Intelligent Traffic Prediction and Control Method and System Based on Dynamic Functional Cell Coupling

Abstract

This invention belongs to the field of intelligent transportation technology, and specifically relates to an intelligent traffic prediction and control method and system based on dynamic functional cell coupling. It aims to solve the problems of low prediction accuracy and poor adaptability of existing traffic planning methods in dynamic networks. This invention extracts high-order structural features of the traffic network based on directed motifs and graph neural networks (GNNs) to dynamically divide urban functional cells; it ensures the continuous evolution of the urban functional cell structure through a time smoothing mechanism; it predicts the mutual feedback strength between urban functional cells using urban functional cell embedding and time-series models (such as LSTM); and it adaptively adjusts the prediction results based on the dynamic states of urban functional cells (birth, expansion, shrinkage, death). Based on dynamically adapting to the time-varying traffic environment, this invention achieves regional traffic flow prediction. When applied to traffic planning, it can optimize regional traffic flow management, congestion early warning, and infrastructure planning, promoting intelligent transportation and green city development.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology and relates to a traffic prediction and control system that integrates a dynamic feedback mechanism of urban spatial organization with a time-series graph neural network. It is particularly suitable for traffic management in large cities with significant human activity rhythm characteristics. Specifically, it relates to an intelligent traffic prediction and control method and system based on dynamic functional cell coupling. Background Technology

[0002] With the continuous improvement of urbanization, the operational efficiency of transportation systems has become one of the important indicators for measuring urban comprehensive governance capabilities and residents' quality of life. Traffic congestion has long plagued urban development, especially in large and medium-sized cities, exhibiting typical phenomena such as periodic peak congestion, structural flow imbalance, and abnormal congestion induced by sudden events. Traditional traffic control methods, such as fixed signal timing strategies and rule-based guidance models, generally rely on static data or manually set control logic, which is difficult to meet the demands of modern urban transportation networks characterized by high time-varying nature, high complexity, and multi-scale linkage. Therefore, improving the intelligence, dynamism, and regional adaptability of urban traffic prediction and control systems has become one of the core objectives of Intelligent Transportation Systems (ITS) research.

[0003] In recent years, with the rapid development of IoT, big data, and AI technologies, deep learning-based traffic flow prediction methods have made significant progress in terms of accuracy and generalization ability. In particular, models such as Graph Neural Networks (GNN) and Long Short-Term Memory (LSTM) networks, due to their excellent spatiotemporal modeling capabilities, have been widely applied to time-series prediction tasks of complex traffic flows. However, despite existing technologies, such as the Chinese patent CN118675324A which discloses a traffic flow prediction system and its application method based on deep learning and dynamic network analysis, proposing a traffic flow prediction system that combines IoT technology, graph convolution mechanisms, and time-series prediction networks, this system constructs a spatiotemporal feature matrix reflecting the evolution of traffic networks by sensing multi-source traffic data, and trains it using graph attention mechanisms and time-series models to achieve prediction of traffic flow evolution trends. This system integrates multiple algorithmic frameworks, including graph attention mechanisms and convolutional neural networks (CNNs), to enhance its ability to characterize traffic flow patterns and, to some extent, compensate for the shortcomings of traditional static models in dealing with dynamic traffic scenarios. However, it still has significant limitations in the following aspects, making it difficult to fully adapt to the evolutionary trends of modern urban transportation systems:

[0004] First: Insufficient consideration of spatial organization and form.

[0005] Current deep learning technologies for traffic prediction and management rely heavily on surface-level correlations of historical traffic data and direct connections between nodes in graph structures, neglecting the structural element of urban spatial organization—a factor with long-term impacts on traffic evolution. In real cities, the spatial distribution of traffic flow is closely related to the layout of urban functional zones. Differences in population density, land use, and economic activity among different functional zones result in stable or cyclical traffic patterns. Urban spatial organization, as a core planning dimension of intelligent transportation systems, essentially reflects the spatial configuration and collaborative relationships of urban functional modules. The rigid time-segmentation strategies used in traditional traffic control systems are ill-suited to the multi-scale complexity of modern urban traffic networks and the time-varying coupling effects emerging between functional units. Existing methods fail to incorporate the spatial coupling relationships of urban functional modules into their modeling framework, resulting in predictions lacking explanatory power regarding spatial organization at the macro level.

[0006] Second: Lack of dynamic feedback mechanism.

