Time-space correlation traffic flow prediction method based on deep learning

Abstract

The invention relates to the technical field of intelligent traffic prediction, in particular to a time-space correlation traffic flow prediction method based on deep learning, which comprises the following steps: S1, collecting traffic flow data and related external environment data in real time, and establishing an initial traffic network topology structure; s2, calculating the dynamic weight of each node; s3, dynamically adjusting the weight of each edge in the traffic network according to the node weight change in the S2 so as to update the topological structure of the traffic network; s4, performing feature extraction on the spatial dependency relationship among the traffic nodes, and extracting time evolution features of the traffic flow data at the same time; s5, constructing an adaptive space-time fusion model; s6, traffic flow prediction is carried out, and adaptive space-time fusion model parameters are optimized; according to the method, through dynamic fusion of the spatial-temporal characteristics and adaptive adjustment of the model parameters, the precision and real-time performance of traffic flow prediction are effectively improved, and the adaptability and generalization ability of the model in a complex dynamic traffic scene are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent traffic prediction technology, and in particular to a spatiotemporal correlation traffic flow prediction method based on deep learning. Background Art

[0002] With the continuous acceleration of urbanization, traffic flow management and prediction has become one of the core issues in intelligent transportation systems. Accurate traffic flow prediction can not only optimize traffic signal control and resource allocation, but also reduce congestion and improve road use efficiency. At present, deep learning technology is gradually being applied to traffic flow prediction, especially by combining models with spatiotemporal data to better capture complex traffic flow trends. Among them, graph convolutional networks can effectively extract spatial dependencies in traffic networks, while time series models such as long short-term memory networks can capture the evolution of traffic flow over time. However, existing traffic flow prediction methods often process spatial and temporal features separately, making it difficult to accurately respond to real-time traffic changes and emergencies.

[0003] Existing technologies have significant limitations in fusing spatiotemporal features and processing real-time dynamic traffic changes, resulting in insufficient prediction accuracy. Specifically, current models struggle to effectively handle the complex spatial dependencies and time series evolution of traffic networks, and are unable to dynamically adjust feature weights to accommodate varying spatiotemporal conditions. Furthermore, in practical applications, existing models lack adaptability to changes in traffic flow, making them inflexible in dynamically changing scenarios such as emergencies and road congestion. Therefore, developing a traffic flow prediction method that can dynamically fuse spatiotemporal features and adaptively adjust model parameters has become essential for addressing traffic management issues. Summary of the Invention

[0004] Based on the above objectives, the present invention provides a time-space correlation traffic flow prediction method based on deep learning.

[0005] A deep learning-based spatiotemporal correlation traffic flow prediction method includes the following steps:

[0006] S1: Based on the sensor network deployed at traffic nodes, traffic flow data and related external environment data are collected in real time, and the initial traffic network topology structure including traffic nodes and edges is established based on the collected data;

[0007] S2: Using the traffic flow data and historical data collected in S1, the dynamic weight of each node is calculated based on the node's traffic flow, geographical location, and road capacity.

[0008] S3: Based on the changes in node weights in S2, combined with real-time traffic conditions and related events, the weights of each edge in the traffic network are dynamically adjusted through the graph neural network to update the traffic network topology.

[0009] S4: Based on the dynamic traffic network topology constructed in step S3, a graph convolutional network is used to extract features of the spatial dependencies between traffic nodes. At the same time, a variational autoencoder is used to model the time series data and extract the temporal evolution characteristics of traffic flow data.

[0010] S5: The spatial features extracted in S4 are fused with the time series features through an adaptive weighting mechanism to construct an adaptive spatiotemporal fusion model;

[0011] S6: Based on the adaptive spatiotemporal fusion model in S5, real-time traffic data is input to perform traffic flow prediction, and the parameters of the adaptive spatiotemporal fusion model are optimized according to the difference between the prediction results and the actual traffic flow.

[0012] Optionally, the S1 specifically includes:

[0013] S11: Deploy a sensor network at traffic nodes to collect traffic flow data, including vehicle speed, vehicle density, and road occupancy rate. The sensor network collects traffic flow data at preset time intervals and uploads the data to a data center.

[0014] S12: Real-time collection of external environmental data, including weather information, road construction conditions, and accident occurrences. The external environmental data is obtained by connecting with a meteorological monitoring system, a road construction management system, and a traffic accident alarm system;

[0015] S13: Preprocessing the data collected in steps S11 and S12, including removing abnormal data and formatting to ensure data integrity and consistency;

[0016] S14: Based on the processed data and the physical connection relationship between traffic nodes and roads, an initial traffic network topology structure of traffic nodes and edges is established. The topology structure uses traffic nodes as vertices of the graph and roads as edges of the graph to reflect the flow path of traffic flow between different nodes.

[0017] Optionally, the S2 specifically includes:

[0018] S21: Acquire the real-time traffic flow data and historical traffic flow data collected in S1;

[0019] S22: Using the geographic location information of the transportation nodes, calculate the spatial relationship of each node relative to other nodes. The spatial relationship is based on the distance, connection mode and traffic flow direction between the transportation nodes, combined with the geographic characteristics of the nodes to form a spatial dependency matrix;

[0020] S23: Weighting the road attributes of each node according to the road capacity, wherein the road capacity includes road width, number of lanes, traffic signal control, and road restrictions, and specifically using weighting factors to perform comprehensive calculations to generate a preliminary weight for each node;

[0021] S24: The data generated in S21, S22, and S23 are dynamically adjusted through a deep neural network model. The deep neural network model dynamically calculates the weight value of each node by combining and analyzing real-time traffic flow and historical data. The weight value is used to reflect the traffic pressure and importance of the node at the current moment, and is updated in real time as the traffic flow and environment change.

