Traffic flow multistage federation prediction method based on vehicle and road cloud integrated network

Abstract

The invention belongs to the technical field of intelligent traffic system data privacy protection, and discloses a traffic flow multi-stage federated prediction method based on a vehicle-road cloud integrated network, which comprises the following steps: defining a vehicle-road cloud integrated network architecture, and constructing a local model, a regional model and a global model based on LSTM (Long Short Term Memory) based on each layer of nodes of the vehicle-road cloud integrated network architecture to form a federated model; and training the federal model by adopting a clustering method according to the spatial distribution of the nodes, and then performing traffic flow prediction. According to the invention, based on the vehicle-road cloud collaborative framework, the hierarchical federal prediction model is designed, and the distribution calculation of the model parameters between the client node and the server is carried out, so that the efficient model training can be realized, and the requirements of large-scale system deployment are met; according to the vehicle-road cloud integrated multistage prediction model provided by the invention, a personalized local model can be obtained, and a global model with higher universality can also be obtained, so that the traffic flow prediction is guaranteed to have higher availability and accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent transportation system data privacy protection, and particularly relates to a traffic flow multi-level federated prediction method based on a vehicle-road cloud integrated network. BACKGROUND

[0002] An intelligent transportation system can not only effectively improve traffic efficiency, but also optimize traffic resource allocation, and therefore has become a hotspot in current application research and practice fields. Traffic flow prediction is a key technology for intelligent transportation system management and services. Traditional traffic flow prediction analyzes historical traffic data, adopts statistical and classical machine learning methods for modeling and prediction, and has limitations such as difficulty in feature extraction of the model and relatively high computational complexity.

[0003] The rapid development of artificial intelligence technologies such as deep learning and large models has laid a solid foundation for constructing intelligent traffic flow prediction models. Domestic and foreign scholars have carried out a series of research and achieved important results. Typical methods include: In 2020, Jianyu Huang et al. studied the application of convolutional neural network (CNN), recurrent neural network (RNN) and its variants in traffic flow prediction in the literature “Traffic Flow Forecasting with Deep Learning: A Survey”. In 2021, Yi Li et al. proposed a graph convolutional network framework for traffic flow prediction in “Graph Convolutional Networks for Traffic Flow Forecasting”, and verified it on multiple urban traffic data sets, demonstrating its strong ability in traffic flow prediction. In 2021, Liu Ming et al. proposed an MVF-LSTM model that improves the input layer of the LSTM model using arithmetic mean filtering MVF in “Short-term Traffic Flow Prediction Based on Improved LSTM Algorithm”. By increasing the dimension of the LSTM hidden layer input, the average relative error is used to adjust the iteration number, time step and number of hidden layer neurons to improve the prediction accuracy of the model. In 2022, Xiang Li et al. proposed an interpretable AI (XAI) method for traffic flow prediction in “Interpretable Traffic Flow Prediction with Deep Learning”. Through LIME and SHAP technologies, the trust and decision-making ability of traffic managers are enhanced. However, these traffic flow prediction models all use data integration analysis methods such as data collection, integration, establishment of a data center for model training, and have the disadvantage of low model construction efficiency.

