A distributed traffic flow prediction method and system based on time-space decoupling

Abstract

The present invention discloses a distributed traffic flow prediction method and system based on time-space decoupling, the method includes: partitioning the urban road network, deploying an edge cloud in each area, and the sensors deployed on the road upload the collected traffic information to the edge cloud corresponding to the area; the edge cloud and the core cloud jointly predict the future traffic conditions through the neural network training model of time-space decoupling; wherein the edge cloud stores the original road traffic data, the core cloud stores the road network topology relationship, and the time-space feature extraction is decoupled and deployed on the edge cloud and the core cloud respectively, the edge cloud obtains the time feature of the traffic information, the core cloud obtains the spatial feature of the traffic information, and the edge cloud collaborates to embed the spatial information in the edge cloud to fuse the time-space feature. Compared with the traditional method, the present invention reduces the transmission delay and computing pressure under the mechanism of edge-cloud collaboration, fully extracts and fuses the time-space feature, and can improve the timeliness and accuracy of the prediction.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation technology and relates to a distributed traffic flow prediction method and system based on edge-cloud collaboration and time-space decoupling. Background Art

[0002] As the number of vehicles in cities continues to increase and the road network structure becomes increasingly complex, the Intelligent Transportation System (ITS) has emerged and developed rapidly. Short-term traffic flow prediction is the core of ITS and provides data support for it. Timely and accurate traffic flow prediction results can help traffic management departments implement traffic management, scientifically guide travelers' routes, and avoid congestion; individual users can also reasonably arrange travel plans based on real-time prediction information and reduce travel time. Therefore, an efficient and accurate traffic flow prediction method is crucial to alleviating traffic congestion, making it one of the current research hotspots in the field of intelligent transportation technology.

[0003] The current traffic flow prediction method is implemented as follows: data is gathered from traffic nodes distributed throughout the traffic network to a core node, which trains the model based on the collected original speed data of the road network and predicts future traffic conditions. This traffic prediction method can produce relatively accurate prediction results and provide assistance for traffic management, but it also has the following defects: it cannot adapt to business scenarios with massive data and low latency; it cannot solve the problem of spatiotemporal feature fusion caused by different spatiotemporal feature dimensions; it cannot flexibly obtain spatiotemporal dynamic correlation. Based on the development of intelligent transportation and edge computing, traditional traffic prediction methods have gradually been improved and replaced by other traffic prediction methods, so as to achieve more real-time and accurate traffic predictions in the technical environment of intelligent transportation and edge cloud collaboration, and provide assistance for more efficient urban traffic management.

[0004] After investigating the existing literature, it was found that the articles entitled “T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction” and “Traffic Flow Prediction Based on Deep Learning in Internet of Vehicles” published in IEEE Transactions on Intelligent Transportation Systems in 2020 and 2021 both proposed a method for traffic flow prediction based on neural networks. Although these methods can accurately predict traffic conditions to a certain extent, these centralized algorithms will bring large communication and computing loads, large latency, and affect the timeliness of prediction information, making it difficult to meet business scenarios with massive data and low latency. At the same time, the extraction of spatiotemporal correlation of traffic information is not sufficient, and there is still room for improvement in the accuracy of traffic flow prediction.

[0005] The search also found that in order to reduce training time, reduce communication costs, and protect data privacy, the articles entitled "Privacy-Preserving Traffic Flow Prediction: A Federated Learning Approach" and "Short-Term Traffic Flow Prediction Based on Graph Convolutional Networks and Federated Learning" published in 2020 and 2022 in the IEEE Internet of Things Journal and IEEE Transactions on Intelligent Transportation Systems, both combined federated learning with neural networks to perform distributed traffic predictions, thereby protecting privacy and reducing latency. However, these methods can only extract one of the spatiotemporal features of traffic information, which makes it impossible for the model to fully integrate the spatiotemporal features of the data.