[0007] Existing deep learning techniques, when modeling interactions between traffic nodes, fail to incorporate the ability to depict dynamic feedback mechanisms between functional regions. In real urban traffic networks, human activities exhibit significant temporal fluctuations, such as commuting tides and leisure travel waves. Traffic interactions between different regions display nonlinear and time-varying coupling characteristics. For example, commuting tides between residential and work areas, and weekend traffic surges between commercial and tourist areas, all demonstrate clear rhythmicity, predictability, and mutual driving forces. Existing methods fail to identify these macroscopic feedback mechanisms stemming from human activity patterns, instead simplifying traffic relationships between regions into isolated edge-weighted evolutions. This limits the model's ability to reflect the effects of traffic linkages between regions, thus affecting its sensitivity to abnormal traffic fluctuations and the accuracy of predictions.

[0008] Furthermore, urban transportation networks often undergo structural changes due to unforeseen events (such as traffic accidents, extreme weather, and large-scale events) or functional zone adjustments (such as the construction of new areas or road reconstruction), leading to the failure of existing traffic patterns. Existing deep learning-based traffic prediction methods are primarily based on fixed model parameters built during the training period, lacking online update and structural adjustment mechanisms. This makes it difficult to achieve real-time response to sudden changes in the traffic network and adaptive adjustment of prediction strategies, thus affecting the timeliness and effectiveness of predictions.

[0009] Therefore, how to effectively integrate the temporal characteristics of human activities with traffic dynamics to achieve accurate prediction and adaptive control is a technical challenge that urgently needs to be overcome. Summary of the Invention

[0010] In view of this, the present invention aims to solve the technical problem that traditional traffic control methods are unable to meet the high time-varying, high complexity, and multi-scale linkage requirements of modern urban traffic networks. It provides an intelligent traffic prediction and control method and system based on dynamic functional cell coupling. This method implicitly divides the city into several urban functional cells based on vehicle trajectory data. Through steps such as constructing a dynamic urban network, extracting motif features, dividing urban functional cells, and predicting the changes in the mutual feedback intensity between cells, it calculates the mutual feedback intensity between functional cells to predict traffic changes between functional cells. This achieves accurate prediction and adaptive control of urban traffic, thereby improving the operating efficiency of the traffic system and reducing congestion.

[0011] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0012] The first objective of this application is to disclose an intelligent traffic prediction and control system and method based on dynamic functional cell coupling, comprising the following steps:

[0013] S1: Obtain trajectory data of taxis, shared cars and other vehicles, and construct an initial static urban network with taxi areas as the main geographical unit;

[0014] S2: Based on the static network, construct a dynamic city network given a time interval Δ;

[0015] S3: By extracting motif features and using a GNN model to aggregate dynamic embeddings of geographic locations, dynamic urban cell structures are identified;

[0016] S4: Based on the dynamic urban cell structure, the interaction features between dynamic urban cells are extracted on the basis of the LSTM model to predict the change in the interaction intensity of urban cells.

[0017] S5: Use forecast results to guide urban transportation planning.

[0018] S1 is specifically as follows:

[0019] Initial static city network G0 construction: Trajectory data from taxis, shared vehicles, and other similar vehicles are collected. Taxi areas are used as the primary geographical unit, forming the node set V0 in the city network. Directed edges E0 are supplemented with the trajectories of all vehicles, with a focus on constructing the trajectories between regions, thus building the static city network G0.<V0,E0> .

[0020] S2 is specifically as follows:

[0021] Constructing a dynamic city network G: Given a time interval Δ, a 24-hour day can be divided into a discrete set of time intervals {Δ, 2Δ, ..., tΔ}. From this set, the dynamic city network can be represented as a set G = {G0, G1, ..., G...} consisting of city snapshots over multiple time intervals. t The static city network G in the t-th time interval. t = <V t E t >, where V t E t Let represent the set of city network nodes and the set of edges in the t-th time interval, respectively. Each node Representing various geographical locations within the city, each edge This represents the interaction relationships between geographical locations (if cells can reach each other, there is an edge; if they cannot reach each other, there is no edge). Each edge has a weight. This represents the nodes in the city network during the t-th time interval. and The intensity of interaction between them. Meanwhile, G t It is composed of multiple city cells Because urban networks are dynamic, the urban cells also change dynamically accordingly. Each urban cell... All contain G t A set of nodes and edges in a cell, with relatively strong interactions inside the cell and relatively weak interactions outside the cell (i.e., between cells).