[0022] Optionally, the S24 specifically includes:

[0023] S241: Segmenting the collected real-time traffic flow data and historical traffic flow data according to a predetermined time window to provide smooth data input for subsequent processing;

[0024] S242: combining the spatial dependency matrix with the normalized preliminary node weight to form an input feature vector of the node, wherein the feature vector includes the real-time traffic, historical traffic, spatial dependency, and preliminary weight of the node;

[0025] S243: Inputting the input feature vector into a preset deep neural network model, wherein the deep neural network model includes several hidden layers, and each layer performs nonlinear transformation based on the output of the previous layer and the model weight and bias to form a new output;

[0026] S244: After multiple layers of nonlinear transformation, the final dynamic weight value of the node is generated;

[0027] S245: Standardize the generated dynamic weight value to generate a standardized node dynamic weight.

[0028] Optionally, the S3 specifically includes:

[0029] S31: Obtain the dynamic weight of each traffic node calculated in S2, as well as the real-time collected road condition information and related event data, wherein the road condition information includes road congestion level and traffic accident status, and the related event data includes road construction status and weather changes;

[0030] S32: Based on the dynamic weights of the transportation nodes and the connection relationships between the nodes, construct an initial transportation network adjacency matrix, where the elements in the adjacency matrix represent the direct connection relationships between the nodes, and the element values ​​are initialized to 1;

[0031] S33: Mapping the real-time traffic information and related event data into initial weights of edges. The initial weights are quantified based on the actual traffic capacity, congestion level, and impact level of the road. A higher value indicates a stronger traffic capacity.

[0032] S34: The dynamic weights of the nodes and the initial weights of the edges are input as input features into a preset graph neural network model. The graph neural network model calculates the dynamic weights of each edge through information propagation and aggregation mechanisms, and uses multi-layer graph convolution operations to iteratively update the features of the nodes and edges.

[0033] S35: Using the dynamic weights of the edges calculated in S34, the topology of the traffic network is updated to generate an updated adjacency matrix, so that the weights of the edges reflect the traffic capacity of the roads in the current traffic network and the impact of real-time events.

[0034] Optionally, the S34 specifically includes:

[0035] S341: Initialize the feature value of each node. The node features include its dynamic weight, traffic flow data, and connection information with other nodes. At the same time, initialize the initial weight value of each edge.

[0036] S342: In each layer of the graph convolution operation, a node receives information from its neighboring nodes. The features of the neighboring nodes are weighted according to the connection strength and weight between the nodes and then aggregated and transmitted to the target node.

[0037] S343: After receiving the information from the neighboring nodes, the target node aggregates the received information with its own eigenvalues. The aggregated eigenvalues ​​are transformed nonlinearly through an activation function to update the node's features.

[0038] S344: updating the edge weights, where the edge weights are calculated based on the updated features of the connected nodes, thereby reflecting the latest traffic conditions between the nodes;

[0039] S345: Repeat the above graph convolution operation in the multi-layer network. Each layer will update the feature values ​​of the nodes and edges, and finally output the features of the nodes and the dynamic weights of each edge.

[0040] Optionally, the S4 specifically includes:

[0041] S41: Based on the dynamic traffic network topology structure constructed in step S3, define the node characteristics and edge characteristics in the traffic network. The node characteristics include traffic flow and road capacity, and the edge characteristics include the dynamic weights between nodes, forming the initial characteristic matrix of the traffic network.

[0042] S42: Use graph convolutional networks to extract spatial features from node and edge features. Specifically, by aggregating information about each node and its neighboring nodes, the spatial dependency relationships between nodes in the transportation network are extracted.

[0043] S43: In each layer of graph convolutional network operation, the node features and the features of its neighboring nodes are weighted according to the dynamic weight of the edge, and the spatial features between the nodes are extracted through the convolution operation. The feature values ​​of the nodes are updated layer by layer, and finally the spatial feature vector F representing the global traffic network is generated. F(L) ;

[0044] S44: Model the time evolution of traffic flow, use variational autoencoders to process the time series data of traffic nodes, and encode the high-dimensional time series data into a low-dimensional latent variable space representation; let the low-dimensional latent variable be Z t , the specific formula is: Z t =μ t +∈ t ·σ t , where μ t is the mean output of the encoder; σ t is the standard deviation output of the encoder; ∈ t is the noise sampled from the standard normal distribution;

[0045] S45: In the latent variable space, the decoder of the variational autoencoder is used to reconstruct the time series data. The reconstruction loss of the decoder is used to extract the temporal variation characteristics of the traffic flow data to capture the short-term and long-term evolution trends of the traffic flow.

[0046] Optionally, the S5 specifically includes:

[0047] S51: normalize the spatial features and temporal features extracted in S4 to ensure that the two features are in the same value range;

[0048] S52: Calculate the weight coefficient θ of spatial and temporal features based on the real-time dynamic situation of the current traffic network s and θ t , the weight coefficient is used to control the contribution of spatial features and temporal features in the model;

[0049] S53: Perform weighted fusion of spatial features and temporal features according to the weights, and let the fused spatiotemporal feature vector be F st , the formula is: Fst =θ s ·F (L) +θ t ·Z t , where F st Represents the fused spatiotemporal feature vector;

[0050] S54: Based on the spatiotemporal feature vector F fused in S53 st , an adaptive spatiotemporal fusion model is constructed through a deep neural network. The adaptive spatiotemporal fusion model adopts a multi-layer network structure, in which each layer performs weighted processing and nonlinear transformation according to the input features.

[0051] Optionally, the S54 specifically includes:

[0052] S541: First, the fused spatiotemporal feature vector is input into the first layer of the deep neural network. The first layer of the network performs a linear transformation on the input features and performs nonlinear mapping through an activation function to generate a new feature representation.

[0053] S542: Input the features generated by the first layer into the second hidden layer, continue to perform linear and nonlinear processing on the features, extract the complex patterns and relationships in the spatiotemporal features layer by layer, and enhance the ability to capture spatiotemporal changes;

[0054] S543: Introducing the attention mechanism into deep neural networks. The attention mechanism dynamically adjusts the contribution of each feature to the final output by learning the importance weights of spatiotemporal features, ensuring that the model can adaptively handle different traffic scenarios.

