A traffic status completion prediction method based on multi-source data pre-training

Abstract

A traffic state completion prediction method based on multi-source data pre-training belongs to the technical field of deep learning algorithms and solves the problem of how to improve the complex data processing capability of traditional traffic models. The invention integrates data from multiple fields and clusters these data through a k-means algorithm to ensure that the data in each cluster are numerically similar. Adjacent data pairs are selected from each cluster in chronological order to capture the correlation in the time dimension. These data pairs are then used as inputs of the model. A variety of data types are used in the pre-training stage, while the formal training stage focuses on traffic data, making full use of the rich features and patterns learned from different data sources in the pre-training stage to achieve "zero-start" training of traffic data. The invention can significantly improve the model's ability to understand and predict traffic data characteristics, especially when faced with sparse data or new domain tasks, and can accelerate the learning and adaptation process of the model.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning algorithms and relates to a traffic status completion prediction method based on multi-source data pre-training. Background Art

[0002] With the continuous acceleration of urbanization and the increasing complexity of transportation systems, people's demand for traffic flow prediction, congestion management, and transportation planning is also growing. However, traditional traffic models have certain limitations in processing massive and complex data, adapting to specific scenarios and data characteristics, and improving model performance and accuracy.

[0003] Existing traffic models often have difficulty in effectively coping with the increasing volume and diversity of traffic data, especially when faced with time-series and discontinuous data, and their performance is often limited. At the same time, existing models also face challenges in adapting to specific scenarios and data characteristics, because the characteristics of different cities, different time periods, and different traffic networks may vary greatly, and traditional models are difficult to respond flexibly. In addition, although there have been some attempts to optimize the model, methods to improve model performance and accuracy still need to be further improved.

[0004] Therefore, to address these challenges, a new type of traffic model is urgently needed. This model should have the following characteristics: it can efficiently process large-scale traffic data, including support for time series and discontinuous data; it has the ability to fine-tune to improve model performance; and it has the ability to adapt to different scenarios and quickly adapt to new scenarios to cope with the diversity and rapid changes of urban transportation systems. Summary of the invention

[0005] The technical solution of the present invention is mainly used to solve the problem of how to improve the complex data processing capability of traditional traffic models.

[0006] The present invention solves the above technical problems through the following technical solutions:

[0007] A traffic state completion prediction method based on multi-source data pre-training, comprising: a pre-training stage and a formal training stage;

[0008] Pre-training stage: collect data from multiple fields and integrate the data, including format standardization and scale standardization; use the k-means algorithm to cluster the data, and select temporally adjacent data pairs from each cluster in chronological order; input the data pairs into the deep fusion MixerLayer module and pre-train the deep fusion MixerLayer module;

[0009] Formal training phase: For the traffic data set, the sliding window method is used to divide the data in the window into Data at the current time point After Data at future time points ; Construct node embedding vectors and generate adjacency matrices, preprocess the data at the current time point through the input fully connected layer, input the preprocessed data into the multi-layer hybrid graph convolutional network for training, and input it into the output fully connected layer after training to obtain the final Prediction results for future time points , and calculate the difference between the predicted results and the data at future time points.

[0010] The present invention integrates data from multiple fields such as transportation, economy, medical care and energy (including water, gas, and heat); clusters these data by applying the k-means algorithm to ensure that the data in each cluster are numerically similar; further, adjacent data pairs are selected from each cluster in chronological order to capture their correlation in the time dimension; these data pairs are then used as input to the model. In the present invention, a variety of data types are used in the pre-training phase, while the formal training phase focuses on traffic data. The purpose of this is to make full use of the rich features and patterns learned from different data sources in the pre-training phase to achieve "zero-start" training of traffic data. This strategy can significantly improve the model's understanding and prediction capabilities of traffic data characteristics, especially when faced with sparse data or new domain tasks, and can accelerate the model's learning and adaptation process.

[0011] Furthermore, the method of clustering data using the k-means algorithm is specifically as follows:

[0012] (1) Initialize the center point by calculating the mean value of each source in the integrated multi-source data set;

[0013] (2) Establish an objective function. The expression of the objective function is as follows:

[0014]

[0015] in, is the number of clusters, It is The set of data points in a cluster, It is a cluster The center point of It is a cluster A data point in Represents data points To center point The square of the Euclidean distance;

[0016] (3) Update the cluster center point. The center point of each cluster is updated by calculating the mean of all data points in the cluster. The specific expression is:

[0017]

[0018] in, Representation Cluster The number of data points in .

