph convolution operation, and They represent the first The first support matrix and the second support matrix of the term, represents trainable parameters; The formula for the specific operation in the graph convolution gated recurrent unit GCRU is as follows: in, Represents a graph convolution operation; express Time series data input at the moment; express The hidden state of the moment; Respectively The update gate, reset gate and candidate state at the moment, Represent the parameters of the update gate, reset gate and candidate state respectively, Represent the bias of update gate, reset gate and candidate state respectively, express Activation function, express Activation function, represents element-wise multiplication, express The hidden state of the moment data, express The hidden state of the moment data.

[0012] The advantages of the present invention are: The present invention calculates the spatial similarity matrix by introducing actual geographic location information. The actual geographic location information of the node is fully considered, deeper features are extracted, and geographic features are innovatively deeply integrated with graph structure learning. Integrate geographic information into the spatiotemporal modeling unit GCRU: GCN is particularly good at capturing the inherent spatial features of graph structure data, while GRU is good at extracting temporal patterns. The gated recurrent unit (GRU) is used in combination with GCN to construct the spatiotemporal modeling unit GCRU. On this basis, geographic spatial information and convolutional layers are introduced into the spatiotemporal modeling unit to enhance the model's perception of spatial heterogeneity and complex relationships. Dynamic fusion mechanism: The spatial similarity matrix W and the matrix G obtained by meta-graph learning are weighted summed using trainable parameters α and β. The sum weights α and β are continuously updated and iterated during the training process, and the most appropriate weights are finally obtained; the prediction accuracy of the model is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0014] Figure 1 The figure is a flow chart of the steps of the method of the present invention. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] Example 1 In this embodiment, Figure 1 As shown, the present invention provides a traffic flow prediction method based on a spatial meta-graph convolutional recurrent neural network, and the specific steps include: S1. Dynamic meta-graph construction: Build a meta-node library, where parameters are initialized by random variables that obey a standard normal distribution, and generate a meta-graph through the meta-node library; Specifically, initialize a meta node library ,in Represents the number of memory items, d represents the dimension of each memory item, the hidden layer features are input into a fully connected layer to obtain the query vector, and then the query vector and the memory items in the meta-node library are used to calculate the attention, and finally the reconstructed representation of all nodes is obtained. The formula is as follows: , , , in, Indicates The hidden features of nodes, represents the learnable weights of the fully connected layer, represents the multiplication operation, represents the learnable bias of the fully connected layer, Represents the query vector, scalar Representation vector With memory item The attention score between Represents the first Transpose of rows, meta-node vector Indicates The reconstructed representation of nodes is No. OK, Indicates the number of nodes; The meta-graph is obtained by the following formula : , , , , in, denote the first embedding vector and the second embedding vector respectively, denote the transpose of the first embedding vector and the second embedding vector respectively, represents the meta-node library, g1 and g2 represent the first fully connected layer and the second fully connected layer, Represent the weights of the first fully connected layer and the second fully connected layer, Indicates the number of memory items; represents the first element graph, represents the second element graph, represents the softmax activation function, Represents the relu activation function.

[0017] S2. Geospatial similarity calculation: Each node in the dataset corresponds to a sensor on the road in the real world, and a similarity matrix is ​​generated based on the geographic latitude and longitude coordinates of the node. ; The data set records the sensor's geographic location information and the time series data collected by the sensor; Specifically, the geospatial similarity is calculated as: The geopy library in Python is used to calculate the straight-line distance between nodes through the longitude and latitude of the nodes. , all nodes form a distance matrix D, The weight is obtained by inputting the threshold Gaussian kernel function. The mathematical expression of the threshold Gaussian kernel function is as follows: , in, Represents the weight matrix No. Row and The elements in the column, Representation Node and The distance between represents the distribution coefficient, Represents the sparse coefficients, which are used to control the matrix The distribution and sparsity of the weight matrix are set to 10 and 0.5 respectively. The zero value of is assigned to 1e-9, and then the weight matrix Each row of is transformed by applying the Softmax function to obtain the similarity matrix .

[0018] S3. Convolutional spatial feature enhancement: the similarity matrix The deep spatial features are extracted through the convolution module to obtain the feature-enhanced similarity matrix ; Specifically, the convolution module includes a convolution layer and a softmax activation function; the convolution kernel size of the convolution layer is 3×3, the number of input and output channels is 1, and the padding is set to 1; Similarity Matrix After processing by the convolution module, the feature-enhanced similarity matrix is ​​obtained , the formula is as follows: , in, Represents the operation of a convolutional layer.