[0055] S544: Using the attention weights to weight the features output by the second hidden layer, generating adaptively adjusted feature representations to optimize the fusion effect of spatiotemporal features;

[0056] S545: The weighted features are input to the output layer, and the output layer generates a prediction result of future traffic flow based on the processed spatiotemporal features.

[0057] Optionally, the S6 specifically includes:

[0058] S61: Calculate the difference between the model prediction result and the actual traffic flow. For the traffic flow at each moment, calculate the error between the predicted value output by the adaptive spatiotemporal fusion model and the actual measured value.

[0059] S62: Based on the size of the error, the model optimizes the parameters through the backpropagation algorithm, and calculates the gradient of the loss function relative to the weights and biases of each layer of the network;

[0060] S63: Update the model parameters using the gradient descent algorithm. Specifically, starting from the output layer, calculate the gradient of weights and biases layer by layer, and update the weight matrix and bias vector of each layer.

[0061] S64: During the optimization process, the weighted coefficients of the spatiotemporal features are adjusted in real time according to the contribution of different spatiotemporal features to the error. The spatial weight coefficient θ is dynamically adjusted through error feedback. s and time weight coefficient θ t ;

[0062] S65: Through multiple rounds of training iterations, the error between the model prediction value and the actual value is continuously minimized, and all parameters in the adaptive spatiotemporal fusion model are gradually optimized.

[0063] Beneficial effects of the present invention:

[0064] The present invention, by introducing an adaptive weighting mechanism and combining graph convolutional networks and variational autoencoders, achieves the dynamic fusion of spatial and temporal features, fully considering the spatial dependencies between traffic nodes and the temporal evolution trend of traffic flow. By dynamically adjusting the weighting coefficients of the features, the present invention can flexibly balance the contribution of spatial and temporal features to the prediction results according to real-time traffic conditions, effectively improving the accuracy of traffic flow prediction. At the same time, the combination of deep neural networks and attention mechanisms ensures that the model can accurately capture important features in traffic flow when facing complex and dynamic traffic environments.

[0065] The present invention realizes adaptive optimization of model parameters through error feedback and back propagation algorithms. It can continuously adjust the weights, biases, and weighted coefficients of spatiotemporal characteristics according to the differences between the actual prediction results and real traffic data, ensuring high-precision predictions under complex and changeable traffic conditions. Through multiple rounds of training iterations, it has strong generalization ability and real-time performance, can effectively respond to complex traffic scenarios such as emergencies and road congestion, and significantly improve the efficiency of traffic flow management. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0067] Figure 1 A schematic diagram of a traffic flow prediction method according to an embodiment of the present invention;

[0068] Figure 2 The figure is a schematic diagram of the process of constructing an adaptive spatiotemporal fusion model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0070] It should be noted that references in the specification to "one embodiment," "an embodiment," "an exemplary embodiment," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment necessarily includes such specific features, structures, or characteristics. In addition, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0071] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0072] like Figure 1-Figure 2 As shown, a spatiotemporal correlation traffic flow prediction method based on deep learning includes the following steps:

[0073] S1: Based on the sensor network deployed at traffic nodes, traffic flow data and related external environment data (such as weather, road construction, emergencies, etc.) are collected in real time, and the initial traffic network topology structure containing traffic nodes and edges is established based on the collected data;

[0074] S2: Using the traffic flow data and historical data collected in S1, the dynamic weight of each node is calculated based on the node's traffic flow, geographical location, and road capacity. The weight is used to measure the traffic pressure and importance of different nodes;

[0075] S3: Based on the changes in node weights in S2, combined with real-time traffic conditions and related events, the weights of each edge in the traffic network are dynamically adjusted through the graph neural network to update the traffic network topology.

[0076] S4: Based on the dynamic traffic network topology constructed in step S3, a graph convolutional network is used to extract features of the spatial dependencies between traffic nodes. At the same time, a variational autoencoder is used to model the time series data and extract the temporal evolution characteristics of traffic flow data.

[0077] S5: The spatial features extracted in S4 are fused with the time series features through an adaptive weighting mechanism to construct an adaptive spatiotemporal fusion model;

[0078] S6: Based on the adaptive spatiotemporal fusion model in S5, real-time traffic data is input to perform traffic flow prediction, and the adaptive spatiotemporal fusion model parameters are optimized according to the difference between the predicted results and the actual traffic flow to improve the long-term prediction accuracy of the model.

[0079] S1 specifically includes:

[0080] S11: Deploy sensor networks at traffic nodes to collect traffic flow data, including vehicle speed, vehicle density, and road occupancy. The sensor networks collect traffic flow data at preset time intervals and upload it to the data center.

[0081] S12: Real-time collection of external environmental data, including weather information, road construction conditions, and accident occurrences. External environmental data is obtained by connecting with the meteorological monitoring system, road construction management system, and traffic accident alarm system to ensure the real-time and accuracy of the data;

[0082] S13: Preprocessing the data collected in steps S11 and S12, including removing abnormal data and formatting to ensure data integrity and consistency;

[0083] S14: Based on the processed data and the physical connection relationship between traffic nodes and roads, an initial traffic network topology structure of traffic nodes and edges is established. The topology structure uses traffic nodes as vertices and roads as edges to reflect the flow paths of traffic flow between different nodes. By deploying sensor networks at traffic nodes and collecting traffic flow data and external environment data in real time, the accuracy and timeliness of the data can be guaranteed. The established initial traffic network topology structure can effectively reflect the actual situation of traffic flow and provide a reliable data basis for subsequent dynamic traffic flow prediction.

