Road traffic flow prediction method based on graph convolutional recurrent neural network considering data missing

Abstract

The present invention discloses a graph convolutional recursive neural network road traffic flow prediction method considering data missing, which relates to the field of intelligent transportation technology, including: data collection and preprocessing; constructing node features based on Gaussian mixture model; designing graph convolution layer considering data loss; establishing and optimizing spatiotemporal dynamics model; predicting future traffic status; establishing a model, using stochastic gradient descent algorithm, continuously optimizing model hyperparameters through training, and obtaining a traffic status prediction scheme; the present invention can handle data missing problems caused by factors such as sparse sensor deployment, data loss and hardware failure, and can effectively capture the spatiotemporal correlation of traffic data, effectively improving the accuracy and reliability of traffic status prediction.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to a graph convolutional recursive neural network road traffic flow prediction method taking into account data missing. Background Art

[0002] Predicting future traffic status is crucial for developing intelligent transportation systems and smart cities, which promote improved traffic efficiency and reduced fuel consumption. However, unlike other time series prediction problems, traffic status prediction is challenging due to its big data collection and high dimensionality. In addition, traffic forecasting reveals significant spatial (such as neighboring areas) and temporal (such as seasonal effects) correlations, further increasing the complexity of prediction. Benefiting from the recent booming artificial intelligence, deep learning (DL) models and neural networks (NN) have been applied to the study of this problem.

[0003] Models based on the integration of graph neural networks (GNNs) and recurrent neural networks (RNNs) are often used to model unstructured traffic sequences and spatial correlations. For example, the Diffused Convolutional Recurrent Neural Network (DCRNN) first proposed to model the spatial dependency of traffic as a diffusion process on a directed graph, and then adopted gated recurrent units (GRUs) to model temporal dependencies and combined with a sequence-to-sequence (Seq2Seq) architecture for multi-step prediction. Based on the pure convolutional structure of spectral graph convolution, a study proposed a spatiotemporal graph convolutional network (STGCN) model to simultaneously extract the spatial and temporal features of traffic flow. In addition, recent studies have introduced a spatiotemporal attention mechanism combined with a graph convolutional network (GCN) to capture dynamic spatial and temporal correlations.

[0004] However, due to detection failures or errors, the integrity of traffic data cannot be guaranteed in some cases, which significantly limits the performance of traffic prediction. Therefore, related research focuses on transfer learning and generative adversarial networks (GANs) to solve the problem of missing data. For example, one study explored a deep learning model combined with transfer learning techniques for short-term traffic prediction when training data is insufficient. In addition, another graph neural network (TEEPEE) combined with transfer learning was introduced to predict data-sparse road network traffic by learning from rich data from other cities. Nevertheless, most transfer learning-based studies focus on the data missing problem of a single road segment or the entire road network, and rarely consider the data missing problem of the internal segments of the entire network. In addition, GANs are proposed together with parallel data to enhance traffic data interpolation, and certain modifications are made to the algorithm's training difficulties. For example, a new GAN framework learns data on valid adjacent links, where the spatiotemporal correlation is represented by graph embedding and deep walks, and Wasserstein GAN is deployed to generate missing traffic status data. However, the performance of GANs depends heavily on the spatial correlation of the target road segment and is limited by the distribution of real data.

[0005] In summary, the existing methods have the following defects:

[0006] 1. Existing traffic status prediction methods rarely consider missing data, and existing data completion methods rarely consider the spatial connection relationship between road networks: Existing traffic status prediction methods have obvious shortcomings in dealing with the problem of missing data. Most traditional methods assume that the data is complete, but in actual applications, factors such as sensor failure and communication interruption often lead to missing traffic data. If these missing data cannot be effectively completed, it will seriously affect the accuracy of the prediction model. In addition, some current data completion methods often ignore the spatial connection relationship between nodes in the road network, only focus on local data information processing, and fail to fully utilize the topological structure and spatial correlation between roads. This method of ignoring spatial relationships may result in low quality of completed data, which in turn affects the accuracy of traffic status prediction.