[0004] With the development of communication technology, the vehicle-road-cloud integrated network is widely used in the field of transportation. Through the vehicle-road-cloud collaborative control system, the human-vehicle-road-cloud is connected together, and the vehicle is no longer an isolated individual. It can not only interact with other vehicles on the road, but also exchange information with roadside facilities and cloud platforms. Based on the vehicle-road-cloud integrated network, a distributed computing-based traffic flow prediction model can overcome the shortcomings of traditional centralized data integration. Scholars at home and abroad have proposed a series of model methods. Typical methods include: In 2020, X. Chen et al. discussed in the paper "Cloud-Edge Integrated Traffic Flow Prediction Using Distributed Deep Learning Networks" how to use distributed deep learning methods to predict traffic flow under the architecture of vehicle-road-cloud integrated network. By distributing deep learning tasks to cloud and edge nodes, the system can reduce the pressure of a single computing node and improve the real-time and accuracy of prediction. In 2022, Q. Li et al. introduced in the paper "Big Data-driven Traffic Flow Prediction in V2X Networks Using Cloud Computing and Edge Intelligence" how to use big data analysis, cloud computing and edge intelligence to realize traffic flow prediction in V2X networks. It can predict traffic flow in real time in different geographical areas and respond quickly to traffic flow emergencies. In 2023, H. Chen et al. proposed a real-time traffic flow prediction framework based on cloud-edge collaborative computing in V2X networks in the paper "Real-Time Traffic Flow Prediction Based on Cloud-Edge Collaborative Computing in V2X Networks". It uses the computing resources in the vehicle-road-cloud integrated network to distribute traffic flow data processing. However, the existing traffic flow prediction model based on vehicle-road-cloud integrated network distributed computing needs to share the original traffic flow data among different network computing nodes, which may cause the leakage of personal privacy security of vehicle owners. As a new type of distributed machine learning technology for privacy protection, federated learning can make multiple participants jointly train the model without leaking the original data, effectively reducing the privacy risk brought by data aggregation.As patent application CN114881371A discloses a traffic flow prediction method based on federated learning, introduces a model construction method based on a client-server architecture, and points out that the problem of data privacy protection can be solved; but this method uses a single client-server architecture for model training, and needs to aggregate the massive model parameters of numerous clients on the server node, which may exist a bottleneck problem of calculation efficiency; in addition, this method based on federated learning can only train a unified model, and cannot meet the demand of traffic flow grading and regional prediction, and usually has the problem of low prediction accuracy. SUMMARY

[0005] To solve the above technical problems, the present application provides a traffic flow multi-level federated prediction method based on a vehicle-road cloud integrated network, which effectively solves the calculation efficiency bottleneck problem and the problem of low prediction accuracy of traditional methods.

[0006] The traffic flow multi-level federated prediction method based on the vehicle-road cloud integrated network provided by the present application comprises: defining a vehicle-road cloud integrated network architecture, including edge nodes, regional nodes, center nodes, and communication links connecting each node for transmission; based on the edge nodes, regional nodes and center nodes, constructing a local traffic flow prediction model, a regional traffic flow prediction model and a global traffic flow prediction model based on LSTM to form a federated model; training the federated model according to the spatial distribution of the nodes using a clustering method; performing traffic flow prediction based on the trained federated model.

[0007] Further, the edge nodes collect traffic flow information through roadside infrastructure, train a local traffic flow prediction model, exchange model parameters with regional nodes to update the local traffic flow prediction model, and then perform road traffic flow prediction; The regional nodes are responsible for exchanging model parameters with the edge nodes and the center nodes, training a regional traffic flow prediction model, and performing traffic flow prediction based on the regional traffic flow prediction model; The center nodes are responsible for exchanging model parameters with the regional nodes, training a global traffic flow prediction model, and performing traffic flow prediction based on the global traffic flow prediction model.

[0008] Further, the local traffic flow prediction model, the regional traffic flow prediction model and the global traffic flow prediction model based on LSTM are respectively composed of a long short-term memory network LSTM, including a forgetting gate, an input gate, a cell state and an output gate. The forgetting gate The forgetting or retaining degree of the traffic flow information is controlled by the formula: (1), wherein, represents a sigmoid activation function, represents a weight matrix of the forgetting gate, represents a bias matrix of the forgetting gate, represents the input traffic flow information at the current moment, represents the output traffic flow information at the last moment; the input gate and the candidate cell state The input of the traffic flow information into the cell state is controlled by the formula: (2), (3), wherein, is a weight matrix of the input gate, is a bias matrix of the input gate, is a weight of the candidate input gate, is a bias matrix of the candidate input gate; the cell state The historical traffic flow information is stored by the formula: (4), wherein, the value range of is (0, 1); the output gate integrates the historical information and the information at the current moment by the formula: (5), (6), wherein, is a weight matrix of the output gate, is a bias matrix of the input gate.