[0019] Furthermore, the basis for selecting the temporally adjacent data pairs is as follows:

[0020]

[0021] in, and It is a cluster After sorting the consecutive data points in time, ensure that each pair of data is adjacent in time. Represents a pair of temporally adjacent data.

[0022] Furthermore, the method for pre-training the deep fusion MixerLayer module is as follows:

[0023] (1) Processing the temporal dimension characteristics of time series data: Use dynamic linear layers to adjust parameters to adapt to the temporal characteristics of input data, apply DropOut operations to reduce overfitting, and use ReLU activation functions to increase nonlinearity; use linear layers and fully connected layers to further integrate and transform features;

[0024] (2) Processing the feature dimensions of the data: Use dynamic linear layers to adjust parameters to adapt to different feature dimensions, apply DropOut operations and ReLU activation functions to increase the generalization ability and nonlinearity of the model, comprehensively process features through linear layers and fully connected layers, and prepare outputs;

[0025] (3) The processing result before output is added to the input of the module to form a residual connection, thereby alleviating the vanishing gradient;

[0026] (4) Use the loss function to measure the error between the model output and the actual observed data.

[0027] Furthermore, the formula of the loss function is as follows:

[0028]

[0029] in, Representation model based on time point Data The generated prediction output, is the actual observed value.

[0030] Furthermore, the method of constructing a node embedding vector and generating an adjacency matrix is ​​as follows:

[0031] (1) Convert each type of data in the traffic dataset into an independent one-hot vector, concatenate all the one-hot encoded vectors to form the embedding vector of each intersection node ;

[0032] (2) After the node embedding vector is constructed, To generate the adjacency matrix ,Right now ;in, is the embedding vector of all intersection nodes The matrix of Transpose it, yes The transposed matrix of .

[0033] Furthermore, the formula for preprocessing the data at the current time point by inputting it into the fully connected layer is as follows:

[0034]

[0035] in, represents the original input data, and are the weight matrix and bias of the fully connected layer, is the activation function, It is the preprocessed data, which is used as the input of the multi-layer hybrid graph convolutional network.

[0036] Furthermore, the operation of each layer of the multi-layer hybrid graph convolutional network is expressed as:

[0037]

[0038] in, Indicates The node features of the layer, for the first layer, , is the weight matrix of this layer, is the adjacency matrix computed from the node embedding vectors.

[0039] Furthermore, the operation of the output fully connected layer of the multi-layer hybrid graph convolutional network is expressed as:

[0040]

[0041] in, represents the output of a multi-layer hybrid graph convolutional network, represents the last layer of a multi-layer hybrid graph convolutional network, and are the weight matrices and biases of the output fully connected layer.

[0042] Furthermore, the formula for calculating the difference between the prediction result and the data at the future time point is:

[0043]

[0044] in, represents the parameters of the deep graph convolutional network, Represents a set of intersections.

[0045] The advantages of the present invention are:

[0046] 1) Traffic models perform well in processing massive and complex data, and support the effective processing of sequential and discontinuous traffic data. The present invention provides a method for efficiently processing massive and complex data, especially for sequential and discontinuous data. Through innovative data preprocessing and model architecture design, it can effectively process irregular data, thereby improving the efficiency and accuracy of traffic data processing.

[0047] 2) The fine tuning strategy improves the model performance, enables it to adapt to specific application scenarios and data characteristics, and improves the prediction accuracy and model generalization ability; the present invention introduces a fine tuning strategy, which combines the advantages of pre-training and specific task training, and can significantly improve the model performance; through this strategy, the model can better adapt to specific application scenarios and data characteristics, improve the prediction accuracy and model generalization ability, and thus has obvious advantages in the field of traffic data processing;

[0048] 3) The introduction of spatial dimensions and zero-start capability makes the model more adaptable and can quickly adapt to new scenarios, providing more reliable data support for traffic management and planning; this design takes into account the spatial distribution and relationship of the data, enhances the model's understanding of the data structure characteristics, and improves prediction accuracy and generalization ability; at the same time, the zero-start capability enables the model to quickly adapt to new scenarios, providing more reliable data support for traffic management and planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a model structure diagram of the traffic status completion prediction method based on multi-source data pre-training of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are 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.

