Traffic flow prediction and signal timing integrated optimization method based on IVCPS

Abstract

The invention relates to a traffic flow prediction and signal timing integrated optimization method based on IVCPS, and belongs to the field of intelligent traffic. According to the method, road network-level large-scale traffic prediction is completed in an information space through combination of local-global spatial-temporal characteristics, and traffic flow prediction and dynamic traffic association degree are combined in a physical space to realize coordination and unification of regional signal lamp optimization and traffic global optimization. Extracting local space-time features on the traffic prediction model in combination with adaptive graph convolution, extracting overall and overall long-time dependence features by using a Transform coding layer, and fusing and outputting a road network-level prediction result; the regional saturation is calculated based on a prediction result, the multi-index correlation degree is combined, traffic sub-regions are dynamically divided through an improved Louvain algorithm, an optimal signal timing scheme is solved by adopting an improved multi-target particle swarm algorithm under a differentiation optimization target, regional signal lamp cooperative control is realized, and the global road network traffic efficiency is effectively optimized while congestion is evacuated.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and relates to an integrated optimization method for traffic flow prediction and signal timing based on intelligent vehicle cyber-physical systems (IVCPS). Background Technology

[0002] In urban traffic control, traffic flow prediction and signal control are two key means to improve road efficiency. The ideal operating state of an Intelligent Vehicle Cyber-Physical System (IVCPS) is to dynamically adjust intersection signal control strategies based on accurate predictions of future traffic conditions, enabling signal timing to proactively adapt to changes in traffic demand, thereby reducing delays and improving the overall operational efficiency of the road network.

[0003] Currently, the main approaches to integrating traffic flow prediction and signal control optimization are as follows: (1) Fixed signal control based on historical traffic data. In existing urban traffic signal control, fixed or time-segmented signal timing schemes based on historical traffic flow statistical characteristics are often adopted. This type of method pre-sets the signal cycle, phase, and green ratio based on the average flow and traffic patterns in previous time periods. In practical applications, this method cannot respond to short-term fluctuations in traffic flow, sudden events, or special traffic patterns such as holidays, often leading to increased vehicle queue length and accumulated delays, and the control effect is significantly reduced under peak or abnormal conditions.

[0004] (2) Single-intersection adaptive control based on real-time detection and short-term prediction. To improve the flexibility of signal control, some systems introduce real-time traffic detection data and combine it with short-term traffic flow prediction to dynamically adjust the signal timing parameters of a single intersection. This type of method usually aims to reduce vehicle delays at the current intersection and corrects the next signal cycle based on the prediction results. However, since the prediction and control scope is limited to a single intersection or local area, it does not fully consider the traffic correlation between upstream and downstream road segments and adjacent intersections, which can easily lead to queues shifting in the road network and affect the overall coordination and stability of traffic operation.

[0005] (3) Traffic optimization methods that separate prediction and signal control. In recent years, with the development of big data and artificial intelligence technologies, traffic flow prediction models based on machine learning or deep learning (such as RNN, GCN, etc.) have been widely used in urban traffic systems. These methods can learn complex spatiotemporal evolution patterns from multi-source traffic data and improve prediction accuracy. However, due to the lack of a deep coupling mechanism between prediction and control, the prediction results fail to fully consider the counter-effects of signal control strategies on traffic flow evolution, and the signal control decisions do not incorporate prediction errors, uncertainties, and prediction feedback into the optimization process. This one-way information transmission mode limits the improvement of the overall system performance and makes it difficult to achieve closed-loop collaborative optimization of prediction and control.

[0006] Therefore, a new control method is urgently needed to solve the problem of congestion during special periods and on special road sections. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide an integrated optimization method based on traffic flow prediction and signal timing, which uses IVCPS technology to solve the problem of congestion relief during special periods and on special road sections, so as to achieve the problem of synergy between global scheduling and local dynamic optimization of the road network, thereby systematically improving urban traffic efficiency.

[0008] To achieve the above objectives, the present invention provides the following technical solution: An integrated optimization method for traffic flow prediction and signal timing based on IVCPS (Integrated Traffic Flow Prediction and Signal Timing) achieves large-scale traffic prediction at the road network level by combining local and global spatiotemporal features in the information space, and in the physical space, it combines traffic flow prediction with dynamic traffic correlation to achieve coordinated and unified regional signal light optimization and global traffic optimization. Specifically, it includes the following steps: S1: Obtain historical traffic flow data and topology of the urban traffic network, and construct a traffic network topology map; S2: Construct a traffic prediction model that combines local and global spatiotemporal features, namely the ACTTGCNformer model, to predict future road network traffic flow data; the ACTTGCNformer model includes a local spatiotemporal feature extraction module, a global spatiotemporal feature extraction module, and a feature fusion and prediction module; The local spatiotemporal feature extraction module extracts detailed features between local spatial nodes from a local perspective using an improved graph convolutional network, ACT-GCN. The output of ACT-GCN is then used as the input of GRU to extract local spatial detailed features and short-term temporal features of traffic data. Here, GRU stands for Gated Recurrent Unit. The global spatiotemporal feature extraction module introduces a combined embedding layer in the Transformer encoding layer to extract long-term dependencies and global spatial correlation features of traffic data; the combined embedding layer includes feature embedding, periodic embedding, and global spatial embedding. The feature fusion and prediction module adopts a "late fusion" approach, which concatenates the features output by the local spatiotemporal feature extraction module with the features output by the global spatiotemporal feature extraction module, and outputs the final traffic flow prediction result through a multilayer perceptron (MLP). S3: Calculate regional saturation based on predicted traffic flow results, combine multiple correlation indicators between road network nodes, and use the improved Louvain algorithm to dynamically divide the traffic network into regions to obtain traffic sub-regions with different saturation states. S4: For the different traffic sub-zones, determine the differentiated signal timing optimization objectives based on their saturation status. Under the overall constraints, use the improved multi-objective particle swarm optimization algorithm (IMOPSO) to solve the optimal signal timing scheme and realize personalized and coordinated control of regional traffic lights.

[0009] Further, in step S2, the ACT-GCN includes a cascaded fully connected layer, ACT-GC, and an activation function; the ACT-GC represents adaptive channel topology graph convolution, and its construction process includes: firstly, the input features are linearly transformed through a fully connected layer, thus transforming the input features... X Transform into a higher-level feature representation Secondly, through correlation modeling functions... To model the correlation between vertices; based on the established correlation modeling function, a linear transformation is used. Increasing the channel dimension increases feature diversity and yields channel-specific feature factors. In shared topology A d Introducing correlation Learn a channel-specific refined topology for each channel. T d Each channel's topology represents the relationships between nodes across different dimensions of traffic data. Finally, feature aggregation is performed on the channel topologies corresponding to different traffic dimensions to obtain the final output. Z This enables the extraction of detailed features of the local spatial topology of the road network.

[0010] Furthermore, in step S2, the feature embedding is obtained by using a fully connected layer to obtain the feature embedding of the sequence; the periodic embedding includes weekly embedding and daily embedding, and by connecting and broadcasting them, the traffic data time embedding is obtained; the global spatial embedding adopts basic spatial embedding, and by connecting the feature embedding and periodic embedding, the final combined embedding is obtained, that is, a high-dimensional dense traffic data representation.

[0011] Furthermore, in step S2, the Transformer encoding layer includes a time-dimensional Transformer layer and a spatial-dimensional Transformer layer, which are connected in series and utilize a multi-head attention mechanism to mine the global correlation between time and space respectively.