[0006] In summary, the problems of the existing technologies are: (1) Most traffic prediction methods are centralized algorithms, which are not suitable for business scenarios with massive data and low latency; (2) Distributed traffic models do not fully integrate the spatiotemporal characteristics of traffic data; (3) The correlation between different times and different nodes is insufficient. The significance of solving the above technical problems is: Based on the current progress of intelligent transportation technology and edge computing technology, a distributed traffic flow prediction method is designed. Through the decoupling of spatiotemporal information and the collaboration of edge and cloud, more real-time and accurate traffic prediction results are achieved, which helps traffic management departments implement traffic management. Individual users can also reasonably arrange travel plans based on real-time prediction information, providing new ideas for future urban road planning and traffic management. Summary of the invention

[0007] Purpose of the invention: In view of the problems existing in the prior art, the purpose of the present invention is to provide a distributed traffic flow prediction method and system based on spatiotemporal decoupling on the basis of edge-cloud collaboration, so as to adapt to the needs of massive data and low-latency services, and fully integrate the spatiotemporal characteristics of traffic data to improve the timeliness and accuracy of the prediction.

[0008] Technical solution: To achieve the above-mentioned invention object, the technical solution adopted by the present invention is a distributed traffic flow prediction method based on time-space decoupling, comprising the following steps:

[0009] Step 1: Divide the urban road network into different areas and deploy an edge cloud in each area. The sensors deployed on the road will upload the collected traffic information to the edge cloud corresponding to the area.

[0010] Step 2: The edge cloud and the core cloud jointly use a time-space decoupled neural network training model to predict future traffic conditions; the edge cloud stores the original road traffic data collected by sensors, and the core cloud stores the topological relationships of all road networks; the time-space decoupled neural network decouples the time-space feature extraction and is deployed on the edge cloud and the core cloud respectively, the edge cloud obtains the time characteristics of the traffic information, the core cloud obtains the spatial characteristics of the traffic information, and uses the edge cloud to collaborate with the edge cloud to embed spatial information in the edge cloud to fuse the time-space features.

[0011] Furthermore, the specific steps of edge-cloud collaborative prediction of traffic flow conditions in step 2 include:

[0012] Step 2.1: Edge Cloud E n Processing sensor node V n,l The reported data can be used to obtain the traffic timing information of each node. n,k,1 ,I n,k,2 ,…,I n,k,t}, and use this time series information as the input of the neural network; where n∈[1,N], k∈[1,K], t represents the time, N is the number of urban traffic road network partitions, and K is the number of road network nodes managed by the edge cloud;

[0013] Step 2.2: Obtain the attention matrix at different times through the attention mechanism, dynamically adjust the input, and then extract the time characteristics of traffic flow information through the encoder module composed of multiple gated recurrent neural networks to obtain the node V n,k The intermediate state vector x n,k ;

[0014] Step 2.3: To embed spatial information, the edge cloud E n Set the node V n,k The intermediate state vector x n,k Upload to the core cloud and match it with the road network topology on the core cloud; state vector x n,k The spatial attention matrix of each node is obtained through the attention mechanism. After dynamically adjusting the input, the spatial features are extracted through the graph convolutional neural network to obtain the new state vector x′ n,k And send it to the corresponding node V n,k ;

[0015] Step 2.4: Edge Cloud E n The state vector x′ of the received embedded spatial information n,k With the original state vector x n,k Merge to [x n,k ; x′ n,k ], realizing the fusion of spatiotemporal characteristics, and taking it as the initial state vector input of the decoder for training to obtain the prediction result.

[0016] Furthermore, the time-space decoupled neural network includes the time module of the edge cloud, the space module of the core cloud, and the time-space fusion mechanism:

[0017] 1) Time module: multiple gated recurrent neural networks (GRUs) are used to extract time features;

[0018] 2) Spatial module: Use graph convolutional neural network (GCN) to extract spatial features;

[0019] 3) Spatiotemporal fusion mechanism: The encoder-decoder framework and edge-cloud collaboration are used to achieve the fusion of spatiotemporal features. The encoder and decoder are both composed of GRU; the encoder converts the input traffic time series information into an intermediate vector x n,k , use this intermediate vector to extract spatial features on the core cloud and obtain a new intermediate vector x′ that embeds spatial information n,kAnd transmit it back to the edge cloud, where the two are merged to obtain

[0020] [x n,k ; x′ n,k ], integrating spatial features.

[0021] Furthermore, an attention mechanism is added in time and space to obtain the dynamic correlation between nodes and improve the prediction accuracy.

[0022] Furthermore, the collected raw road traffic data and time model are located on the edge cloud, and the topological relationship and spatial model are located on the core cloud. Different optimization strategies are adopted for the models on the edge cloud and the core cloud using the federated learning method:

[0023] 1) Edge cloud: First calculate the node V n,k Based on the difference between the local model and the global model, the nodes are sorted from small to large according to the difference calculation results to obtain the node candidate set participating in the model optimization; the node models with small differences in the node candidate set are selected according to a certain proportion and uploaded to the core cloud to participate in the model optimization, so as to reduce the impact of the non-independent and identically distributed data on the model accuracy, and further reduce the uploaded model parameters and reduce the transmission delay; the core cloud sends the weighted averaged global model to the edge cloud, and the nodes update the model parameters.