[0007] 2. Existing traffic status prediction methods are difficult to capture the spatiotemporal characteristic data of road traffic status: A core challenge of traffic status prediction is to effectively capture the spatiotemporal characteristics of road traffic status. Traditional prediction methods usually find it difficult to simultaneously consider the complex interactive relationships between time and space, which are crucial in traffic flow. In the temporal dimension, traffic flow has obvious periodicity and trend, and is significantly affected by external factors such as weather and holidays. In the spatial dimension, traffic status is highly correlated between adjacent sections, and the traffic conditions of a section of road will directly affect the roads around it. Existing methods are often unable to deal with these complex spatiotemporal characteristics and cannot accurately model and predict dynamically changing traffic conditions. Although emerging technologies such as deep learning and graph neural networks have improved this problem to a certain extent, they still face many challenges in specific applications, such as the complexity of the model, the demand for training data, and the high requirements for computing resources. Summary of the invention

[0008] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a graph convolutional recurrent neural network road traffic flow prediction method that takes data missing into account. The present invention can handle data missing problems caused by factors such as sparse sensor deployment, data loss and hardware failure, and can effectively capture the spatiotemporal correlation of traffic data, effectively improving the accuracy and reliability of traffic status prediction.

[0009] To achieve the above object, the technical solution adopted by the present invention is: a graph convolutional recurrent neural network road traffic flow prediction method considering data missing, comprising the following steps:

[0010] Step 1: Data collection and preprocessing: According to the experimental design, data missing data are processed, data set is divided, normalized, and detector maps are constructed to obtain historical traffic data characteristics on the road network;

[0011] Step 2: Construct node features based on Gaussian mixture model: In the input of the first layer of neural network, the input node features are represented by Gaussian mixture model to integrate the missing processing mechanism with the prediction mechanism;

[0012] Step 3: Design a graph convolution layer that takes data loss into consideration: For the neural network layer whose input contains missing data, the node features based on the Gaussian mixture model are fused into the graph convolution network to construct a graph convolution layer.

[0013] Step 4: Establish and optimize the spatiotemporal dynamics model: The temporal correlation of traffic data is captured by a variant gated recurrent unit of a recurrent neural network. Using the Seq2Seq architecture based on a recurrent neural network, the multi-layer perceptron layer of the variant gated recurrent unit is replaced by a graph convolution layer to capture the spatial correlation. Thus, a graph convolutional gated recurrent unit that takes data loss into account is constructed. Both the encoder and the decoder are composed of graph convolutional gated recurrent units. At the same time, a node parameter learning module is attached to the graph convolutional network.

[0014] Step 5: Predict future traffic states: During training, historical time series data is fed into the encoder and its final state is used to initialize the decoder; the decoder then generates predictions based on ground observations.

[0015] Step 6: Build a model based on steps 1 to 5, use the stochastic gradient descent algorithm, continuously optimize the model hyperparameters through training, and obtain a traffic status prediction solution.

[0016] As a further improvement of the present invention, in step 1, the missing data is processed as follows:

[0017] The missing values ​​of the original data are filled by linear interpolation; time and traffic flow are selected as the input features for traffic status prediction. It is assumed that the time characteristics of all nodes in the road network can be obtained. According to the missing rate parameter, some node data are removed, and the nodes with missing data are randomly selected by uniform probability.

[0018] As a further improvement of the present invention, in step 1, the data set is divided and normalized as follows:

[0019] The dataset is divided into training set, validation set and test set in chronological order, and the Z-Score method is used to normalize the dataset.

[0020] As a further improvement of the present invention, in step 1, the detector map is constructed as follows:

[0021] By calculating the road network distance between detectors, a weighted adjacency matrix A of the nodes is established to represent the detector graph:

[0022]

[0023] In the formula, side a n,n' The weight of d n,n' Decision;d n,n' represents the road network distance between detectors n and n'; ∈ is the standard deviation of the distance, and δ is the threshold.