[0009] Further, the federated model is trained by a clustering method according to the spatial distribution of the nodes, including: the central node performs iterative clustering on all regional nodes according to the spatial position information of the regional nodes to obtain different regional node clusters; each regional node performs iterative clustering on all edge nodes under the jurisdiction of the regional node according to the spatial position information of the edge nodes to obtain different edge node clusters; Model parameters are sent down step by step in the order of central nodes, regional node clusters, regional nodes, edge node clusters, and edge nodes for model training and parameter update. The updated parameters are then uploaded step by step for hierarchical aggregation.

[0010] Furthermore, model parameters are distributed step by step in the order of central node, regional node cluster, regional node, edge node cluster and edge node, including: 1) In the initial training phase, the central node converts the LSTM gradient parameters in the global model Randomly initialized to , and Issued to regional node clusters; among them, Forget Gate , input gate , output gate , unit status The weight matrix, Forget Gate , input gate , output gate , unit status The deviation matrix of 2) Each regional node cluster is randomly selected from its constituent members Regional node forwarding model parameters ; 3) Regional nodes receive the forwarded model parameters , which is further forwarded to the edge node clusters; 4) Each edge node cluster is randomly selected from its constituent members Edge node forwarding model parameters .

[0011] Furthermore, the model is trained and parameters are updated, including: The edge node receives the forwarded model parameters ,The traffic flow federated model is trained by denoising, normalizing, and missing value processing of the traffic flow data collected by the ,sensors, and dividing the data into training set, validation set, and test set; For edge nodes, by setting an upper limit on the number of iteration rounds And the total number of batches of data , for each batch in each round The data in the model are analyzed using the stochastic gradient descent method. The training update formula is: (7), in, is the global model parameter issued, is the learning rate of the edge nodes, represents the calculation of gradient, is the loss function of the edge node, are the local model parameters of the previous round of joint training, The updated value of the model parameters in this round of training.

[0012] Furthermore, the updated parameters are uploaded level by level and aggregated hierarchically, including: 1) The local model parameters that the edge node will update Upload to the edge node cluster to which it belongs; 2) The edge node cluster receives updated model parameters from all edge nodes and performs aggregate calculations of the model parameters. The formula is as follows: (8), in, is the number of edge nodes contained in the edge node cluster, is the number of the edge node included in the edge node cluster, For the The model parameters updated by edge nodes; 3) The edge node cluster will aggregate the model parameters Further upload to the regional node to which it belongs; 4) Regional nodes receive model parameters aggregated from all edge node clusters , and perform aggregate calculation of model parameters, the formula is as follows: (9), in, is the number of edge node clusters contained in the regional node, is the number of the edge node cluster, For the The model parameters updated by the edge node cluster; 5) Model parameters that regional nodes will aggregate Further upload to the regional node cluster to which it belongs; 6) The regional node cluster receives the model parameters aggregated by all regional nodes , and perform aggregate calculation of model parameters, the formula is as follows: (10), in, is the number of regional nodes contained in the regional node cluster, The number of the region node that contains the region node in the region node cluster, For the Model parameters for regional node updates; 7) The regional node cluster aggregates the model parameters further upload to the center node; 8) The center node receives all the regional node cluster aggregated model parameters and performs the aggregated calculation of the model parameters, as follows: (11), wherein, is the number of regional node clusters contained by the center node, is the number of the regional node cluster, is the model parameter updated by the regional node cluster; 9) The center node distributes the aggregated model parameters to the regional node cluster, and repeats 1) - 8) until the local model training of all edge nodes converges to the specified accuracy or reaches the set number of training rounds.

[0013] The beneficial effects of the present application are: the vehicle-road cloud integrated multi-level prediction model based on the vehicle-road cloud collaborative framework, the hierarchical federated prediction model is designed, the traffic flow prediction has higher prediction accuracy, and the usability of the model is guaranteed; the vehicle-road cloud integrated multi-level prediction model of the present application realizes the simultaneous training of the personalized local model and the universal global model through the step-by-step distribution of model parameters, local model training and parameter updating, step-by-step uploading and hierarchical aggregation of model parameters, and meets the needs of large-scale system deployment. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is a flowchart of the method of the present application; Figure 2 is a vehicle-road cloud integrated network framework diagram; Figure 3 is an LSTM model schematic diagram; Figure 4 is a regional node clustering process diagram; Figure 5 is an edge node clustering process diagram; Figure 6 is a model parameter distribution schematic diagram; Figure 7 is a model parameter uploading and aggregation calculation schematic diagram; Figure 8 is a traffic flow prediction schematic diagram based on the edge node local prediction model; Figure 9 is a traffic flow prediction schematic diagram based on the regional node prediction model and its subordinate edge node cluster model; Figure 10is a traffic flow prediction schematic diagram based on a global model of a center node and a cluster model of regional nodes under the center node. DETAILED DESCRIPTION

[0015] In order to make the content of the present application more easily understood, the present application will be further described in detail below according to specific embodiments and in conjunction with the accompanying drawings.

