A multi-scale controlled traffic flow prediction method and system for urban road network

Abstract

This invention discloses a multi-scale controlled traffic flow prediction method and system for urban road networks, belonging to the field of intelligent transportation and traffic information engineering technology. It addresses the technical problems in existing technologies, such as one-sided quantification of micro-node features, easy dilution of bottleneck features in multi-scale dimensionality reduction mapping, and severe mismatch between the spatiotemporal granularity of micro-control actions and macro-evolutionary trends. The method includes: estimating the traffic situation of the control subnet to obtain a meso-level subnet-level state vector; performing dimensionality reduction processing on the meso-level subnet-level state vector of the entire road network using subnet features related to gating mechanisms; aggregating meso-level subnet-level control features for micro-level traffic light parameter control; concatenating the meso-level subnet-level control features with spatial features after spatial congestion diffusion effects to obtain the current network hidden layer state; and performing meso-scale multi-step prediction and macro-scale full-network prediction on the current network hidden layer state to obtain the overall macro-level traffic situation index.

Description

Technical Field

[0001] This application relates to the fields of intelligent transportation and traffic information engineering technology, and in particular to a multi-scale controlled traffic flow prediction method and system for urban road networks. Background Technology

[0002] With the rapid development of next-generation traffic information collection technologies such as video surveillance and the Internet of Things, the in-depth mining and systematic application of massive urban traffic big data resources have provided crucial data support for the refined control and management of urban traffic. However, urban road traffic systems inherently possess characteristics of discontinuity, uncontrollability, unmeasurability, and strong nonlinearity. In the research and system development process targeting the dynamic evolution mechanism of traffic flow in large-scale urban road networks, existing technologies generally suffer from the following technical shortcomings: (1) Insufficient decoupling of multidimensional topology and traffic characteristics of nodes. Most existing prediction methods treat each intersection equally or only perform rough data statistics. There is a lack of a method that comprehensively quantifies and calculates the physical complexity, topological connectivity and time-varying importance of nodes based on the physical topology of the road network and dynamic travel OD information. This makes it impossible for the system to accurately locate the key "eyes" nodes that dominate congestion spread and cascading failure in the road network.

[0003] (2) Local bottleneck conditions are easily "diluted" and masked in dimensionality reduction mapping. Traditional dimensionality reduction mapping mechanisms that transform micro-node conditions into macro-road network conditions often use simple arithmetic averages or static weighted summations. In actual traffic operations, this linear projection method is very likely to cause severe oversaturation or queuing overflow characteristics of local core hub nodes, which are diluted and masked by the normal data of most smooth nodes in the overall road network, making it impossible for management departments to capture the early signs of local cascading collapses in a timely manner at the macro level.

[0004] (3) There is a serious mismatch between the spatiotemporal granularity of micro-control actions and macro-evolutionary trends. Traffic signal control schemes for single nodes (intersections) (such as cycle, green ratio, and phase difference) are micro-level, multi-directional time allocation parameters, while the road network status is a macro-level, regional spatial trend indicator. Existing spatiotemporal network models lack an alignment and embedding mechanism for multi-scale control features, and cannot seamlessly integrate single-point signal control actions as disturbance parameters into the dynamic evolution equation of macro-traffic flow. Therefore, it is difficult to pre-determine the spatiotemporal intervention and dissipation effect of the timing strategy on the future traffic flow status of the entire network before the actual timing scheme is issued. Summary of the Invention

[0005] This application provides a multi-scale controlled traffic flow prediction method and system for urban road networks to solve the following technical problems: the quantification of micro-node features in the prior art is one-sided, the multi-scale dimensionality reduction mapping easily dilutes bottleneck features, and the spatiotemporal granularity of micro-control actions and macro-evolutionary trends is seriously mismatched.

[0006] The embodiments of this application adopt the following technical solutions: On the one hand, embodiments of this application provide a multi-scale controlled traffic flow prediction method for urban road networks, including: calculating multi-dimensional indicators related to node complexity, node connectivity, and node importance for the graph network structure corresponding to the urban road network to obtain multi-dimensional feature indicators of nodes, and generating dynamic node weights based on the traffic state of the entire road network; estimating the traffic state of the control subnetwork based on the spatial homogeneity of traffic state according to the node dynamic weights, and obtaining a meso-level subnetwork state vector for traffic state estimation; and performing subnetwork feature dimensionality reduction processing under relevant gating mechanisms on the meso-level subnetwork state vector in the entire road network to obtain a global macro-level control subnetwork. The system analyzes network state indicators; based on the node dynamic weights and signal control scheme vectors, it aggregates meso-level sub-network control features for micro-level traffic light parameter control; based on the meso-level sub-network state vectors in the global macro-network state indicators, it concatenates the meso-level sub-network control features with spatial features after spatial congestion diffusion effects, and obtains the current network hidden layer state based on gated cyclic units; through the sub-network routing gating weights in the global macro-network state indicators, it performs meso-scale multi-step prediction and macro-scale whole-network prediction on the current network hidden layer state to obtain the overall macro-level traffic situation indicators of the entire road network in the graph network structure.