[0051] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments:

[0052] Embodiment 1

[0053] like Figure 1 As shown, a traffic status completion prediction method based on multi-source data pre-training in this embodiment includes the following contents:

[0054] 1. Pre-training stage

[0055] In this phase, the hybrid processing technology of MixerLayer is used to conduct in-depth learning specifically for time series features and their cross-relationships. The model structure includes fully connected (FC), linear (Linear), and dynamic linear (DyLinear) layers, with ReLU and Dropout operations. The historical values ​​in the selected adjacent time pairs are input into MixerLayer for processing, and the difference between the output of the model and the future values ​​in the adjacent time pairs is used to calculate the loss, and then the parameters of MixerLayer are iteratively optimized. A variety of data types are used in the pre-training phase, while the formal training phase focuses on traffic data. The purpose of this is to make full use of the rich features and patterns learned from different data sources in the pre-training phase to achieve "zero-start" training of traffic data. This strategy can significantly improve the model's understanding and prediction capabilities of traffic data characteristics, especially when faced with sparse data or new domain tasks, and can accelerate the model's learning and adaptation process.

[0056] 1.1 K-means clustering algorithm

[0057] In the project, the effective integration and processing of multi-source data from different fields is a key step to ensure the accuracy and effectiveness of the analysis. The following is a detailed description of the process of data integration and initial center point selection of the K-means algorithm:

[0058] (1) Data integration

[0059] In order to ensure that data from different fields (such as transportation, economy, health care and energy) can be analyzed together, the following integration steps are first performed:

[0060] Format standardization: unify the format of all input data. This includes unifying the representation of date and time, the format of numerical data, etc., to ensure that all data is consistent in form and easy to process and analyze.

[0061] Scaling: Normalizing or standardizing values ​​from different data sources. This step adjusts the data to the same scale, eliminating the effects of differences in original units and magnitudes, so that data can be directly compared and calculated.

[0062] (2) Initialize the center point

[0063] When executing the K-means clustering algorithm, the selection of the initial center point has an important impact on the effect and stability of clustering. The initial center point is determined by calculating the data mean of each source in the integrated multi-source data set. This method is based on the statistical characteristics of actual data and can improve the efficiency and stability of clustering compared to random selection.

[0064] (3) Objective function

[0065] The goal of the K-means algorithm is to minimize the sum of distances within the cluster. The optimized objective function is expressed as follows:

[0066]

[0067] in, is the number of clusters, It is The set of data points in a cluster, It is a cluster The center point of It is a cluster A data point in Represents data points To center point The square of the Euclidean distance.

[0068] (4) Update cluster center

[0069] The center point of each cluster is updated by calculating the mean of all data points in the cluster. The specific expression is:

[0070]

[0071] in, Representation Cluster The number of data points in .

[0072] Through the above steps, K-means clustering helps simplify complex multi-source data into several manageable groups. The data points within each group are similar in key features, which facilitates in-depth analysis and further data processing. Especially in multivariate data environments, this method not only improves the efficiency of analysis, but also enhances the accuracy and reliability of data-driven decision-making.

[0073] 1.2. Constructing data input

[0074] Using K-means clustering algorithm to group data into clusters Afterwards, the data within each cluster is time-sorted. This process is performed based on the timestamp of each data point, with the goal of making the dataset appear sequential in time. This time ordering is critical because it helps capture the temporal continuity and dynamic changes between data points, which is key to understanding and predicting time-dependent data.

[0075] In each cluster After the internal data points are sorted by time, the temporally adjacent data pairs are defined. These data pairs are the basis for analyzing the dynamics of the time series and are crucial for model training and prediction. The selection of temporally adjacent data pairs is based on the following definition:

[0076]

[0077] Here, and It is a cluster After sorting the consecutive data points in , we ensure that each pair of data is adjacent in time.

[0078] 1.3. Deep Fusion MixerLayer

[0079] (1) Time Mixing Module: This module mainly processes the time dimension characteristics of time series data. It processes data through the following steps: Use the dynamic linear layer (DyLinear) to adjust parameters to adapt to the temporal characteristics of the input data. Apply DropOut to reduce overfitting. The ReLU activation function is used to increase nonlinearity. The linear layer (Linear) and the fully connected layer (FC) further integrate and transform the features.