[0024] 2) Core cloud: Since the core cloud only stores the topological structure of the entire road network and cannot directly use traffic data for model optimization, the model is optimized by aggregating the gradient information uploaded by the edge cloud.

[0025] Based on the same inventive concept, the present invention provides a distributed traffic flow prediction system based on time-space decoupling, comprising:

[0026] The data collection module is used to partition the urban road network. An edge cloud is deployed in each area. The sensors deployed on the road upload the collected traffic information to the edge cloud corresponding to the area.

[0027] And an edge-cloud collaborative prediction module, in which the edge cloud and the core cloud jointly use a spatiotemporal decoupled neural network training model to predict future traffic conditions; the edge cloud stores the original road traffic data collected by sensors, and the core cloud stores the topological relationships of all road networks; the spatiotemporal decoupled neural network decouples the spatiotemporal feature extraction and is deployed on the edge cloud and the core cloud respectively, the edge cloud obtains the temporal features of the traffic information, the core cloud obtains the spatial features of the traffic information, and uses edge-cloud collaboration to embed spatial information in the edge cloud to fuse the spatiotemporal features.

[0028] Based on the same inventive concept, the present invention provides another distributed traffic flow prediction system based on time-space decoupling, including multiple edge cloud computing devices and core cloud computing devices, each edge cloud computing device corresponds to a partition of the urban traffic road network, receives and stores traffic information collected by sensors on the roads in the partition; the core cloud computing device communicates with each edge cloud computing device and stores the topological relationship of all road networks; the edge cloud computing device includes a memory, a processor and a computer program stored in the memory and executable on the processor, and when the computer program is loaded into the processor, the processing steps of the edge cloud in the above-mentioned distributed traffic flow prediction method based on time-space decoupling are implemented; the core cloud computing device includes a memory, a processor and a computer program stored in the memory and executable on the processor, and when the computer program is loaded into the processor, the processing steps of the core cloud in the above-mentioned distributed traffic flow prediction method based on time-space decoupling are implemented.

[0029] Beneficial effects: Based on the development of intelligent transportation and edge computing technologies, the present invention provides a distributed traffic flow prediction method and system based on time-space decoupling, which provides a solution for achieving more real-time and accurate distributed traffic management under massive data. The present invention partitions the road network according to actual needs, and the original data and models are stored and managed by each area; according to the different dimensions of time-space characteristics, the time-space modules are decoupled and deployed on the edge cloud and the core cloud respectively, and the edge-cloud collaboration is used for time-space fusion; an attention mechanism is added to further enhance the acquisition of time-space correlation. Compared with the prior art, the present invention has the following advantages: (1) The use of a distributed traffic prediction method can delegate the computing pressure to the edge nodes. Under the edge-cloud collaboration mechanism, the uploaded data and models are reduced, the transmission delay is reduced, and the difference comparison strategy is further added to select some nodes to participate in the model optimization, reduce the communication load, and ensure the real-time nature of the prediction results. (2) The decoupling of the spatiotemporal modules makes the traffic prediction method no longer restricted by the separation caused by the different dimensions of spatiotemporal characteristics. The edge cloud extracts the temporal characteristics of the traffic information, the core cloud extracts the spatial characteristics of the traffic information, and uses the edge cloud to collaborate with the edge cloud to embed spatial information in the edge cloud for the fusion of spatiotemporal characteristics. The spatiotemporal characteristics are fully extracted and fused, and the attention mechanism is further integrated to improve the timeliness and accuracy of the prediction. (3) The present invention can provide decision-making references for the traffic management departments in each region, scientifically guide traffic, avoid congestion, and provide users with real-time prediction information so that users can reasonably arrange travel plans and reduce travel time. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is the distributed traffic scene graph provided by the present invention.

[0031] Figure 2It is a flow chart of a distributed traffic flow prediction method based on time-space decoupling provided by the present invention.

[0032] Figure 3 It is a model optimization flow chart provided by the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the embodiments of the present invention are described in detail below in conjunction with the accompanying drawings: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given. It should be understood that the specific examples described here are only used to explain the present invention, but the protection scope of the present invention is not limited to the following embodiments.