[0007] This application's embodiments accurately anchor key bottleneck nodes that dominate the spread of congestion in road networks by constructing a node phase conflict topological entropy model and a queuing overflow cascade risk index. Simultaneously, it introduces an adaptive dynamic partitioning mechanism for control subnets, a generalized p-norm nonlinear adaptive pooling operator, and a Gumbel-Softmax dynamic routing gating mechanism to open up a three-layer, multi-scale adaptive dimensionality reduction mapping channel between micro-nodes, meso-subnets, and the macro-network, thereby focusing on solving the dynamic evolution mechanism of traffic flow states in large-scale road networks. Furthermore, it utilizes a collaborative update mechanism between spatial domain graph convolutional networks and temporal domain recurrent neural networks, combined with a hierarchical aggregation method from micro-level traffic light parameters to meso- and macro-level dimensions, to seamlessly integrate single-point control timing schemes into the evolutionary deduction of meso- and macro-level features, ultimately achieving high-precision rolling evolution prediction of the future discrete-time step state of the road network under controlled conditions.

[0008] In one feasible implementation, a multi-dimensional index calculation of node complexity, node connectivity, and node importance is performed on the graph network structure corresponding to the urban road network to obtain multi-dimensional node feature indicators, and dynamic node weights based on the traffic status of the entire road network are generated. Specifically, this includes: quantifying and scoring the number of approach lanes, the number of phase settings, and the number of traffic conflict points at road intersections in the graph network structure to obtain the node complexity; determining the node connectivity based on the number of road segments and traffic capacity between nodes in the graph network structure; calculating the number of shortest paths passing through the current node and the actual traffic flow carried by the current node according to the road network OD matrix of the graph network structure to obtain the node importance; integrating the node complexity, node connectivity, and node importance into multi-dimensional features to obtain the multi-dimensional node feature indicators; and performing nonlinear mapping calculation on the normalized multi-dimensional node feature indicators using a Softmax function with a built-in time window to obtain dynamic node weights based on the traffic status of the entire road network.

[0009] In one feasible implementation, before estimating the traffic situation of the control subnet based on the spatial homogeneity of the traffic situation according to the node dynamic weights, and obtaining the meso-level subnet state vector for traffic situation estimation, the method further includes: according to Obtain the adjacency matrix of the static physical space view. ;in, Represents a node With nodes The shortest path length between; based on The adjacency matrix of the dynamic traffic semantic view is obtained. ;in, for Time Node The dynamic travel demand feature vector, and The elements in the table represent the real-time traffic from node i to the other nodes; To quantify the dynamic correlation between node i and node j in the direction of travel flow; To quantify the dynamic correlation between node i and node m in the direction of travel flow; For smoothing coefficients; according to The adjacency matrix of the two views is obtained. ;in, These are the weighting coefficients; according to ,get Time Node With nodes Traffic situation affinity matrix between ;in, For nodes With nodes exist The Euclidean distance between the eigenvectors at time step; It is based on the inherent variance of normalized traffic state fluctuations statistically derived from historical data of the entire road network; through spectral clustering algorithm, the traffic situation affinity matrix is ​​subjected to real-time graph trimming, and the global road network containing multiple nodes in the graph network structure is dynamically divided into multiple sets of non-overlapping control subnets.

[0010] In one feasible implementation, based on the node dynamic weights, traffic situation estimation is performed on the control subnets based on the spatial homogeneity of traffic situation to obtain a meso-level subnet-level state vector for traffic situation estimation. Specifically, this includes: based on each control subnet in the set of control subnets, according to... This yields the meso-level subnet-level state vector used to estimate the traffic situation of each control subnet. ;in, It represents the Hadamardi (or Hadama) stack; It is a nonlinear exponent that adaptively adjusts according to the overall congestion level of the control subnet; The dynamic weight of the node; Let be the three-dimensional feature vector at time t, node i, used to represent the traffic state vector; Let be the set of control subnets, and let contain k control subnets.