[0012] Furthermore, in step S3, the dynamic regional division of the traffic network using the improved Louvain algorithm specifically includes: (1) Combining the weight division of correlation degree: The intersection is abstracted as a network node, and an index system including signal cycle correlation degree, traffic flow correlation degree, density correlation degree and distance correlation degree is constructed. By integrating each correlation degree and assigning different weight parameters, the edge weight based on correlation degree is obtained. (2) Local adjustment based on modularity optimization: After obtaining the edge weights based on the degree of association, the modularity is optimized, that is, by adjusting the community structure in the network, the modularity is maximized. (3) Global adjustment based on community aggregation: Calculate the current average saturation of the region and the predicted average saturation at the next adjustment time. If the two do not belong to the same saturation range, dynamically adjust the weight coefficients of each correlation index and re-divide the region.

[0013] Furthermore, in step S3, when optimizing the modularity, a regularization penalty term is used to suppress the formation of ultra-large regions, expressed as:

[0014] in, For the optimized modularity, for k area, for k The total number of nodes in the region N This represents the total number of nodes across all regions. To control the intensity of punishment;

[0015] in, For modularity, Represents a node and nodes Edge weights between them; Representing nodes respectively and nodes The degree (i.e., the sum of the weights of the edges connected to that node). This represents the sum of the weights of all edges in the network; Represents an indicator function, when the node and nodes If they belong to the same community, the value is 1; otherwise, it is 0.

[0016] Furthermore, in step S4, the overall constraint condition is:

[0017] in, For the common signal period; The total lost time includes the sum of the yellow light time and the all-red light time for all phases, which is the time during which vehicles cannot pass. It is the sum of the flow rate ratios of each phase, and the maximum value of the flow rate to saturation flow rate ratio among all phases of the critical intersection; n This represents the number of intersections within a sub-region of the road network. Minimum green light time, It is the minimum green light time for pedestrians; Traffic flow saturation; This represents the phase difference between two adjacent intersections.

[0018] Furthermore, in step S4, determining the differentiated signal timing optimization objective based on its saturation state includes: (1) In areas with low traffic flow saturation, minimizing total vehicle delay and minimizing total exhaust carbon emissions are taken as optimization objectives; (2) In the saturation area of ​​traffic flow, the optimization objectives are to minimize vehicle delay, minimize the number of stops, and maximize traffic capacity; (3) In areas with high traffic flow saturation, maximizing traffic capacity will be the optimization objective.

[0019] Furthermore, in step S4, the improved multi-objective particle swarm optimization algorithm (IMOPSO) includes: (1) Dynamic adjustment of inertia weight: The distribution entropy of the Pareto front is quantified to reflect the degree of solution dispersion; the current particle population diversity is judged based on the historical maximum and minimum entropy values ​​up to the current iteration, and the inertia weight is dynamically adjusted accordingly; when the population diversity is high, the weight is reduced to enhance local exploration capability; when the population diversity is low, the weight is increased to enhance global exploration capability and avoid premature convergence.

[0020] (2) Dynamic adjustment of learning factors: The average normalized distance between the current population and the reference Pareto front is used as the convergence index; an exponential decay mechanism is introduced to dynamically adjust individual learning factors, and a linear interpolation mechanism is introduced to dynamically adjust group learning factors, so as to balance the influence of individual experience and group cooperation.

[0021] The beneficial effects of this invention are as follows: (1) This invention addresses the challenge of achieving coordinated regional traffic scheduling and traffic light optimization control at road network intersections at a single scale by designing a dual-scale solution. Specifically, at the large-scale road network level, this invention designs a regional division algorithm based on the state feedback and traffic flow prediction of regional traffic flow, transforming the traffic scheduling of the road network into personalized traffic control for the region; at the small-scale regional level, this invention designs a corresponding signal timing optimization algorithm based on personalized traffic objectives and traffic conditions, realizing real-time regional traffic scheduling under road network-level traffic planning.

[0022] (2) To address the challenge of co-modeling short-term local features and long-term global features of the road network during specific time periods and road segments in traffic flow prediction, this invention proposes a traffic prediction model (ACTTGCNformer) that combines local and global spatiotemporal features. Specifically, from a local perspective, an adaptive channel topology graph convolution is designed to extract local spatial details and short-term features of traffic flow; from a global perspective, a combined embedding Transformer module is designed to extract long-term dependencies and long-distance spatial correlation features of traffic flow by introducing a combination of feature, time, and space embedding. Multi-dimensional feature learning replaces single-dimensional feature learning, and the dimensions of the global and local feature layers are spliced ​​together to form a traffic flow fusion perception scheme that effectively complements spatiotemporal local details and global regular features.

[0023] (3) In view of the problem that regional signal timing is difficult to meet the dynamic personalized traffic needs, the present invention designs a regional signal timing optimization scheme for dynamic traffic correlation. Specifically, the present invention uses the traffic flow saturation sensed by local and global collaboration to divide personalized traffic status categories, and dynamically adjusts the weight of node correlation indicators such as signal cycle correlation degree and traffic flow correlation degree in combination with traffic status. Based on the traffic status and node correlation degree, the Louvain algorithm is used to perform personalized dynamic division of the region. Targeted optimization objectives and signal timing schemes are established for different personalized regions and solved by the improved multi-objective particle swarm algorithm to realize personalized collaborative control of regional traffic lights, thereby systematically improving urban traffic efficiency.

[0024] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the traffic network topology; Figure 2 The ACTTGCNformer model framework; Figure 3 This describes the channel topology feature extraction process. Figure 4 An improved graph convolutional network structure; Figure 5 This is a schematic diagram of the GRU structure; Figure 6 A structure for extracting spatiotemporal features of traffic data from a local perspective; Figure 7 This is a schematic diagram of the combined embedding; Figure 8 This is a schematic diagram of the spacetime structure of the Transformer. Figure 9 This is a schematic diagram of the MLP structure; Figure 10 This indicates the direction of traffic flow at adjacent intersections of the MLP. Figure 11 Flowchart for region division; Figure 12 Diagram showing the obstruction at the front of the convoy. Figure 13 A diagram illustrating the obstruction at the rear of the convoy; Figure 14 This is a schematic diagram of three-phase and four-phase intersections. Detailed Implementation

[0026] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0027] Example 1: Please see Figures 1 to 14 This embodiment provides an integrated optimization method for traffic flow prediction and signal timing based on IVCPS. It designs a strategy that meets the personalized signal timing requirements of different regions through three stages: traffic flow prediction, dynamic area division, and signal timing optimization. This provides a more accurate, efficient, and intelligent solution for urban road network management.

[0028] 1. Traffic flow prediction The purpose of traffic flow forecasting is to use historical traffic flow data, combined with topology maps and other external information, to build a predictive model to accurately predict traffic flow over a future period.

[0029] Prediction Principle: In regional traffic prediction tasks, each intersection within the region is used as a key node, and the traffic network is abstracted into a topological graph structure, such as... Figure 1 As shown.

[0030] Define the urban transportation network as an unweighted graph. This indicates that the unweighted graph encompasses the relationships between network nodes. Among them, This represents the set of sensor information points in the road network. n It refers to the number of sensor information acquisition nodes; This represents the set of edges between any two road nodes. The adjacency matrix represents the connection relationship between nodes in the road network. The specific expression of the set of sensor information collection points in the matrix is ​​shown in Equation (1).

[0031] (1) Based on the constructed traffic network topology, an improved GCN-Transformer composite model is established. The overall model architecture is as follows: Figure 2 As shown. The designed model includes the following modules: (1) A local spatiotemporal feature extraction module is composed of a graph convolutional network (GCN) and a gate recurrent unit (GRU). An adaptive channel topology graph convolution is introduced in this module to improve the dynamic mining capability of spatiotemporal features of traffic network from a local perspective.