[0022] S3 is specifically as follows:

[0023] The following steps will be used to achieve the dynamic urban cell structure at each moment. Identification:

[0024] Step 1: Dynamic motif feature extraction.

[0025] (1) For each node The number of motifs it participates in is calculated using the following formula:

[0026]

[0027] Where A is G t The directed weighted adjacency matrix,

[0028] (2) For a static city network G0, obtain the set of all motif instances in the network. For each node Obtain the motif instance M that it participates in, and compute... Weighted motif strength:

[0029]

[0030] in, It is a node The product of the weights of all edges in M, and the time-smoothed motif strength. Where α is the first smoothing factor.

[0031] (3) Construct the weighted motif matrix W M , Represents a node Weighted connection strength in directed motifs is used to capture The association strength in directed weighted motifs enhances the identification of urban cell structure, as shown in the following formula:

[0032]

[0033] To avoid differences in weight scales, normalize W M The formula is as follows:

[0034]

[0035] in Normalized W M It facilitates GNN aggregation and balances the contributions of different nodes.

[0036] Time-smoothed motif matrix:

[0037]

[0038] γ is the second smoothing factor.

[0039] To reduce the time complexity of directed weighted motif feature extraction, a random walk-based motif sampling algorithm is used to approximately estimate N. m (v) and

[0040] Step 2: GNN Feature Aggregation:

[0041] The node features are aggregated using GNN, combined with the directed weighted adjacency matrix A and the weighted motif matrix W. M Generate node embeddings.

[0042] (1) Initialize features: node features Including out-degree, in-degree, and smooth motif intensity.

[0043] (2) GNN aggregation:

[0044]

[0045] in I is the identity matrix, which retains node information by adding self-loops. H (l) It is the node embedding of the l-th layer. σ is the learnable weight, and σ is the activation function. yes The out-degree matrix is ​​a combination matrix used to fuse edge weight information and motif information, where λ∈[0,1] is a hyperparameter.

[0046] Based on this, the embedded smooth Z is obtained. t The formula is as follows:

[0047]

[0048] Where β∈[0,1] is the third smoothing factor.

[0049] Step 3: Dynamic city cell identification and adjustment.

[0050] A weighted K-means method is used to dynamically identify urban functional cells by utilizing the geometric distance and node weights in the embedding space. The centroids of urban functional cells at time t-1 are referenced when initializing the cluster centers. The optimization objective is as follows:

[0051]

[0052] in, It is a node The weights are combined with motif intensity and degree. μ t,i It is the functional cell of the city The weighted centroid is calculated using the following formula:

[0053]

[0054] When t = 0, randomly initialize k0 centroids. If t >= 1, use... The center of mass μ t-1,i As initial values, add or delete centroids to match k. t .

[0055] Identify [the problem] by calculating the weighted similarity between dynamic urban cells. and The cell state.

[0056]

[0057] Dynamic urban cell birth: if With all of but For newly formed urban cells;

[0058] Dynamic urban cell death: if With all of but die;

[0059] Dynamic city cell body persistence: if and satisfy and but yes The continuation;

[0060] Dynamic urban cell expansion: if and but merge

[0061] Dynamic urban cell shrinkage: If and The city's cells shrank.

[0062] S4 is specifically as follows:

[0063] (1) Cell-level feature extraction: embedding Z at the obtained nodes t Based on this, each cell body is generated. The embedding formula is as follows:

[0064]

[0065] (2) Construction of cell body feedback characteristics

[0066] First, for each cell body Constructing eigenvectors The formula is as follows:

[0067]

[0068] in, It is the sum of the motif intensities of all nodes within the cell body. These are the sum of the out-degree and in-degree of nodes within the cell body, respectively.

[0069] Subsequently, the mutual feedback features between cell body pairs are constructed. The formula is as follows:

[0070]

[0071] in This indicates that at time t, the cell body... The formula for the mutual feedback strength between them is as follows:

[0072]

[0073] in, It is the edge Weight at time t.

[0074] (3) Temporal modeling

[0075] LSTM is used to model the time series of urban cell-level feedback intensity, with historical feedback features as input. The LSTM model is constructed as follows:

[0076] h t ,c t =LSTM(X) i,j ,h t-1 ,c t-1 )

[0077]

[0078] Where h t ,c t These are the hidden states and cell states of the LSTM; W out It is the weight of the output layer, b out This is the output layer bias.