[0084] S2 specifically includes:

[0085] S21: Obtain the real-time traffic flow data and historical traffic flow data collected in S1. The traffic flow data includes the vehicle traffic volume, vehicle speed, and road occupancy rate at the node, ensuring that the traffic flow data covers different time periods and traffic flow changes;

[0086] S22: Using the geographic location information of the transportation nodes, the spatial relationship of each node relative to other nodes is calculated. The spatial relationship is based on the distance between the transportation nodes, the connection mode and the direction of traffic flow, combined with the geographic characteristics of the nodes to form a spatial dependency matrix;

[0087] S23: Weight the road attributes of each node based on road capacity. Road capacity includes road width, number of lanes, traffic signal control, and road restrictions (such as one-way travel and speed limits). Specifically, weighted factors are used for comprehensive calculation to generate a preliminary weight for each node. Specifically, each factor corresponds to a different weight coefficient. The specific formula is as follows: Among them, PW i is the initial weight of the node; W i Indicates the width of the road connecting the nodes; L i Indicates the number of lanes on the road; T i Represents the traffic signal control time, and the reciprocal is taken to reflect that the greater the delay, the lower the weight; R i represents the road restriction coefficient, taking into account the impact of speed limits, one-way traffic, and other conditions on road traffic; α, β, γ, and δ are weighting coefficients, which are adjusted based on statistical analysis of historical data and actual conditions to ensure that the influence of each factor in the weight calculation is reasonably distributed;

[0088] S24: The data generated in S21, S22, and S23 are dynamically adjusted through a deep neural network model. The deep neural network model dynamically calculates the weight value of each node by combining and analyzing real-time traffic flow and historical data. The weight value is used to reflect the traffic pressure and importance of the node at the current moment, and is updated in real time as the traffic flow and environment change. The above steps comprehensively consider factors such as traffic flow, the geographical location of the node, and the road capacity, and use the deep neural network model to dynamically adjust the weight of each node. This can accurately reflect the dynamic changes of traffic nodes, ensure the timeliness and accuracy of weight calculation, and provide a more accurate data basis for subsequent traffic flow predictions.

[0089] S24 specifically includes:

[0090] S241: Segmenting the collected real-time traffic flow data and historical traffic flow data according to a predetermined time window to ensure clear traffic characteristics in each time period and provide smooth data input for subsequent processing;

[0091] S242: combining the spatial dependency matrix with the normalized preliminary node weight to form an input feature vector of the node. The feature vector includes the node's real-time traffic, historical traffic, spatial dependency, and preliminary weight, providing a comprehensive description of the node.

[0092] S243: Inputting the input feature vector into a preset deep neural network model. The deep neural network model includes several hidden layers. Each layer performs nonlinear transformation based on the output of the previous layer and the model weights and biases to form a new output.

[0093] S244: After multiple layers of nonlinear transformation, the final dynamic weight value of the node is generated. The dynamic weight value reflects the traffic pressure and importance of the node at the current moment and changes dynamically according to the input characteristics of the node;

[0094] S245: Standardizing the generated dynamic weight values ​​to ensure that the weight values ​​of different nodes are comparable, and generating standardized node dynamic weights.

[0095] The calculation steps for dynamically calculating the weight value of each node are as follows:

[0096] The real-time traffic flow data X collected in step S21 r (t) and historical traffic flow data X h (t) is used for input processing. First, the real-time data and historical data are segmented into time windows. The time window size is Δt. The processed data is: Among them, X r (t) represents the real-time traffic flow data of node i at time t, X h (t) represents the historical traffic flow data of node i, and Δt represents the size of the time window;

[0097] The spatial dependence matrix A generated in step S22 s The node preliminary weight PW generated in step S23 i Combine to form the input feature vector F i , the formula is as follows: Among them, A s is the spatial dependency matrix between transportation nodes;

[0098] The eigenvector F i Input deep neural network model, the model contains several hidden layers, the specific calculation process of the model is: H (l) =σ(W (l) ·H (l-1) +b (l) ), H (l) Represents the output vector of the lth layer; W (l) represents the weight matrix of the lth layer; b (l) represents the bias vector of the lth layer; H (l-1) is the output of the l-1th layer; σ is the activation function, using the ReLU activation function, and the calculation formula is: σ(x) = max(0, x), where x represents the input vector;

[0099] After passing through all hidden layers, the final dynamic weight value D of the output node i , represents the dynamic weight of node i at the current moment, and the calculation formula is: D i =W (L) ·H (L-1) +b (L) , where W (L) is the weight matrix of the output layer; b (L) is the bias vector of the output layer; H (L-1) is the output of the last hidden layer;

[0100] Finally, the dynamic weight D of the output i Perform standardization, and the standardized node weight D i ' is calculated by the following formula: Among them, D i ′ is the normalized weight of node i; D max is the maximum value of all node dynamic weights; D min The minimum value among all the dynamic weights of nodes; by inputting real-time traffic flow, historical data, spatial dependency matrix and preliminary weights into the deep neural network model, and undergoing nonlinear transformation of multi-layer neural networks, the weight of each node is finally dynamically calculated, ensuring the dynamic and accuracy of the node weight.

[0101] S3 specifically includes:

[0102] S31: Obtain the dynamic weight of each traffic node calculated in S2, as well as the real-time collected road condition information and related event data. The road condition information includes road congestion level and traffic accident status, and the related event data includes road construction status and weather changes.

[0103] S32: Based on the dynamic weights of the traffic nodes and the connection relationships between the nodes, an initial traffic network adjacency matrix is ​​constructed. The elements in the adjacency matrix represent the direct connection relationships between the nodes, and the element values ​​are initialized to 1;

[0104] S33: Mapping real-time traffic information and related event data into initial edge weights. The initial weights are quantified based on the actual road capacity, congestion level, and impact level. A higher value indicates a higher capacity.

[0105] S34: The dynamic weights of the nodes and the initial weights of the edges are used as input features and fed into the preset graph neural network model. The graph neural network model calculates the dynamic weights of each edge through information propagation and aggregation mechanisms, and uses multi-layer graph convolution operations to iteratively update the features of the nodes and edges.