[0032] (2) A global spatiotemporal feature extraction module, mainly composed of co-embedded embedding and a Transformer encoder layer. A co-embedded embedding layer is introduced into the Transformer encoder layer, replacing complex model design with effective representation of the data itself, to achieve exploration of spatiotemporal relationships from a global perspective, including feature embedding, periodic embedding, and global spatial embedding. This improves the modeling of spatiotemporal regularities from a global perspective. (3) Feature fusion and prediction module: The "late fusion" method is used to fuse the features extracted by different modules. This method only fuses the final output features of each model, which can retain the independence of each model's features while avoiding mutual interference of gradients during training. Then, it is combined with multilayer perceptron (MLP) for traffic flow prediction.

[0033] 1.1) Local Spatial Feature Extraction In the local spatial feature extraction module, based on the significant advantages of GCN in topology construction and spatial local feature extraction, the same feature aggregation approach as scalable semi-supervised learning graph convolutional neural networks (GCNs) is adopted. To address the problem of insufficient feature detail representation between local topology nodes, an adaptive channel topology graph convolution is introduced. This convolution differs from ordinary graph convolution structures and can simultaneously learn the shared topology structure of the road network and the topology features specific to different channels, so as to achieve adaptive detailed feature extraction of local traffic topology structure.

[0034] A shared topology built upon an urban road network can be understood as a fixed adjacency matrix in graph convolution. It serves as a topological prior for all layers, using the same topological structure across different channels for feature aggregation and updating, and providing general correlations between vertices. Typical topological shared graph convolution utilizes weights... W Perform feature transformation, and simultaneously through a ij right V i Aggregate the neighboring vertices and update their representations. z i , as shown in equation (2).

[0035] (2) in, for The neighborhood representation, x j It is a node v j The features are represented on the surface.

[0036] In static diagram a ij Often, these methods rely on predefined parameters from the input topology graph, thus lacking the ability to extract detailed features between different topology nodes. Therefore, to improve the ability to adaptively extract detailed features between nodes in the traffic topology graph, an Adaptive Channel Topology Graph Convolution (ACT-GC) is constructed by learning shared topological structures and channel-specific topological features. The specific process is as follows: First, a fully connected layer is used to linearly transform the input features, thus transforming the input features... X Transform into a higher-level feature representation As shown in equation (3): (3) Secondly, through the correlation modeling function To model the correlation between vertices. To save computational costs, the input features are sent to... Previously, using linear transformation p and q To reduce feature dimensionality, a simple and effective correlation modeling function is employed. It is represented as shown in equation (4).

[0037] (4) in, It is an activation function; The modeling principle is based on calculations. and The distance between them, and the nonlinear transformation of these distances as v i and v j Channel-specific topology relationships between them; p and q It is a 2-in-1 convolution kernel with a kernel size of 1. d Convolution, then average over the dimensions.

[0038] Based on the established correlation modeling function, linear transformation is used. Increasing the channel dimension increases feature diversity and yields channel-specific feature factors. As shown in equation (5).

[0039] (5) in, yes A vector in reflects and The topological relationships between them. To increase the flexibility of related modeling, It is defined as a non-forced symmetric correlation factor, i.e. .

[0040] By sharing topology A d Introducing correlation This allows learning a channel-specific refined topology for each channel. The topology of each channel represents the relationship between nodes under different dimensions of traffic data, as shown in equation (6).

[0041] (6) in, It is a trainable scalar with adjustable refinement intensity.

[0042] Finally, feature aggregation is performed on the channel topology corresponding to different traffic dimensions to obtain the final output. Z This enables the extraction of detailed features of the local spatial topology of the road network, as shown in equation (7).

[0043] (7) in, It's a connection operation. The specific process for extracting local spatial channel topology features is as follows: Figure 3 As shown.

[0044] Based on Adaptive Channel Topology Graph Convolution (ACT-GC), an improved graph convolutional network, ACT-GCN, is constructed by concatenating fully connected layers, improved graph convolutional modules, and activation functions. Figure 4 As shown, a fully connected layer is introduced after the input features to perform a linear transformation on the input features. For high-dimensional input data, it is mapped to an initial representation with lower dimensions for forward propagation. Before the output, batch normalization (BN) and ReLU activation function operations are performed on the obtained convolution. Finally, residual connections are introduced to obtain the output of the improved convolution module, as shown in equations (8) and (9). (8) (9) 1.2) Short-term time feature extraction After extracting the locally refined spatial features, GRU (Gated Recurrent Unit) is used to extract short-term temporal correlations. GRU has excellent extraction capabilities for short-term features due to its simple structure and high computational efficiency.

[0045] The core of GRU lies in two gating mechanisms: the Update Gate and the Reset Gate. Figure 5 shows the specific structure of GRU. The Update Gate determines whether to save the information from the previous time step in the current hidden state and whether to update the amount of information. The Reset Gate determines the degree to which the information from the previous time step should be ignored in the current hidden state. The formulas are shown in equations (10) and (11).

[0046] (10) (11) in, This indicates that the gate status is being updated at the current time. ; This indicates that the door status is being reset at the current time. ; This represents the input at the current moment; This indicates the hidden state at the previous moment; , These are applied to the current input of the update gate and the reset gate, respectively. The weight matrix; , These are respectively used to update the door and reset the door's hidden state from the previous moment. The weight matrix; , These are the offsets for updating the door and resetting the door, respectively. express sigmoid Activation function; , , , These are the parameters that the GRU module needs to learn.

[0047] The GRU learns short-term features in two main steps. The first step is to generate candidate hidden states by combining the input gate and the reset gate, as shown in equation (12).

[0048] (12) in, This represents the candidate hidden state at the current moment; , These are the input features and the weight matrices of the hidden states, respectively. Indicates the output of the reset gate. Hidden state compared to the previous time step Multiply element by element.

[0049] Finally, the final output of the current hidden state is determined by updating the gate, as shown in equation (13).

[0050] (13) In the short-term time feature extraction module, each GRU layer considers the output of the previous GRU layer, thus effectively capturing the short-term time features of local traffic flow time series data.

[0051] From a local perspective, an improved Graph Convolutional Network (ACT-GCN) is used to extract detailed features between local spatial nodes. The output of this improved GCN is then used as the input to a Generative Routine Analyzer (GRU). Leveraging the GRU's strength in extracting short-term temporal features, this approach effectively models the dynamic spatiotemporal features of traffic data, thereby improving the predictive model's dynamic adaptability to changes in traffic conditions. The structure for extracting spatiotemporal features of traffic data from a local perspective is as follows: Figure 6 As shown.

[0052] 1.3) Combined Embedded Design To gain a more comprehensive understanding of the global spatiotemporal characteristics of traffic data, an embedding layer is introduced into the input. Compared to complex model structure design, embedding simplifies the model structure while improving the effective understanding of the traffic data itself. The designed embedding layer is a co-embedded layer, composed of feature embedding, periodic embedding, and global spatial embedding, such as... Figure 7 As shown below. The following is a detailed introduction to combined embeddings.

[0053] (1) Feature embedding: In order to preserve the original information in the original traffic data, a fully connected layer is used to obtain the feature embedding of the sequence, as shown in Equation (14): (14) in, , This represents a fully connected layer.

[0054] (2) Periodic embedding: To better perceive the periodicity in time series. Periodic embedding consists of three parts, namely, week embedding ( ) and embedded in Japan ( By concatenating and broadcasting these data, the temporal embeddings of traffic data can be obtained. The formulas are shown in equations (15), (16), (17), and (18).

