Road network traffic control method and electronic device

Abstract

This application provides a road network traffic control method and electronic device, relating to the field of intelligent traffic control technology. The method includes: acquiring multi-source data of a target area road network; generating low-dimensional spatiotemporal features based on the multi-source data of the target area road network; inputting the low-dimensional spatiotemporal features into a multimodal large model, where the multimodal large model extracts features layer by layer from the low-dimensional spatiotemporal features based on a built-in multi-layer cascaded structure to generate high-dimensional spatiotemporal features; constructing a collaborative control probability matrix based on the high-dimensional spatiotemporal features; generating a control topology map based on the collaborative control probability matrix; generating a control strategy for the target area road network based on the control topology map; and performing traffic control on the target area road network based on the control strategy. This application constructs a control topology map adapted to the current traffic situation based on actual changes in traffic flow and generates an adapted control strategy for traffic control, effectively solving the problem of mismatch between control strategies and actual traffic demand in existing technologies.

Description

Technical Field

[0001] This application relates to the field of intelligent traffic control technology, and more specifically, to a road network traffic control method and electronic equipment. Background Technology

[0002] Traffic control is a key element in ensuring the efficient operation of urban transportation and smooth travel for residents. The effectiveness of traffic control directly determines the effectiveness of traffic congestion management and the efficiency of road network traffic.

[0003] Currently, traffic management in urban road networks still relies on heuristic rules such as human experience, static boundaries, or distance thresholds, which cannot adapt to the time-varying characteristics of traffic flow. For example, it cannot adapt to the dynamic changes in traffic flow during morning and evening rush hours, off-peak hours, and holidays, resulting in a serious mismatch between management strategies and actual traffic needs. Summary of the Invention

[0004] The purpose of this application is to provide a road network traffic control method and electronic device to address the shortcomings of the prior art, thereby solving the problem of a serious mismatch between the control strategies and actual traffic demands in the prior art.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, embodiments of this application provide a road network traffic management method, the method comprising: Acquire multi-source data on the road network of the target area; Based on the multi-source data of the road network in the target area, low-dimensional spatiotemporal features are generated. The low-dimensional spatiotemporal features are input into a multimodal large model. The multimodal large model then extracts features from the low-dimensional spatiotemporal features layer by layer based on its built-in multi-layer cascaded structure, generating high-dimensional spatiotemporal features. Based on these high-dimensional spatiotemporal features, a collaborative control probability matrix is ​​constructed. The collaborative control probability matrix is... The matrix, Let be the number of intersections in the target area's road network, and be the th intersection. Line number The element values ​​of the column are used to characterize the first element in the road network of the target area. The intersection and the first The probability that coordinated management is needed between intersections; Based on the aforementioned collaborative control probability matrix, a control topology graph is generated; Based on the control topology map, a control strategy for the road network in the target area is generated, and traffic control is performed on the road network in the target area based on the control strategy.

[0006] As one possible implementation, generating low-dimensional spatiotemporal features based on multi-source data of the target area road network includes: Based on the multi-source data, an original spatiotemporal tensor is constructed, and the traffic dimension features in the original spatiotemporal tensor are projected to obtain traffic features; Encode the temporal features in the original spatiotemporal tensor to generate temporally encoded features; A digital model of the road network in the target area is performed to obtain the physical adjacency matrix of the road network in the target area. The physical adjacency matrix is ​​as follows: The matrix, The number of road segments in the target area road network, the first... Line number The element values ​​of the column are used to represent the first... The section and the first Are the road sections connected? The physical adjacency matrix is ​​normalized to generate a Laplacian matrix, and the Laplacian matrix is ​​decomposed to generate spatial topological coding features. The traffic features, the temporal coding features, and the spatial topology coding features are fused together to generate the low-dimensional spatiotemporal features.

[0007] As one possible implementation, each layer of the multimodal large model includes a temporal multi-head attention module and a spatial graph mask attention module; The process of extracting features from the low-dimensional spatiotemporal features layer by layer from the multimodal large model based on a built-in multi-layer cascaded structure to generate high-dimensional spatiotemporal features includes: The low-dimensional spatiotemporal features are input into the temporal multi-head attention module of the first layer of the multimodal large model. The temporal multi-head attention module performs temporal encoding on the low-dimensional spatiotemporal features through an upper triangular temporal causal mask to obtain the first intermediate feature. The first intermediate feature is input into the spatial graph mask attention module of the first layer of the multimodal large model. The spatial graph mask attention module performs spatial encoding on the first intermediate feature through a physical graph mask to obtain the first-layer spatiotemporal features. The first-layer spatiotemporal features are input into the next-layer temporal multi-head attention module. The temporal multi-head attention module performs temporal encoding on the first-layer spatiotemporal features through an upper triangular temporal causal mask to obtain the second intermediate features. The second intermediate features are input into the next-layer spatial graph mask attention module. The spatial graph mask attention module performs spatial encoding on the second intermediate features through a physical graph mask to obtain the next-layer spatiotemporal features. The process is executed layer by layer until the last layer of the multimodal large model is completed, resulting in the high-dimensional spatiotemporal features. These high-dimensional spatiotemporal features are used to characterize the high-dimensional latent features of each intersection in the target area road network.

[0008] As one possible implementation, the construction of a collaborative control probability matrix based on the high-dimensional spatiotemporal features includes: The high-dimensional spatiotemporal features are input into the decoder, which calculates the feature interaction value between the high-dimensional latent features corresponding to any two intersections in the road network of the target area. The feature interaction values ​​are fused using the learning parameters of the decoder to obtain the collaborative control mapping values ​​corresponding to the two intersections; By using an activation function to map the collaborative control mapping values ​​corresponding to the two intersections, the probability that collaborative control is required between the two intersections can be obtained. Based on the probability that coordinated management is required between any two intersections in the target area road network, the coordinated management probability matrix is ​​constructed.

[0009] As one possible implementation, the control topology graph includes multiple nodes and multiple edges. Each node represents an intersection in the target area road network, and each edge represents the collaborative control relationship between two intersections corresponding to two nodes. The generation of the control topology map based on the collaborative control probability matrix includes: The collaborative control probability matrix is ​​binarized to obtain a binary adjacency matrix; Each intersection in the target area road network is treated as a node. Based on the probability that collaborative management is required between any two intersections in the target area road network, an edge is constructed between the nodes corresponding to any two intersections. Combined with the binary adjacency matrix, the management topology graph is generated.

[0010] As one possible implementation, the control strategy for generating the road network of the target area based on the control topology map includes: Identify the connected components in the control topology graph, including: linear chain connected components, grid connected components, and independent nodes; For each of the connected components, a control region corresponding to each connected component is determined, and a control strategy corresponding to the control region is generated.

[0011] As one possible implementation, determining the control region corresponding to each of the connected components and generating the control policy corresponding to the control region includes: If the connected component is the linear chain connected component, then the control area corresponding to the linear chain connected component is determined to be the trunk green wave area, and the control strategy corresponding to the control area is determined to be a one-way or two-way green wave coordination strategy. If the connected component is the grid-like connected component, then the control area corresponding to the grid-like connected component is determined to be the coordinated control area, and the control strategy corresponding to the control area is determined to be the collaborative overflow prevention strategy. If the connected component is the independent node, then the independent node is regarded as an independent control area, and the control strategy of the independent control area is determined to be the sensing control strategy.

[0012] As one possible implementation, the multimodal large model is trained based on the following steps: A training dataset is constructed based on historical multi-source data of multiple regional road networks and historical control strategies corresponding to multiple regional road networks. The training dataset includes multiple training samples. Each training sample includes an input feature sample generated based on the historical multi-source data of the regional road network and an expert truth matrix generated based on the historical control strategies corresponding to the regional road network. Input the training dataset into the initial large model to obtain the actual probability matrix output by the initial large model; The loss result of the initial large model is determined based on the expert truth matrix and the actual probability matrix, and the initial large model is iteratively adjusted based on the loss result to obtain the multimodal large model.