[0034] This embodiment adopts Figure 1 In the scenario of urban traffic zoning management, a distributed traffic flow prediction method based on time-space decoupling is proposed. The process is as follows Figure 2 First, in this traffic scenario, the entire urban traffic network is divided into N areas, and an edge cloud E is deployed in area n. n , the sensor k on the road is used as a road network node V n,k By the area E to which it belongs n Unified management, and the speed information of this node collected at time t I n,k,t Upload to edge cloud E n Storage, define edge cloud and nodes respectively

[0035] E n = {V n,1 ,V n,2 ,…,V n,k |n∈[1,N],k∈[1,K]},V n,k = {I n,k,1 ,I n,k,2 ,…,I n,k,t |k∈[1,K],t∈[1,∞]}. Secondly, the core cloud stores all edge clouds {E 1 ,E 2 ,…,E N}The topological structure of the nodes within the network, the road network topology is represented by the adjacency matrix A of all nodes V.

[0036] The basic goal of this embodiment is to achieve more real-time and accurate traffic prediction by utilizing edge-cloud collaboration. According to the different dimensions of spatiotemporal feature acquisition, the neural network in the present invention decouples the spatiotemporal feature extraction and deploys them on the edge cloud and core cloud respectively, and integrates the spatiotemporal features by embedding spatial information in the edge cloud. In order to improve the prediction accuracy, the attention mechanism is added to further obtain the dynamic correlation between nodes. In order to reduce the transmission delay, some nodes are selected to participate in the model optimization considering the differences with the global model.

[0037] This embodiment provides that the speed information of the road network is aggregated in a time interval of 5 minutes. n,k The road speed {I n,k,1 ,…,I n,k,i} as the input of the neural network to predict the road speed in the next j time slots The actual value of the road speed in the future j time slots is expressed as { n,k,1 ,…,O n,k,j}. Node V n,k The loss functions of and core cloud are:

[0038]

[0039]

[0040] The specific steps of using edge-cloud collaboration to predict traffic flow conditions are as follows:

[0041] 1. Edge cloud processing node V n,k The reported data (mainly including time and speed information) is used to obtain the traffic timing information of each node. n, k , 1,…}, and use this time series information as the input of the neural network;

[0042] 2. Since traffic information at different time points is correlated, the attention matrix between the input time series information at different times is first obtained through the attention mechanism to dynamically adjust the input. Then, the encoder module composed of multiple gated recurrent neural networks is used to extract the time features of the traffic information and obtain the node V n,k The intermediate state vector x n,k ;

[0043] 3. In order to embed spatial information, the edge cloud converts the intermediate state vector x of each traffic node n,k Upload to the core cloud and match it with the road network topology on the core cloud. State vector x n,k Similarly, the spatial attention matrix of each node is obtained through the attention mechanism. After dynamically adjusting the input, the spatial features are extracted through the graph convolutional neural network to obtain the new state vector x′ n,kAnd send it to the corresponding node V n,k ;

[0044] 4. The edge cloud receives the state vector x′ of the embedded spatial information n,k With the original state vector x n,k Merge to [x n,k ; x′ n,k ], realizing the fusion of spatiotemporal characteristics of traffic information, and taking it as the initial state vector input of the decoder for training to obtain the prediction result.

[0045] In the present invention, an encoder-decoder module composed of multiple GRUs is used to extract temporal features. Gated recurrent neural network (GRU) is a type of recurrent neural network (RNN), which can solve problems such as long-term memory and gradient in back propagation. It is expressed as:

[0046] u i =σ(W u ·[h i-1 ,I n, k ,i ]),

[0047] r i =σ(W r ·[h i-1 ,I n, k ,i ]),

[0048]

[0049]

[0050] Among them, I n,k,i Represents node V n,k The speed at time i, h i-1 、h i represents the output hidden vector at time i-1 and i; u i is the update gate at time i, which controls the extent to which the state information of the previous time is brought into the current state; r i is the reset gate at time i, which controls how much information of the previous state is written into the current candidate set Up; W u , W r , W h are the corresponding weight matrices respectively, and σ represents the activation function.