[0055] (15) (16) (17) in, , These are the weekly embedding matrix and the daily embedding matrix, respectively. Indicates the dimension of the week embedding (number of days in a week); This indicates the dimension of the daily embedding (the number of timestamps per day). Indicates a specific time interval (within one hour); They represent time steps respectively The corresponding weekly index and timestamp index.

[0056] (3) Global Spatial Embedding: In order to capture the spatially inherent dependencies in the global road network, such as the interaction of traffic flow at different geographical locations, the spatial embedding adopts the basic spatial embedding, as shown in Equation (18). (18) in, N This represents the number of sensors in the road network.

[0057] In summary, by connecting the above embeddings, the final combined embedding expression is obtained as shown in equation (19).

[0058] (19) The sum of hidden dimensions .

[0059] By fusing three embedding methods to form co-embedding, a high-dimensional and dense representation of traffic data is generated, thereby effectively supporting the extraction and analysis of global spatiotemporal features of traffic data.

[0060] 1.4) Extraction of global spatiotemporal features of road network Building upon enhanced traffic feature representation through embedding, a Transformer multi-head attention mechanism is employed to map spatiotemporal features into a new feature space, represented as queries (Q), keys (K), and values ​​(V). Furthermore, a multi-dimensional (temporal and spatial) Transformer hierarchical concatenation process is used to achieve in-depth mining of the global spatiotemporal features of the traffic system. Figure 8 As shown.

[0061] In the time dimension Transformer layer, the query matrix, key matrix, and value matrix are obtained. Using a multi-head attention mechanism, the attention of each head is calculated independently first, and finally the outputs of all heads are concatenated and linearly transformed to obtain the (temporal relationship in different spatial nodes). The outputs of the time dimension Transformer layer are shown in equations (20), (21), and (22).

[0062] (20) (twenty one) (twenty two) in, It is the output projection matrix of the Transformer layer in the time dimension. LayerNorm This involves residual joins and layer normalization operations. The expressions for the query matrix, key matrix, and value matrix are as follows: , Given input, with T time steps and N spatial nodes, , , It is the first in the time dimension of the Transformer h The projection matrix of the head, Dimensions for each head.

[0063] Similarly, by taking the output of the time-dimensional Transformer layer as the input of the spatial-dimensional Transformer layer and passing it through the same structure as the time-dimensional Transformer layer, the global spatiotemporal correlation output can be obtained, as shown in equations (23) and (24).

[0064] (twenty three) (twenty four) in, This is the output projection matrix of the Transformer layer in the spatial dimension. The expressions for the query matrix, key matrix, and value matrix are as follows: .

[0065] Therefore, from a global traffic perspective, by combining the advantages of Transformer in extracting long-distance / long-sequence features and integrating multiple embeddings, the ability to extract spatial long-distance features and long-term regular features can be effectively improved, and it can complement the dynamic spatiotemporal feature extraction from a local perspective.

[0066] 1.5) Feature fusion and prediction output After extracting local and global spatiotemporal features from the road network traffic data, a "Late Fusion" method is used to achieve feature fusion. The core of "Late Fusion" is to fuse only the final output features of two independently trained models. This approach preserves the independence of each model while avoiding gradient interference during training, and achieves simple and efficient complementarity through feature concatenation, while ensuring that the prediction model effectively takes into account the dynamics, stability, and accuracy of traffic changes.

[0067] Since the last layer of ACTT-GCN and Transformer respectively reflects the key features of local dynamic spatiotemporal characteristics and global long-term regularity characteristics, the features of the last layer of ACTT-GCN module and Transformer module are extracted respectively. This effectively preserves the unique features of each specific model while improving the overall predictive ability of the model.

[0068] The feature extraction of the last layer of the ACTT-GCN module graph convolutional layer is represented as follows: ; The feature extraction of the last layer of the self-attention layer in the Transformer module is represented as follows: ; in, T For time steps, N This represents the number of nodes in the transportation network. and These are the feature dimensions of the ACTT-GCN module and the Transformer module, respectively.

[0069] Finally, the two features are concatenated along the feature dimension to obtain the fused feature, as shown in equation (25): (25) in, , For connection operations.

[0070] This fusion approach retains the unique advantages of both models: ACTT-GCN captures local dynamic features, while Transformer captures global long-term dependency features. At the same time, it achieves complementary enhancement between features of the two models through the concatenation operation.

[0071] The final prediction module uses a multi-layer perceptron (MLP) as the prediction model to predict traffic flow. A schematic diagram of MLP operation is shown below. Figure 9 As shown.

[0072] When making predictions using an MLP, traffic flow features are passed from the MLP input layer through a series of hidden layers to the output layer. Each neuron receives input from the neurons in the previous layer and performs a weighted and nonlinear transformation of these inputs through an activation function. Each node in the MLP contains a neuron with a nonlinear activation function, which can effectively fuse temporal and spatial features and decode them into the traffic flow for the next time step.

[0073] 2. Dynamic Region Division The Louvain algorithm is a community detection algorithm based on modularity optimization. Its main goal is to identify community structures in a network by maximizing modularity. The algorithm has low time complexity and its model structure is easy to improve (e.g., similarity models based on traffic flow features).

[0074] Therefore, this method improves the Louvain community algorithm by introducing a traffic correlation index, adding a qualitative description of the state of the divided areas, and dynamically adjusting the correlation coefficient based on traffic flow prediction data to achieve a more accurate dynamic division of the traffic network. The following is a detailed description of the division process.

[0075] 2.1) Weighting based on relevance In road network topology, intersection nodes exhibit complex interrelationships, and the static edge weights of the modularity in the Louvain algorithm are insufficient to effectively represent these intrinsic interactions. Therefore, to achieve more accurate dynamic division of traffic areas, intersections are abstracted as network nodes, and an index system is constructed that includes signal cycle correlation, traffic flow correlation, density correlation, and distance correlation. The connection weights between nodes are dynamically quantified to achieve a rich representation of node relationships.

[0076] Modularity is a core indicator for measuring the quality of region partitioning, used to assess the tightness of community structure in a network. Its basic idea is that a good community partitioning should make the connections within communities as tight as possible, while the connections between communities should be as sparse as possible. The formula for calculating modularity is shown in expression (26).

[0077] (26) in, Represents a node and nodes Edge weights between them; Representing nodes respectively and nodes The degree (i.e., the sum of the weights of the edges connected to that node). This represents the sum of the weights of all edges in the network; Represents an indicator function, when the node and nodes If they belong to the same community, the value is 1; otherwise, it is 0.

[0078] To achieve more accurate dynamic partitioning, in static... Based on edge weights, traffic correlation is introduced to enrich the representation of node relevance, and different correlation weights are assigned according to different states. The following is a detailed introduction to the correlation.

[0079] (1) Signal periodicity correlation Signal cycle correlation is an indicator that quantifies the degree of signal coordination between two adjacent intersections by measuring the difference in signal cycles. Adjacent intersections typically use similar signal cycles to maintain a stable phase difference. If the difference in signal cycles between intersections is too large, delays will occur at the intersection with the shorter signal cycle. Therefore, the smaller the cycle difference, the more similar the traffic characteristics of the two adjacent intersections are, and the more suitable they are to be included in the same control zone. The defined signal cycle correlation is shown in equation (27).

[0080] (27) In the formula,

[0081] (2) Traffic flow correlation The correlation of traffic flow at intersections refers to the quantitative expression of the interaction between traffic flows at adjacent intersections. The larger the correlation value, the higher the correlation degree. Its expression is shown in equation (28).