[0013] As one possible implementation, determining the loss result of the initial large model based on the expert truth matrix and the actual probability matrix includes: Based on the actual probability matrix and the expert truth matrix, the expert cloning loss is determined; Based on the actual probability matrix and the transpose of the actual probability matrix, the symmetry constraint loss is determined. Based on the actual probability matrix and the number of intersections in the regional road network, the sparsity regularization loss is determined. The expert cloning loss, the symmetry constraint loss, and the sparsity regularization loss are weighted and fused to obtain the loss result of the initial large model.

[0014] Secondly, embodiments of this application provide a road network traffic control device, the device comprising: The acquisition module is used to acquire multi-source data of the road network in the target area; The first generation module is used to generate low-dimensional spatiotemporal features based on multi-source data of the road network in the target area; The construction module is used to input the low-dimensional spatiotemporal features into a multimodal large model. The multimodal large model then extracts features from the low-dimensional spatiotemporal features layer by layer based on its built-in multi-layer cascaded structure, generating high-dimensional spatiotemporal features. Based on these high-dimensional spatiotemporal features, a collaborative control probability matrix is ​​constructed. The collaborative control probability matrix is... The matrix, Let be the number of intersections in the target area's road network, and be the th intersection. Line number The element values ​​of the column are used to characterize the first element in the road network of the target area. The intersection and the first The probability that coordinated management is needed between intersections; The second generation module is used to generate a control topology map based on the collaborative control probability matrix; The control module is used to generate a control strategy for the road network of the target area based on the control topology map, and to perform traffic control on the road network of the target area based on the control strategy.

[0015] As one possible implementation, the first generation module is specifically used for: Based on the multi-source data, an original spatiotemporal tensor is constructed, and the traffic dimension features in the original spatiotemporal tensor are projected to obtain traffic features; Encode the temporal features in the original spatiotemporal tensor to generate temporally encoded features; A digital model of the road network in the target area is performed to obtain the physical adjacency matrix of the road network in the target area. The physical adjacency matrix is ​​as follows: The matrix, The number of road segments in the target area road network, the first... Line number The element values ​​of the column are used to represent the first... The section and the first Are the road sections connected? The physical adjacency matrix is ​​normalized to generate a Laplacian matrix, and the Laplacian matrix is ​​decomposed to generate spatial topological coding features. The traffic features, the temporal coding features, and the spatial topology coding features are fused together to generate the low-dimensional spatiotemporal features.

[0016] As one possible implementation, each layer of the multimodal large model includes a temporal multi-head attention module and a spatial graph mask attention module; the construction module is specifically used for: The low-dimensional spatiotemporal features are input into the temporal multi-head attention module of the first layer of the multimodal large model. The temporal multi-head attention module performs temporal encoding on the low-dimensional spatiotemporal features through an upper triangular temporal causal mask to obtain the first intermediate feature. The first intermediate feature is input into the spatial graph mask attention module of the first layer of the multimodal large model. The spatial graph mask attention module performs spatial encoding on the first intermediate feature through a physical graph mask to obtain the first-layer spatiotemporal features. The first-layer spatiotemporal features are input into the next-layer temporal multi-head attention module. The temporal multi-head attention module performs temporal encoding on the first-layer spatiotemporal features through an upper triangular temporal causal mask to obtain the second intermediate features. The second intermediate features are input into the next-layer spatial graph mask attention module. The spatial graph mask attention module performs spatial encoding on the second intermediate features through a physical graph mask to obtain the next-layer spatiotemporal features. The process is executed layer by layer until the last layer of the multimodal large model is completed, resulting in the high-dimensional spatiotemporal features. These high-dimensional spatiotemporal features are used to characterize the high-dimensional latent features of each intersection in the target area road network.

[0017] As one possible implementation, the building module is specifically used for: The high-dimensional spatiotemporal features are input into the decoder, which calculates the feature interaction value between the high-dimensional latent features corresponding to any two intersections in the road network of the target area. The feature interaction values ​​are fused using the learning parameters of the decoder to obtain the collaborative control mapping values ​​corresponding to the two intersections; By using an activation function to map the collaborative control mapping values ​​corresponding to the two intersections, the probability that collaborative control is required between the two intersections can be obtained. Based on the probability that coordinated management is required between any two intersections in the target area road network, the coordinated management probability matrix is ​​constructed.

[0018] As one possible implementation, the control topology graph includes multiple nodes and multiple edges, where each node represents an intersection in the target area road network, and each edge represents the collaborative control relationship between two intersections corresponding to two nodes; the second generation module is specifically used for: The collaborative control probability matrix is ​​binarized to obtain a binary adjacency matrix; Each intersection in the target area road network is treated as a node. Based on the probability that collaborative management is required between any two intersections in the target area road network, an edge is constructed between the nodes corresponding to any two intersections. Combined with the binary adjacency matrix, the management topology graph is generated.

[0019] As one possible implementation, the control module is specifically used for: Identify the connected components in the control topology graph, including: linear chain connected components, grid connected components, and independent nodes; For each of the connected components, a control region corresponding to each connected component is determined, and a control strategy corresponding to the control region is generated.

[0020] As one possible implementation, the control module is specifically used for: If the connected component is the linear chain connected component, then the control area corresponding to the linear chain connected component is determined to be the trunk green wave area, and the control strategy corresponding to the control area is determined to be a one-way or two-way green wave coordination strategy. If the connected component is the grid-like connected component, then the control area corresponding to the grid-like connected component is determined to be the coordinated control area, and the control strategy corresponding to the control area is determined to be the collaborative overflow prevention strategy. If the connected component is the independent node, then the independent node is regarded as an independent control area, and the control strategy of the independent control area is determined to be the sensing control strategy.

[0021] As one possible implementation, the road network traffic control device further includes a training module, which is specifically used for: A training dataset is constructed based on historical multi-source data of multiple regional road networks and historical control strategies corresponding to multiple regional road networks. The training dataset includes multiple training samples. Each training sample includes an input feature sample generated based on the historical multi-source data of the regional road network and an expert truth matrix generated based on the historical control strategies corresponding to the regional road network. Input the training dataset into the initial large model to obtain the actual probability matrix output by the initial large model; The loss result of the initial large model is determined based on the expert truth matrix and the actual probability matrix, and the initial large model is iteratively adjusted based on the loss result to obtain the multimodal large model.

[0022] As one possible implementation, the training module is specifically used for: Based on the actual probability matrix and the expert truth matrix, the expert cloning loss is determined; Based on the actual probability matrix and the transpose of the actual probability matrix, the symmetry constraint loss is determined. Based on the actual probability matrix and the number of intersections in the regional road network, the sparsity regularization loss is determined. The expert cloning loss, the symmetry constraint loss, and the sparsity regularization loss are weighted and fused to obtain the loss result of the initial large model.

[0023] Thirdly, embodiments of this application provide an electronic device, including: a processor, a storage medium, and a bus. The storage medium stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the road network traffic control method as described in any of the first aspects above.

[0024] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the road network traffic control method as described in any of the first aspects above.