[0079] The loss function is:

[0080]

[0081] N represents the number of cell pairs in the city.

[0082] Identify [the problem] by calculating the weighted similarity between dynamic urban cells. and The cell state is used to dynamically adjust the intensity of urban cell feedback.

[0083]

[0084] Dynamic urban cell birth: if For newly formed cells, the predicted feedback strength of the new cells is initialized based on the average feedback strength at the current time step:

[0085]

[0086] Dynamic urban cell death: if Died at time t+h

[0087] Dynamic urban cell expansion and contraction: If and Matching, adjusting according to size changes:

[0088]

[0089] in This represents the basic mutual feedback strength of the urban functional cell pairs at that moment.

[0090] This represents the predicted value of the interaction strength.

[0091] Compared with existing technologies, the intelligent traffic prediction and control method and system based on dynamic functional cell coupling described in this invention have the following advantages:

[0092] (1) This invention captures the directed weighted high-order structure of the traffic network (such as the cyclic flow pattern reflected by the directed triangle) through a motif-enhanced GNN, and extracts regional features based on dynamic urban functional cells. Combined with the LSTM time series model, it predicts the mutual feedback intensity between urban functional cells (such as inter-regional vehicle flow). This can more accurately predict the flow between urban functional cells, support more scientific traffic light optimization, route planning and congestion warning, reduce traffic delays and improve urban traffic efficiency.

[0093] (2) This method introduces a time smoothing mechanism (including smoothing of motif intensity, node embedding, and interaction intensity) to ensure the continuous evolution of urban functional cell structure and interaction intensity; at the same time, it dynamically updates the urban functional cell identification and prediction results by incremental motif counting and reusing GNN weights. Dynamic adaptability supports real-time traffic management, such as quickly responding to traffic changes caused by new area development or emergencies (such as large-scale events) and optimizing traffic resource allocation.

[0094] (3) This method uses weighted similarity matching and state determination algorithms to clearly identify the dynamic state of traffic communities (such as the “birth” of new commercial areas and the “expansion” of main road communities during peak hours), and adaptively adjusts the interaction intensity prediction according to the state. This can support traffic planners to adjust their strategies in a timely manner, such as planning new bus routes for new urban functional cells or increasing road capacity for expanding urban functional cells, thus promoting the sustainable development of urban transportation networks. Attached Figure Description

[0095] Figure 1 This is a flowchart of the intelligent traffic prediction and control method based on dynamic functional cell coupling as described in an embodiment of the present invention;

[0096] Figure 2 This is a schematic diagram of the urban functional cell feedback time-series prediction system in an embodiment of the present invention. Detailed Implementation

[0097] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0098] It should be noted that, in this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0099] Existing traffic control systems largely rely on rigid time-segmentation strategies, which struggle to cope with the multi-scale complexity of modern urban traffic networks and the time-varying coupling effects between functional units. With the development of ubiquitous sensing technology, the acquisition of big data on human activities provides a new methodological foundation for the quantitative analysis of the dynamic characteristics of urban functional units, making this analysis possible. However, existing traffic control systems fail to incorporate the nonlinear time-varying correlations between functional units into the control decision framework, resulting in a lack of dynamic feedback. Therefore, intelligent traffic prediction systems based on dynamic functional units and time-series graphical neural networks have emerged, offering new ideas and solutions for addressing urban traffic problems.

[0100] like Figures 1-2 As shown, this application discloses an intelligent traffic prediction and control method based on dynamic functional cell coupling, including:

[0101] S1: Obtain the trajectory data of urban vehicles, and construct an initial static urban network with taxi areas as the main geographical unit. The nodes of the static urban network represent urban geographical locations, and the edges represent traffic interaction relationships between nodes, and have direction and weight.

[0102] S2: Construct a dynamic urban network from a static urban network. The dynamic urban network consists of multiple snapshots of the urban network at different time segments, which are used to characterize the time-varying structure of the urban transportation network.

[0103] S3: In each network snapshot of a time segment, motif features are extracted, a motif feature matrix is ​​constructed, and a GNN model is used to aggregate the dynamic embedding of geographic locations to divide urban functional cells. Combined with the cell state of the previous time segment, the birth, death, expansion, shrinkage or continuous state of the cell is identified.