[0016] As shown in Figure 1 The traffic flow multi-level federal prediction method based on the vehicle-road cloud integrated network according to the present application comprises: defining a vehicle-road cloud integrated network architecture, including edge nodes, regional nodes, a center node, and communication links connecting the nodes for transmission; based on the edge nodes, the regional nodes, and the center node, constructing a local traffic flow prediction model, a regional traffic flow prediction model, and a global traffic flow prediction model based on LSTM, forming a federal model; training the federal model according to the spatial distribution of the nodes using a clustering method; performing traffic flow prediction based on the trained federal model.

[0017] Stage 1, vehicle-road cloud integrated network architecture construction; The vehicle-road cloud integrated network is the core of building a logically collaborative and physically dispersed vehicle-road cloud collaborative control system, and is also the basis for building a federal traffic flow federal prediction model, including edge nodes, regional nodes, and a center node, and communication links connecting the three types of nodes for information transmission, as shown in Figure 2 .

[0018] Among them, the edge node collects traffic flow information through roadside infrastructure (such as cameras, sensors, roadside units, etc.), trains a local traffic flow prediction model, exchanges model parameters with the regional node to update the local model, and then performs road traffic flow prediction; the regional node is responsible for exchanging model parameters with the edge node and the center node, training a regional traffic flow prediction model, and performing traffic flow prediction based on the regional model. The center node is responsible for exchanging model parameters with the regional node, training a global traffic flow prediction model, and performing traffic flow prediction based on the global model.

[0019] Stage 2, definition of traffic flow prediction model based on LSTM; Traffic flow belongs to time series data, and the long short-term memory (LSTM) in the time cycle neural network has flexible information processing capacity, can better control information flow when processing sequence data, and has good network learning and reasoning ability. Therefore, the application adopts LSTM as the public structure of the local model, the regional model and the global model based on the vehicle-road cloud integrated network collaborative training, and the core structure thereof comprises a forgetting gate, an input gate, a unit state and an output gate, as shown in Figure 3 .

[0020] Specifically, (1) the forgetting gate is used for controlling the forgetting or retaining degree of traffic flow information, and the formula is: (1), wherein, denotes a sigmoid activation function, denotes a weight matrix of the forgetting gate, denotes a bias matrix of the forgetting gate, denotes the input traffic flow information at the current moment, denotes the output traffic flow information at the last moment; (2) the input gate and the candidate unit state are used for controlling the inflow degree of the input traffic flow information into the unit state, and the formula is: (2), (3), wherein, is a weight matrix of the input gate, is a bias matrix of the input gate, is a weight of the candidate input gate, is a bias matrix of the candidate input gate; (3) the unit state is used for storing historical traffic flow information, and the formula is: (4), wherein, the value range of is (0, 1) ; (4) the output gate integrates the historical information and the current moment information , and the formula is: (5), (6), wherein, is the weight matrix of the output gate, is the bias matrix of the input gate.

[0021] Phase 3: Federated model training based on clustering fusion; Phase 3-1: Clustering of computing nodes based on spatial location information; (1) Central node ( ) for all regional nodes ( ) based on its spatial location information ( ), and the K-Means algorithm is used to perform iterative clustering to obtain Different regional node clusters ,like Figure 4 As shown; (2) Each regional node controls all edge nodes under its jurisdiction ( ) based on its spatial location information ( ), and the K-Means algorithm is used to perform iterative clustering to obtain Different edge node clusters ,like Figure 5 shown.