[0106] S35: Using the dynamic weights of the edges calculated in S34, the traffic network topology is updated to generate an updated adjacency matrix so that the weights of each edge reflect the current traffic capacity of the roads in the traffic network and the impact of real-time events. By combining the node weights with real-time road conditions and event data through the graph neural network, the weights of each edge in the traffic network are dynamically adjusted, realizing the dynamic update of the traffic network topology, which can more accurately reflect the actual situation of traffic flow and provide data support with higher real-time and accuracy for traffic flow prediction.

[0107] S34 specifically includes:

[0108] S341: Initializing the feature value of each node. The node features include its dynamic weight, traffic flow data, and connection information with other nodes. At the same time, the initial weight value of each edge is initialized. The edge weight is determined by the connection relationship between nodes and real-time traffic information.

[0109] S342: In each layer of the graph convolution operation, a node receives information from its neighboring nodes. The features of the neighboring nodes are weighted according to the connection strength and weight between the nodes and then aggregated and transmitted to the target node.

[0110] S343: After receiving the information from the neighboring nodes, the target node aggregates the received information with its own eigenvalues. The aggregated eigenvalues ​​are transformed nonlinearly through an activation function to update the node's features.

[0111] S344: updating the edge weights, where the edge weights are calculated based on the updated features of the connected nodes, thereby reflecting the latest traffic conditions between the nodes;

[0112] S345: Repeat the above graph convolution operation in a multi-layer network. Each layer will update the feature values ​​of the nodes and edges, and finally output the features of the nodes and the dynamic weights of each edge. The dynamic weights reflect the traffic capacity and impact of each road in the traffic network at the current moment.

[0113] The specific steps to calculate the dynamic weight of each edge are as follows:

[0114] Initialize the node feature matrix H (0) , the initial features of each node are composed of the dynamic weight of the node and the information of the surrounding neighbor nodes, and the feature value of the node Including the node's traffic flow, historical weight and connectivity with other nodes;

[0115] Define the initial weight matrix E of the edge (0), where the initial weight value of each edge is based on the connectivity of the node and the relevant real-time traffic information. The initial weight is determined by information such as the traffic capacity and congestion status of the road section. In the l-th layer graph convolution operation, information propagation is performed, and node i receives information from its neighbor node j. The formula is as follows: in, represents the information received by node i from its neighbor nodes at layer l; N(i) is the set of neighbor nodes of node i; d i and d j are the degrees of node i and node j respectively (i.e. the number of connected edges); W (l) is the weight matrix of the lth layer; is the eigenvalue of neighbor node j at layer l-1; b (l) is the bias vector of the lth layer;

[0116] After the information is propagated, node i aggregates the received information with its own characteristics and updates the node’s characteristic value. The updated characteristic value Calculated by the following formula: Among them, σ is the activation function (usually ReLU), which performs nonlinear transformation on the updated features; is the weight matrix of the node itself;

[0117] After each layer of graph convolution operation is completed, the weight value of the edge is updated The edge weight connecting i and j is adjusted by aggregating the features of node i and node j. The formula is as follows: in, represents the weight of the edge between node i and node j in layer l; is the weight matrix of the l-th layer edge; Represents the concatenation operation of feature vectors;

[0118] After multi-layer graph convolution operations, the features of nodes and edges will be continuously updated iteratively, and finally the updated features of each node will be obtained. and dynamic weights of edges Where L is the total number of layers in the graph neural network, and the output edge weight is the final dynamic weight. The above steps use the graph neural network model to dynamically update the characteristics of nodes and edges using information propagation and aggregation mechanisms. This can more accurately capture the real-time status of the traffic network, especially in the face of changing road conditions and emergencies. It dynamically adjusts the weights in the traffic network and improves the accuracy of traffic flow prediction.

[0119] S4 specifically includes:

[0120] S41: Based on the dynamic traffic network topology structure constructed in step S3, define the node characteristics and edge characteristics in the traffic network. The node characteristics include traffic flow and road capacity, and the edge characteristics include the dynamic weights between nodes, forming the initial characteristic matrix of the traffic network.

[0121] S42: A graph convolutional network is used to extract spatial features from node and edge features. Specifically, by aggregating the information of each node and its neighboring nodes, the spatial dependency relationship between nodes in the traffic network is extracted to reflect the connectivity and mutual influence between traffic nodes. For example, in the convolution operation of the first layer, the feature of node i is calculated by aggregating the features of its neighboring nodes. The updated feature of node i in the first layer is expressed as: in, Represents the feature vector of node i in layer l; is the weight matrix of the lth layer (s represents the parameters related to spatial features); is the bias vector; N(i) is the set of neighbor nodes of node i; d i and d j Represent the degree of node i and node j respectively; σ is the activation function, usually the ReLU function;

[0122] S43: In each layer of graph convolutional network operation, the node features and the features of its neighboring nodes are weighted according to the dynamic weight of the edge, and the spatial features between the nodes are extracted through the convolution operation. The feature values ​​of the nodes are updated layer by layer, and finally the spatial feature vector F representing the global traffic network is generated. (L) ;

[0123] S44: Model the time evolution of traffic flow, use variational autoencoder (VAE) to process the time series data of traffic nodes, and encode the high-dimensional time series data into a low-dimensional latent variable space representation; let the low-dimensional latent variable be Z t , the specific formula is: Z t =μ t +∈ t ·σ t , where μ t is the mean output of the encoder; σ t is the standard deviation output of the encoder; ∈ t is the noise sampled from the standard normal distribution;

[0124] S45: In the latent variable space, the decoder of the variational autoencoder is used to reconstruct the time series data. The reconstruction loss of the decoder is used to extract the temporal variation characteristics of the traffic flow data to capture the short-term and long-term evolution trends of the traffic flow. The above steps extract the spatial dependencies between nodes through the graph convolutional network. Combined with the modeling of the time series by the variational autoencoder, it can comprehensively capture the spatial characteristics and temporal variation characteristics of the traffic flow, ensure that the prediction model can accurately predict the dynamic change trend of the traffic flow, and improve the accuracy and robustness of the prediction.