[0011] In one feasible implementation, the mesoscopic sub-network level state vectors in the entire network are subjected to sub-network feature dimensionality reduction processing under a relevant gating mechanism to obtain global macroscopic network state indices, specifically including: according to The original scores of each control subnet's contribution to the risk of network-wide paralysis are obtained. ;in, This refers to the state vector at the mesoscopic subnet level; These are parameters of a multilayer perceptron; It is a multilayer perceptron; according to Obtain the subnet routing gating weights ;in, This represents random noise that follows a standard Gumbel distribution; These are the annealing temperature parameters; To control the number of subnets, n is a mathematical constant; It is random noise; according to The global macroscopic network state index is obtained by dimensionality reduction and aggregation of the control subnet features.

[0012] In one feasible implementation, based on the node dynamic weights and signal control scheme vectors, meso-level subnet-level control features for micro-level traffic light parameter control are aggregated, specifically including: according to , obtain node exist The signal control scheme vector at time t is ;in, This refers to the traffic light signal cycle; For the current road intersection Real-time green light ratio for each independent release phase; based on The meso-level subnet-level control characteristics are obtained by aggregating the control actions of all nodes within the control subnet. ;in, To control the set of subnets; The dynamic weight of the node; according to This yields the macroscopic global control matrix, which is formed by aggregating the control characteristics of all control subnets; where, To control the number of subnets; k is a mathematical constant; To control the subnet routing gating weights of the subnet.

[0013] In one feasible implementation, before concatenating the meso-level control features with the spatial features resulting from the spatial congestion diffusion effect based on the meso-level sub-network state vector in the global macro-network state index, the method further includes: according to The adjacency matrix of the meso-adaptive subnet is obtained. ;in, and Control subnets state vectors and control subnets at the meso-level The mesoscopic subnet-level state vector; This indicates vector concatenation; This indicates the control subnet within the control subnet set. Geometric center and control subnet The Euclidean distance between the geometric centers, The learnable topology mapping weight matrix is ​​represented; based on the meso-adaptive subnet adjacency matrix, the control subnet set, and the control subnet connection edges, a meso-level dynamic interactive topology graph is constructed.

[0014] In one feasible implementation, based on the meso-level subnet-level state vector in the global macro-network state index, the meso-level subnet-level control features are concatenated with the spatial features after the spatial congestion diffusion effect, and the current network hidden layer state is obtained based on a gated cyclic unit. Specifically, this includes: according to... The spatial feature vector after the spatial congestion diffusion effect is obtained. ;in, This is the spatial feature transformation matrix; It is a non-linear activation function; This refers to the mesoscopic adaptive subnet adjacency matrix in the mesoscopic subnet hierarchical dynamic interaction topology graph. and Control subnets state vectors and control subnets at the meso-level The mesoscopic subnet-level state vector; To control the number of subnets; according to The updated current network hidden layer state is obtained. ;in, This refers to the hidden layer state of the historical network. The meso-level subnet-level control features are used for micro-level traffic light parameter control.

[0015] In one feasible implementation, by using the subnet routing gating weights in the global macro network state index, the current network hidden layer state is predicted using mesoscale multi-step prediction and macroscale whole-network prediction to obtain the overall macro traffic situation index of the entire road network in the graph network structure. Specifically, this includes: based on... This yields the future mesoscale subnet-level state vector for each control subnet at multiple future time steps based on multi-step prediction at the mesoscale; where, This refers to the current hidden layer state of the network; For the future The time-based meso-level subnet-level state vector; by controlling the subnet routing gating weights of the subnets, the future meso-level subnet-level state vector of each controlled road network is weighted and summed based on the traffic situation results to obtain the overall macro-level traffic situation index based on the entire road network.

[0016] On the other hand, embodiments of this application also provide a multi-scale controlled traffic flow prediction system for urban road networks, characterized in that the multi-scale controlled traffic flow prediction system for urban road networks includes instructions executable by at least one processor, so that at least one processor can execute a multi-scale controlled traffic flow prediction method for urban road networks as described in any of the above embodiments.