[0051] Encoder-decoder is a common framework in deep learning. This embodiment uses GRU as its internal structure. In the present invention, encoder-decoder is mainly used to achieve the fusion of spatiotemporal features. The encoder converts the input sequence {I n,k,1 ,…,I n,k,i} is transformed into an intermediate vector x n,k , use this intermediate vector to extract spatial features on the core cloud and obtain a new intermediate vector x′ that embeds spatial information n,k And transmit it back to the edge cloud, where the two are merged to obtain [x n,k ; x′ n,k ], integrating spatial features.

[0052] In the present invention, GCN is used to extract the spatial features of traffic information. Graph Convolutional Networks (GCN) is a neural network architecture that operates on graph data and is used to extract features of non-Euclidean data such as road networks, which can be expressed as:

[0053]

[0054] Among them, x n,k Node V representing the edge cloud upload n,k The intermediate state vector, x′ n,k represents the intermediate state vector after embedding spatial information, σ represents the activation function, W is the learnable weight matrix, A is the adjacency matrix of the road network graph structure, I is the unit matrix, for The degree matrix of .

[0055] In order to obtain the dynamic correlation between road network nodes, attention mechanisms are added to the edge cloud and core cloud respectively, which can be expressed as:

[0056] B=softmax((W q X) T W a (W k X)),

[0057] X′=BW v X,

[0058] Among them, X is the input of the attention mechanism, and the input of the temporal attention mechanism is the traffic timing information of the node {I n,k,1 ,…,I n,k,i}, the input of the spatial attention mechanism is the intermediate state vector {x 1,1 ,…,x N,K}, X′ represents the dynamic adjustment of the input relevance using the attention matrix B, softmax(·) represents normalization, and W q , W k , W v , W a is the learnable weight matrix.

[0059] The collected raw road traffic data and time model are located on the edge cloud, and the topological relationship and spatial model are located on the core cloud. Different optimization strategies are adopted for the models on the edge cloud and the core cloud using the federated learning method, such as Figure 3 As shown:

[0060] Edge cloud: Affected by factors such as location, the speed data collected by different road network nodes are distributed differently, that is, the original data of each node is not independent and identically distributed. In order to reduce the impact of non-independent and identically distributed data on model accuracy, and further reduce the uploaded model parameters and reduce transmission delay, this embodiment calculates the node model of the r-1th round With the global model The node candidate set Hr for the rth round of model optimization is obtained by sorting the differences from small to large. The difference is calculated as follows:

[0061]

[0062] in, is the global model of the r-1th round, is the node V in the r-1th round n,k The parameter model of

[0063] ‖·‖ F Represents the F-norm of the matrix.

[0064] Select node models with small differences in the node candidate set Hr at a certain ratio Upload to the core cloud to participate in model optimization. The core cloud performs weighted average of the received model parameters to obtain the global model parameters, and then sends the aggregated global model to the edge cloud. Each road network node updates its parameters based on the global model stored in the edge cloud in this area. Model aggregation is represented as follows:

[0065]

[0066] in, Node V participating in the rth round of model optimization n,f The parameter model of is the global model of the rth round, n is the amount of data of all nodes, n f is the amount of data at node f.

[0067] 2) Core cloud: Since the core cloud only stores the topological structure of the entire road network and cannot directly use traffic data for model optimization, it is necessary to store the nodes V on the edge cloud. n,k Gradient of round r Upload to the core cloud, and then use the chain rule to get the gradient The core cloud aggregates the gradients of each node for model optimization, which is expressed as:

[0068]

[0069]

[0070] in, They represent the model parameters of the core cloud in the rth round and the r+1th round respectively, is the local gradient of the core cloud in round r, represents the global gradient of the core cloud in the rth round, which is used to update the core cloud model, and η represents the learning rate.

[0071] Based on the same inventive concept, an embodiment of the present invention discloses a distributed traffic flow prediction system based on time-space decoupling, including: a data acquisition module, used to partition the urban road network, deploy an edge cloud in each area, and the sensors deployed on the road upload the collected traffic information to the edge cloud corresponding to the area; and an edge-cloud collaborative prediction module, in which the edge cloud and the core cloud jointly use a time-space decoupled neural network training model to predict future traffic conditions; the edge cloud stores the original road traffic data collected by the sensors, and the core cloud stores the topological relationship of all road networks; the time-space decoupled neural network decouples the time-space feature extraction and is deployed on the edge cloud and the core cloud respectively, the edge cloud obtains the time characteristics of the traffic information, the core cloud obtains the spatial characteristics of the traffic information, and the edge cloud collaborates to embed spatial information in the edge cloud to fuse the time-space features.