[0019] S4. Adaptive feature fusion: Introducing two trainable parameters and As a weight, the similarity matrix for feature enhancement With the first element graph , Second Element Graph Perform weighted summation to obtain two support matrices ; Specifically, the feature-enhanced similarity matrix and the first element graph Perform weighted summation to obtain the first support matrix , the formula is as follows: , in, represents the first learnable weight coefficient, Represents the first support matrix; similarly, the similarity matrix for feature enhancement and the first element graph Perform weighted summation to obtain the second support matrix , the formula is as follows: , in, represents the second learnable weighting coefficient.

[0020] S5. Spatiotemporal joint prediction: The spatiotemporal dependency is modeled through an encoder-decoder architecture consisting of graph convolutional gated recurrent units (GCRUs). Both the encoder and decoder contain one layer of GCRU units, which supports the matrix Input the GCRU unit and perform graph convolution operation; the encoder inputs the past The decoder outputs the data of the future The prediction results at each time point.

[0021] Specifically, the collected raw time series data is input into the encoder-decoder architecture composed of graph convolution gated recurrent units GCRU, and the past time step data Input into the encoder for processing to obtain the hidden state of the data at time t ; The hidden state of the data and the reconstructed representation of the node After splicing, the splicing hidden state is obtained , input the concatenated hidden state into the decoder for prediction, and get the future Prediction results of time nodes .

[0022] Specifically, the graph convolution gated recurrent unit GCRU is: Common graph convolution operations in graph convolution gated recurrent units (GCRUs) The formula is as follows: , in, represents the input of the graph convolution operation, represents the graph convolution operation, Represents the parameters of the graph convolution operation, and They represent the first The first support matrix and the second support matrix of the term, represents trainable parameters; The formula for the specific operation in the graph convolution gated recurrent unit GCRU is as follows: , in, Represents a graph convolution operation; express Time series data input at the moment; express The hidden state of the moment; Respectively The update gate, reset gate and candidate state at the moment, Represent the parameters of the update gate, reset gate and candidate state respectively, Represent the bias of update gate, reset gate and candidate state respectively, express Activation function, express Activation function, represents element-wise multiplication, express The hidden state of the moment data, express The hidden state of the moment data.

[0023] Example 2 In this example, the method of the present invention is applied to the PEMS-BAY dataset and the METR-LA dataset, and compared and verified with other model methods.

[0024] Dataset Description: The PEMS-BAY dataset and the METR-LA dataset are two datasets widely used in spatiotemporal graph prediction tasks, especially in the field of traffic prediction. METR-LA Dataset: METR-LA is a traffic speed dataset collected from loop detectors on the Los Angeles County road network. It contains data from 207 sensors from March 2012 to June 2012, a period of 4 months. Traffic information is recorded at a speed every 5 minutes.

[0025] PEMS-BAY Dataset: PEMS-BAY is a traffic speed dataset collected by the Performance Measurement System (PeMS) of the California Transportation Agency (Cal-Trans). It contains data from 325 sensors in the Bay Area from January to June 2017, a period of six months. Traffic information is recorded at a speed every five minutes.

[0026] Experimental settings: For METRLA and PEMSBAY, each RNN layer in the encoder and decoder has 64 units, and the memory group has 20 64-dimensional meta nodes. The observation step size λ and the prediction horizon µ are both set to 12. Adam is used as the optimizer, where the learning rate is set to 0.01 and the batch size is set to 64. The adaptive parameters α and β are both initialized to 0.5. The input and output dimensions of the convolutional layer are set to 1, and a convolution kernel of size 3 × 3 is used. Padding is applied to ensure that the shape of the output data is the same as the shape of the input. All experiments are performed with a GeForce RTX 4090D GPU.

[0027] Experiments on the dataset: To evaluate the performance of our Geospatially Enabled Meta-Graph Convolutional Recurrent Neural Network (GEM-GCRN), we compared it with several baseline models, which are grouped according to their underlying model architectures: Models based on Graph Neural Networks (GNN): GWNET model: A graph-based model that applies graph wavelet transform to capture spatial dependencies in traffic flow data.

[0028] MTGNN model: An extension of GWNet, combining adaptive graph learning to improve the model's capabilities.