[0025] According to the road network traffic control method and electronic device of this application embodiment, based on the multi-layer cascaded structure of a multimodal large model, deep feature mining is performed on multi-source data of the target area road network. The original low-dimensional spatiotemporal features are transformed into high-dimensional spatiotemporal features rich in traffic semantics. The high-dimensional features are then used to construct a collaborative control probability matrix. The element values ​​in the collaborative control probability matrix represent the probability that collaborative control is required between two corresponding intersections. Based on the collaborative control probability matrix, a control topology map is dynamically generated, and a corresponding control strategy is generated according to the control topology map. This application embodiment accurately quantifies the real-time collaborative needs between intersections by using high-dimensional features through feature extraction and dimensional transformation. This enables the construction of a control topology map adapted to the current traffic situation based on the actual changes in traffic flow, and the generation of an appropriate control strategy for traffic control. This effectively solves the problem in the prior art where the control strategy cannot adapt to the time-varying traffic flow characteristics such as morning and evening peak hours due to reliance on manual experience or static distance thresholds, thus causing a mismatch between the control strategy and the actual traffic demand. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A schematic flowchart of a road network traffic control method provided in an embodiment of this application is shown; Figure 2 A flowchart illustrating a low-dimensional spatiotemporal feature generation method provided in an embodiment of this application is shown. Figure 3 A flowchart illustrating a collaborative control probability matrix construction method provided in an embodiment of this application is shown. Figure 4A flowchart illustrating a control strategy generation method provided in an embodiment of this application is shown; Figure 5 The diagram shows a flowchart of a training method for a multimodal large model provided in an embodiment of this application. Figure 6 This paper shows a schematic diagram of the structure of a road network traffic control device according to an embodiment of this application; Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0029] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0030] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0031] Figure 1 A flowchart illustrating a road network traffic control method according to an embodiment of this application is shown. (Refer to...) Figure 1 As shown, the method specifically includes the following steps: S101. Obtain multi-source data of the road network in the target area.

[0032] Optionally, the multi-source data includes GPS trajectory data of urban floating vehicles within the target area's road network, real-time data collected by roadside radar or video detectors, and expert strategy log data from the traffic management platform.

[0033] Optionally, after acquiring multi-source data of the road network in the target area, the multi-source data is first subjected to preprocessing operations such as data cleaning, spatiotemporal alignment, normalization, and missing value completion, so as to construct low-dimensional spatiotemporal features based on the preprocessed multi-source data.

[0034] S102. Generate low-dimensional spatiotemporal features based on multi-source data of the road network in the target area.

[0035] Optionally, after preprocessing the multi-source data of the target area road network, an original spatiotemporal tensor is constructed based on the processed multi-source data. A multilayer perceptron is then used to perform nonlinear upscaling projection on the traffic dimension features in the original spatiotemporal tensor to obtain traffic features. Sine position coding is then applied to the temporal dimension of the original spatiotemporal tensor to generate temporal coding features. Furthermore, a physical topology map of the target area road network is digitally modeled to obtain the physical adjacency matrix. This physical adjacency matrix is ​​then binarized to obtain the Laplacian matrix. Subsequently, feature decomposition is performed on the Laplacian matrix to extract spatial topology coding features. Finally, the traffic features, temporal coding features, and spatial topology coding features are superimposed and fused to generate a low-dimensional spatiotemporal feature that incorporates prior physical knowledge of the road network. .

[0036] S103. Input the low-dimensional spatiotemporal features into the multimodal large model. The multimodal large model extracts features from the low-dimensional spatiotemporal features layer by layer based on the built-in multi-layer cascade structure to generate high-dimensional spatiotemporal features. Based on the high-dimensional spatiotemporal features, construct the collaborative control probability matrix.

[0037] Optionally, the multimodal large model has a multi-layered cascaded structure, with each layer including a temporal multi-head attention module and a spatial graph mask attention module. Based on the built-in multi-layered cascaded structure of the multimodal large model, low-dimensional spatiotemporal features are first input into the temporal multi-head attention module of the first layer of the multimodal large model, and then the temporal multi-head attention module passes the features through an upper triangular temporal causal mask. Temporal encoding is performed on the low-dimensional spatiotemporal features to obtain the first intermediate feature. This first intermediate feature is then input into the spatial graph masking attention module of the first layer of the multimodal large model. The spatial graph masking attention module then processes the data through a physical graph mask. Spatial encoding is performed on the first intermediate feature to obtain the first-layer spatiotemporal feature. Then, the first-layer spatiotemporal feature is input into the next-layer temporal multi-head attention module, which uses an upper triangular temporal causality mask. Temporal encoding is performed on the first-layer spatiotemporal features to obtain the second intermediate features. The second intermediate features are then input into the next-layer spatial graph mask attention module, which processes them through the physical graph mask. Spatial encoding is performed on the second intermediate feature to obtain the next layer of spatiotemporal features. This process is repeated layer by layer until the final layer of the multimodal large model is completed, resulting in high-dimensional spatiotemporal features. These high-dimensional spatiotemporal features are used to characterize the high-dimensional latent features of each intersection in the road network of the target area.

[0038] For example, taking a multimodal large model with a six-layer cascaded structure, each layer is completely identical. Based on this, low-dimensional spatiotemporal features are... The first layer of the input multimodal large model is processed by the temporal multi-head attention module and the spatial graph mask attention module of the first layer, and the spatiotemporal features of the first layer are output. Then, the first layer of spatiotemporal features The second layer of the input multimodal large model, after being processed by the temporal multi-head attention module and the spatial graph mask attention module of the second layer, outputs the spatiotemporal features of the second layer. And so on, layers 3 through 5 are executed sequentially, with the spatiotemporal characteristics of layer 3 recorded as follows: The fourth layer of spatiotemporal characteristics are The fifth layer of spatiotemporal characteristics are And the fifth layer of spatiotemporal characteristics are The input is sent to the last layer, which is the sixth layer, and the output is the sixth layer's spatiotemporal features. That is, the final high-dimensional spacetime characteristics .

[0039] It is worth noting that the embodiments of this application will use low-dimensional spatiotemporal features The high-dimensional spatiotemporal features are extracted from the input multimodal large model, based on the multi-layer cascaded structure built into the multimodal large model and combined with the attention mechanism, and finally output high-dimensional spatiotemporal features at the intersection level. From low-dimensional spatiotemporal features Transformation into high-dimensional spatiotemporal features The process is actually performing temporal compression. For example, using this low-dimensional spatiotemporal feature... For example, based on the multi-layer cascaded structure built into the multimodal large model, the temporal features of 12 time steps are aggregated layer by layer, and the scattered short temporal features are fused into a global temporal representation, achieving effective compression of the temporal dimension T. The compressed output is a high-dimensional spatiotemporal feature at the intersection level. That is, the high-dimensional spatiotemporal characteristics Only the high-dimensional latent features of each intersection in the target area road network are retained to adapt to the input requirements of subsequent calculation of the collaborative control probability matrix and the generation of the control topology map.

[0040] Optionally, high-dimensional spatiotemporal features The input is fed into the decoder of a multimodal large model. For any two intersections in the target area road network, the feature interaction value between the high-dimensional latent features corresponding to the two intersections is calculated. Then, the learned parameters of the decoder are used to linearly fuse these feature interaction values ​​to obtain a collaborative control mapping value. An activation function is then used to transform this collaborative control mapping value into the probability that collaborative control is required between the two intersections. In this way, the calculated collaborative control probabilities of any two intersections are combined to construct a collaborative control probability matrix representing the global intersection collaborative relationship. .

[0041] S104. Generate a control topology map based on the collaborative control probability matrix.

[0042] Optionally, control topology map It includes multiple nodes and multiple edges. Each node represents an intersection in the road network of the target area, and each edge represents the collaborative management relationship between two intersections corresponding to two nodes. This collaborative management relationship refers to whether collaborative management is required between the two intersections.