[0024] As a further improvement of the present invention, the step 2 is specifically as follows:

[0025] Assume that the feature data X of the node n ∈R C From the C-dimensional probability distribution, choose the mixed Gaussian model as the probability distribution model, then:

[0026]

[0027]

[0028]

[0029] Where I is the number of components, p i is a constrained mixing parameter, and ∑ i p i =1,μ [i] and∑ [k] denote the mean and covariance matrix of the Gaussian components, respectively. is the covariance matrix of the (c, c′) term; for each component, we introduce the mean matrix and the variance matrix

[0030]

[0031]

[0032] In the formula, and Representing the mean and variance of nodes with complete data, the data of node characteristics are formulated as:

[0033]

[0034] As a further improvement of the present invention, in step 3, a generalized neuron activation function is used to process the parameters of the Gaussian mixture model through a neural network, thereby constructing a graph convolutional network that takes into account data missing, as follows:

[0035] First, we calculate the expected activation of the neurons in the first hidden layer, keeping the rest of the network structure unchanged; assuming the input data, the generalized response of the activation function is expressed as:

[0036] X n，c ~N(μ,∑ 2 )ReLU

[0037]

[0038]

[0039]

[0040] Assuming that the model input follows the node features represented by the Gaussian mixture model, and given the aggregation matrix L and the weight matrix W, we have:

[0041]

[0042]

[0043]

[0044] Where ⊙ is element-by-element multiplication;

[0045] The node parameter learning module is used to enhance the graph convolution layer, and the weight matrix W is expressed as:

[0046]

[0047] In the formula, is the node embedding matrix, d is the embedding dimension (d<<N); is the weight pool;

[0048] but and Rewritten as:

[0049]

[0050]

[0051] The graph convolution layer utilizes the desired activation function in the first graph convolution layer to process the node data following the Gaussian mixture model and keeps the other layer unchanged.

[0052] As a further improvement of the present invention, in step 4, the graph convolution gated recursive unit considering data loss is calculated as follows:

[0053]

[0054]

[0055]

[0056]

[0057] Where, X (t) ,h (t) represents the input and output at time t, r (t) ,u (t) They are the reset gate and update gate at time t, represents the defined graph convolution, Θ r , Θ u , Θ C are the parameters of the corresponding filter.

[0058] As a further improvement of the present invention, it also includes: evaluating the model effect according to existing evaluation indicators.

[0059] Traditional research has studied prediction and completion separately. The present invention aims to combine missing data processing and traffic prediction in a single neural network. Some scholars have proposed an innovative inductive graph neural kriging (1GNNK) model, which focuses on reconstructing the data of unsampled sensors in the network. By introducing diffuse graph convolutional networks (DGCNs), it can accurately characterize the random characteristics of spatial and directional dependencies. In addition, in order to handle diverse data reconstruction tasks and predict traffic status, another research team has developed a spatiotemporal data recovery network (STAR). STAR not only has the ability to capture long-term patterns, but also incorporates an adaptive memory-based attention model through its residual gated temporal convolutional network to learn and strengthen spatial correlation. However, although these inductive methods have achieved remarkable results on local subgraph data, they often ignore the overall correlation from the global perspective of the network. In order to overcome this limitation, the present invention is committed to developing a new type of missing data interpolation technology and closely integrating it with advanced deep learning traffic prediction models. Through this innovative integrated solution, we aim to build a comprehensive neural network that can handle both missing data and traffic prediction tasks, thereby further improving prediction accuracy and data recovery efficiency.

[0060] The present invention aims to solve the common problem of missing data in traffic status prediction, which is one of the key challenges in urban traffic control and management. Due to factors such as sparse sensor deployment, data loss and hardware failure, missing data often appear in nodes in the road network, which brings uncertainty and limitations to traffic prediction; to address this problem, the present invention uses graph convolutional recurrent neural networks to effectively process missing data and make accurate traffic status predictions. This method not only integrates the processing mechanism of missing data, but also combines the advantages of graph learning, and can perform comprehensive data integration and analysis within a single neural network framework.

[0061] The beneficial effects of the present invention are:

[0062] 1. The present invention proposes a method for GCNmd to handle traffic flow prediction with missing data: the missing data is represented by a Gaussian mixture model, the first GCNmd layer uses the expected activation function, and the rest of the network structure remains unchanged. This method integrates the processing of missing node data and the learning of spatiotemporal relationships of the graph into a single neural network architecture.

[0063] 2. The present invention proposes a road traffic flow prediction model: In order to better capture the characteristics of each node in the case of missing data, the present invention adds a node parameter learning module to GCNmd. At the same time, the present invention uses a simple but powerful RNN variant - the gated recurrent unit (GRU), combined with a sequence-to-sequence (Seq2Seq) model to capture the spatiotemporal relationship of traffic status and perform flexible multi-step prediction.