[0082]

[0083] in, This represents the normalized travel time of a vehicle from the upstream intersection to the downstream intersection; This indicates the number of traffic branches (such as left-turn, straight, and right-turn lanes) flowing from the upstream intersection to the downstream intersection. Indicates the first Traffic flow of each branch; This represents the largest single-branch traffic flow among all branches.

[0084] The direction of one-way traffic flow at adjacent intersections is as follows: Figure 10 As shown in the figure. Since the traffic flow is bidirectional, the expression for the traffic flow correlation degree after considering bidirectional traffic flow is shown in equation (29).

[0085] (29) in, Indicators representing the correlation between two-way traffic flows; , They represent the intersections. i arrive j , j arrive i Traffic flow correlation in direction.

[0086] (3) Density correlation Traffic density on a road segment can directly reflect the congestion status of that segment; therefore, density correlation is also an important indicator of the relationship between road network nodes. A higher density correlation indicates a stronger congestion correlation between road segments along the path. The expression for density correlation is shown in equation (30).

[0087] (30) The expression after considering the two-way density is shown in equation (31).

[0088] (31) in, Indicates an intersection i and intersection j Density correlation between them Indicates from the intersection i To the intersection j Total traffic volume. These represent the number of lanes and the length of the road segment in the corresponding direction, respectively.

[0089] (4) Node distance correlation Studies have shown that the interaction between intersections is most significant within 800 meters. This interaction includes the transmission of traffic flow, the propagation of congestion, and the coordinated control of traffic lights. Therefore, the distance between adjacent nodes i and j is used as the node distance correlation degree, as shown in Equation (32).

[0090] (32) By combining the correlation between nodes, signal cycle correlation, traffic flow correlation, and traffic density correlation, and assigning different weight parameters, edge weights based on correlation are obtained. The expression is shown in equation (33).

[0091] (33) Among them, Weight parameters during region partitioning Dynamic weighting is performed based on regional saturation.

[0092] 2.2) Local Adjustment Based on Modularity Optimization After obtaining the edge weights that combine the correlation degree, the modularity is optimized by adjusting the community structure in the network to improve the modularity. When the modularity M reaches its maximum value, it signifies that the Louvain algorithm has completed the optimization of the region structure at the current level, and the partitioning at this point is optimal under the modularity index.

[0093] Specifically, this means treating each node as a separate community; computing each node... The module degree increment after moving to the community of adjacent nodes (preferably nodes with high similarity) Determine if the maximum modularity increment of the node is greater than 0. If it is, place the node in the community of the neighboring node with the largest modularity increment; otherwise, do not move it. Iterate the above process until the regions to which all vertices belong no longer change. The specific formula for the modularity increment is shown in equation (34).

[0094] (34) in, The sum of the weights of the edges within the target community. For nodes The sum of edge weights to nodes within the target community; It is the sum of edge weights of all nodes in the target community.

[0095] When optimizing modularity, a few super-large regions are often formed, resulting in an uneven distribution of traffic areas. Therefore, by introducing a regularization term, the balance of region size is constrained while optimizing modularity. The regularization penalty term suppresses the formation of super-large regions and retains small-scale regions with key functions, thereby improving the rationality and practicality of the partitioning. The expression is shown in Equation (35).

[0096] (35) in, fork area, for k The number of all nodes in the region. N This represents the total number of nodes across all regions. To control the intensity of punishment.

[0097] Assuming the node From the region A Move to area B The change of the regularization term is shown in equation (36).

[0098] (36) If the area is moved B Increased scale ( Then, regularization penalty Enhance, This reduces the tendency to merge, thereby effectively suppressing the formation of super-large regions and resulting in a more balanced and reasonable zoning scale.

[0099] 2.3) Global Adjustment Based on Community Aggregation After optimizing the modularity using the Louvain algorithm, the entire traffic network has been initially divided. To obtain a more reasonable partitioning result globally, the optimized traffic regions need to be merged into a new supernode, and the edge weights of the supernode need to be recalculated. This process is then iterated until the modularity no longer increases significantly.

[0100] Traffic conditions are dynamic, therefore, region partitioning should also be a dynamic process. Since the static weight design of the Louvain algorithm is ill-suited to the dynamic spatiotemporal changes of the traffic network, this paper proposes a method to dynamically update the weight coefficients of the correlation index between regional nodes by comparing the current saturation level of a region with the saturation level at the next adjustment time. This allows for dynamic adjustment of region partitioning. To prevent abrupt adjustments from causing system instability, existing research generally recommends that the region adjustment cycle be controlled within 15-30 minutes.

[0101] The average saturation of all intersections within a region is defined as the regional average saturation, and the expression for the regional average saturation is shown in equation (37).

[0102] (37) in, This indicates taking the average value; Indicates an intersection Traffic flow, Indicates an intersection Maximum traffic capacity. The regional saturation at the next adjustment time.

[0103] if and If they belong to the same saturation range, maintain the original partitioning results; if and If they do not belong to the same saturation range, the weighting coefficients are adjusted and the regions are redefined.

[0104] The average saturation of a region reflects the traffic conditions of that region. Generally speaking, the higher the saturation, the more severe the traffic congestion and the worse the traffic conditions. Conversely, the lower the saturation, the better the traffic conditions. Table 1 shows the road traffic conditions corresponding to different saturation values.

[0105] To achieve a more reasonable dynamic partitioning, SPASS principal component analysis was used to determine the weights of each indicator. The weight percentages were determined by analyzing the principal component contribution rates of signal cycle correlation, flow correlation, and density correlation at key intersections under three saturation levels. Table 2 shows the contribution rates of each indicator obtained from SPASS principal component analysis.

[0106] Table 1 Correspondence between Saturation and Traffic Condition

[0107] Table 2 Principal Component Contribution Rate

[0108] As can be seen, under low saturation, the road is in a smooth state, and the signal cycle and traffic flow are the core of coordinated control, so the density correlation accounts for a small proportion; under medium saturation, the road becomes congested, and the influence of density correlation begins to become significant, but the influence of signal cycle and traffic flow is still important; under high saturation, regional congestion is severe, so density correlation becomes the most critical influencing factor.

[0109] Based on the analysis in Tables 1 and 2, the correlation weight coefficients under different saturation levels can be obtained according to the principal component contribution rates of each correlation degree. The weight model is shown in Equation (38).

[0110] (38) Based on the actual road network division, in order to obtain a more reasonable partitioning, the partitioning of this model follows the principles below: (1) Principle of Regional Independence: To ensure the effective implementation of traffic management strategies, this model strictly adheres to the constraint of unique road segment affiliation, meaning that any road segment can only be assigned to a single traffic sub-region, eliminating cross-sub-region overlap. Under this principle, it is conducive to the reasonable classification of all road segments in the road network and ensures that there are no omissions or duplicate assignments.

[0111] (2) Dynamic Adaptability Principle: By comparing the current sub-region saturation with the sub-region saturation at the next adjustment time, the weight coefficients of different correlation indicators are updated to ensure that the traffic sub-region division can adapt to the time-varying characteristics of traffic conditions. This improves the model's adaptability to dynamic traffic networks.

[0112] (3) Principle of zoning stability: Set a reasonable adjustment cycle for sub-zones to prevent "oscillation effect" caused by sudden adjustment, so as to balance the responsiveness and stability of the traffic system.

[0113] (4) Spatial proximity principle: Although some road segments that are not directly connected and are far apart may be classified into the same sub-zone due to similar traffic patterns, their overall spatial layout should maintain a clustered characteristic and avoid a discrete distribution. This principle is conducive to optimizing the physical manageability of sub-zone division.