[0072] Specifically, the edge-cloud collaborative prediction module includes: a time feature extraction unit for edge cloud E n Processing sensor node V n,k The reported data is used to obtain the traffic timing information of each node, and this timing information is used as the input of the neural network; the attention matrix at different times is obtained through the attention mechanism, and the input is dynamically adjusted. Then, the encoder module composed of multiple gated recurrent neural networks is used to extract the time characteristics of the traffic flow information and obtain the node V n,k The intermediate state vector x n,k ; Spatial feature extraction unit for edge cloud E n Set the node V n,k The intermediate state vector x n,k Upload to the core cloud and match it with the road network topology on the core cloud; state vector xn,k The spatial attention matrix of each node is obtained through the attention mechanism. After dynamically adjusting the input, the spatial features are extracted through the graph convolutional neural network to obtain the new state vector x′ n,k And send it to the corresponding node V n,k ; and spatiotemporal fusion prediction unit for edge cloud E n The state vector x′ of the received embedded spatial information n,k With the original state vector x n,k Merge to [x n,k ; x′ n,k ], realize the fusion of spatiotemporal characteristics, and use it as the initial state vector input of the decoder for training to obtain the prediction result. The specific implementation of each unit refers to the above method embodiment, which will not be repeated here.

[0073] Based on the same inventive concept, an embodiment of the present invention discloses a distributed traffic flow prediction system based on spatiotemporal decoupling, comprising a plurality of edge cloud computing devices and a core cloud computing device, each edge cloud computing device corresponding to a partition of an urban traffic road network, receiving and storing traffic information collected by sensors on roads within the partition; the core cloud computing device communicates with each edge cloud computing device and stores the topological relationship of all road networks; the edge cloud computing device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and when the computer program is loaded onto the processor, the processing steps of the edge cloud in the aforementioned distributed traffic flow prediction method based on spatiotemporal decoupling are implemented; the core cloud computing device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and when the computer program is loaded onto the processor, the processing steps of the core cloud in the aforementioned distributed traffic flow prediction method based on spatiotemporal decoupling are implemented.

[0074] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A distributed traffic flow prediction method based on spatiotemporal decoupling, characterized in that: The following steps are involved: Step 1: Divide the urban road network into different areas and deploy an edge cloud in each area. The sensors deployed on the road will upload the collected traffic information to the edge cloud corresponding to the area. Step 2: The edge cloud and the core cloud jointly train a model through a time-space decoupled neural network to predict future traffic conditions; the edge cloud stores the original road traffic data collected by sensors, and the core cloud stores the topological relationships of all road networks; the time-space decoupled neural network decouples the time-space feature extraction and is deployed on the edge cloud and the core cloud respectively, the edge cloud obtains the time features of the traffic information, the core cloud obtains the spatial features of the traffic information, and the edge cloud collaborates to embed spatial information in the edge cloud to fuse the time-space features; The specific steps of edge-cloud collaborative prediction of traffic flow conditions in step 2 include: Step 2.1: Edge Cloud E n Processing sensor node V n，k The reported data can be used to obtain the traffic timing information of each node. n，k，1 , I n,k,2 , ..., I n，k，t }, taking this time series information as the input of the neural network; where n∈[1, N], k∈[1, K], t represents the time, N is the number of urban traffic road network partitions, and K is the number of road network nodes managed by the edge cloud; Step 2.2: Obtain the attention matrix at different times through the attention mechanism, dynamically adjust the input, and then extract the time characteristics of traffic flow information through the encoder module composed of multiple gated recurrent neural networks to obtain the node V n，k The intermediate state vector x n,k ; Step 2.3: To embed spatial information, the edge cloud E n Set the node V n，k The intermediate state vector x n,k Upload to the core cloud and match it with the road network topology on the core cloud; state vector x n,k The spatial attention matrix of each node is obtained through the attention mechanism. After dynamically adjusting the input, the spatial features are extracted through the graph convolutional neural network to obtain the new state vector x′ n，k And send it to the corresponding node V n，k ; Step 2.4: Edge Cloud E n The state vector x′ of the received embedded spatial information n，k With the original state vector x n,k Merge to [x n,k ; x′ n，k ], realize the fusion of spatiotemporal characteristics, and use it as the initial state vector input of the decoder for training to obtain the prediction result; The time-space decoupled neural network includes the time module of the edge cloud, the space module of the core cloud, and the time-space fusion mechanism; The time module uses a combination of multiple gated recurrent neural networks (GRUs) to extract time features, which can be expressed as: u i =σ(W u ·[h i-1 ,I n，k，i ]), r i =σ(W r ·[h i-1 ,I n，k，i ]), Among them, I n，k，i Represents node V n，k The speed at time i, h i-1 、h i represents the output hidden vector at time i-1 and i; u i is the update gate at time i, which controls the extent to which the state information of the previous time is brought into the current state; r i is the reset gate at time i, which controls how much information of the previous state is written into the current candidate set Up; W u , W r , W h are the corresponding weight matrices respectively, σ represents the activation function; The spatial module uses a graph convolutional neural network (GCN) to extract spatial features, which can be expressed as: Among them, x n,k Node V representing the edge cloud upload n，k The intermediate state vector, x′ n,k represents the intermediate state vector after embedding spatial information, σ represents the activation function, W is the learnable weight matrix, A is the adjacency matrix of the road network graph structure, I is the unit matrix, for The degree matrix of The spatiotemporal fusion mechanism uses the encoder-decoder framework and edge-cloud collaboration to achieve the fusion of spatiotemporal features. The encoder and decoder are both composed of GRU. The encoder converts the input sequence {I n，k，1 , ..., I n，k，i } is transformed into an intermediate vector x n,k , use this intermediate vector to extract spatial features on the core cloud and obtain a new intermediate vector x′ that embeds spatial information n，k And transmit it back to the edge cloud, where the two are merged to obtain [x n,k ; x′ n，k ], integrating spatial features.