[0029] STGCN model: A spatiotemporal graph convolutional network that captures spatial and temporal dependencies for traffic prediction by using graph convolution and temporal convolution.

[0030] Models based on Recurrent Neural Networks (RNN): AGCRN model: A model combining graph convolution and gated recurrent units (GRU) to capture temporal and spatial dependencies in traffic flow prediction.

[0031] DCRNN model: A diffused convolutional recurrent neural network that predicts traffic flow by learning spatiotemporal patterns using diffusion-based graph convolution and RNN.

[0032] MegaCRN model: A variant of CRN (Convolutional Recurrent Network) that improves prediction accuracy by extending additional features such as hypernetwork learning.

[0033] Model based on attention mechanism: GTS model: A model that learns the probability of each edge from long-term historical data and focuses on the relationship between each node and its connected nodes through an attention mechanism.

[0034] Other models: – STID model: A model that integrates spatiotemporal interdependencies and dynamic features, using an attention mechanism and a graph-based approach to predict traffic flow.

[0035] – STWave model: a spatial feature extraction model based on wavelet transform combined with temporal modeling to capture complex traffic patterns.

[0036] In order to comprehensively evaluate the performance of our model (GEM-GCRN) in capturing spatial and temporal dependencies in traffic flow data, we compare the model with these baselines. Each benchmark uses a different approach to solve the traffic prediction problem, such as graph neural networks, recurrent networks, and attention mechanisms. By analyzing the performance of GEM-GCRN relative to these baselines, we demonstrate that geo-augmented meta-graph learning is effective in improving traffic prediction accuracy.

[0037] Table 1 Comparison of experimental data on the METR-LA dataset As can be seen from Table 1, on the METR-LA dataset, the model of the present invention shows excellent performance in all evaluation indicators. When the prediction step is 3 steps, the model of the present invention GEM-GCRN (without convolutional module) outperforms the current state-of-the-art model MegaCRN in three indicators: MAE (mean absolute error), RMSE (relative mean absolute error), and MAPE (mean absolute percentage error). When the prediction step is 6 steps, GEM-GCRN without convolutional module performs 0.6% better than MegaCRN in MAE and 0.1% better in MAPE. When the prediction step is 12 steps, GEM-GCRN is 0.5% lower than STAEformer and 0.5% lower than MegaCRN. It can be observed that when the prediction range is 15 minutes and 30 minutes (i.e., 3 prediction steps and 6 prediction steps), the proposed method performs well in all indicators and outperforms other methods. However, when the prediction time span is extended to 60 minutes (i.e., 12 prediction steps), its performance decreases slightly. After adding the convolutional module to extract features from the similarity matrix, the model performance was significantly improved within a 60-minute prediction time.

[0038] Table 2 Experimental comparison on the PEMS-BAY dataset As shown in Table 2, on the PEMS-BAY dataset, the proposed method continues to show effectiveness, and after adding the convolution module, the proposed model performs better than other methods in almost all evaluation indicators.

[0039] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A spatiotemporal data prediction method based on spatial meta-graph convolutional recurrent neural network, characterized in that: The following steps are involved: S1. Dynamic meta-graph construction: Build a meta-node library, where parameters are initialized by random variables that obey a standard normal distribution, and generate a meta-graph through the meta-node library; S2. Geospatial similarity calculation: Each node in the dataset corresponds to a sensor on the road in the real world, and a similarity matrix is ​​generated based on the geographic latitude and longitude coordinates of the node. ; The data set records the sensor's geographic location information and the time series data collected by the sensor; S3. Convolutional spatial feature enhancement: the similarity matrix The deep spatial features are extracted through the convolution module to obtain the feature-enhanced similarity matrix ; S4. Adaptive feature fusion: Introducing two trainable parameters and As a weight, the similarity matrix for feature enhancement With the first element graph , Second Element Graph Perform weighted summation to obtain two support matrices ; S5. Spatiotemporal joint prediction: The spatiotemporal dependency is modeled through an encoder-decoder architecture consisting of graph convolutional gated recurrent units (GCRUs). Both the encoder and decoder contain one layer of GCRU units, which supports the matrix Input the GCRU unit and perform graph convolution operation; the encoder inputs the past The decoder outputs the data of the future The prediction results at each time point.