[0125] S5 specifically includes:

[0126] S51: Normalize the spatial features and temporal features extracted in S4 to ensure that the two features are in the same value range so that they can be reasonably weighted in the subsequent fusion process;

[0127] S52: Calculate the weight coefficient θ of spatial and temporal features based on the real-time dynamic situation of the current traffic network s and θ t , the weight coefficient is used to control the contribution of spatial and temporal features in the model, and is calculated using the following formula: Among them, θ s is the weight of spatial features; θ t is the weight of the time feature; s is the weighting parameter related to spatial features, λ t is a weighting parameter related to time features; λ is dynamically adjusted through the training process s and λ t , so that θ s and θ t Adaptive change;

[0128] S53: Perform weighted fusion of spatial features and temporal features according to the weights, and let the fused spatiotemporal feature vector be F st , the formula is: F st =θ s ·F (L) +θ t ·Z t , where F st Represents the fused spatiotemporal feature vector, which contains the spatial dependency information of traffic nodes and the time series evolution trend;

[0129] S54: Based on the spatiotemporal feature vector F fused in S53 st, an adaptive spatiotemporal fusion model is constructed through a deep neural network. The adaptive spatiotemporal fusion model adopts a multi-layer network structure, in which each layer performs weighted processing and nonlinear transformation according to the input features; the above steps dynamically fuse spatial features and temporal features through adaptive weighting, and construct an adaptive spatiotemporal fusion model through a deep neural network, which can effectively balance the impact of spatial and temporal factors on traffic flow, and improve the flexibility and accuracy of the prediction model when facing complex dynamic traffic conditions.

[0130] The construction of the adaptive spatiotemporal fusion model in S54 specifically includes:

[0131] S541: First, the fused spatiotemporal feature vector is input into the first layer of the deep neural network. The first layer performs a linear transformation on the input features and performs nonlinear mapping through an activation function to generate a new feature representation, ensuring that the spatiotemporal features are effectively processed.

[0132] S542: Input the features generated by the first layer into the second hidden layer, continue to perform linear and nonlinear processing on the features, extract the complex patterns and relationships in the spatiotemporal features layer by layer, and enhance the ability to capture spatiotemporal changes;

[0133] S543: Introducing the attention mechanism into deep neural networks. The attention mechanism dynamically adjusts the contribution of each feature to the final output by learning the importance weights of spatiotemporal features, ensuring that the model can adaptively handle different traffic scenarios.

[0134] S544: Using the attention weights to weight the features output by the second hidden layer, generating adaptively adjusted feature representations to optimize the fusion effect of spatiotemporal features;

[0135] S545: The weighted features are input to the output layer. The output layer generates a prediction result of future traffic flow based on the processed spatiotemporal features, and provides a specific prediction value of traffic flow in the future time period.

[0136] The specific steps to build an adaptive spatiotemporal fusion model are as follows:

[0137] First, the fused spatiotemporal feature vector F st Input the first layer of the deep neural network. The input of the first layer of the network is the spatiotemporal feature vector, which contains the adaptively weighted spatial feature F (L) and time feature Z t The calculation formula of the first layer is: H1=σ(W1·F st +b1), where H1 is the output of the first hidden layer; W1 is the weight matrix of the first layer; b1 is the bias vector of the first layer; σ(x) is the activation function;

[0138] Then, the output H1 of the first layer is input to the second hidden layer. The second layer continues to perform linear transformation and nonlinear processing on the input features, and the output is H2. The calculation formula of this layer is: H2 = σ(W2·H1+b2), where H2 is the output of the second hidden layer; W2 is the weight matrix of the second layer; b2 is the bias vector of the second layer;

[0139] Next, in order to enable the model to adaptively handle different traffic flow characteristics, the model introduces an attention mechanism. The attention mechanism dynamically adjusts the impact of features on the output by learning the importance weights of spatiotemporal features. The specific formula is: st =softmax(W a H2+b a ), where A st is the attention weight of spatiotemporal features; W a is the weight matrix of the attention layer; b a is the bias vector of the attention layer; softmax(x) is the normalized activation function that converts the output into a probability distribution;

[0140] Next, the attention weight A st Applied to the output of the second layer to generate the weighted feature representation H att , the formula is: H att =A st H2: The weighted feature representation can adaptively adjust the importance of different spatiotemporal features, ensuring that the model can flexibly respond to different traffic scenarios.

[0141] Finally, the weighted feature H att Input to the output layer to generate the predicted value of future traffic flow. The calculation formula of the output layer is: pred =W o ·H att +b o , where Y pred is the future traffic flow predicted by the model; W o is the weight matrix of the output layer; b o is the bias vector of the output layer; the above steps introduce the attention mechanism and combine it with the deep neural network to perform weighted processing and adaptive adjustment on the fused spatiotemporal features. The adaptive spatiotemporal fusion model constructed can more accurately capture the complex spatiotemporal variation characteristics in traffic flow, improve the model's predictive ability in complex traffic scenarios, and enhance its flexibility and generalization ability.

[0142] The optimization of the adaptive spatiotemporal fusion model parameters in S6 specifically includes:

[0143] S61: Calculate the difference between the model prediction result and the actual traffic flow. For the traffic flow at each moment, calculate the error between the predicted value output by the adaptive spatiotemporal fusion model and the actual measured value. Specifically, the mean square error (MSE) is used as the evaluation indicator. The calculation formula of the mean square error is:

[0144] Among them, Y pred,i is the predicted value of the model at the i-th moment; Y true,i is the actual traffic flow value at the i-th moment; n is the total number of samples;

[0145] S62: Based on the size of the error, the model optimizes the parameters through the backpropagation algorithm. For the weights and biases of each layer of the network, the gradient of the loss function with respect to these parameters is calculated to understand the contribution of these parameters to the error.