[0104] S4: Based on the dynamic urban functional cell structure, construct the mutual feedback feature vector between urban functional cells to form the mutual feedback time series between cell pairs. Use a long short time memory network (LSTM) to model the mutual feedback time series between cell pairs and predict the change in mutual feedback intensity between urban functional cells in subsequent time segments.

[0105] S5: Provide support for urban traffic planning based on forecast results, including any one or more of the following: traffic light timing, route navigation, traffic warning, and regional dispatching recommendations.

[0106] The intelligent traffic prediction and control method described in this application constructs an initial static traffic network by collecting driving trajectory data covering vehicles such as taxis and shared vehicles, and using taxi operating areas as the primary reference to divide the city into several taxi operating areas. In this model, each node represents a specific location in the city, and the edges between nodes represent the traffic flow relationships between these locations. These edges not only have directionality but also carry weights to represent the magnitude of traffic flow. Subsequently, the static network is divided into a series of network snapshots that reflect time-varying characteristics at time intervals. Each snapshot reflects the urban traffic conditions at a specific point in time, thereby capturing the time-varying characteristics of the traffic network. In each time-segment network snapshot, specific patterns (m) in the network are analyzed. This model utilizes OTIF features and graph neural network (GNN) technology to incorporate the dynamic characteristics of geographical location into the model, thereby dividing the city into different functional cells. These cells represent areas with similar functions or traffic characteristics within the city. Simultaneously, by combining the cell state from the previous moment, the model updates the state of urban functional cells to identify cell additions, disappearances, and structural changes (expansion, shrinkage, or persistence). Based on these dynamically changing functional cell structures, the model statistically analyzes the mutual feedback strength between each pair of functional cells and generates a mutual feedback time series. Finally, this series is input into an LSTM model for time series prediction to obtain changes in the flow intensity between functional cells in future time periods. This provides quantitative support for traffic signal timing, travel route recommendation, congestion warning, and regional resource scheduling. The core of the intelligent traffic prediction and control method based on dynamic functional cell coupling described in this application is as follows: First, by organically combining motif analysis and graph neural networks, high-level structural features of the network are extracted and spatiotemporal embeddings of nodes are generated to accurately reflect the strength changes of traffic connections between regions and outside the region during dynamic partitioning. Then, a weighted clustering algorithm is used to abstract homogeneous regions into functional cells. Next, a mutual feedback time series is constructed based on historical flow data between functional cells, and LSTM is used to perform time series modeling and prediction on the series, thereby achieving accurate prediction of the future traffic flow intensity between functional cells. Finally, the prediction results are transformed into executable traffic management strategies such as signal timing optimization, route guidance, and congestion warning.

[0107] This application addresses the neglect of regional coupling relationships in traditional traffic forecasting by introducing high-order network features based on motifs and spatiotemporal clustering of functional cells. Furthermore, it enhances the perception of traffic network structure evolution and temporal changes through the organic integration of dynamic graph neural networks and LSTM. This enables more accurate capture of nonlinear feedback effects between regions in complex scenarios such as morning and evening rush hours and holiday traffic peaks, and generates real-time traffic control strategies accordingly. Consequently, it significantly improves traffic light timing efficiency, optimizes route navigation, enhances congestion warning accuracy, and strengthens the system's adaptive response to sudden events and network structure changes.

[0108] As a preferred example of this application, in step S1, the construction of the initial static city network includes the following specific steps: collecting trajectory data of taxis, shared cars, and other vehicles, and using taxi areas as the primary geographical unit as the node set V0 in the city network; supplementing the city network with the directed edge set E0 of all vehicle trajectories; and focusing on constructing the trajectories between regions, thereby forming a directed graph containing the node set and edge set, i.e., the static city network G0 =<V0,E0> In the example of this application, the initial static urban network is constructed using a combination of multi-source data fusion and regional abstraction. First, trajectory data of various urban vehicles, including taxis and shared cars, is collected through vehicle positioning systems, urban traffic management platforms, or third-party travel platforms. This data includes elements such as time, location, speed, and travel path for each vehicle, possessing high timeliness and geographical accuracy. Then, the city is divided according to taxi operating areas, and these areas are abstracted as nodes in the network. Each node corresponds to a specific geographical location and represents a type of traffic functional area, forming the node set V0 in the urban network. Next, the trajectory paths of all vehicles are analyzed, and the directed connections between nodes are determined by the starting and ending areas traversed by the trajectories. The number of flows between each pair of areas within a specific time period is counted, thereby constructing a set of directed edges E0. Each edge is assigned a weight representing the flow intensity, and finally, these are integrated to form a directed graph G0.<V0,E0> As a static urban network model, it provides structured support for subsequent dynamic network generation, traffic cell segmentation, and temporal modeling.