[0017] This application provides a multi-scale controlled traffic flow prediction method and system for urban road networks. Compared with the prior art, the embodiments of this application have the following beneficial technical effects: This application's embodiments accurately anchor key bottleneck nodes that dominate the spread of congestion in road networks by constructing a node phase conflict topological entropy model and a queuing overflow cascade risk index. Simultaneously, it introduces an adaptive dynamic partitioning mechanism for control subnets, a generalized p-norm nonlinear adaptive pooling operator, and a Gumbel-Softmax dynamic routing gating mechanism to open up a three-layer, multi-scale adaptive dimensionality reduction mapping channel between micro-nodes, meso-subnets, and the macro-network, thereby focusing on solving the dynamic evolution mechanism of traffic flow states in large-scale road networks. Furthermore, it utilizes a collaborative update mechanism between spatial domain graph convolutional networks and temporal domain recurrent neural networks, combined with a hierarchical aggregation method from micro-level traffic light parameters to meso- and macro-level dimensions, to seamlessly integrate single-point control timing schemes into the evolutionary deduction of meso- and macro-level features, ultimately achieving high-precision rolling evolution prediction of the future discrete-time step state of the road network under controlled conditions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart of a multi-scale controlled traffic flow prediction method for urban road networks provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of a multi-scale controlled traffic flow prediction device for urban road networks, provided as an embodiment of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0020] This application provides a multi-scale controlled traffic flow prediction method for urban road networks, such as... Figure 1 As shown, the multi-scale controlled traffic flow prediction method for urban road networks specifically includes steps S101-S106: S101. Calculate multi-dimensional indicators related to node complexity, node connectivity, and node importance for the graph network structure corresponding to the urban road network, obtain multi-dimensional feature indicators of nodes, and generate dynamic node weights based on the traffic status of the entire road network.

[0021] It should be noted that the urban road network is first abstracted as a graph network structure. , where the set of nodes Represents the urban road network An intersection; a set of edges Indicates the physical road segments connecting nodes; Given an adjacency matrix, if the nodes With nodes Physically connected, then ,otherwise Additionally, the definition The travel demand matrix at any time is , of which matrix elements Indicates starting from the origin To the finish line Traffic flow.

[0022] Specifically, it is necessary to quantify and score the number of approach lanes, the number of phase settings, and the number of traffic conflict points at road intersections in the graph network structure to obtain the node complexity.

[0023] In one embodiment, a quantitative score is calculated based on the number of approach lanes, the number of phase settings, and the number of traffic conflict points at the intersection; this is mainly used to characterize the physical topology and spatial channelization features of the intersection itself; the calculation formula is as follows: ;in, This represents the total number of approach lanes at the intersection. This represents the number of signal phases. This represents the number of internal traffic conflict points. , , These are the weighting coefficients.

[0024] Furthermore, based on the number of road segments and traffic capacity between nodes in the graph network structure, the node connectivity is determined.

[0025] In one embodiment, the degree centrality of a node is calculated based on graph theory, i.e., the number of road segments directly connected to the node and their capacity; this is mainly used to measure the node's connectivity with the surrounding road network; the calculation formula is: ;in, Represents nodes The set of directly adjacent nodes. Represents a node To the node Theoretical accessibility.

[0026] Furthermore, based on the road network OD matrix of the graph network structure, the number of shortest paths passing through the current node and the actual traffic flow are calculated to obtain the node importance.

[0027] In one embodiment, the road network OD matrix is ​​combined with The number of shortest paths passing through a node (between centrality) and the actual traffic flow it carries are calculated to reflect its position in the global travel chain of the road network. The calculation formula is as follows: ;in, Indicates from node To the node The number of shortest paths, express Passing through nodes The number of paths, express Time Node To the node Traffic.

[0028] Furthermore, node complexity, node connectivity, and node importance are integrated into a multi-dimensional feature to obtain a multi-dimensional feature index for nodes.

[0029] Furthermore, by using the Softmax function with its built-in time window, nonlinear mapping calculations are performed on the normalized multidimensional feature indicators of the nodes to obtain the dynamic weights of the nodes based on the traffic conditions of the entire road network.

[0030] In one embodiment, based on the three metrics calculated above, node weights also need to be calculated. This application uses a Softmax function with a time window to perform a nonlinear mapping function to obtain the node weights. exist Dynamic weights of nodes based on the real-time traffic conditions of the entire road network The specific calculation formula is as follows: and ;in, , , Let represent the normalized node complexity, node connectivity, and node importance, respectively. , , The feature fusion weights are obtained from the model learning.

[0031] S102. Based on the dynamic weights of the nodes, the traffic situation is estimated by calculating the control subnet based on the spatial homogeneity of the traffic situation, and the meso-level state vector for traffic situation estimation is obtained.

[0032] It should be noted that, Time, node The traffic state vector is 3D feature vector In this application, ,Right now ,in, Indicates average delay. Indicates the queue length. Indicates saturation.