[0043] Optionally, after constructing the collaborative control probability matrix Then, based on the preset adaptive threshold and indicator functions For the collaborative control probability matrix Binarization is performed to obtain a binary adjacency matrix. Specifically, a binary adjacency matrix The probability matrix for collaborative management is also about to be established. Each element value in the data is compared with a preset adaptive threshold. Perform size comparisons one by one, using indicator functions during the comparison process. As a judgment rule, if the element value is strictly greater than the adaptive threshold If the indicator function outputs 1, it means there is a collaborative control relationship between the two intersections. If the element value is less than or equal to the adaptive threshold... If the output of the indicator function is 0, it means there is no cooperative relationship between the two intersections. Thus, by traversing the cooperative control probability matrix... All elements in the matrix will ultimately coordinate and control the probability matrix. Convert to a discrete binary adjacency matrix consisting only of 0s and 1s This completes the mapping from the probability space to the topological graph structure.

[0044] S105. Generate a control strategy for the road network in the target area based on the control topology map, and implement traffic control on the road network in the target area based on the control strategy.

[0045] Optionally, by identifying the connected components in the control topology map, intersections that establish cooperative relationships with each other are divided into independent control areas. Based on the specific topology of each connected component, such as linear chain, grid or independent node, corresponding differentiated control strategies are generated. The generated differentiated control strategies are then used to implement targeted and precise traffic control in each control area of ​​the target area road network.

[0046] Based on this, the road network traffic control method according to the embodiments of this application, based on the multi-layer cascaded structure of a multimodal large model, performs deep feature mining on the multi-source data of the target area road network, transforming the original low-dimensional spatiotemporal features into high-dimensional spatiotemporal features rich in traffic semantics, and using the high-dimensional features to construct a collaborative control probability matrix. The element values ​​in the collaborative control probability matrix represent the probability that collaborative control is required between two corresponding intersections. Based on the collaborative control probability matrix, a control topology map is dynamically generated, and corresponding control strategies are generated according to the control topology map. The embodiments of this application, through feature extraction and dimensional transformation, accurately quantify the real-time collaborative needs between intersections using high-dimensional features, thereby enabling the construction of a control topology map adapted to the current traffic situation based on the actual changes in traffic flow, and generating adapted control strategies for traffic control. This effectively solves the problem in the prior art where the reliance on manual experience or static distance thresholds leads to control strategies that cannot adapt to time-varying traffic flow characteristics such as morning and evening peak hours, resulting in a mismatch between control strategies and actual traffic demand.

[0047] Figure 2 A flowchart illustrating a low-dimensional spatiotemporal feature generation method provided in an embodiment of this application is shown. (Refer to...) Figure 2 As shown, step S102 above generates low-dimensional spatiotemporal features based on multi-source data of the target area's road network, specifically including the following steps: S201. Construct the original spatiotemporal tensor based on multi-source data, and project the traffic dimension features in the original spatiotemporal tensor to obtain traffic features.

[0048] Optionally, before constructing the original spatiotemporal tensor based on multi-source data, the multi-source data is first subjected to preprocessing operations such as data cleaning, spatiotemporal alignment, normalization, and missing value completion. Among them, data cleaning includes removing abnormal data, such as vehicle speed of 0 but no queue, traffic surge without trajectory support, invalid data due to equipment failure, and empty signal message data. Spatiotemporal alignment includes unifying the collection timestamps and spatial coordinates of all multi-source data, for example, matching the identification IDs of signalized intersections in the road network with a fixed time step of 5 minutes. Normalization includes normalizing features with different dimensions such as traffic flow, vehicle speed, and queue length to the [0,1] interval to eliminate the interference of dimension differences on the multimodal large model. Missing value completion includes using interpolation of the mean of adjacent time steps to complete the data for occasional equipment disconnection, so as to ensure the continuity of the time series.

[0049] Optionally, after preprocessing the multi-source data, an original spatiotemporal tensor is constructed based on the processed multi-source data. This original spatiotemporal tensor is a three-dimensional tensor, including a temporal dimension T, a spatial node dimension N, and a feature dimension F, i.e., the original spatiotemporal tensor. For example, taking a target area road network including 1200 signalized intersections and 3500 road segments as an example, the spatial node dimension N is 1200. Multi-source data of the target area road network for one hour is collected at a fixed time step of 5 minutes, that is, the time series dimension T is 12 and the feature dimension F is 6. This mainly includes six standardized traffic dimension features, which are all core control indicators in the field of traffic control, including hourly traffic flow at intersections, average vehicle speed on road segments, intersection travel delay, number of stops at intersections, queue length at approach lanes, and lane occupancy rate. Among them, hourly traffic flow at intersections refers to the total traffic flow through all approach lanes of a single intersection; average vehicle speed on road segments refers to the average speed of upstream and downstream road segments associated with the intersection; travel delay at intersections refers to the average delay time for vehicles to pass through the intersection; number of stops at intersections refers to the average number of times vehicles wait at the intersection; queue length at approach lanes refers to the maximum queue length of vehicles at each approach lane; and lane occupancy rate refers to the percentage of time vehicles occupy a lane in a road segment. Thus, the original spatiotemporal tensor is constructed. .

[0050] Furthermore, due to the original spacetime tensor The traffic dimension features in the model are low-dimensional, artificial features, which cannot directly meet the high-dimensional feature calculation requirements of multimodal large models. Therefore, a multi-layer perceptron (MLP) is needed to process the original spatiotemporal tensor. The traffic dimension features in the model are nonlinearly upscaled. Specifically, the network structure of the Multilayer Perceptron (MLP) consists of two fully connected layers, an activation function, and a standard feature projection structure with layer normalization, which projects the original spatiotemporal tensor into a nonlinear dimension. The six-dimensional traffic features are input into the first fully connected layer of the Multilayer Perceptron (MLP) to increase their dimensionality, resulting in 64-dimensional traffic features. After introducing nonlinear feature association through an activation function, these 64-dimensional traffic features are input into the second fully connected layer of the MLP to increase their dimensionality again, resulting in 128-dimensional traffic features. Finally, after layer normalization, the standardized projected feature tensor, i.e., the projected traffic features, is obtained. This traffic feature .

[0051] It is worth noting that the original spacetime tensor It only contains surface traffic numerical features and lacks nonlinear correlation mining capabilities. However, the traffic features after nonlinear transformation by multilayer perceptron (MLP) have been used to mine the implicit correlations between the original six-dimensional traffic features, such as the coupling relationship between traffic flow and queue length, and vehicle speed and delay. This solves the problems of insufficient representation ability and weak feature correlation of the original spatiotemporal tensor as a low-dimensional feature.

[0052] S202. Encode the temporal features in the original spatiotemporal tensor to generate temporal encoded features.

[0053] Optionally, the original spacetime tensor The temporal dimension T in the model is a temporal feature used to characterize the temporal sequence and evolution of traffic characteristics. It corresponds to the temporal causality of traffic flow dynamic changes and is the basic dimension for short-term traffic prediction and dynamic information control. It is different from the spatial node dimension N, which characterizes spatial location, and the feature dimension F, which characterizes feature attributes.

[0054] For example, continuing to collect multi-source data of the target area's road network for one hour using the fixed time step of 5 minutes (i.e., time series dimension T is 12), corresponding to 12 consecutive, equally spaced time steps, with a single step interval of 5 minutes, the complete time series sequence is arranged as [t0, t1, t2...t11], strictly sorted from oldest to newest time, where t0 is historical data from 1 hour ago, and t11 is the latest real-time data. This time series sequence completely covers the short-term fluctuations, peak changes, congestion accumulation, and dissipation processes of traffic flow. In this embodiment, feature mining is performed on the time series sequence to uncover the continuous evolution pattern of traffic flow within 1 hour, such as the time series characteristics of increasing traffic flow during the morning peak, congestion dissipation during the evening peak, and stable fluctuations during off-peak hours, and sinusoidal time position coding is used. Encoding of temporal features, sinusoidal time position encoding Based on the periodicity of sine and cosine trigonometric functions, positional encoding is performed on 12 time steps to distinguish the order of different time steps and ensure temporal causality. Specifically, the sine time step positional encoding is used. By iterating through 12 time steps, temporal location vectors are generated one by one, and the dimensions are aligned to 128 to obtain temporal coding features. .