[0080] (2) Feature Mixing Module: This module processes the feature dimensions of the data. Its steps include: using the dynamic linear layer (DyLinear) to adjust the parameters to adapt to different feature dimensions. Applying DropOut and ReLU activation functions to increase the generalization ability and nonlinearity of the model. Comprehensively processing features through the linear layer (Linear) and the fully connected layer (FC) and preparing the output.

[0081] (3) Residual connection: In each module, the processing result before the output is added to the input of the module to form a residual connection, which helps alleviate the gradient vanishing problem and improve the training efficiency of the network. This connection method allows the network to learn the difference between input and output, thereby paying more attention to useful changes and improving the overall performance of the model.

[0082] (4) Pre-training loss function: Using the selected time-adjacent data pairs, the model is then pre-trained to optimize the model's ability to understand and predict time series data. During the pre-training phase, a loss function is used to measure the error between the model output and the actual observed data:

[0083]

[0084] in, Representation model based on time point Data The generated prediction output, are the actual observed values, which serve as a true comparison benchmark for the model's predictive performance.

[0085] By minimizing The predicted output and its actual subsequent value The loss function is designed to improve the accuracy of the model in predicting future data points. Through this method, from the time sorting of data to the selection of adjacent data pairs to the pre-training of the model, the entire process not only deepens the understanding of the temporal attributes of the data, but also enhances the model's ability to handle complex time series data. Such a process provides a solid foundation for in-depth analysis and prediction of time-dependent data.

[0086] 2. Formal training stage

[0087] This stage is specifically for traffic data sets, using a sliding window method to divide the data in the window into Data at the current time point After Data at future time points The current time point data input to the model is trained after feature mapping through a fully connected layer, and the difference between the predicted result and the actual future value is used to calculate the loss. The model structure mainly includes a hybrid graph convolutional network (M-GCN), in which the GCN part embeds information such as events, environment, and location of intersection nodes, which is used to construct the adjacency matrix, while the parameters of MixerLayer are fixed during pre-training.

[0088] 2.1 Node Embedding Vector

[0089] In order to effectively process and analyze urban traffic networks, node embedding vectors are constructed and adjacency matrices are generated. The whole process begins by collecting information from a variety of data sources, including traffic management systems, social media, news reports, weather station data, urban planning departments, and traffic monitoring equipment. The types of data of interest include event information (such as traffic accidents, construction, and traffic control), environmental information (such as weather conditions and time conditions), geographic coordinates of intersections, traffic flow data, and other information such as intersection design type and surrounding facilities.

[0090] All collected data is vectorized through one-hot encoding. Each type of information such as event type, weather conditions, time conditions, intersection design type and surrounding facilities is converted into an independent one-hot vector. All one-hot encoded vectors are concatenated to form a comprehensive embedding vector for each intersection node. Such integration not only ensures that data from different information sources are treated equally, but also plays an effective role in the model.

[0091] 2.2. Construction of adjacency matrix

[0092] After the node embedding vector is constructed, To generate the adjacency matrix ,in is a matrix containing the embedding vectors of all intersection nodes, This calculation method enables the adjacency matrix to capture and express the relationships and similarities between nodes. These relationships are based on the inner product of the node embedding vectors, which provides the basis for the next step of the graph convolutional network:

[0093]

[0094] in, is a matrix whose column vectors consist of the embedding vectors of all intersection nodes. composition, yes This approach allows the adjacency matrix to capture and express the relationships and similarities between nodes, which are calculated based on the inner product of the node embedding vectors.

[0095] 2.3 M-GCN Structure

[0096] The input data of the model is first preprocessed by the fully connected layer (FC layer) to adjust the data format to meet the input requirements of the graph convolutional network (GCN). The mathematical expression of this preprocessing is:

[0097]

[0098] in, represents the original input data, and are the weight matrix and bias of the fully connected layer, is the activation function, It is the preprocessed data, which is used as the input of the multi-layer hybrid graph convolutional network.