[0114] In summary, the flowchart for the traffic network area division based on dynamic correlation is as follows: Figure 11 As shown.

[0115] 3. Signal timing optimization Traffic demand varies significantly across different times and regions, such as congestion relief during peak hours and environmental protection needs during off-peak hours. To address this issue, after identifying regions with different saturation levels using a traffic flow prediction-based regional segmentation model, personalized optimization objectives will be designed for different types of regions based on their differentiated needs. Furthermore, signal control parameters such as intersection signal cycles and green light ratios will be adjusted and allocated more rationally under different conditions based on these optimization objectives, thereby improving road network resource utilization while resolving congestion.

[0116] 3.1) Overall Constraints When constructing complex mathematical models, the rigorous setting of constraints is equally important. In signal control systems, signal period, phase, saturation, and other parameters are crucial indicators. Therefore, when designing appropriate signal timing schemes for different regions, relevant constraint design is necessary.

[0117] (1) Common signal cycle: Even at different intersections in the same area, the signal cycles still differ. Without a unified common signal cycle as a benchmark, problems such as disordered traffic flow, increased intersection conflicts, and limited priority of public transportation will occur. Therefore, when designing regional signal timing strategies, it is necessary to prioritize the development of a unified common signal cycle for sub-regions of the road network to ensure coordinated operation among nodes.

[0118] The intersection with the highest traffic volume is designated as the critical intersection in the area, and the signal cycle of the critical intersection is taken as the common cycle. The common cycle is estimated using the classic Webster formula, as shown in equation (39).

[0119] (39) in, The total lost time includes the sum of the yellow light time and the all-red light time for all phases, which is the time during which vehicles cannot pass. The sum of the flow rate ratios for each phase is the maximum value of the ratio of flow rate to saturation flow rate across all phases of the intersection. Generally, this represents the sum of the flow rate ratios for each phase. .

[0120] Based on experience in traffic engineering, the minimum signal light cycle should not be less than the sum of the minimum green light times for each phase, but it should not exceed 200 seconds.

[0121] (2) Traffic flow saturation: In order to effectively solve the problem of regional optimization control, it is defined that there is no oversaturation state at the intersection. Therefore, the range of intersection saturation setting is as shown in equation (40).

[0122] (40) (3) Minimum Green Light Time: The minimum green light time is a key parameter in traffic signal timing, referring to the shortest green light duration that a certain phase must maintain within a signal cycle. Essentially, it is a lower limit constraint on the green light ratio, ensuring that vehicles and pedestrians can pass through intersections safely and orderly. The minimum green light design of this invention refers to the "Specifications for the Application of Traffic Signal Control Methods on Urban Roads". The minimum green light time for the pedestrian phase is used as the overall minimum green light time, covering the minimum green light time for the motor vehicle phase during design. The straight-ahead phase for motor vehicles is not less than 15 seconds, and the left-turn phase is not less than 10 seconds. The expressions are as shown in equations (41) and (42). (41) (42) in, It is the minimum green light time for pedestrians (including safety redundancy), L is the pedestrian crossing distance (meters), and 1.2 (meters / second) is the pedestrian walking speed.

[0123] (4) Phase difference Phase difference is the time difference between the start times of the green light at two adjacent intersections. Since the research object is area signal control, phase difference is one of the most important constraints to consider. The expression for the phase difference between two adjacent intersections is shown in equation (43).

[0124] (43) In summary, the overall constraints are shown in equation (44).

[0125] (44) 3.2) Design of Regional Optimization Objectives Based on the constraints mentioned above, and according to the traffic characteristics of different saturation levels, more suitable personalized optimization targets are designed for each region.

[0126] (1) Signal optimization objectives in areas with low traffic flow saturation When the control area is a low-saturation sub-region, it indicates that traffic in that area is flowing smoothly. Since the traffic flow in low-saturation sub-regions is relatively small, based on the national green development strategy and actual traffic scheduling needs, the optimization objective is primarily to improve vehicle mobility while reducing unnecessary delays and carbon emissions. This invention uses total vehicle delays and total carbon emissions as the optimization objectives for low-saturation sub-regions.

[0127] ① Total vehicle delays Traffic flow delays within a region can be categorized into two scenarios based on the time it takes for vehicles to pass through intersections and the phase difference between intersections. The first scenario is when the front of the traffic flow arrives at the intersection at a red light; the second scenario is when the rear of the traffic flow arrives at the intersection at a red light. Because the front and rear of a traffic queue cannot be blocked simultaneously, the calculation of total vehicle delay is divided into two cases: The front of the convoy was blocked, such as Figure 12 As shown, when the convoy reaches the stop line, the intersection light is red, and the phase green light has not yet been turned on. The vehicle delay formulas at this time are shown in equations (45) and (46).

[0128]

[0129]

[0130] in, This indicates the time (in seconds) remaining until the red light ends when the head of the convoy arrives. , These represent the distance between the two intersections, the phase difference, and the average speed of vehicles traveling on the road segment, respectively. When the head of the convoy is obstructed... Indicates the first n The capacity of each intersection (veh / h); Indicates the first n Traffic volume at each intersection (veh / h).

[0131] The rear of the convoy was blocked, such as Figure 13 As shown, the convoy arrives some time after the green light at the intersection begins, and the light turns red when the tail of the convoy reaches the intersection. The vehicle delay formula at this time is (47).

[0132]

[0133] in, For the first n The duration of the red light at each intersection; At that time, the rear of the convoy was blocked.

[0134] In summary, assuming there are k intersections in the area, the total vehicle delay is expressed as equation (48).

[0135]

[0136] ②Total carbon emissions from exhaust gas To better respond to the goals of carbon peaking and carbon neutrality, the MOVES model is incorporated into the exhaust emission objective function to calculate the overall carbon emissions of the region. The open database in the MOVES model stores the basic emission factors (micro-units Bin) of vehicles under different vehicle types, speeds, and power-to-weight ratios, as well as correction coefficients for other relevant influencing factors, exhibiting strong scalability. The specific expression of the exhaust emission objective function is shown in equation (49).

[0137] (49) in, The emission factor is obtained by calling the MOVES database. Let i be the number of vehicles of type i. The average mileage (in kilometers) of vehicle type i within the sub-area.

[0138] (2) Signal optimization objectives in saturation areas of traffic flow When the control area is classified as a moderately saturated region, it indicates that the traffic in that area is congested. At this point, the traffic flow is high, making congestion and queuing likely. Therefore, in this situation, vehicle delays, number of stops, and traffic capacity are all important optimization indicators that need to be considered.

[0139] ① Total vehicle delays Given that the modeling process of total vehicle delay has been described in detail in the previous section, the optimization objective expression is directly given in this invention as shown in equation (50).

[0140]

[0141] ② Number of parking sessions The average number of stops a vehicle makes while waiting in line at an intersection is called the "number of stops". To ensure smooth traffic flow at intersections, the number of secondary stops should be minimized, and all vehicles should be allowed to pass within one signal cycle. Referring to the Akcelik model, the expression is shown in equation (51).

[0142]

[0143] in, Indicates an intersection n In time period rThe average number of vehicles queuing within the area; For the first n Signal cycle duration (seconds) at each intersection; Time period r Internal path selection p Through the first n Traffic flow at each intersection; path p In the n Saturation at each intersection; This is the parking correction factor (default 0.9).

[0144] Based on this, assuming that there are a total of k At the intersection, vehicles during the time period r The set of driving routes within is P The expression for the number of parking times in the area is shown in equation (52).