[0055] S203. Perform digital modeling of the road network in the target area to obtain the physical adjacency matrix of the road network in the target area.

[0056] Optionally, multiple signalized intersections and road segments in the target area's road network are digitally modeled to generate a physical topology map of the target area's road network. The physical topology graph includes multiple nodes and multiple edges. Each node represents an intersection, and each edge represents a road segment connecting two nodes at the intersection. (Physical topology graph based on the target area road network) The physical connectivity relationships of the indicated multiple intersections construct the physical adjacency matrix of the road network in the target area. The physical adjacency matrix for The matrix, The number of road segments in the target area's road network, the first Line number The element values ​​of the column are used to represent the first... The section and the first Whether the road segments are connected, specifically, if the first... Line number If the element value of a column is greater than 0, it indicates that the first element is the first element. The section and the first If the first road segment has a physical road connection, then the second road segment has a physical road connection. Line number If the element value of a column is less than or equal to 0, it indicates that the first element is the first element. The section and the first The road sections are not connected.

[0057] S204. Normalize the physical adjacency matrix to generate the Laplacian matrix, and perform eigenvalue decomposition on the Laplacian matrix to generate spatial topological coding features.

[0058] Optionally, the physical adjacency matrix of the road network in the target area. After normalization, the Laplacian matrix of the road network in the target area is obtained. Then, eigenvalue decomposition is performed on the Laplacian matrix, and the eigenvectors corresponding to the first K non-zero eigenvalues ​​are taken as the Laplacian location codes. K can take the value 8, and Laplace position encoding is used. Encoding the global structural attributes of the road network to capture the spatial topological relationships between intersections, thereby generating spatial topological coding features. This spatial topological coding feature .

[0059] Optionally, the global structural attributes of the road network refer to the inherent, static, and unchanging set of topological features of a signalized intersection within the urban road network system. These attributes do not change with traffic flow and specifically include topological location attributes, connectivity attributes, road network centrality attributes, and structural association attributes. Among these, the topological location attribute refers to the intersection's position at the level of arterial road, branch road, commercial district, or core area; the connectivity attribute refers to the number of upstream and downstream connecting intersections, the number of branch road connections, and the intersection level; the road network centrality attribute refers to the intersection's traffic weight, hub level, and radiation range within the overall road network; and the structural association attribute refers to the linear, grid, or isolated topological form of the road network to which the intersection belongs.

[0060] S205. The traffic features, temporal coding features, and spatial topological coding features are fused to generate low-dimensional spatiotemporal features.

[0061] Optionally, traffic characteristics Time-series coding features and spatial topological coding features Low-dimensional spatiotemporal features can be generated by pixel-by-pixel overlay and fusion. This low-dimensional spatiotemporal feature And this low-dimensional spatiotemporal feature .

[0062] Based on this, the embodiments of this application encode the original spatiotemporal tensor to generate low-dimensional spatiotemporal features that integrate the physical priors of the road network, transforming the static structural features of the physical road network into features that the model can recognize, thus solving the problem that traditional models cannot adapt to the road network topology.

[0063] Figure 3 A flowchart illustrating a collaborative control probability matrix construction method provided in an embodiment of this application is shown. (Refer to...) Figure 3 As shown, the above steps, based on high-dimensional spatiotemporal features, construct a collaborative control probability matrix, specifically including the following steps: S301. Input the high-dimensional spatiotemporal features into the decoder. The decoder calculates the feature interaction value between the high-dimensional latent features corresponding to any two intersections in the road network of the target area.

[0064] Optionally, high-dimensional spatiotemporal features This high-dimensional spatiotemporal feature Only the high-dimensional latent features of each intersection within the target area's road network are retained. This is done during the generation of high-dimensional spatiotemporal features. Subsequently, a decoder based on a multimodal large model, such as a bilinear decoder, is used to process high-dimensional spatiotemporal features. It is mapped to a collaborative control probability matrix for inter-intersection collaborative control.

[0065] For example, for any two intersections in the target area road network, such as intersections and intersection From the high-dimensional spatiotemporal characteristics Obtain the intersections separately and intersection The high-dimensional hidden features are denoted as follows: and Then calculate the intersection High-dimensional hidden features Intersection High-dimensional hidden features The feature interaction value between them, the feature interaction value is ,in, It represents the Hadamardi (or Hadama) stack.

[0066] S302. The feature interaction values ​​are fused using the learning parameters of the decoder to obtain the collaborative control mapping values ​​corresponding to the two intersections.

[0067] Optionally, the learning parameters of the decoder are denoted as... , After calculating the intersection and intersection Feature interaction values ​​between the high-dimensional latent features corresponding to these two intersections Then, the learned parameters of the decoder are used. , For feature interaction values The fusion process is performed, and the computational logic for the fusion process is as follows: You can then find the intersection. and intersection The corresponding collaborative control mapping values ​​for these two intersections .

[0068] S303. Use the activation function to map the collaborative control mapping values ​​corresponding to the two intersections to obtain the probability that collaborative control is required between the two intersections.

[0069] Optionally, an activation function can be used. At the intersection and intersection The corresponding collaborative control mapping values ​​for these two intersections By mapping, the intersection can be obtained. and intersection The probability that coordinated control is required between these two intersections is given by the following expression (1): (1) in, Indicates intersection and intersection The probability that coordinated management and control are required between them. This represents the activation function. Indicates intersection High-dimensional hidden features, Indicates intersection High-dimensional hidden features, It represents the Hadamah accumulation. , Indicates the learning parameters of the decoder. This indicates transpose.

[0070] S304. Based on the probability that collaborative control is required between any two intersections in the target area road network, construct a collaborative control probability matrix.

[0071] Optionally, for any two intersections in the target area's road network. and Any two intersections can be calculated based on the method shown in expression (1) above. and The probability that coordinated management is required between them. Thus, a collaborative control probability matrix is ​​constructed. .

[0072] Optionally, the collaborative control probability matrix for The matrix. Where, The number of intersections in the target area's road network, the first Line number The element values ​​of the column are used to characterize the first element in the road network of the target area. The intersection and the first The probability that coordinated management is needed between intersections.

[0073] Based on this, the embodiments of this application use a decoder to perform deep nonlinear interaction and mapping on the high-dimensional latent features of each intersection in the target area road network. This can accurately capture the potential and complex spatiotemporal dependencies between any two intersections in the target area road network, thereby generating a collaborative control probability matrix. This lays a solid data foundation for the subsequent generation of a control topology map adapted to the current traffic conditions.

[0074] Figure 4 A flowchart illustrating a control strategy generation method provided in an embodiment of this application is shown. (Refer to...) Figure 4 As shown, the above steps generate a control strategy for the road network in the target area based on the control topology map, specifically including the following steps: S401. Identify and manage connected components in the topology graph.

[0075] Alternatively, connected components refer to the control topology graph. A set of nodes that establish collaborative control edges is constructed within the set. All intersections within this set are timed and collaboratively controlled, while intersections outside the set operate independently. In this embodiment, the connected components include: linear chain-like connected components, grid-like connected components, and independent nodes. Each connected component is an independent control area. Specifically, a linear chain-like connected component refers to multiple intersections connected linearly along a single main road, forming a chain-like collaborative structure with no branch intersections. A grid-like connected component refers to multiple horizontal and vertical roads intersecting to form a regional grid node cluster, with multiple intersections interconnected and influencing each other. An independent node refers to an independent intersection without any collaborative control edges and no collaborative relationship with any surrounding intersections.