[0099] Preprocessed data Input to the multi-layer graph convolutional network (M-GCN). In M-GCN, the operation of each layer can be defined as:

[0100]

[0101] in, Indicates The node features of the layer, for the first layer, , is the weight matrix of this layer, is the adjacency matrix calculated from the node embedding vector. In this way, the graph convolution operation at each layer not only updates the feature representation of the node, but also enhances the spatial dependency reflected by the similarity of node embeddings.

[0102] Output after multi-layer M-GCN processing ,in represents the last layer of M-GCN, which is then input to another fully connected layer. The purpose of this layer is to transform the complex features of the deep graph convolutional network into the final output, which is suitable for specific tasks such as traffic flow prediction. The operation of the output layer can be expressed as:

[0103]

[0104] in, represents the output of a multi-layer hybrid graph convolutional network, represents the last layer of a multi-layer hybrid graph convolutional network, and are the weight matrices and biases of the output fully connected layer.

[0105] Through such a structural setting, the M-GCN model can not only process and analyze the complex relationships of the input data, but also deeply learn the intrinsic characteristics of the data through a multi-layer structure, thereby improving the accuracy of prediction and the generalization ability of the model.

[0106] The formula for calculating the difference between the predicted result and the data at a future time point is:

[0107]

[0108] in, represents the parameters of the deep graph convolutional network, Represents a set of intersections.

[0109] 2.4. MixerLayer with fixed parameters

[0110] In the formal training phase, MixerLayer and M-GCN layers are applied alternately to enhance the model's ability to process time series and feature dimensions. Specifically, each M-GCN layer is followed by a MixerLayer. The M-GCN layer processes graph structure data, while the MixerLayer deeply mines time and feature dimension information. This alternating layer structure not only optimizes the spatial and temporal feature learning of data, but also enhances the model's performance when processing complex data.

[0111] The innovation of the present invention lies in its ability to efficiently process massive amounts of complex data, and it has the following key features: First, the system has the ability to process sequential and discontinuous traffic data, and can effectively process the data even if it is discontinuous or irregular in time. Secondly, the present invention includes special data preprocessing and model architecture, which improves the efficiency and accuracy of the system's understanding and analysis of traffic data. Most importantly, the present invention overcomes the shortcomings of traditional methods by significantly improving learning results, allowing the system to perform better when processing complex data. Overall, the present invention integrates efficient processing methods for massive amounts of complex data, including support for sequential and discontinuous data, special data preprocessing and model architecture design, and improvements to traditional methods, which improves data processing efficiency and accuracy.

[0112] The innovation of the present invention is to propose a fine-tuning strategy to significantly improve the performance of the model. The strategy mainly includes a secondary training strategy, that is, a method that combines pre-training and fine-tuning. First, the basic language comprehension ability is acquired through the pre-training model. Then, the model is retrained using a fine-tuning method to better adapt to specific application scenarios and data characteristics, fully tap the potential of the model, and improve performance and accuracy. This strategy is not just about fine-tuning parameters, but also combines the advantages of pre-training and specific task training, providing an efficient model optimization method for different application scenarios and significantly improving performance.

[0113] The innovation of the present invention lies in the integration of spatial dimension and zero-start capability, which gives the model stronger adaptability. First, the addition of spatial dimension considerations enables the model to deeply learn the spatial distribution and relationship of the data, better understand the data structure characteristics, and improve prediction accuracy and generalization capabilities. Secondly, the introduction of zero-start capability enables the model to quickly adapt to new scenarios, that is, to quickly learn and adapt to new tasks or environments without pre-training data. By combining spatial dimensions and zero-start capabilities, the present invention gives the model stronger adaptability, enabling it to understand data more accurately and quickly achieve accurate predictions in new scenarios, which has important application potential and market competitive advantages.

[0114] Embodiment 2

[0115] An electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the traffic state completion prediction method based on multi-source data pre-training in Example 1, and the processor is configured to execute the program stored in the memory.

[0116] Embodiment 3

[0117] A storage medium stores a computer program, which, when executed by a processor, executes the steps of the traffic state completion prediction method based on multi-source data pre-training in embodiment 1.