[0064] 3. The present invention improves the accuracy of traffic status prediction. Experiments show that the proposed model is significantly better than other models in terms of MAE and RMSE indicators.

[0065] 4. Through the improved traffic prediction method, the present invention helps traffic management departments to more accurately predict traffic flow changes, take measures in advance to alleviate traffic congestion, and improve road utilization efficiency. It can not only reduce vehicle waiting time and fuel consumption, but also reduce carbon emissions and promote the development of sustainable urban transportation. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flow chart of an embodiment of the present invention;

[0067] Figure 2 This is a diagram of the architecture of a graph convolutional recursive neural network system taking into account data missing in an embodiment of the present invention;

[0068] Figure 3 This is an architecture diagram of the first layer in the GCNdm model in an embodiment of the present invention;

[0069] Figure 4 Schematic diagram of the comparison of prediction performance of different prediction time ranges on the PeMSD4 dataset in an embodiment of the present invention, where (a) is the MAE index and (b) is the RMSE index;

[0070] Figure 5 Schematic diagram of the influence of model parameters in an embodiment of the present invention, wherein (a) is the influence of the number of GMM components, and (b) is the influence of the embedding dimension;

[0071] Figure 6 : The relationship between MAE loss and training rounds in the embodiment of the present invention, where (a) is a prediction time range of 15 minutes, and (b) is a prediction time range of 60 minutes;

[0072] Figure 7 Schematic diagram of the prediction performance of a single node in the PeMSD4 dataset for one day in an embodiment of the present invention, where (a) is the prediction effect after completing data, and (b) is the prediction effect with missing data. DETAILED DESCRIPTION

[0073] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0074] Example

[0075] like Figure 1 As shown in the figure, a graph convolutional recurrent neural network road traffic flow prediction method considering data missing is used. GCRNNdm can complete the historical missing data while predicting traffic flow, thereby reducing the prediction deviation caused by data quality problems and making the prediction results more accurate. Traffic flow data is collected and preprocessed through step 1, and the graph convolutional recurrent neural network model considering data missing is constructed through steps 2 to 4. The result is processed and analyzed through step 5, as shown in the figure below:

[0076] Step 1: Data collection and preprocessing: According to the experimental design, the data is preprocessed by missing data, data set division, normalization, and detector map construction to obtain the historical traffic data characteristics on the road network;

[0077] Step 2: Based on the Gaussian mixture model feature construction, in the input of the first layer of the neural network, the input node features need to be represented by the GMM model to reasonably integrate the missing processing mechanism with the prediction mechanism;

[0078] Step 3: Design of graph convolutional layer (GCNdm) considering data loss. The spatial correlation of traffic data is captured by GCN. For the design of neural network layer with missing data input, this embodiment combines the GMM node features constructed in step 2 and fuses them into GCN to construct the GCNdm layer;

[0079] Step 4: Spatiotemporal dynamics modeling (GCRNNdm) and optimization, the temporal correlation of traffic data is captured by the RNN variant gated recurrent unit (GRU). In order to achieve multi-step prediction, this embodiment uses the Seq2Seq architecture based on recurrent neural networks. The multi-layer perceptron (MLP) layer of GRU is replaced by GCNdm, thereby constructing a graph convolutional gated recurrent unit (GCGRUdm) that considers data loss. Both the encoder and decoder are composed of GCGRUdm; at the same time, in order to better capture the characteristics of a single node in the case of data loss, a node parameter learning module is attached to GCN;

[0080] Step 5: Prediction of future traffic status. In order to achieve multi-step prediction, this embodiment adopts the Seq2Seq architecture based on recurrent neural network. Both encoder and decoder use GCGRUdm module. During the training process, historical time series data is input into the encoder and its final state is used to initialize the decoder. Then, the decoder generates predictions based on ground observations;

[0081] Step 6: Model solution: Establish a model according to the above steps, use the stochastic gradient descent algorithm, continuously optimize the model hyperparameters through training, and obtain the traffic status prediction solution;

[0082] Step 7: Model effect evaluation: evaluate the effect of the model proposed in the present invention according to existing evaluation indicators.