[0022] Phase 3-2: Collaborative training based on hierarchical clustering of computing nodes; (1) Model parameters are distributed step by step; 1) Initial model training phase, central node The LSTM gradient parameters in the global model Randomly initialized to , and Issued to Regional node clusters ;in, Forget Gate , input gate , output gate , unit status The weight matrix, Forget Gate , input gate , output gate , unit status The deviation matrix of During the model training phase, the rounds of collaborative interaction between central nodes, regional node clusters, regional nodes, edge node clusters, and edge nodes are calculated. Perform global model gradient parameter Iterative updates; 2) Each regional node cluster , randomly selected from its constituent members Regional nodes Forwarding model parameters ; 3) Regional nodes Receive forwarded model parameters , which is further forwarded to the Edge node clusters ; 4) Each edge node cluster , randomly selected from its constituent members edge nodes Forwarding model parameters ,like Figure 6 shown.

[0023] (2) Model training and parameter updating; edge nodes Receive forwarded model parameters The traffic flow federation model is trained by denoising, normalizing, and processing missing values ​​of traffic flow data (vehicle speed, traffic volume, vehicle density, etc.) collected by sensors (such as cameras, GPS, radar, etc.), and dividing the data into training sets, validation sets, and test sets.

[0024] For edge nodes , by setting an upper limit on the number of iterations And the total number of batches of data , for each batch in each round The data in the model are analyzed using the stochastic gradient descent method. The training update formula is: (7), in, is the global model parameter issued, It is an edge node The learning rate, represents the calculation of gradient, For edge nodes The loss function is are the local model parameters of the previous round of joint training, The updated value of the model parameters in this round of training.

[0025] (3) Model parameter upload and layered aggregation; 1) Edge Node The local model parameters that will be updated Upload to the edge node cluster to which it belongs; 2) Edge node cluster Receive updated model parameters from all edge nodes and perform aggregate calculations on the model parameters. The formula is as follows: (8), wherein, is the number of edge nodes contained in the edge node cluster, is the number of edge nodes contained in the edge node cluster, is the model parameter updated by the th edge node; 3) edge node cluster aggregates the model parameters further uploads the aggregated model parameters to the regional node to which it belongs; ; 4) regional node receives all the model parameters aggregated by the edge node cluster , and performs aggregated calculation of the model parameters, as follows: (9) wherein, is the number of edge node clusters contained in the regional node, is the number of edge node clusters, is the model parameter updated by the th edge node cluster; 5) regional node aggregates the model parameters further uploads the aggregated model parameters to the regional node cluster to which it belongs; ; 6) regional node cluster receives all the model parameters aggregated by the regional node , and performs aggregated calculation of the model parameters, as follows: (10), wherein, is the number of edge node clusters contained in the regional node, is the number of edge node clusters, is the model parameter updated by the th edge node cluster; 7) regional node cluster aggregates the model parameters further uploads the aggregated model parameters to the center node; ; 8) center node receives all the model parameters aggregated by the regional node cluster , and performs aggregated calculation of the model parameters, as follows: (11), wherein, is the number of regional node clusters contained in the center node, is the number of regional node clusters, For the first Regional node cluster updates model parameters; 9) Central node The aggregated model parameters , issued to Regional node cluster , repeat 1) -8), until the local model training of all edge nodes converges to the specified accuracy, or reaches the set number of training rounds. As Figure 7 Shown.

[0026] In the present application, the input of the local model is the collected traffic flow data and the issued parameters of the global model, which exchanges parameters with the regional model of the upper level in the training process, and the regional model further exchanges parameters with the global model.

[0027] Stage 4, traffic flow prediction based on federal model;

[0028] After federal collaborative training, each edge node Get local prediction model ; Regional node Get regional prediction model ; Central node Get global prediction model Based on hierarchical model, traffic flow prediction of corresponding level can be carried out, as Figure 8 Shown, specifically:

[0029] (1) Edge node Based on the stored local prediction model , provide road-level real-time and weak real-time traffic flow prediction for connected cars, as Figure 8 Shown, wherein Figure 8 The actual traffic flow and predicted traffic flow information of edge node 4 in Table 1.