[0033] Specifically, to avoid distortion of macroscopic indicators due to the spatiotemporal heterogeneity of a large-scale road network, this application does not directly aggregate all nodes into the overall network state. Instead, it dynamically divides the road network into several control subnetworks with "homogeneous internal traffic states and clear boundary characteristics." That is, Represents the adjacency matrix of a static physical space view, based on Obtain the adjacency matrix of the static physical space view. .in, Represents a node With nodes The shortest path length between them.

[0034] Furthermore, express A dynamic traffic semantic view adjacency matrix for each moment. It is used to capture implicit synchronous congestion associations that are not physically adjacent but share the same commuter traffic flow. Its calculation method is as follows: 1) Assume Time Node The dynamic travel demand feature vector is Its composition is , of which elements This indicates that at this moment, the node... Flow to Node Real-time traffic.

[0035] 2) Calculate the cosine similarity between nodes. First, quantize the nodes. With nodes Dynamic correlation in travel flows: .

[0036] 3) Spatial probability normalization based on Softmax To eliminate the influence of the absolute size of global traffic and ensure probabilistic comparability of semantic relevance across the entire network, the Softmax function is used to row-normalize the similarity matrix, resulting in the final dynamic semantic adjacency matrix elements. That is, according to The adjacency matrix of the dynamic traffic semantic view is obtained. .in, for Time Node The dynamic travel demand feature vector, and The elements in the table represent the real-time traffic from node i to the other nodes; To quantify the dynamic correlation between node i and node j in the direction of travel flow; To quantify the dynamic correlation between node i and node m in the direction of travel flow; This is the smoothing coefficient.

[0037] Furthermore, This is a dual-view fusion adjacency matrix. That is, based on... The adjacency matrix of the two views is obtained. ;in, These are the weighting coefficients.

[0038] Further calculation Time Node With nodes Traffic situation affinity matrix between That is: according to ,get Time Node With nodes Traffic situation affinity matrix between .in, For nodes With nodes exist The Euclidean distance between the eigenvectors at time step; It is the inherent variance of normalized traffic condition fluctuations, which is statistically derived from historical data of the entire road network.

[0039] Furthermore, it is necessary to use spectral clustering algorithms to perform real-time graph trimming on the traffic situation affinity matrix and dynamically divide the global road network containing multiple nodes into a set of multiple non-overlapping control subnets.

[0040] In one embodiment, a spectral clustering algorithm can be used to analyze the traffic situation affinity matrix. Perform real-time graph cropping to transform the global road network containing N nodes. Dynamic subdivision A set of mutually non-overlapping control subnets: .

[0041] Furthermore, in each divided control subnet Internally, using arithmetic averages to calculate mesoscopic traffic conditions can obscure local extreme congestion disturbances in spatially heterogeneous traffic flows, making it impossible for the system to effectively identify abnormal fluctuations in the "weakest link" within the subnet. Therefore, to accurately capture bottleneck intersections where queuing overflow begins within the subnet, this application employs an adaptive generalized p-norm nonlinear pooling operator to calculate the subnet. meso-level subnet-level state vector Based on each control subnet in the set of control subnets, according to This yields the meso-level subnet-level state vector used to estimate the traffic situation of each control subnet. .in, It represents the Hadamardi (or Hadama) stack; It is a nonlinear exponent that adaptively adjusts according to the overall congestion level of the control subnet; Dynamic weights for nodes; Let be the three-dimensional feature vector at time t, node i, used to represent the traffic state vector; Let K be a set of control subnets, containing k control subnets.

[0042] In one embodiment, It is a nonlinear index that adaptively adjusts according to the overall congestion level of the subnet, and its calculation formula is: When all intersections within the subnet are open, the saturation is... Very low, then Approximately equal to 1, when a core intersection within the subnet experiences severe oversaturation, the saturation level is... Approximately equal to 1, then Increase.

[0043] S103. Perform dimensionality reduction processing on the sub-network features of the meso-level sub-network in the entire network under the relevant gating mechanism to obtain the global macro-network state index.

[0044] Specifically, a multilayer perceptron is first used to calculate the raw score of each subnetwork's contribution to the overall network's risk of failure. Then according to The original scores of each control subnet's contribution to the risk of network-wide paralysis are obtained. .in, This represents the state vector at the mesoscopic subnet level. These are parameters of a multilayer perceptron; It is a multilayer perceptron.

[0045] Furthermore, in order to achieve discretized outlier focusing on high-risk subnets while maintaining end-to-end differentiability of the model, Gumbel-Softmax is used to calculate the subnet routing gating weights. That is: according to Obtain the subnet routing gating weights .in, This represents random noise that follows a standard Gumbel distribution; These are the annealing temperature parameters; To control the number of subnets, n is a mathematical constant; It is random noise.