[0146] S63: Update the model parameters using the gradient descent algorithm. Starting from the output layer, calculate the gradient of weights and biases layer by layer, and update the weight matrix and bias vector of each layer. The update formula is as follows: Among them, W (l) is the weight matrix of the lth layer; b (l) is the bias vector of the lth layer; η is the learning rate, which is used to control the step size of parameter update; and are the gradients of the loss function with respect to weights and biases, respectively;

[0147] S64: During the optimization process, the weighted coefficients of the spatiotemporal features are adjusted in real time according to the contribution of different spatiotemporal features to the error. The spatial weight coefficient θ is dynamically adjusted through error feedback. s and time weight coefficient θ t , the update formula is:

[0148] S65: Through multiple rounds of training iterations, the error between the model's predicted value and the actual value is continuously minimized, and all parameters in the adaptive spatiotemporal fusion model are gradually optimized to ensure that the model can dynamically adjust and adapt to different traffic flow patterns in complex traffic scenarios; through the above steps, the weights, biases and spatiotemporal feature weighting coefficients in the adaptive spatiotemporal fusion model can be dynamically optimized, so that the model can gradually improve its prediction accuracy in actual traffic scenarios, ensuring that the model parameters can be automatically adjusted when facing complex and dynamically changing traffic environments, thereby improving the model's adaptability and generalization capabilities.

[0149] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0150] 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 principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A spatiotemporal correlation traffic flow prediction method based on deep learning, characterized in that: The following steps are involved: S1: Based on the sensor network deployed at the traffic nodes, the traffic flow data and related external environment data are collected in real time, and the initial traffic network topology structure including traffic nodes and edges is established based on the collected data; S2: Using the traffic flow data and historical data collected in S1, the dynamic weight of each node is calculated based on the node's traffic flow, geographical location, and road capacity; S3: According to the changes in node weights in S2, combined with real-time traffic conditions and related events, the weights of each edge in the traffic network are dynamically adjusted through the graph neural network to update the topological structure of the traffic network; S4: Based on the dynamic traffic network topology structure constructed in step S3, a graph convolutional network is used to extract features of the spatial dependency relationship between traffic nodes. At the same time, a variational autoencoder is used to model the time series data and extract the time evolution characteristics of traffic flow data. S5: The spatial features extracted in S4 are fused with the time series features through an adaptive weighting mechanism to construct an adaptive spatiotemporal fusion model; S6: Based on the adaptive spatiotemporal fusion model in S5, real-time traffic data is input to perform traffic flow prediction, and the adaptive spatiotemporal fusion model parameters are optimized according to the difference between the prediction result and the actual traffic flow.

2. According to claim 1, a method for predicting traffic flow based on spatiotemporal correlation based on deep learning is characterized in that: The S1 specifically includes: S11: deploying a sensor network at a traffic node to collect traffic flow data, wherein the traffic flow data includes vehicle speed, vehicle density, and road occupancy rate. The sensor network collects the data at a preset time interval and uploads the data to a data center; S12: Real-time collection of external environmental data, including weather information, road construction conditions, and accident conditions. The external environmental data is obtained by connecting with a meteorological monitoring system, a road construction management system, and a traffic accident alarm system; S13: pre-processing the data collected in steps S11 and S12, including removing abnormal data and formatting the data to ensure the integrity and consistency of the data; S14: According to the processed data, based on the physical connection relationship between the traffic nodes and the roads, an initial traffic network topology structure of traffic nodes and edges is established. The topology structure uses traffic nodes as vertices of the graph and roads as edges of the graph to reflect the flow path of traffic flow between different nodes.

3. The method for predicting traffic flow based on spatiotemporal correlation based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21: Acquire the real-time traffic flow data and historical traffic flow data collected in S1; S22: using the geographical location information of the traffic nodes, calculating the spatial relationship of each node relative to other nodes, wherein the spatial relationship is based on the distance, connection mode and traffic flow direction between the traffic nodes, combined with the geographical characteristics of the nodes to form a spatial dependency matrix; S23: weighting the road attributes of each node according to the road capacity, wherein the road capacity includes road width, number of lanes, traffic signal control, and road restrictions, and specifically using weighted factors for comprehensive calculation to generate a preliminary weight for each node; S24: The data generated in S21, S22, and S23 are dynamically adjusted through a deep neural network model. The deep neural network model dynamically calculates the weight value of each node by combining and analyzing real-time traffic flow and historical data. The weight value is used to reflect the traffic pressure and importance of the node at the current moment, and is updated in real time as the traffic flow and environment change.

4. The method for predicting traffic flow based on spatiotemporal correlation based on deep learning according to claim 3 is characterized in that: The S24 specifically includes: S241: Processing the collected real-time traffic flow data and historical traffic flow data in sections according to a predetermined time window to provide smooth data input for subsequent processing; S242: combining the spatial dependency matrix with the normalized preliminary node weight to form an input feature vector of the node, wherein the feature vector includes the real-time flow, historical flow, spatial dependency relationship and preliminary weight of the node; S243: Inputting the input feature vector into a preset deep neural network model, wherein the deep neural network model includes a plurality of hidden layers, and each layer performs nonlinear transformation according to the output of the previous layer and the model weight and bias to form a new output; S244: After multiple layers of nonlinear transformation, the final dynamic weight value of the node is generated; S245: Standardize the generated dynamic weight value to generate a standardized node dynamic weight.

5. The method for predicting traffic flow based on spatiotemporal correlation based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31: Obtain the dynamic weight of each traffic node calculated in S2, as well as the road condition information and related event data collected in real time, wherein the road condition information includes the road congestion degree and traffic accident conditions, and the related event data includes road construction conditions and weather changes; S32: constructing an initial traffic network adjacency matrix based on the dynamic weights of the traffic nodes and the connection relationships between the nodes, wherein the elements in the adjacency matrix represent the direct connection relationships between the nodes, and the element values ​​are initialized to 1; S33: Mapping the real-time traffic information and related event data into initial weights of edges, wherein the initial weights are quantified according to the actual traffic capacity, congestion level, and impact level of the road, and a higher value indicates a stronger traffic capacity; S34: The dynamic weight of the node and the initial weight of the edge are input as input features into a preset graph neural network model. The graph neural network model calculates the dynamic weight of each edge through information propagation and aggregation mechanisms, and uses multi-layer graph convolution operations to iteratively update the features of the nodes and edges. S35: Using the dynamic weights of the edges calculated in S34, the topological structure of the traffic network is updated to generate an updated adjacency matrix, so that the weight of each edge reflects the capacity of the roads in the current traffic network and the impact of real-time events.