[0030] Table 1 Actual traffic flow and predicted traffic flow information of edge node 4

[0031] The present application uses accuracy (Acc), precision (Prec), recall (Rec) and F1 score to evaluate the traffic flow prediction accuracy, and the calculation formula of each index is as follows: (12), (13), (14), (15)。

[0032] The prediction accuracy and convergence time of the edge node 4 are shown in Table 2.

[0033] Table 2 Prediction accuracy and convergence time of edge node 4

[0034] (2) Regional node Based on the stored regional prediction model , the regional node is provided for the transportation and traffic management department, and a plurality of road edge nodes constitute a spatial sub-region, and a weak real-time or non-real-time traffic flow prediction for traffic regulation is provided, such as Figure 9 shown, wherein Figure 9 The information comparison information of the actual traffic flow and the predicted traffic flow of the regional node is shown in Table 3.

[0035] Table 3 Information comparison of actual traffic flow and predicted traffic flow of regional node

[0036] The prediction accuracy and convergence time of the regional node are shown in Table 4.

[0037] Table 4 Prediction accuracy and convergence time of regional node

[0038] (3) Center node Based on the stored global prediction model , the center node is provided for the city traffic decision-making department, and a traffic flow prediction for macroscopic traffic data analysis is provided, such as Figure 10 shown, the information comparison information of the actual traffic flow and the predicted traffic flow of the center node is shown in Table 5.

[0039] Table 5 Information comparison of actual traffic flow and predicted traffic flow of center node

[0040] The prediction accuracy and convergence time of the center node are shown in Table 6.

[0041] Table 6 Prediction accuracy and convergence time of center node

[0042] The prediction accuracy and convergence speed in Tables 2, 4, and 6 can be seen that the present application effectively solves the calculation efficiency bottleneck problem and the low prediction accuracy problem existing in the traditional method, and achieves the goal of traffic flow prediction with higher usability and accuracy.

[0043] The above merely describes the preferred embodiments of the present application, and is not intended to further limit the present application, and any equivalent changes made according to the content of the present application and the drawings are within the protection scope of the present application.

Claims

1. A multi-level federated traffic flow prediction method based on a vehicle-road-cloud integrated network, characterized by: include: Define the integrated vehicle-road-cloud network architecture, including edge nodes, regional nodes, central nodes, and the communication links connecting each node for transmission; Based on edge nodes, regional nodes, and central nodes, we build LSTM-based local traffic flow prediction models, regional traffic flow prediction models, and global traffic flow prediction models to form a federated model. Training the federated model using a clustering method based on the spatial distribution of nodes; Traffic flow prediction is performed based on the trained federated model.

2. The multi-level federated traffic flow prediction method based on the vehicle-road-cloud integrated network according to claim 1 is characterized in that: The edge nodes collect traffic flow information through roadside infrastructure, train local traffic flow prediction models, and exchange model parameters with regional nodes to update the local traffic flow prediction models, thereby performing road traffic flow prediction; The regional nodes are responsible for exchanging model parameters with edge nodes and central nodes, training regional traffic flow prediction models, and performing traffic flow prediction based on the regional traffic flow prediction models; The central node is responsible for exchanging model parameters with regional nodes, training the global traffic flow prediction model, and performing traffic flow prediction based on the global traffic flow prediction model.

3. The multi-level federated traffic flow prediction method based on the vehicle-road-cloud integrated network according to claim 1 is characterized in that: The LSTM-based local traffic flow prediction model, regional traffic flow prediction model, and global traffic flow prediction model are respectively composed of a long short-term memory network LSTM, including a forget gate, an input gate, a unit state, and an output gate; The forget gate The formula used to control the degree of forgetting or retaining traffic flow information is: (1)， in, represents the sigmoid activation function, represents the weight matrix of the forget gate, represents the bias matrix of the forget gate, Indicates the traffic flow information input at the current moment, Indicates the traffic flow information output at the last moment; The input gate and candidate cell status It is used to control the degree of traffic flow information input into the unit state. The formula is: (2)， (3)， in, is the weight matrix of the input gate, is the bias matrix of the input gate, is the weight of the candidate input gate, is the bias matrix of the candidate input gate; The unit status Used to store historical traffic flow information, the formula is: (4)， in, The value range of is (0,1); The output gate Historical information and current time information To integrate, the formula is: (5)， (6)， in, is the weight matrix of the output gate, is the bias matrix of the input gate.