[0046] Furthermore, subnet features are dimensionality-reduced and aggregated into global macro-network state indicators through a network-wide gating matrix. according to This yields a global macroscopic network state index formed by dimensionality reduction and aggregation of the control subnet features.

[0047] S104. Based on the node dynamic weights and signal control scheme vectors, aggregate the meso-level subnet-level control features for micro-level traffic light parameter control.

[0048] Specifically, nodes exist The signal control scheme vector at time t is In this application, its specific physical meaning is determined by the signal phase structure of the intersection. According to... , obtain node exist The signal control scheme vector at time t is .in, This refers to the traffic light signal cycle; For the current road intersection Real-time green light ratio for each independent release phase.

[0049] Furthermore, using the node dynamic weights calculated above... Subnet The control actions of all internal nodes are aggregated into meso-level subnet control characteristics, thereby realizing the meso-level subnet control tensor. The aggregation. That is, based on This yields the meso-level subnet-level control characteristics, which are the aggregated control actions of all nodes within the control subnet. .in, To control the set of subnets; For the dynamic weight of the node.

[0050] Furthermore, the subnet routing gating weights calculated using the Gumbel-Softmax algorithm described above are used. The control characteristics of all subnets in the entire network are aggregated into a macro-level global control matrix, thereby realizing a macro-level network-wide control tensor. The aggregation. That is, based on This yields the macroscopic global control matrix, which is formed by aggregating the control characteristics of all control subnets. Among them, To control the number of subnets; k is a mathematical constant; To control the subnet routing gating weights of the subnet.

[0051] S105. Based on the meso-level subnet-level state vector in the global macro-network state index, the meso-level subnet-level control features and the spatial features after the spatial congestion diffusion effect are concatenated, and the current network hidden layer state is obtained based on the gated cyclic unit.

[0052] Specifically, in order to characterize the macroscopic traffic flow transfer between subnets, the system, based on the division... Each subnet is used to construct a meso-level subnet topology diagram. Thus, the adjacency matrix of the meso-adaptive subnet is defined. Then according to The adjacency matrix of the meso-adaptive subnet is obtained. .in, and Control subnets state vectors and control subnets at the meso-level The mesoscopic subnet-level state vector; This indicates vector concatenation; This indicates the control subnet within the control subnet set. Geometric center and control subnet The Euclidean distance between the geometric centers, This represents the learnable topology mapping weight matrix. Finally, based on the meso-adaptive subnet adjacency matrix, the control subnet set, and the control subnet connection edges, a dynamic interactive topology graph of the meso-subnet hierarchy is constructed.

[0053] Furthermore, regarding the control subnet In terms of traffic situation Based on this, the state information of neighboring subnets is collected, and the spatial feature vector after the spatial congestion diffusion effect is calculated. That is: according to The spatial feature vector after the spatial congestion diffusion effect is obtained. .in, This is the spatial feature transformation matrix; It is a non-linear activation function; This refers to the mesoscopic adaptive subnet adjacency matrix in the mesoscopic subnet hierarchical dynamic interactive topology graph. and Control subnets state vectors and control subnets at the meso-level The mesoscopic subnet-level state vector; To control the number of subnets. That is, subnets... The current state of a network consists of its own state and the adjacency matrix of its nearest neighbor subnets. It is determined by the incoming traffic flow.

[0054] Furthermore, in order to forcibly introduce the active intervention effect of the traffic light timing scheme into the feature evolution, the system will incorporate the aforementioned spatial features. Mesoscale subnet-level control characteristics aligned with the steps above Feature concatenation is performed, and the data is input into the gated recurrent unit. The current hidden layer state is then updated by incorporating historical memory. It is also necessary to act promptly according to Get the updated current network hidden layer state. .in, This refers to the hidden layer state of the historical network. This refers to the meso-level subnet-level control characteristics used for micro-level traffic light parameter control. When the traffic light timing scheme... When changes occur, the current hidden layer state of the network It will spontaneously evolve in the direction of congestion dissipation.

[0055] S106. By using the subnet routing gating weights in the global macro network state index, perform meso-scale multi-step prediction and macro-scale whole-network prediction on the current network hidden layer state to obtain the overall macro traffic situation index of the entire road network in the graph network structure.