6. The method for predicting traffic flow based on spatiotemporal correlation based on deep learning according to claim 5, characterized in that: The S34 specifically includes: S341: Initialize the feature value of each node, the node's features include its dynamic weight, traffic flow data, and connection information with other nodes, and initialize the initial weight value of each edge; S342: In each layer of the graph convolution operation, a node receives information from its neighbor nodes. The features of the neighbor nodes are weighted according to the connection strength and weight between the nodes and summarized and transmitted to the target node. S343: After receiving the information of the neighboring nodes, the target node aggregates the received information with its own feature value, and the aggregated feature value is nonlinearly transformed through an activation function, thereby updating the node's features; S344: updating the edge weights, where the edge weights are calculated based on the updated features of the connected nodes, thereby reflecting the latest traffic conditions between the nodes; S345: Repeat the above graph convolution operation in a multi-layer network. Each layer will update the feature values ​​of the nodes and edges, and finally output the features of the nodes and the dynamic weights of each edge.

7. The method for predicting traffic flow based on spatiotemporal correlation based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41: Based on the dynamic traffic network topology structure constructed in step S3, the node characteristics and edge characteristics in the traffic network are defined. The node characteristics include traffic flow and road capacity, and the edge characteristics include dynamic weights between nodes, so as to form an initial characteristic matrix of the traffic network. S42: Use graph convolutional networks to extract spatial features from node and edge features. Specifically, the spatial dependency relationship between nodes in the traffic network is extracted by aggregating the information of each node and its neighboring nodes. S43: In each layer of graph convolution network operation, the node features and the features of its neighboring nodes are weighted according to the dynamic weight of the edge, and the spatial features between the nodes are extracted through the convolution operation. The feature values ​​of the nodes are updated layer by layer, and finally the spatial feature vector F representing the global transportation network is generated. (L) ; S44: Model the time evolution of traffic flow, use variational autoencoders to process the time series data of traffic nodes, and encode the high-dimensional time series data into a low-dimensional latent variable space representation; let the low-dimensional latent variable be Z t , the specific formula is: Z t =μ t +∈ t ·σ t , where μ t is the mean output of the encoder; σ t is the standard deviation output of the encoder; ∈ t is the noise sampled from a standard normal distribution; S45: In the latent variable space, the decoder of the variational autoencoder is used to reconstruct the time series data. The reconstruction loss of the decoder is used to extract the temporal variation characteristics of the traffic flow data to capture the short-term and long-term evolution trends of the traffic flow.

8. The method for predicting traffic flow based on spatiotemporal correlation based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51: Standardize the spatial features and temporal features extracted in S4 to ensure that the two features are in the same value range; S52: Calculate the weight coefficient θ of spatial and temporal features based on the real-time dynamic situation of the current traffic network s and θ t , the weight coefficient is used to control the contribution of spatial features and temporal features in the model; S53: Perform weighted fusion of spatial features and temporal features according to the weights, and let the fused spatiotemporal feature vector be F st , the formula is: F st =θ s ·F (L) +θ t ·Z t , where F st Represents the fused spatiotemporal feature vector; S54: Based on the spatiotemporal feature vector F fused in S53 st , an adaptive spatiotemporal fusion model is constructed through a deep neural network. The adaptive spatiotemporal fusion model adopts a multi-layer network structure, in which each layer performs weighted processing and nonlinear transformation according to the input features.

9. The method for predicting traffic flow based on spatiotemporal correlation based on deep learning according to claim 8, characterized in that: The S54 specifically includes: S541: First, the fused spatiotemporal feature vector is input into the first layer of the deep neural network. The first layer of the network performs a linear transformation on the input feature and performs a nonlinear mapping through an activation function to generate a new feature representation. S542: Input the features generated by the first layer into the second hidden layer, continue to perform linear and nonlinear processing on the features, extract the complex patterns and relationships in the spatiotemporal features layer by layer, and enhance the ability to capture spatiotemporal changes; S543: Introducing the attention mechanism into the deep neural network, the attention mechanism dynamically adjusts the contribution of each feature to the final output by learning the importance weights of spatiotemporal features, ensuring that the model can adaptively handle different traffic scenarios; S544: using the attention weight to weight the features output by the second hidden layer, generating an adaptively adjusted feature representation, and optimizing the fusion effect of the spatiotemporal features; S545: The weighted features are input to the output layer, and the output layer generates a prediction result of future traffic flow based on the processed spatiotemporal features.

10. The method for predicting traffic flow based on spatiotemporal correlation based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61: Calculate the difference between the model prediction result and the actual traffic flow, and for the traffic flow at each moment, calculate the error between the prediction value output by the adaptive spatiotemporal fusion model and the actual measurement value; S62: Based on the size of the error, the model optimizes the parameters through the back propagation algorithm, and calculates the gradient of the loss function relative to these parameters for the weights and biases of each layer of the network; S63: Update the model parameters using the gradient descent algorithm, starting from the output layer, calculating the gradients of weights and biases layer by layer, and updating the weight matrix and bias vector of each layer; S64: During the optimization process, the weighting coefficients of the spatiotemporal features are adjusted in real time according to the contribution of different spatiotemporal features to the error. The spatial weight coefficient θ is dynamically adjusted through error feedback. s and the time weight coefficient θ t ; S65: Through multiple rounds of training iterations, the error between the model prediction value and the actual value is continuously minimized, and all parameters in the adaptive spatiotemporal fusion model are gradually optimized.

Images