[0145]

[0146] ③Traffic capacity In this invention, traffic capacity refers to the maximum number of vehicles that can pass through a sub-area per unit time when vehicles travel at a reasonable and safe speed under specific road facilities and traffic management measures. The expression is shown in equation (53).

[0147]

[0148] in, Intersection n saturation flow rate; The effective green light time at the intersection.

[0149] (3) Signal optimization objectives in areas with high traffic flow saturation When the controlled area is a high-saturation area, it indicates that the traffic in that area is congested. Therefore, alleviating congestion becomes the most urgent and critical need under high-saturation conditions. Thus, maximizing traffic capacity becomes the sole objective under these conditions, as expressed in expression (54).

[0150]

[0151] 3.3) IMOPSO Algorithm After designing corresponding personalized objectives for different regions, in order to generate reasonable and suitable signal timing schemes based on the congestion characteristics of different regions and achieve coordinated control of the regions, it is necessary to optimize and solve multiple objectives.

[0152] Multi-objective optimization differs from single-objective optimization in that it optimizes multiple objectives simultaneously, striving to improve the core objective's performance while ensuring that the performance of other objectives is not reduced.

[0153] Therefore, to improve the multi-objective optimization performance of traffic signal timing, and addressing the problem that MOPSO is prone to getting trapped in a local Pareto front due to premature convergence of the particle swarm during the multi-objective solution process, a dynamic weighting strategy is defined to improve the optimization performance of the algorithm, resulting in the Improved Multi-Objective Particle Swarm Optimization (IMOPSO) algorithm. The core of this strategy lies in dynamically adjusting key parameters such as inertia weights and learning factors based on the real-time state of the population and the search progress.

[0154] (1) First, the dynamic adjustment expression of the inertia weight is shown in equations (55) and (56).

[0155]

[0156]

[0157] in, It is the number of hypercube mesh divisions; N Population size; For the first i The proportion of particles within each grid cell to the total number of particles. In previous studies, and These are the upper limit of the weight in the global search phase and the lower limit of the weight in the local development phase, respectively, and are usually defined as 0.8 and 0.4. This formula still uses this definition.

[0158] The specific approach is as follows: quantify the distribution entropy of the Pareto front. To reflect the degree of dispersion of solutions; and then according to The diversity of particle populations is used to dynamically adjust the inertial weights. When diversity is high, Approaching At this time, reduce To enhance local development capabilities; when diversity is low, Approaching ,improve This enhances global exploration capabilities and avoids premature convergence.

[0159] (2) Secondly, the dynamic adjustment expressions of the learning factor are shown in equations (58), (59), and (60).

[0160]

[0161]

[0162]

[0163] The specific approach is as follows: the average normalized distance between the current population and the reference Pareto front is used as a convergence metric, defined as follows: In general particle swarm optimization algorithms, 、 The fixed value is 2. Therefore =2.5 Slightly higher than the baseline value, maintaining strong individual exploration ability in the early stages. = 0.5, to ensure that the group effect is weakened in the later stages of convergence.

[0164] In individual learning factors In terms of dynamic adjustment, an exponential decay mechanism is introduced, which is applied when the population is far from the Pareto front ( Larger), lower To reduce the influence of individual experience and promote group cooperation to accelerate convergence; when the population approaches the frontier ( (smaller), increase To enhance individual experience. Among them, This is the attenuation coefficient (default is 1). For the first j The maximum and minimum values ​​of the objective function in the current archive. m The number of objective functions.

[0165] In group learning factor For dynamic adjustment, a linear interpolation mechanism is introduced to dynamically adjust the group cooperation strength based on the current convergence state. When the convergence is relatively large, Approaching the maximum value enhances information sharing within the group and accelerates the global search; when convergence is small... Approaching the minimum value reduces the influence of the group and avoids over-reliance on historical best values.

[0166] Since the established multi-objective optimization model for traffic light timing is a problem of finding the optimal value with constraints, an initial population needs to be generated in conjunction with the constraints during the algorithm initialization phase. The optimized intersections only consider two phase types: three-phase and four-phase. For example... Figure 14 The diagram shows the intersection scenarios for three-phase and four-phase.

[0167] The particle encoding design employs real-number coding. Specifically, the particle length is: l That is, the particle dimension, which is equal to the sum of the phase numbers of each signal-controlled intersection; the position of the particle in each dimension: [ ], that is, the green light time for each phase (including three-phase and four-phase); the range of particle positions in each dimension: [ [This refers to the range of green light time variations in each phase; the maximum velocity of each particle in each dimension:] Dynamic adjustment range of inertia weight: [ ].

[0168] In summary, the specific steps of the improved multi-objective particle swarm optimization algorithm are as follows: Step 1: Initialize algorithm-related parameters, including population size N and upper and lower limits of inertia weight. w min , w max Individual learning factor c 1,base upper and lower limits of group learning factor c 2,max c 2,min External archive, length of each particle, and velocity of each particle in each dimension. v i and location x i .

[0169] Step 2: Multi-objective fitness calculation, which means calculating the fitness of each particle across all objective functions. Compare the dominance relationships between particles and add non-dominated solutions from the current population to the archive.

[0170] Step 3: Calculate the distribution entropy and convergence index. Dynamically adjust the inertia weight. w And learning factors c1 and c2.

[0171] Step 4: Determine the individual historical best position for each particle. Pbest If the current position dominates its historical best position, then update. Pbest If they do not control each other, they will be randomly selected to be retained or updated.

[0172] Step 5: Quantize the distribution density of solutions in the Archive using adaptive grid or crowding distance.

[0173] Step 6: Based on density difference information, a roulette wheel strategy is used to select the global guiding particle from the external archive. gbest Prefer solutions from sparse regions.

[0174] Step 7: Update the velocity of all particles in the swarm and location .

[0175] Step 8: Update the Archive by merging the non-dominated solutions of the new generation of particle swarm optimization with the historical archive. Filter out the new non-dominated solutions and update the Archive.

[0176] Step 9: Termination condition judgment: If the preset number of iterations is reached and the termination condition is met, output the result and the particle swarm algorithm ends; otherwise, return to step 2 and continue to execute the improved multi-objective particle swarm algorithm.

[0177] By applying the multi-objective particle swarm optimization algorithm to coordinate the optimization of regional traffic lights, the optimal solution set on the Pareto front can be obtained, thereby generating personalized timing schemes that take into account the multi-objective needs of different regions. This can efficiently alleviate local congestion while achieving dynamic and balanced allocation of road network resources.

[0178] Example 2: This example provides a traffic signal control system applying the method described in Example 1, comprising: Data acquisition module: used to collect real-time traffic flow data of the road network; Prediction Calculation Module: Runs the traffic prediction model and outputs the prediction results of future traffic flow; Region division module: Runs the dynamic region division algorithm to divide the road network into several independent control sub-regions; Timing optimization module: Runs an improved multi-objective particle swarm optimization algorithm (IMOPSO) for each sub-region to generate signal timing schemes and send them to the signal controllers.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A traffic flow prediction and signal timing integrated optimization method based on IVCPS, characterized in that, The method specifically includes the following steps: S1: Obtain historical traffic flow data and topology of the urban traffic network, and construct a traffic network topology map; S2: Construct a traffic prediction model that combines local and global spatiotemporal features, namely the ACTTGCNformer model, to predict future road network traffic flow data; the ACTTGCNformer model includes a local spatiotemporal feature extraction module, a global spatiotemporal feature extraction module, and a feature fusion and prediction module; The local spatiotemporal feature extraction module extracts detailed features between local spatial nodes from a local perspective using an improved graph convolutional network, ACT-GCN. The output of ACT-GCN is then used as the input of GRU to extract local spatial detailed features and short-term temporal features of traffic data. Here, GRU stands for Gated Recurrent Unit. The global spatiotemporal feature extraction module introduces a combined embedding layer in the Transformer encoding layer to extract long-term dependencies and global spatial correlation features of traffic data; the combined embedding layer includes feature embedding, periodic embedding, and global spatial embedding. The feature fusion and prediction module adopts a "late fusion" approach, which concatenates the features output by the local spatiotemporal feature extraction module with the features output by the global spatiotemporal feature extraction module, and outputs the final traffic flow prediction result through a multilayer perceptron. S3: Calculate regional saturation based on predicted traffic flow results, combine multiple correlation indicators between road network nodes, and use the improved Louvain algorithm to dynamically divide the traffic network into regions to obtain traffic sub-regions with different saturation states. S4: For the different traffic sub-zones, determine the differentiated signal timing optimization objectives based on their saturation status. Under the overall constraints, use the improved multi-objective particle swarm optimization algorithm to solve the optimal signal timing scheme and realize personalized collaborative control of regional traffic lights.