[0056] It should be noted that the network hidden layer state, which integrates spatiotemporal features and control actions, is obtained. Subsequently, the system directly achieves discrete multi-step rolling prediction outputs at both the meso- and macro-scales through a fully connected multilayer perceptron.

[0057] Specifically, the macroscopic situation of each subnet is directly mapped and output using a multilayer perceptron for the next 1 to T time steps: according to This yields the future mesoscale subnet-level state vector for each control subnet at multiple future time steps based on multi-step prediction at the mesoscale. Wherein, This represents the current state of the network's hidden layer. For the future The mesoscopic subnet-level state vector of time quantity.

[0058] Furthermore, by controlling the subnet routing gating weights of the subnets, the future meso-level state vectors of each controlled road network are subjected to weighted summation of relevant traffic situation results, and finally the overall macro-level traffic situation index based on the entire road network is obtained.

[0059] In one embodiment, at each future prediction time step (assuming ), ,in, The subnet routing gating weights calculated in the above steps are still used. The prediction results of each control subnet are weighted and summed to obtain the final overall macro traffic situation index for the entire network. : .

[0060] In addition, embodiments of this application also provide a multi-scale controlled traffic flow prediction system for urban road networks, such as... Figure 2 As shown, the multi-scale controlled traffic flow prediction system 200 for urban road networks can be executed by at least one processor 201 using instructions 202, enabling at least one processor 201 to perform the following: Multidimensional indicators related to node complexity, node connectivity, and node importance are calculated for the graph network structure corresponding to the urban road network to obtain multidimensional feature indicators of nodes, and dynamic node weights based on the traffic status of the entire road network are generated. Based on the dynamic weights of the nodes, the traffic situation is estimated by calculating the control subnet based on the spatial homogeneity of the traffic situation, and a meso-level subnet state vector for traffic situation estimation is obtained. The meso-level sub-network state vectors in the entire road network are subjected to sub-network feature dimensionality reduction processing under relevant gating mechanisms to obtain global macro-network state indices. Based on the node dynamic weights and signal control scheme vectors, meso-level subnet-level control features for micro-level traffic light parameter control are aggregated. Based on the meso-level subnet-level state vector in the global macro-network state index, the meso-level subnet-level control features and the spatial features after the spatial congestion diffusion effect are concatenated, and the current network hidden layer state is obtained based on the gated cyclic unit. By using the subnet routing gating weights in the global macro network state index, the current network hidden layer state is predicted in multiple steps at the meso scale and in the entire network at the macro scale, resulting in the overall macro traffic situation index of the entire road network in the graph network structure.

[0061] This application's embodiments accurately anchor key bottleneck nodes that dominate the spread of congestion in road networks by constructing a node phase conflict topological entropy model and a queuing overflow cascade risk index. Simultaneously, it introduces an adaptive dynamic partitioning mechanism for control subnets, a generalized p-norm nonlinear adaptive pooling operator, and a Gumbel-Softmax dynamic routing gating mechanism to open up a three-layer, multi-scale adaptive dimensionality reduction mapping channel between micro-nodes, meso-subnets, and the macro-network, thereby focusing on solving the dynamic evolution mechanism of traffic flow states in large-scale road networks. Furthermore, it utilizes a collaborative update mechanism between spatial domain graph convolutional networks and temporal domain recurrent neural networks, combined with a hierarchical aggregation method from micro-level traffic light parameters to meso- and macro-level dimensions, to seamlessly integrate single-point control timing schemes into the evolutionary deduction of meso- and macro-level features, ultimately achieving high-precision rolling evolution prediction of the future discrete-time step state of the road network under controlled conditions.

[0062] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0063] The foregoing has described specific embodiments of this application. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0064] The above description is merely an embodiment of this application and is not intended to limit this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of this application should be included within the specification of this application.

Claims

1. A method for multi-scale controlled traffic flow prediction for urban road networks, characterized in that, The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises:

2. The method of claim 1, wherein, The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises:

3. The method of claim 1, wherein, The method comprises: According to , a static physical space view adjacency matrix is obtained; wherein, represents the shortest path length between node and node ; According to , a dynamic traffic semantic view adjacency matrix is obtained; wherein, is the dynamic travel demand feature vector of the node at the moment, and the element in the matrix represents the real-time flow from node i to the remaining nodes; is the dynamic correlation between node i and node j in the travel direction; is the dynamic correlation between node i and node m in the travel direction; is the smoothing coefficient; According to , a dual-view fusion adjacency matrix is obtained; wherein, is a weight coefficient; according to ,get Time Node With nodes Traffic situation affinity matrix between ;in, For nodes With nodes exist The Euclidean distance between the eigenvectors at time step; It is the inherent variance of normalized traffic condition fluctuations, statistically derived from historical data of the entire road network. The method comprises:

4. The method of claim 3, wherein, The method comprises: Based on each control subnetwork in the control subnetwork set, according to , the meso-subnetwork level state vector for estimating the traffic situation of each control subnetwork is obtained ; wherein, represents Hadamard product; is a nonlinear index that is adaptively adjusted according to the overall congestion degree of the control subnetwork; is the node dynamic weight; is a three-dimensional feature vector for representing the traffic state vector at time t, node i; is the control subnetwork set, and contains k control subnetworks.