[0076] S402. For each connected component, determine the control area corresponding to each connected component and generate the control strategy corresponding to the control area.

[0077] Optionally, if the connected component is a linear chain-like connected component, the control area corresponding to the linear chain-like connected component is determined as the trunk green wave area, and the control strategy corresponding to the control area is determined as a unidirectional or bidirectional green wave coordination strategy. If the connected component is a grid-like connected component, the control area corresponding to the grid-like connected component is determined as a coordinated control area, and the control strategy corresponding to the control area is determined as a cooperative anti-overflow strategy. If the connected component is an independent node, the independent node is taken as an independent control area, and the control strategy for the independent control area is determined as an inductive control strategy.

[0078] For example, the one-way or two-way green wave coordination strategy is adapted to linear chain-like connected components. Its applicable scenarios are long-distance linear arterial roads such as urban expressways and main roads, where traffic flow is mainly straight-through, with high traffic demand and strong continuity. These are the core channels for urban commuting. The control strategy logic is to lock the signal cycle of all intersections along the entire arterial road, accurately calculate the phase difference of each intersection based on the arterial road's design speed and intersection spacing, and unify the green signal ratio of the core straight-through phase at each intersection to form a continuous and stable green wave band along the arterial road. This allows vehicles to pass through multiple intersections continuously at a constant speed without stopping or waiting, significantly reducing the number of stops and travel delays on the arterial road.

[0079] For example, the collaborative overflow prevention strategy is adapted to grid-like connected components. Its applicable scenarios are dense grid areas such as urban business districts, residential areas, and transportation hubs, which have problems such as small intersection spacing, complex traffic flow weaving, easy congestion, and easy queuing overflow. The control strategy logic is to achieve overall traffic balance in the area, abandon the optimal timing of a single intersection, unify the control logic of all intersections in the area, dynamically fine-tune the green light ratio and phase duration of each intersection, balance the traffic flow distribution in the area, so as to avoid excessive queuing at a single intersection overflowing to adjacent intersections, block the spread of congestion, achieve traffic balance in the area's road network, and improve the overall traffic carrying capacity of the region.

[0080] For example, the sensing control strategy is adapted to independent nodes. Its applicable scenarios are isolated intersections with low traffic volume, low priority, and no collaborative value, such as urban branch roads, community intersections, and secondary roads. The control strategy logic is not to coordinate with any surrounding intersections, but to rely entirely on its own real-time traffic flow data to adaptively adjust the green light ratio and timing of each phase. When there are many vehicles, the green light is extended; when there are few vehicles, the green light is shortened; and when there are no vehicles, the main road is allowed to pass. This is to adapt to the scattered traffic flow needs of branch roads while ensuring the traffic efficiency of the main road, reduce the invalid green light time, and avoid resource waste.

[0081] Based on this, the embodiments of this application implement differentiated traffic control such as trunk line green wave, regional overflow prevention, or single-point sensing through topology structure identification and precise strategy matching. This not only effectively improves the traffic efficiency of trunk lines and the overall carrying capacity of the regional road network, but also avoids the waste of resources at low-traffic intersections, thereby maximizing the release of the traffic potential of the urban road network and ensuring the smooth and safe operation of traffic.

[0082] Figure 5 This diagram illustrates a flowchart of a training method for a multimodal large model provided in an embodiment of this application. (Refer to...) Figure 5 As shown, the method specifically includes the following steps: S501. Based on historical multi-source data of multiple regional road networks and historical control strategies corresponding to multiple regional road networks, a training dataset is constructed.

[0083] Optionally, the training dataset includes multiple training samples, each of which includes input feature samples generated from historical multi-source data of the regional road network and expert ground truth matrices generated from historical control strategies corresponding to the regional road network.

[0084] For example, the process of generating input feature samples based on historical multi-source data of the regional road network is similar to the aforementioned process of generating low-dimensional spatiotemporal features based on multi-source data of the target regional road network, and will not be repeated here.

[0085] For example, the expert truth matrix For R The matrix, where R is the number of intersections in the regional road network, and the th... Line number The element values ​​of the column are used to characterize the first element in the regional road network. The intersection and the first Whether coordinated control has been implemented between intersections; specifically, if the element value is 1, it indicates that the expert has coordinated control over the intersections. and intersection Collaborative control was implemented. If the element value is 0, it means that the expert did not manage the intersection. and intersection Implement collaborative management and control.

[0086] S502. Input the training dataset into the initial large model to obtain the actual probability matrix output by the initial large model.

[0087] Optionally, the training dataset is input into the initial large model, which extracts features from the input feature samples layer by layer based on its built-in multi-layer cascaded structure, generating corresponding high-dimensional spatiotemporal features, and then generating the actual probability matrix based on the high-dimensional spatiotemporal features. The actual probability matrix The specific generation process is the same as the aforementioned collaborative control probability matrix. The generation process is similar and will not be described in detail here.

[0088] S503. Determine the loss result of the initial large model based on the expert truth matrix and the actual probability matrix, and iteratively adjust the initial large model based on the loss result to obtain the multimodal large model.

[0089] Optionally, based on the actual probability matrix and the expert truth matrix, the expert cloning loss is determined; based on the actual probability matrix and its transpose, the symmetry constraint loss is determined; based on the actual probability matrix and the number of intersections in the regional road network, the sparsity regularization loss is determined; and the expert cloning loss, symmetry constraint loss, and sparsity regularization loss are weighted and fused to obtain the loss result of the initial large model.

[0090] For example, expert cloning loss The following expression (2) is used to calculate: (2) in, Indicates expert cloning loss, Indicates the weight of positive samples. Representing the expert truth matrix The Middle Line number The element value of the column, which is used to characterize the first element. The intersection and the first Have coordinated control measures been implemented between the intersections? Represents the actual probability matrix The Middle Line number The element value of the column, which is used to characterize the first element. The intersection and the first The probability that coordinated management is needed between intersections.

[0091] For example, symmetry constraint loss The following expression (3) is used to calculate: (3) in, Represents the loss due to symmetry constraints. Represents the actual probability matrix. Indicates transpose. This represents the transpose of the actual probability matrix. This represents the Frobenius norm.

[0092] For example, sparsity regularization loss The following expression (4) is used to calculate: (4) in, This represents the sparsity regularization loss. Represents the actual probability matrix The Middle Line number The element values ​​of the column, This indicates the number of intersections in the regional road network.

[0093] For example, the expert cloning loss is based on the following expression (5). Symmetry constraint loss Sparsity regularization loss We perform weighted fusion to obtain the loss result of the initial large model: (5) in, Indicates the result of loss. Indicates expert cloning loss, Represents the loss due to symmetry constraints. This represents the sparsity regularization loss. , These represent the weighting coefficients.

[0094] For example, embodiments of this application construct a method that includes expert clone loss. Symmetry constraint loss Sparsity regularization loss The joint loss function of these three types of constraints is used to iteratively adjust the initial large model, thereby obtaining a multimodal large model adapted to the time-varying traffic flow scenario.

[0095] Based on this, the embodiments of this application optimize the initial large model by constructing a joint loss function that includes expert cloning loss, symmetry constraint loss and sparsity regularization loss, so that the trained multimodal large model can accurately learn and fit historical expert experience, significantly improving the adaptability of the multimodal large model in complex and ever-changing traffic flow time-varying scenarios.