[0118] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A traffic status completion prediction method based on multi-source data pre-training, characterized in that: include: Pre-training phase and formal training phase; Pre-training stage: collect data from multiple fields and integrate the data, including format standardization and scale standardization; The k-means algorithm is used to cluster the data, and temporally adjacent data pairs are selected from each cluster in chronological order; the data pairs are input into the deep fusion MixerLayer module, and the deep fusion MixerLayer module is pre-trained; The method for pre-training the deep fusion MixerLayer module is as follows: (1) Processing the temporal dimension characteristics of time series data: Use dynamic linear layers to adjust parameters to adapt to the temporal characteristics of input data, apply DropOut operations to reduce overfitting, and use ReLU activation functions to increase nonlinearity; use linear layers and fully connected layers to further integrate and transform features; (2) Processing the feature dimensions of the data: Use dynamic linear layers to adjust parameters to adapt to different feature dimensions, apply DropOut operations and ReLU activation functions to increase the generalization ability and nonlinearity of the model, comprehensively process features through linear layers and fully connected layers, and prepare outputs; (3) The processing result before output is added to the input of the module to form a residual connection, thereby alleviating the vanishing gradient; (4) Use the loss function to measure the error between the model output and the actual observed data; The formula of the loss function is as follows: in, Representation model based on time point Data The generated prediction output, is the actual observed value; Formal training phase: For the traffic data set, the sliding window method is used to divide the data in the window into Data at the current time point After Data at future time points ; Construct node embedding vectors and generate adjacency matrices. Preprocess the data at the current time point through the input fully connected layer. The preprocessed data is input into the multi-layer hybrid graph convolutional network for training. After training, it is input into the output fully connected layer to obtain the final Prediction results for future time points , and calculate the difference between the predicted results and the data at future time points.

2. The traffic status prediction method based on multi-source data pre-training according to claim 1 is characterized in that: The method of clustering data using the k-means algorithm is as follows: (1) Initialize the center point by calculating the mean value of each source in the integrated multi-source data set; (2) Establish an objective function. The expression of the objective function is as follows: in, is the number of clusters, It is The set of data points in a cluster, It is a cluster The center point of It is a cluster A data point in Represents data points To center point The square of the Euclidean distance; (3) Update the cluster center point. The center point of each cluster is updated by calculating the mean of all data points in the cluster. The specific expression is: in, Representation Cluster The number of data points in .

3. The traffic status prediction method based on multi-source data pre-training according to claim 2 is characterized in that: The basis for selecting the temporally adjacent data pairs is as follows: in, and It is a cluster After sorting the consecutive data points in time, ensure that each pair of data is adjacent in time. Represents a pair of temporally adjacent data.

4. The traffic status completion prediction method based on multi-source data pre-training according to claim 1 is characterized in that: The method of constructing a node embedding vector and generating an adjacency matrix is ​​as follows: (1) Convert each type of data in the traffic dataset into an independent one-hot vector, concatenate all the one-hot encoded vectors to form the embedding vector of each intersection node ; (2) After the node embedding vector is constructed, To generate the adjacency matrix ,Right now ;in, is the embedding vector of all intersection nodes The matrix of Transpose it, yes The transposed matrix of .

5. The traffic status completion prediction method based on multi-source data pre-training according to claim 1 is characterized in that: The formula for preprocessing the data at the current time point by inputting the fully connected layer is as follows: in, represents the original input data, and are the weight matrix and bias of the fully connected layer, is the activation function, It is the preprocessed data, which is used as the input of the multi-layer hybrid graph convolutional network.

6. The traffic status completion prediction method based on multi-source data pre-training according to claim 1 is characterized in that: The operation of each layer of the multi-layer hybrid graph convolutional network is expressed as: in, Indicates The node features of the layer, for the first layer, , is the weight matrix of this layer, is the adjacency matrix computed from the node embedding vectors.

7. The traffic status completion prediction method based on multi-source data pre-training according to claim 6 is characterized in that: The operation of the output fully connected layer of the multi-layer hybrid graph convolutional network is expressed as: in, represents the output of a multi-layer hybrid graph convolutional network, represents the last layer of a multi-layer hybrid graph convolutional network, and are the weight matrices and biases of the output fully connected layer.

8. The traffic status completion prediction method based on multi-source data pre-training according to claim 1 is characterized in that: The formula for calculating the difference between the predicted result and the data at a future time point is: in, represents the parameters of the deep graph convolutional network, Represents a set of intersections.

Images