2. The traffic flow prediction and signal timing integrated optimization method according to claim 1, characterized in that, In step S2, the ACT-GCN includes a cascaded fully connected layer, ACT-GC, and an activation function; the ACT-GC represents adaptive channel topology graph convolution, and its construction process includes: firstly, the input features are linearly transformed through a fully connected layer, thus transforming the input features... X Transform into a higher-level feature representation Secondly, through correlation modeling functions... To model the correlation between vertices; based on the established correlation modeling function, a linear transformation is used. Increasing the channel dimension yields channel-specific feature factors. In shared topology A d Introducing correlation Learn a channel-specific refined topology for each channel. T d Each channel's topology represents the relationships between nodes across different dimensions of traffic data. Finally, feature aggregation is performed on the channel topologies corresponding to different traffic dimensions to obtain the final output. Z This enables the extraction of detailed features of the local spatial topology of the road network.

3. The traffic flow prediction and signal timing integrated optimization method according to claim 1, characterized in that, In step S2, the feature embedding is obtained by using a fully connected layer to obtain the feature embedding of the sequence; the periodic embedding includes weekly embedding and daily embedding, and by connecting and broadcasting them, the time embedding of traffic data is obtained. The global spatial embedding adopts basic spatial embedding. By connecting feature embedding and periodic embedding, the final combined embedding is obtained, which is a high-dimensional dense representation of traffic data.

4. The traffic flow prediction and signal timing integrated optimization method according to claim 1, characterized in that, In step S2, the Transformer encoding layer includes a time-dimensional Transformer layer and a spatial-dimensional Transformer layer, which are connected in series and use a multi-head attention mechanism to mine the global correlation between time and space respectively.

5. The traffic flow prediction and signal timing integrated optimization method according to claim 1, characterized in that, Step S3, which involves using the improved Louvain algorithm to dynamically divide the traffic network into regions, specifically includes: (1) Combining the weight division of correlation degree: The intersection is abstracted as a network node, and an index system including signal cycle correlation degree, traffic flow correlation degree, density correlation degree and distance correlation degree is constructed. By integrating each correlation degree and assigning different weight parameters, the edge weight based on correlation degree is obtained. (2) Local adjustment based on modularity optimization: After obtaining the edge weights based on the degree of association, the modularity is optimized, that is, by adjusting the community structure in the network, the modularity is maximized. (3) Global adjustment based on community aggregation: Calculate the current average saturation of the region and the predicted average saturation at the next adjustment time. If the two do not belong to the same saturation range, dynamically adjust the weight coefficients of each correlation index and re-divide the region.

6. The traffic flow prediction and signal timing integration optimization method according to claim 5, characterized in that, In step S3, the methods for calculating each correlation degree are as follows: (1) Signal periodicity correlation This refers to an indicator that quantifies the degree of signal coordination between two adjacent intersections by measuring the difference in signal cycles. Its expression is: In the formula, , They represent the intersections. i , j The duration of the cycle; (2) Correlation of traffic flow at intersections This refers to a quantitative expression of the traffic flow interaction between adjacent intersections; the higher the correlation value, the higher the correlation. Its expression is: in, This represents the normalized travel time of a vehicle from the upstream intersection to the downstream intersection; This indicates the number of traffic branches flowing from the upstream intersection to the downstream intersection; Indicates the first Traffic flow of each branch; This represents the maximum single-branch traffic flow among all branches; Since traffic flow is bidirectional, the expression for traffic flow correlation considering bidirectional traffic flow is: in, Indicators representing the correlation between two-way traffic flow; , They represent the intersections. i arrive j , j arrive i Traffic flow correlation in terms of direction; (3) Expression for density correlation considering bidirectional density: in, Indicates an intersection i and intersection j Density correlation between them Indicates from the intersection i To the intersection j Total traffic volume; These represent the number of lanes and the length of the road segment in the corresponding direction, respectively. (4) Node distance correlation degree represents the distance between adjacent nodes.

7. The traffic flow prediction and signal timing integrated optimization method according to claim 5, characterized in that, In step S3, when optimizing the modularity, a regularization penalty term is used to suppress the formation of ultra-large regions. The expression is: in, For the optimized modularity, for k area, for k The total number of nodes in the region N This represents the total number of nodes across all regions. To control the intensity of punishment; in, For modularity, Represents a node and nodes Edge weights between them; Representing nodes respectively and nodes The degree; This represents the sum of the weights of all edges in the network; Represents an indicator function, when the node and nodes If they belong to the same community, the value is 1; otherwise, it is 0.

8. The traffic flow prediction and signal timing integrated optimization method according to claim 1, characterized in that, In step S4, the overall constraint condition is: in, For the common signal period; The total lost time includes the sum of the yellow light time and the all-red light time for all phases, which is the time during which vehicles cannot pass. It is the sum of the flow rate ratios of each phase, and the maximum value of the flow rate to saturation flow rate ratio among all phases of the critical intersection; n This represents the number of intersections within a sub-region of the road network. Minimum green light time, It is the minimum green light time for pedestrians; Traffic flow saturation; This represents the phase difference between two adjacent intersections.

9. The traffic flow prediction and signal timing integrated optimization method according to claim 1, characterized in that, In step S4, determining the differentiated signal timing optimization objective based on its saturation state includes: (1) In areas with low traffic flow saturation, minimizing total vehicle delay and minimizing total exhaust carbon emissions are taken as optimization objectives; (2) In the saturation area of ​​traffic flow, the optimization objectives are to minimize vehicle delay, minimize the number of stops, and maximize traffic capacity; (3) In areas with high traffic flow saturation, maximizing traffic capacity will be the optimization objective.

10. The integrated optimization method according to claim 1, characterized in that, In step S4, the improved multi-objective particle swarm optimization algorithm includes: (1) Dynamic adjustment of inertia weight: The distribution entropy of the Pareto front is quantified to reflect the degree of solution dispersion; the current particle population diversity is judged based on the historical maximum and minimum entropy values ​​up to the current iteration, and the inertia weight is dynamically adjusted accordingly; when the population diversity is high, the weight is reduced to enhance local development capability; when the population diversity is low, the weight is increased to enhance global exploration capability. (2) Dynamic adjustment of learning factors: The average normalized distance between the current population and the reference Pareto front is used as the convergence index; an exponential decay mechanism is introduced to dynamically adjust individual learning factors, and a linear interpolation mechanism is introduced to dynamically adjust group learning factors, so as to balance the influence of individual experience and group cooperation.

Images