[0083] The present embodiment is further described below:

[0084] (1) Data missing processing:

[0085] Traffic data is aggregated every 5 minutes, resulting in 288 data points per day. Missing values ​​are filled by linear interpolation.

[0086] This embodiment selects time and traffic flow as input features for traffic state prediction. Assuming that the time features of all nodes in the road network can be obtained, the traffic flow data of the nodes may be completely missing. Based on the previously preprocessed data and according to the missing rate parameter, some node data is removed, and uniform probability is used to randomly select nodes with missing data.

[0087] (2) Dataset division and normalization:

[0088] In this embodiment, the data set is divided into a training set, a validation set and a test set in chronological order. The split ratio of the data set is 7:2:1. The Z-Score method is used to normalize the data set.

[0089] (3 Build the detector map:

[0090] The road network distance between detectors is calculated, and a weighted adjacency matrix A of the nodes is established to represent the detector graph.

[0091]

[0092] In the formula, side a n,n′ The weight of d n,n′ Decision;d n,n′ represents the road network distance between detectors n and n′; ∈ is the standard deviation of the distance, and δ is the threshold.

[0093] The basic principle of constructing node features based on GMM in step 2 is as follows:

[0094] In this embodiment, it is assumed that the feature data X of the node n ∈R C From the C-dimensional probability distribution (such as C = 2), here we choose the mixed Gaussian model as the probability distribution model, then:

[0095]

[0096]

[0097]

[0098] Where I is the number of components, p i is a constrained mixing parameter, and ∑ i p i =1,μ [i] and∑ [k] denote the mean and covariance matrix of the Gaussian components, respectively. is the covariance matrix of the (c, c′) term. In addition, this embodiment introduces a mean matrix for each component and the variance matrix

[0099]

[0100]

[0101] In the formula, and represents the mean and variance of nodes with complete data. The mean and variance of nodes with missing data will be affected by other known features. In summary, the data of node features can be formulated as:

[0102]

[0103] The design principle of the graph convolution layer (GCNdm) considering data loss in step 3 is:

[0104] The layer-by-layer propagation rule of the graph convolutional layer (GCN) generally used for complete graph network data can be expressed as:

[0105] H (l+1) =σ(LH (l) W (l) ) (8)

[0106]

[0107] Where L∈R N×N is the aggregation matrix, is the weighted adjacency matrix of an undirected graph with self-connections added, is the degree matrix of the adjacency matrix with self-connection, indicating the number of edges each node has, including the edges connected to itself, W (l) is the trainable weight matrix of a particular layer and σ(·) is the activation function.

[0108] Combined with the node feature construction method based on GMM in step 2, this embodiment uses a generalized neuron activation function to process GMM parameters through a neural network. Thus, a framework of a graph convolutional network (GCNdm) considering data missing is constructed, such as Figure 2 As shown in Figure 1, the computing architecture of the first layer is as follows: Figure 3 As shown:

[0109] The method first calculates the expected activation of neurons in the first hidden layer and keeps the rest of the network structure unchanged. Assuming the input data, the generalized response of the activation function can be expressed as: n，c ~N(μ,∑ 2 )ReLU;

[0110]

[0111]

[0112]

[0113] Assuming that the model input follows the GMM shown in equation (7) in step 2, and given the aggregation matrix L and the weight matrix W, we have:

[0114]

[0115]

[0116]

[0117] Where ⊙ is the element-wise multiplication.

[0118] As mentioned above, the GCN model transforms features by sharing them among all nodes, which may ignore node-specific patterns. Therefore, this embodiment adopts a node parameter learning module to enhance GCNdm, and the weight matrix W can be expressed as:

[0119]

[0120] In the formula, is the node embedding matrix, d is the embedding dimension (d<<N); It is the weight pool.

[0121] but and can be rewritten as:

[0122]

[0123]

[0124] In summary, the GCNdm model proposed in this embodiment uses the expected activation function in the first graph convolution layer to process the node data following the GMM, and keeps the other layer unchanged.