5. The method of claim 1, wherein, The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method According to , an original score of contribution of each control subnetwork to the risk of paralysis of the whole network is obtained ; wherein, is the mesoscopic subnetwork level state vector; is a multi-layer perception parameter; is a multi-layer perception; According to , the subnet routing gate weight is obtained; wherein, represents a random noise subject to a standard Gumbel distribution; is an annealing temperature parameter; is used to control the number of subnets, and n is a mathematical constant; is a random noise; According to , the global macro network state indicator aggregated from the control subnet features is obtained.

6. The method of claim 1, wherein, Based on the node dynamic weight and the signal control scheme vector, a mesoscopic subnet level control feature for microcosmic traffic light parameter control is aggregated, specifically including: According to , the node The signal control scheme vector at the moment is ; wherein, The red light signal period; The real-time green signal ratio of the independent release phase of the current road intersection According to , the mesoscopic subnetwork level control features aggregated from the control actions of all nodes in the control subnetwork are obtained ; wherein, is a control subnetwork set; is the dynamic weight of the node; According to , a macro global control matrix is obtained by aggregating the control features of all control subnets; wherein, is the number of control subnets; k is a mathematical constant; is the subnet routing gate weight of the control subnet.

7. The method of claim 1, wherein, Before the mesoscopic subnet level control feature is spliced with the spatial feature after the spatial congestion diffusion effect based on the mesoscopic subnet level state vector in the global macroscopic network state index, the method further includes: According to , a mesoscopic adaptive subnetwork adjacency matrix is obtained; wherein, and are the mesoscopic subnetwork level state vectors of the control subnetwork and the mesoscopic subnetwork level state vectors of the control subnetwork respectively; denotes vector splicing; denotes the Euclidean distance between the geometric center of the control subnetwork and the geometric center of the control subnetwork , denotes the learnable topological mapping weight matrix; Based on the mesoscopic adaptive subnet adjacency matrix, the control subnet set and the control subnet connection edge, a mesoscopic subnet level dynamic interaction topology graph is constructed.

8. The method of claim 7, wherein, Based on the mesoscopic subnet level state vector in the global macroscopic network state index, the mesoscopic subnet level control feature is spliced with the spatial feature after the spatial congestion diffusion effect, and based on the gating recurrent unit, a current network hidden layer state is obtained, specifically including: according to The spatial feature vector after the spatial congestion diffusion effect is obtained. ;in, This is the spatial feature transformation matrix; It is a non-linear activation function; This refers to the mesoscopic adaptive subnet adjacency matrix in the mesoscopic subnet hierarchical dynamic interaction topology graph. and Control subnets state vectors and control subnets at the meso-level The mesoscopic subnet-level state vector; To control the number of subnets; According to , an updated current network hidden layer state is obtained; wherein, is a historical network hidden layer state; is the meso-subnet level control feature for microscopic traffic light parameter control.

9. The method of claim 1, wherein, Through the subnet routing gating weight in the global macroscopic network state index, the current network hidden layer state is subjected to mesoscopic scale multi-step prediction and macroscopic scale whole network prediction, and the overall macroscopic traffic situation index of the whole road network in the graph network structure is obtained, specifically including: According to , a future meso-subnet-level state vector of each control subnet at a plurality of future time steps is obtained based on meso-scale multi-step prediction; wherein, is the current network hidden layer state; is a future amount of meso-subnet-level state vectors Through the subnet routing gating weight of the control subnet, the future mesoscopic subnet level state vector of each control road network is subjected to weighted summation processing on the traffic situation result, and the overall macroscopic traffic situation index based on all road networks is obtained.

10. A multi-scale controlled traffic flow prediction system for urban road networks, characterized in that, The multi-scale controlled traffic flow prediction system for urban road network can be executed by at least one processor to enable the at least one processor to perform the multi-scale controlled traffic flow prediction method for urban road network according to any one of claims 1-9.

Images