4. The multi-level federated traffic flow prediction method based on the vehicle-road-cloud integrated network according to claim 1 is characterized in that: The federated model is trained using a clustering method based on the spatial distribution of nodes, including: The central node uses the K-Means algorithm to perform iterative clustering on all regional nodes based on their spatial location information. Different regional node clusters; Each regional node uses the K-Means algorithm to perform iterative clustering on all edge nodes under its jurisdiction based on their spatial location information. Different edge node clusters; Model parameters are sent down step by step in the order of central nodes, regional node clusters, regional nodes, edge node clusters, and edge nodes for model training and parameter update. The updated parameters are then uploaded step by step for hierarchical aggregation.

5. The multi-level federated traffic flow prediction method based on the vehicle-road-cloud integrated network according to claim 4 is characterized in that: Model parameters are distributed step by step by central node, regional node cluster, regional node, edge node cluster, and edge node, including: 1) In the initial training phase, the central node converts the LSTM gradient parameters in the global model Randomly initialized to , and Issued to regional node clusters; among them, Forget Gate , input gate , output gate , unit status The weight matrix, Forget Gate , input gate , output gate , unit status The deviation matrix of 2) Each regional node cluster is randomly selected from its constituent members Regional node forwarding model parameters ; 3) Regional nodes receive the forwarded model parameters , which is further forwarded to the edge node clusters; 4) Each edge node cluster is randomly selected from its constituent members Edge node forwarding model parameters .

6. The multi-level federated traffic flow prediction method based on the vehicle-road-cloud integrated network according to claim 5 is characterized in that: Model training and parameter updates include: The edge node receives the forwarded model parameters ,The traffic flow federated model is trained by denoising, normalizing, and missing value processing of the traffic flow data collected by the ,sensors, and dividing the data into training set, validation set, and test set; For edge nodes, by setting an upper limit on the number of iteration rounds And the total number of batches of data , for each batch in each round The data in the model are analyzed using the stochastic gradient descent method. The training update formula is: (7)， in, is the global model parameter issued, is the learning rate of the edge nodes, represents the calculation of gradient, is the loss function of the edge node, are the local model parameters of the previous round of joint training, The updated values ​​of the model parameters for this round of training.

7. The multi-level federated traffic flow prediction method based on the vehicle-road-cloud integrated network according to claim 6 is characterized in that: Upload the updated parameters step by step and aggregate them layer by layer, including: 1) The local model parameters that the edge node will update Upload to the edge node cluster to which it belongs; 2) The edge node cluster receives updated model parameters from all edge nodes and performs aggregate calculations of the model parameters. The formula is as follows: (8)， in, is the number of edge nodes contained in the edge node cluster, is the number of the edge node included in the edge node cluster, For the Model parameters updated by edge nodes; 3) The edge node cluster will aggregate the model parameters Further upload to the regional node to which it belongs; 4) Regional nodes receive model parameters aggregated from all edge node clusters , and perform aggregate calculation of model parameters, the formula is as follows: (9)， in, is the number of edge node clusters contained in the regional node, is the number of the edge node cluster, For the The model parameters updated by the edge node cluster; 5) Model parameters that regional nodes will aggregate Further upload to the regional node cluster to which it belongs; 6) The regional node cluster receives the model parameters aggregated by all regional nodes , and perform aggregate calculation of model parameters, the formula is as follows: (10)， in, is the number of regional nodes contained in the regional node cluster, The number of the region node that contains the region node in the region node cluster, For the Model parameters for regional node updates; 7) Regional node clusters will aggregate model parameters Further upload to the central node; 8) The central node receives the model parameters aggregated by all regional node clusters , and perform aggregate calculation of model parameters, the formula is as follows: (11)， in, is the number of regional node clusters contained in the central node, is the number of the regional node cluster, For the Model parameters for regional node cluster updates; 9) The central node aggregates the model parameters Issued to Repeat 1)-8) for each regional node cluster until the local model training of all edge nodes converges to the specified accuracy or the set number of training rounds is reached.

Images