[0096] Based on the same inventive concept, this application also provides a road network traffic control device corresponding to the road network traffic control method. Since the principle of the road network traffic control device in this application is similar to that of the road network traffic control method described above, the implementation of the road network traffic control device can refer to the implementation of the road network traffic control method, and the repeated parts will not be described again.

[0097] Reference Figure 6 The diagram shown is a structural schematic of a road network traffic control device provided in an embodiment of this application. The road network traffic control device 600 includes: an acquisition module 601, a first generation module 602, a construction module 603, a second generation module 604, and a control module 605, wherein: The acquisition module 601 is used to acquire multi-source data of the road network in the target area. The first generation module 602 is used to generate low-dimensional spatiotemporal features based on multi-source data of the road network in the target area. Module 603 is used to input low-dimensional spatiotemporal features into the multimodal large model. The multimodal large model extracts features from the low-dimensional spatiotemporal features layer by layer based on its built-in multi-layer cascaded structure, generating high-dimensional spatiotemporal features. Based on the high-dimensional spatiotemporal features, a collaborative control probability matrix is ​​constructed. The collaborative control probability matrix is ​​as follows: The matrix, The number of intersections in the target area's road network, the first Line number The element values ​​of the column are used to characterize the first element in the road network of the target area. The intersection and the first The probability that coordinated management is needed between intersections; The second generation module 604 is used to generate a control topology map based on the collaborative control probability matrix; The control module 605 is used to generate control strategies for the road network of the target area based on the control topology map, and to perform traffic control on the road network of the target area based on the control strategies.

[0098] Based on this, the road network traffic control device according to the embodiments of this application, based on the multi-layer cascaded structure of a multimodal large model, performs deep feature mining on multi-source data of the target area road network, transforming the original low-dimensional spatiotemporal features into high-dimensional spatiotemporal features rich in traffic semantics, and using these high-dimensional features to construct a collaborative control probability matrix. The element values ​​in this collaborative control probability matrix represent the probability that collaborative control is required between two corresponding intersections. Based on this collaborative control probability matrix, a control topology map is dynamically generated, and corresponding control strategies are generated according to the control topology map. The embodiments of this application, through feature extraction and dimensional transformation, accurately quantify the real-time collaborative needs between intersections using high-dimensional features. This enables the construction of a control topology map adapted to the current traffic situation based on actual changes in traffic flow, and the generation of adapted control strategies for traffic control. This effectively solves the problem in existing technologies where reliance on human experience or static distance thresholds leads to control strategies that cannot adapt to time-varying traffic flow characteristics such as morning and evening peak hours, resulting in a mismatch between control strategies and actual traffic demands.

[0099] In one possible implementation, the first generation module 602 described above is specifically used for: The original spatiotemporal tensor is constructed based on multi-source data, and the traffic dimension features in the original spatiotemporal tensor are projected to obtain traffic features. Encode the temporal features in the original spatiotemporal tensor to generate temporally encoded features; A digital model of the road network in the target area is performed to obtain the physical adjacency matrix of the road network in the target area. The physical adjacency matrix is ​​as follows: The matrix, The number of road segments in the target area's road network, the first Line number The element values ​​of the column are used to represent the first... The section and the first Are the road sections connected? The physical adjacency matrix is ​​normalized to generate a Laplacian matrix, and the Laplacian matrix is ​​decomposed to generate spatial topological coding features. Traffic features, temporal coding features, and spatial topological coding features are fused together to generate low-dimensional spatiotemporal features.

[0100] In one possible implementation, each layer of the multimodal large model includes a temporal multi-head attention module and a spatial graph mask attention module; the aforementioned construction module 603 is specifically used for: The low-dimensional spatiotemporal features are input into the temporal multi-head attention module of the first layer of the multimodal large model. The temporal multi-head attention module performs temporal encoding on the low-dimensional spatiotemporal features through the upper triangular temporal causal mask to obtain the first intermediate feature. The first intermediate feature is input into the spatial graph mask attention module of the first layer of the multimodal large model. The spatial graph mask attention module performs spatial encoding on the first intermediate feature through the physical graph mask to obtain the first layer spatiotemporal features. The first-layer spatiotemporal features are input into the next-layer temporal multi-head attention module. The temporal multi-head attention module performs temporal encoding on the first-layer spatiotemporal features through the upper triangular temporal causal mask to obtain the second intermediate features. The second intermediate features are input into the next-layer spatial graph mask attention module. The spatial graph mask attention module performs spatial encoding on the second intermediate features through the physical graph mask to obtain the next-layer spatiotemporal features. The process is executed layer by layer until the last layer of the multimodal large model is completed, resulting in high-dimensional spatiotemporal features. These high-dimensional spatiotemporal features are used to characterize the high-dimensional latent features of each intersection in the road network of the target area.

[0101] In one possible implementation, the aforementioned building module 603 is specifically used for: The high-dimensional spatiotemporal features are input into the decoder, which calculates the feature interaction value between the high-dimensional latent features corresponding to any two intersections in the road network of the target area. The feature interaction values ​​are fused using the learned parameters of the decoder to obtain the collaborative control mapping values ​​corresponding to the two intersections; By using an activation function to map the collaborative control mapping values ​​corresponding to the two intersections, the probability that collaborative control is required between the two intersections can be obtained. Based on the probability that collaborative management is required between any two intersections in the target area's road network, a collaborative management probability matrix is ​​constructed.

[0102] In one possible implementation, the control topology graph includes multiple nodes and multiple edges. Each node represents an intersection in the target area's road network, and each edge represents the collaborative control relationship between two intersections connecting two nodes. The aforementioned second generation module 604 is specifically used for: The collaborative control probability matrix is ​​binarized to obtain a binary adjacency matrix; Treat each intersection in the target area road network as a node. Based on the probability that collaborative management is required between any two intersections in the target area road network, construct the edge between the nodes corresponding to any two intersections. Combine this with a binary adjacency matrix to generate a management topology graph.

[0103] In one possible implementation, the aforementioned control module 605 is specifically used for: Identify and manage connected components in the topology graph. Connected components include: linear chain connected components, grid connected components, and independent nodes. For each connected component, determine the control region corresponding to each connected component and generate the control strategy corresponding to the control region.

[0104] In one possible implementation, the aforementioned control module 605 is specifically used for: If the connected component is a linear chain connected component, then the control area corresponding to the linear chain connected component is determined to be the trunk green wave area, and the control strategy corresponding to the control area is determined to be a one-way or two-way green wave coordination strategy. If the connected component is a grid-like connected component, then the control area corresponding to the grid-like connected component is determined as a coordinated control area, and the control strategy corresponding to the control area is determined as a collaborative overflow prevention strategy. If the connected component is an independent node, then the independent node is treated as an independent control area, and the control strategy for the independent control area is determined to be the sensing control strategy.

[0105] In one possible implementation, the aforementioned road network traffic control device further includes a training module, which is specifically used for: A training dataset is constructed based on historical multi-source data of multiple regional road networks and historical control strategies corresponding to multiple regional road networks. The training dataset includes multiple training samples, each of which includes input feature samples generated based on historical multi-source data of regional road networks and expert truth matrices generated based on historical control strategies corresponding to regional road networks. Input the training dataset into the initial large model to obtain the actual probability matrix output by the initial large model; The loss result of the initial large model is determined based on the expert truth matrix and the actual probability matrix, and the initial large model is iteratively adjusted based on the loss result to obtain the multimodal large model.

[0106] In one possible implementation, the training module described above is specifically used for: Determine the expert cloning loss based on the actual probability matrix and the expert truth matrix; The symmetry constraint loss is determined based on the actual probability matrix and its transpose. Based on the actual probability matrix and the number of intersections in the regional road network, the sparsity regularization loss is determined. The expert cloning loss, symmetry constraint loss, and sparsity regularization loss are weighted and fused to obtain the loss result of the initial large model.