[0125] The spatiotemporal dynamics modeling (GCRNNdm) in step 4 is as follows:

[0126] For traffic state prediction, traffic data exhibits spatial correlation and involves complex temporal correlation. The present invention adopts RNN variant gated recurrent unit (GRU) to model temporal dependency. At the same time, GRU and GCNdm are combined to jointly process the spatiotemporal correlation of traffic data. The multi-layer perceptron (MLP) layer of GRU is replaced by GCNdm, and the graph convolution gated recurrent unit (GCGRUdm) considering data loss is calculated as follows:

[0127]

[0128]

[0129]

[0130]

[0131] Where, X (t) ,h (t) represents the input and output at time t, r (t) ,u (t) They are the reset gate and update gate at time t, represents the graph convolution defined in Equation 8-18, Θ r , Θ u , Θ C are the parameters of the corresponding filters. Similar to GRU, all parameters in GCGRUdm can be trained end-to-end via back-propagation through time.

[0132] To achieve multi-step prediction, this embodiment uses a Seq2Seq architecture based on a recurrent neural network, and both the encoder and decoder are GCGRUdm modules. During training, historical time series data is input to the encoder and its final state is used to initialize the decoder. The decoder then generates predictions based on ground truth observations. During testing, the decoder's input data is replaced by the predicted data generated by the model.

[0133] The GCRNNdm model is formed by combining the above spatiotemporal dynamic modeling and Seq2Seq architecture. The entire network is trained by maximizing the likelihood of generating accurate future predictions using backpropagation over time.

[0134] The main purpose of this embodiment is to establish a graph convolutional recurrent neural network road traffic flow prediction model that takes into account data missing to alleviate traffic congestion. In order to ensure the authenticity of the data, this embodiment uses real traffic flow data collected by the California Transportation Agency Performance Measurement System (PeMS) for prediction. The traffic data is summarized every 5 minutes, generating 288 data points per day. Missing values ​​are filled by linear interpolation. This embodiment selects time and traffic flow as input features for traffic state prediction. Assuming that the time characteristics of all nodes in the road network can be obtained, the traffic flow data of the node may be completely missing. Based on the previously preprocessed data and according to the missing rate parameter, some node data is removed, and uniform probability is used to randomly select nodes with missing data. This embodiment divides the data set into a training set, a validation set, and a test set in chronological order. The diversion ratio of the data set is 7:2:1. The Z-Score method is applied to normalize the data set.

[0135] Secondly, in order to evaluate the effect of this embodiment, the following widely used interpolation method is used as a benchmark model:

[0136] (1) Mean: This method replaces missing data with the mean of the observed node data.

[0137] (2) K-NN: This method samples similar data through the nearest neighbor nodes and then replaces the missing data with the average of these data. In the experiment, k=3 is set.

[0138] (3) Missforest: This is a non-parametric interpolation method that uses random forests to predict missing values.

[0139] (4) GAIN: This is a prediction method based on generative adversarial networks for estimating missing data.

[0140] (5) IGNNK: This method uses stacked graph convolutional layers to build the model and applies a training strategy to generalize missing data interpolation.

[0141] These methods aim to fill in missing values ​​and then make predictions on the complete graph. In addition, prediction models that directly handle incomplete data graphs are compared.

[0142] Furthermore, the proposed graph convolutional recurrent neural network road traffic flow prediction model considering data missing is trained. The proposed model structure is as follows Figure 2 As shown. The main modules in GCRNNdm include initialization, graph convolutional recurrent layer, encoder-decoder. First, the Gaussian mixture model (GMM) parameters are initialized based on historical data, and then the input data is converted to a form that follows a mixture of Gaussian distributions. The initialized historical data is input to the encoder, and then the decoder makes predictions based on the final state of the encoder and the real data (training).

[0143] All neural network-based methods were implemented in Python using Pytorch 1.8.1 and trained using the Adam optimizer with learning rate annealing. Two datasets, PeMSD4 and PeMSD8, were used for performance analysis. The epoch was set to 100, the patience value was set to 10, and the hyperparameters of all models were selected by a grid search strategy on the validation set. The mean absolute error (MAE) and root mean square error (RMSE) were used as evaluation metrics.