2. The distributed traffic flow prediction method based on time-space decoupling according to claim 1 is characterized by: By adding attention mechanisms in time and space to obtain the dynamic correlation between nodes, the prediction accuracy is improved, which is expressed as: B=softmax((W q X) T W a (W k X)), X′=BW v X, Among them, X is the input of the attention mechanism, and the input of the temporal attention mechanism is the traffic timing information of the node {I n，k，1 , ..., I n，k，i }, the input of the spatial attention mechanism is the intermediate state vector {x 1，1 , ..., x N，K }, X′ represents the dynamic adjustment of the input relevance using the attention matrix B, softmax(·) represents normalization, and W q , W k , W v , W a is the learnable weight matrix.

3. The distributed traffic flow prediction method based on time-space decoupling according to claim 1 is characterized by: The collected raw road traffic data and time model are located on the edge cloud, and the topological relationship and spatial model are located on the core cloud. Using the federated learning method, different optimization strategies are adopted for the models on the edge cloud and core cloud.

4. The distributed traffic flow prediction method based on time-space decoupling according to claim 3 is characterized by: The optimization strategy of the edge cloud model is: Calculate the node model for round r-1 With the global model The node candidate set H participating in the r-th round of model optimization is obtained by sorting the differences from small to large. r , the difference is calculated as follows: in, is the global model of the r-1th round, is the node V in the r-1th round n，k The parameter model of ||·|| F represents the F norm of the matrix; Select node candidate set H with a certain ratio r Node model with small differences in the medium Upload to the core cloud to participate in the model optimization. The core cloud performs weighted average of the received model parameters to obtain the global model parameters, and then sends the aggregated global model to the edge cloud. Each road network node updates the parameters according to the global model stored in the edge cloud in the region. Model aggregation is expressed as follows: in, Node V participating in the rth round of model optimization n，f The parameter model of is the global model of the rth round, n is the amount of data of all nodes, n f is the data volume of node f, and F is the number of nodes participating in model optimization.

5. The distributed traffic flow prediction method based on time-space decoupling according to claim 3 is characterized by: The optimization strategy of the core cloud model is: n，k Gradient of round r Upload to the core cloud, and then use the chain rule to get the gradient The core cloud aggregates the gradients of each node for model optimization, which is expressed as: Among them, L n，k , L C Node V n，k and the loss function of the core cloud, They represent the model parameters of the core cloud in the rth round and the r+1th round respectively, is the local gradient of the core cloud in round r, represents the global gradient of the core cloud in the rth round, which is used to update the core cloud model, and η represents the learning rate.