[0107] The processing flow of each module in the device and the interaction flow between each module can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.

[0108] This application also provides an electronic device 700, such as... Figure 7 The diagram shown is a structural schematic of an electronic device 700 provided in an embodiment of this application, including: a processor 701 and a memory 702, and optionally, a bus 703. The memory 702 stores machine-readable instructions executable by the processor 701. When the electronic device 700 is running, the processor 701 and the memory 702 communicate via the bus 703. When the machine-readable instructions are executed by the processor 701, the steps of the road network traffic control method described in any of the preceding claims are performed.

[0109] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the road network traffic control method described in any of the preceding claims.

[0110] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0111] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0112] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A road network traffic control method, characterized in that, include: Acquire multi-source data on the road network of the target area; Based on the multi-source data of the road network in the target area, low-dimensional spatiotemporal features are generated. The low-dimensional spatiotemporal features are input into a multimodal large model. The multimodal large model then extracts features from the low-dimensional spatiotemporal features layer by layer based on its built-in multi-layer cascaded structure, generating high-dimensional spatiotemporal features. Based on these high-dimensional spatiotemporal features, a collaborative control probability matrix is ​​constructed. The collaborative control probability matrix is... The matrix, Let be the number of intersections in the target area's road network, and be the th intersection. Line number The element values ​​of the column are used to characterize the first element in the road network of the target area. The intersection and the first The probability that coordinated management is needed between intersections; Based on the aforementioned collaborative control probability matrix, a control topology graph is generated; Based on the control topology map, a control strategy for the road network in the target area is generated, and traffic control is performed on the road network in the target area based on the control strategy. The step of constructing a collaborative management probability matrix based on the high-dimensional spatiotemporal features includes: inputting the high-dimensional spatiotemporal features into a decoder; the decoder calculating the feature interaction value between the high-dimensional latent features corresponding to any two intersections in the target area road network; fusing the feature interaction value using the learning parameters of the decoder to obtain the collaborative management mapping value corresponding to the two intersections; mapping the collaborative management mapping value corresponding to the two intersections using an activation function to obtain the probability that collaborative management is required between the two intersections; and constructing the collaborative management probability matrix based on the probability that collaborative management is required between any two intersections in the target area road network. The control topology graph includes multiple nodes and multiple edges. Each node represents an intersection in the target area road network, and each edge represents the collaborative control relationship between two intersections corresponding to two nodes. The step of generating the control topology graph based on the collaborative control probability matrix includes: binarizing the collaborative control probability matrix to obtain a binary adjacency matrix; taking each intersection in the target area road network as a node, constructing an edge between nodes corresponding to any two intersections based on the probability that collaborative control is required between any two intersections in the target area road network, and combining it with the binary adjacency matrix to generate the control topology graph.

2. The method according to claim 1, characterized in that, The generation of low-dimensional spatiotemporal features based on multi-source data of the target area road network includes: Based on the multi-source data, an original spatiotemporal tensor is constructed, and the traffic dimension features in the original spatiotemporal tensor are projected to obtain traffic features; Encode the temporal features in the original spatiotemporal tensor to generate temporally encoded features; A digital model of the road network in the target area is performed to obtain the physical adjacency matrix of the road network in the target area. The physical adjacency matrix is ​​as follows: The matrix, The number of road segments in the target area road network, the first... Line number The element values ​​of the column are used to represent the first... The section and the first Are the road sections connected? The physical adjacency matrix is ​​normalized to generate a Laplacian matrix, and the Laplacian matrix is ​​decomposed to generate spatial topological coding features. The traffic features, the temporal coding features, and the spatial topology coding features are fused together to generate the low-dimensional spatiotemporal features.

3. The method according to claim 1, characterized in that, Each layer of the multimodal large model includes a temporal multi-head attention module and a spatial graph mask attention module; The process of extracting features from the low-dimensional spatiotemporal features layer by layer from the multimodal large model based on a built-in multi-layer cascaded structure to generate high-dimensional spatiotemporal features includes: The low-dimensional spatiotemporal features are input into the temporal multi-head attention module of the first layer of the multimodal large model. The temporal multi-head attention module performs temporal encoding on the low-dimensional spatiotemporal features through an upper triangular temporal causal mask to obtain the first intermediate feature. The first intermediate feature is input into the spatial graph mask attention module of the first layer of the multimodal large model. The spatial graph mask attention module performs spatial encoding on the first intermediate feature through a physical graph mask to obtain the first-layer spatiotemporal features. The first-layer spatiotemporal features are input into the next-layer temporal multi-head attention module. The temporal multi-head attention module performs temporal encoding on the first-layer spatiotemporal features through an upper triangular temporal causal mask to obtain the second intermediate features. The second intermediate features are input into the next-layer spatial graph mask attention module. The spatial graph mask attention module performs spatial encoding on the second intermediate features through a physical graph mask to obtain the next-layer spatiotemporal features. The process is executed layer by layer until the last layer of the multimodal large model is completed, resulting in the high-dimensional spatiotemporal features. These high-dimensional spatiotemporal features are used to characterize the high-dimensional latent features of each intersection in the target area road network.

4. The method according to claim 1, characterized in that, The control strategy for generating the road network of the target area based on the control topology map includes: Identify the connected components in the control topology graph, including: linear chain connected components, grid connected components, and independent nodes; For each of the connected components, a control region corresponding to each connected component is determined, and a control strategy corresponding to the control region is generated.

5. The method according to claim 4, characterized in that, The step of determining the control region corresponding to each of the connected components and generating the control policy corresponding to the control region includes: If the connected component is the linear chain connected component, then the control area corresponding to the linear chain connected component is determined to be the trunk green wave area, and the control strategy corresponding to the control area is determined to be a one-way or two-way green wave coordination strategy. If the connected component is the grid-like connected component, then the control area corresponding to the grid-like connected component is determined to be the coordinated control area, and the control strategy corresponding to the control area is determined to be the collaborative overflow prevention strategy. If the connected component is the independent node, then the independent node is regarded as an independent control area, and the control strategy of the independent control area is determined to be the sensing control strategy.

6. The method according to claim 1, characterized in that, The multimodal large model is trained based on the following steps: A training dataset is constructed based on historical multi-source data of multiple regional road networks and historical control strategies corresponding to multiple regional road networks. The training dataset includes multiple training samples. Each training sample includes an input feature sample generated based on the historical multi-source data of the regional road network and an expert truth matrix generated based on the historical control strategies corresponding to the regional road network. Input the training dataset into the initial large model to obtain the actual probability matrix output by the initial large model; The loss result of the initial large model is determined based on the expert truth matrix and the actual probability matrix, and the initial large model is iteratively adjusted based on the loss result to obtain the multimodal large model.

7. The method according to claim 6, characterized in that, The step of determining the loss result of the initial large model based on the expert truth matrix and the actual probability matrix includes: Based on the actual probability matrix and the expert truth matrix, the expert cloning loss is determined; Based on the actual probability matrix and the transpose of the actual probability matrix, the symmetry constraint loss is determined. Based on the actual probability matrix and the number of intersections in the regional road network, the sparsity regularization loss is determined. The expert cloning loss, the symmetry constraint loss, and the sparsity regularization loss are weighted and fused to obtain the loss result of the initial large model.

8. An electronic device, characterized in that, include: The electronic device includes a processor and a memory, the memory storing machine-readable instructions executable by the processor, which, when the electronic device is in operation, are executed by the processor to perform the steps of the road network traffic control method as described in any one of claims 1 to 6.

Images