[0144] Finally, the experimental effect is evaluated. This embodiment mainly conducts the following five experiments to demonstrate the technical effect of the present invention:

[0145] (1) Prediction comparison experiment of different time ranges:

[0146] In order to analyze the impact of the model on short-term and long-term forecasts, experiments were conducted with forecast time ranges of 5, 15, 30, 45, and 60 minutes. Figure 4 As shown in the figure, the MAE and RMSE performance of each of the PeMSD4 dataset under different baseline models when the missing rate is 20%. It can be seen that most methods show smoother performance, and the trend increases slightly with the increase of the time horizon. When the prediction time range increases from 5 minutes to 30 and 60 minutes, the MAE of the GCRNNdm model increases by 8.98% and 9.39%, respectively. Among the baseline methods, the prediction effect of the I GNNK method is more significantly affected by the prediction time, and the average mean life expectancy increases by 17.94% when the time increases from 5 minutes to 60 minutes. This phenomenon proves the effectiveness of recurrent neural networks in dealing with time dependencies and long-term predictions.

[0147] (2) Comparative test of different missing rates:

[0148] Table 1 shows the prediction performance comparison of different methods on the PeMSD4 and PeMSD8 datasets. When the missing rate changes from 10% to 30%, the MAE and RMSE of the GCRNNdm model increase by 14.66% and 17.89%, respectively; the MAE and RMSE of other baseline models increase more significantly, by at least 35.96% and 39.85%, respectively.

[0149] Table 1 Prediction performance of different methods on PeMSD4 and PeMSD8 datasets

[0150]

[0151] It can be concluded that, compared with other benchmark models, the GCRNNdm model proposed in this embodiment is relatively less affected by the missing rate; in addition, the accuracy of traffic prediction is obviously affected by the data node loss rate, and a higher loss rate leads to lower accuracy in predicting traffic status.

[0152] Compared with other baseline models, the GCRNNdm model proposed in this embodiment can obtain the best prediction results under the same missing rate. Compared with the best performance of the baseline model, for missing rates of 10%, 20% and 30%, the MAE of the GCRNNdm model is reduced by approximately 12.46%, 24.96% and 35.51%, respectively.

[0153] (3) Sensitivity analysis of model parameters:

[0154] The sensitivity analysis in the experiment focuses on two key model parameters of GCRNNdm: the number of GMM components and the embedding dimension of the node parameter learning module. Figure 5 As shown, this experiment was conducted with a data loss rate of 20% and a prediction time of 30 minutes.

[0155] Figure 5 (a) shows the results obtained by testing six different numbers of GMM components. It is observed that, as the number of components increases, there is an initial decrease in both MAE and RMSE for traffic state prediction, followed by a slight increase. The best performance in terms of MAE and RMSE is achieved with 8 and 4 GMM components, respectively.

[0156] Figure 5 (b) shows the results of the node parameter learning module with ten different embedding dimensions. It can be observed that as the embedding dimension increases, MAE and RMSE begin to decrease rapidly and then gradually stabilize.

[0157] This shows that by integrating neural networks, a small number of GMM components and parameter learning modules with embedding dimensions can effectively capture the changing trends of node traffic status.

[0158] (4) Performance analysis of loss function during iteration:

[0159] Figure 6 The performance of the loss function in the iterative process is shown. Figure 6 As shown in (a) and (b) of the figure, the effectiveness of the proposed model is analyzed by taking the MAE loss value after 100 iterations with a prediction interval of 15 minutes and 60 minutes and a node loss rate of 20%. It can be seen that the MAE decreases rapidly in the first 20 iterations, and then decreases slowly after 20 iterations until convergence. This shows that the model proposed in this embodiment can quickly fit the trend with a limited number of iterations.

[0160] (5) Visual analysis of results:

[0161] Taking the scenario in PeMSD4 with a missing rate of 20% and a prediction time span of 15 minutes as an example, a node with missing data and a node with complete data are randomly selected in the network. The corresponding predictions are as follows: Figure 7 shown. Figure 7 The results of the complete data node in (a) show that the GCRNNdm method can accurately predict traffic flow trends and produce stable predictions. Figure 7 For the missing data nodes in (b), the GCRNNdm model is able to predict the traffic flow change trend for most of the time of the day, thereby effectively predicting the missing data nodes to a certain extent.