6. A distributed traffic flow prediction system based on time-space decoupling, characterized in that: include: The data collection module is used to partition the urban road network. An edge cloud is deployed in each area. The sensors deployed on the road upload the collected traffic information to the edge cloud corresponding to the area. And the edge cloud collaborative prediction module, the edge cloud and the core cloud jointly use the time-space decoupled neural network training model to predict future traffic conditions; the edge cloud stores the original road traffic data collected by sensors, and the core cloud stores the topological relationship of all road networks; the time-space decoupled neural network decouples the time-space feature extraction and is deployed on the edge cloud and the core cloud respectively, the edge cloud obtains the time features of the traffic information, the core cloud obtains the spatial features of the traffic information, and uses the edge cloud collaboration to embed spatial information in the edge cloud to fuse the time-space features; The edge-cloud collaborative prediction module includes: Temporal feature extraction unit for edge cloud E n Processing sensor node V n，k The reported data can be used to obtain the traffic timing information of each node. n，k，1 , I n，k，2 , ..., I n，k，t }, taking this time series information as the input of the neural network; where n∈[1, N], k∈[1, K], t represents the time, N is the number of urban traffic road network partitions, and K is the number of road network nodes managed by the edge cloud; and the attention matrix at different times is obtained through the attention mechanism, the input is dynamically adjusted, and then the encoder module composed of multiple gated recurrent neural networks is used to extract the time characteristics of traffic flow information and obtain the node V n，k The intermediate state vector x n,k ; Spatial feature extraction unit for edge cloud E n Set the node V n，k The intermediate state vector x n,k Upload to the core cloud and match it with the road network topology on the core cloud; state vector x n,k The spatial attention matrix of each node is obtained through the attention mechanism. After dynamically adjusting the input, the spatial features are extracted through the graph convolutional neural network to obtain the new state vector x′ n，k And send it to the corresponding node V n，k ; And the spatiotemporal fusion prediction unit for edge cloud E n The state vector x′ of the received embedded spatial information n，k With the original state vector x n,k Merge to [x n,k ; x′ n，k ], realize the fusion of spatiotemporal characteristics, and use it as the initial state vector input of the decoder for training to obtain the prediction result; The time-space decoupled neural network includes the time module of the edge cloud, the space module of the core cloud, and the time-space fusion mechanism; The time module uses a combination of multiple gated recurrent neural networks (GRUs) to extract time features, which can be expressed as: u i =σ(W u ·[h i-1 ,I n，k，i ]), r i =σ(W r ·[h i-1 ,I n，k，i ]), Among them, I n，k，i Represents node V n，k The speed at time i, h i-1 、h i represents the output hidden vector at time i-1 and i; u i is the update gate at time i, which controls the extent to which the state information of the previous time is brought into the current state; r i is the reset gate at time i, which controls how much information of the previous state is written into the current candidate set Upper: W u , W r , W h are the corresponding weight matrices respectively, σ represents the activation function; The spatial module uses a graph convolutional neural network (GCN) to extract spatial features, which can be expressed as: Among them, x n,k Node V representing the edge cloud upload n，k The intermediate state vector, x′ n,k represents the intermediate state vector after embedding spatial information, σ represents the activation function, W is the learnable weight matrix, A is the adjacency matrix of the road network graph structure, I is the unit matrix, for The degree matrix of The spatiotemporal fusion mechanism uses the encoder-decoder framework and edge-cloud collaboration to achieve the fusion of spatiotemporal features. The encoder and decoder are both composed of GRU. The encoder converts the input sequence {I n，k，1 , ..., I n，k，i } is transformed into an intermediate vector x n,k , use this intermediate vector to extract spatial features on the core cloud and obtain a new intermediate vector x′ that embeds spatial information n,k And transmit it back to the edge cloud, where the two are merged to obtain [x n,k ; x′ n,k ], integrating spatial features.

7. A distributed traffic flow prediction system based on time-space decoupling, characterized by: The invention comprises a plurality of edge cloud computing devices and a core cloud computing device, each of which corresponds to a partition of an urban traffic road network and receives and stores traffic information collected by sensors on roads within the partition; the core cloud computing device communicates with each edge cloud computing device and stores the topological relationship of all road networks; the edge cloud computing device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and when the computer program is loaded onto the processor, the processing steps of the edge cloud in the distributed traffic flow prediction method based on time-space decoupling according to any one of claims 1 to 5 are implemented; the core cloud computing device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and when the computer program is loaded onto the processor, the processing steps of the core cloud in the distributed traffic flow prediction method based on time-space decoupling according to any one of claims 1 to 5 are implemented.

Images