2. According to claim 1, a spatiotemporal data prediction method based on spatial meta-graph convolutional recurrent neural network is characterized in that: Step S1 specifically includes: Initialize a meta node library ,in, Represents the number of memory items, d represents the dimension of each memory item, the hidden layer features are input into a fully connected layer to obtain the query vector, and then the query vector and the memory items in the meta-node library are used to calculate the attention, and finally the reconstructed representation of all nodes is obtained. The formula is as follows: ， ， ， in, Indicates The hidden features of nodes, represents the learnable weights of the fully connected layer, represents the multiplication operation, represents the learnable bias of the fully connected layer, Represents the query vector, scalar Representation vector With memory item The attention score between Represents the first Transpose of rows, meta-node vector Indicates The reconstructed representation of nodes is No. OK, Indicates the number of nodes; The meta-graph is obtained by the following formula : ， ， ， ， in, denote the first embedding vector and the second embedding vector respectively, denote the transpose of the first embedding vector and the second embedding vector respectively, represents the meta-node library, g1 and g2 represent the first fully connected layer and the second fully connected layer, Represent the weights of the first fully connected layer and the second fully connected layer, represents the first element graph, represents the second element graph, represents the softmax activation function, Represents the relu activation function.

3. The spatiotemporal data prediction method based on spatial meta-graph convolutional recurrent neural network according to claim 2 is characterized in that: The geospatial similarity calculation in step S2 specifically includes: The geopy library in Python is used to calculate the straight-line distance between nodes through the longitude and latitude of the nodes. , all nodes form a distance matrix D, The weight is obtained by inputting the threshold Gaussian kernel function. The mathematical expression of the threshold Gaussian kernel function is as follows: ， in, Represents the weight matrix No. Row and The elements in the column, Representation Node and The distance between represents the distribution coefficient, Represents the sparse coefficients, which are used to control the matrix The distribution and sparsity of the weight matrix The zero value of is assigned to 1e-9, and then the weight matrix Each row of is transformed by applying the Softmax function to obtain the similarity matrix .

4. The method for spatiotemporal data prediction based on spatial meta-graph convolutional recurrent neural network according to claim 3 is characterized in that: Step S3 specifically includes: The convolution module includes a convolution layer and a softmax activation function; the convolution kernel size of the convolution layer is 3×3, the number of input and output channels is 1, and the padding is set to 1; Similarity Matrix After processing by the convolution module, the feature-enhanced similarity matrix is ​​obtained , the formula is as follows: ， in, Represents the operation of a convolutional layer.

5. The method for spatiotemporal data prediction based on spatial meta-graph convolutional recurrent neural network according to claim 4 is characterized in that: Step S4 specifically includes: Similarity matrix for feature enhancement and the first element graph Perform weighted summation to obtain the first support matrix , the formula is as follows: ， in, represents the first learnable weight coefficient, Represents the first support matrix; similarly, the similarity matrix for feature enhancement and the first element graph Perform weighted summation to obtain the second support matrix , the formula is as follows: ， in, represents the second learnable weighting coefficient.

6. The method for spatiotemporal data prediction based on spatial meta-graph convolutional recurrent neural network according to claim 5, characterized in that: In step S5, the collected raw time series data is input into the encoder-decoder architecture composed of graph convolution gated recurrent units GCRU, and the past time step data Input into the encoder for processing to obtain the hidden state of the data at time t ; The hidden state of the data and the reconstructed representation of the node After splicing, the splicing hidden state is obtained , input the concatenated hidden state into the decoder for prediction, and get the future Prediction results of time nodes .

7. The method for spatiotemporal data prediction based on spatial meta-graph convolutional recurrent neural network according to claim 6, characterized in that: The graph convolution gated recurrent unit GCRU is specifically: Common graph convolution operations in graph convolution gated recurrent units (GCRUs) The formula is as follows: ， in, represents the input of the graph convolution operation, represents the graph convolution operation, Represents the parameters of the graph convolution operation, and They represent the first The first support matrix and the second support matrix of the term, represents trainable parameters; The formula for the specific operation in the graph convolution gated recurrent unit GCRU is as follows: in, express Time series data input at each moment; express The hidden state of the moment; Respectively The update gate, reset gate and candidate state at the moment, Represent the parameters of the update gate, reset gate and candidate state respectively, Represent the bias of update gate, reset gate and candidate state respectively, express Activation function, express Activation function, represents element-wise multiplication, express The hidden state of the moment data, express The hidden state of the moment data.

Images