[0162] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A graph convolutional recurrent neural network road traffic flow prediction method considering data missing, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing: According to the experimental design, the data is missing, divided, normalized, and the detector map is constructed to obtain the historical traffic data characteristics on the road network; In step 1, the detector graph is constructed as follows: By calculating the road network distance between detectors, a weighted adjacency matrix A of the nodes is established to represent the detector graph: In the formula, side a n,n′ The weight of d n,n′ Decision;d n,n′ Denotes detector n and n ′ The road network distance between them; ∈ is the standard deviation of the distance, δ is the threshold; Step 2: Construct node features based on Gaussian mixture model: In the input of the first layer of neural network, the input node features are represented by Gaussian mixture model to integrate the missing processing mechanism with the prediction mechanism; Step 3: Design a graph convolution layer that takes data loss into consideration: For the neural network layer whose input contains missing data, the node features based on the Gaussian mixture model are fused into the graph convolution network to construct a graph convolution layer. Step 4: Establish and optimize the spatiotemporal dynamics model: The temporal correlation of traffic data is captured by a variant gated recurrent unit of a recurrent neural network. Using the Seq2Seq architecture based on a recurrent neural network, the multi-layer perceptron layer of the variant gated recurrent unit is replaced by a graph convolution layer to capture the spatial correlation. Thus, a graph convolutional gated recurrent unit that takes data loss into account is constructed. Both the encoder and the decoder are composed of graph convolutional gated recurrent units. At the same time, a node parameter learning module is attached to the graph convolutional network. Step 5: Predict future traffic status: During training, historical time series data is fed into the encoder and its final state is used to initialize the decoder; the decoder then generates predictions based on ground observations. Step 6: Build a model based on steps 1 to 5, use the stochastic gradient descent algorithm, continuously optimize the model hyperparameters through training, and obtain a traffic status prediction solution.

2. The method for predicting road traffic flow using a graph convolutional recurrent neural network considering data loss according to claim 1, characterized in that: In step 1, the missing data is processed as follows: The missing values ​​of the original data are filled by linear interpolation. Time and traffic flow are selected as input features for traffic status prediction to obtain the time characteristics of all nodes in the road network. According to the missing rate parameter, some node data are removed, and nodes with missing data are randomly selected by uniform probability.

3. The graph convolutional recurrent neural network road traffic flow prediction method considering data missing according to claim 2 is characterized in that: In step 1, the data set is divided and normalized as follows: The dataset is divided into training set, validation set and test set in chronological order, and the Z-Score method is used to normalize the dataset.

4. The method for predicting road traffic flow using a graph convolutional recurrent neural network considering data loss according to claim 1, characterized in that: The step 2 is specifically as follows: Node feature data X n ∈R C From the C-dimensional probability distribution, choose the mixed Gaussian model as the probability distribution model, then: Where I is the number of components, p i is a constrained mixing parameter, and ∑ i p i =1,μ [i] and Σ [k] denote the mean and covariance matrix of the Gaussian components, respectively. is (c,c ′ ) term; introduce the mean matrix for each component and the variance matrix In the formula, and Representing the mean and variance of nodes with complete data, the data of node characteristics are formulated as:

5. The graph convolutional recurrent neural network road traffic flow prediction method considering data missing according to claim 4 is characterized in that: In step 3, the parameters of the Gaussian mixture model are processed by a neural network using a generalized neuron activation function, thereby constructing a graph convolutional network that takes into account data missing, as follows: First, we calculate the expected activation of the neurons in the first hidden layer, keeping the rest of the network structure unchanged; the generalized response of the activation function is expressed as: X n,c ~N(μ,Σ 2 )ReLU The model input follows the node features represented by the Gaussian mixture model, and given the aggregation matrix L and the weight matrix W, we have: Where ⊙ is element-by-element multiplication; The node parameter learning module is used to enhance the graph convolution layer, and the weight matrix W is expressed as: In the formula, is the node embedding matrix, d is the embedding dimension (d< <N); is the weight pool; but and Rewritten as: The graph convolution layer utilizes the desired activation function in the first graph convolution layer to process the node data following the Gaussian mixture model and keeps the other layer unchanged.

6. The graph convolutional recurrent neural network road traffic flow prediction method considering data missing according to claim 5 is characterized in that: In step 4, the graph convolution gated recurrent unit considering data loss is calculated as follows: Where, X (t) ,h (t) represents the input and output at time t, r (t) ,u (t) They are the reset gate and update gate at time t, represents the graph convolution defined, Θ r ,Θ u ,Θ C are the parameters of the corresponding filter.

Images