[0109] This application uses trajectory data of multiple types of urban vehicles as input information sources and combines regional division to construct an urban traffic network, which effectively improves the completeness and diversity of traffic network modeling, avoids coverage blind spots or local deviations that may be caused by a single data source, and at the same time, the directed edge set constructed by using trajectory flow relationship can more realistically reflect the directionality and intensity characteristics of traffic flow, thereby improving the accuracy of static network in restoring the real urban traffic conditions.

[0110] As a preferred example of this application, in step S2, the trajectory containing location and time information is taken as input. Given a fixed time interval Δ, based on a 24-hour day, it is divided into a discrete time interval set {Δ, 2Δ, ..., tΔ}. Each time interval represents an independent snapshot of urban traffic. By summarizing and processing the vehicle trajectory data collected in each time interval, the traffic interaction between different geographical locations in the city is extracted. These geographical locations are abstracted as network nodes, which represent various geographical locations in the city, such as traffic intersections, commercial areas, and residential areas. The flow paths between regions are modeled as directed edges, which represent the interaction relationships between these geographical locations, such as vehicle flow and pedestrian movement, thereby forming a dynamic urban network G = {G0, G1, ..., G...}. t The static city network G in the t-th time interval. t = <V t E t >, where V t E t Let these represent the set of city network nodes and the set of edges in the t-th time interval, respectively. Each node... Representing various geographical locations within the city, each edge This represents the interaction between geographical locations (there is an edge between cells if they are reachable, and no edge if they are not reachable; i and j are two reachable cell nodes).

[0111] This application discretizes the time series of a city's day into multiple time periods and constructs static urban network snapshots for each period, thereby achieving a fine-grained characterization of urban traffic conditions. This enables a more accurate capture of the time-varying characteristics of urban traffic conditions and provides more reliable data support for subsequent traffic forecasting and planning.

[0112] As a preferred example of this application, the static urban network consists of multiple urban functional cells. Each urban functional cell All contain G t A set of nodes and edges in a cell, with relatively strong interactions within the cell body and relatively weak interactions outside the cell body (i.e., between cell bodies), G t Each edge in Set a weight This represents the nodes in the city network during the t-th time interval. and The interaction strength between them is normalized to limit the edge weights to the range [0,1], and the urban functional cells are dynamically evolved over time intervals. In constructing a dynamic urban network, to accurately model the traffic connection strength between urban functional cells, the system sets weights for all edges in the urban network snapshot generated at each time interval. For example, in the t-th time interval, weights are assigned to each connected node... and edge The frequency of traffic interactions between the two sides is calculated based on vehicle trajectory data, and the value is assigned to the edge as the initial interaction strength value. Because the nodes, edges, and weights of dynamic urban networks change over time, the distribution of interaction intensity between different time periods and regions can vary by orders of magnitude. Traditional methods often suffer from low processing efficiency or unstable results due to inconsistent data formats or differences in weight ranges. To address this issue, this method normalizes all edge weights in each urban network snapshot, thus normalizing all edge weights. Mapping to the closed interval [0,1] makes the edge weights between different snapshots have a uniform scale and are comparable, eliminating the error or instability caused by uneven distribution of the original data, thereby improving the adaptability and convergence of subsequent model processing.

[0113] As a preferred example of this application, the urban functional cell division in step S3 includes the following steps;

[0114] S31: For a directed weighted dynamic city network G = {G0, G1, ..., G...} t Based on the specified directed motif type, extract the motif features for each time step t and generate a weighted motif intensity. and motif matrix W M And integrate historical information for time smoothing;

[0115] S32: Aggregate node features using a graph neural network, combined with a directed weighted adjacency matrix A. t and smooth motif matrix Generate node embedding Z t And time smoothing is used to make the embedding continuous;

[0116] S33: Z-based node embedding t The weighted K-means algorithm was used to divide the dynamic urban functional cells. Clustering initialization is performed using the cell centroids from the previous time step.

[0117] In the example of this application, during the process of dividing urban functional cells, representative directed motif structures in the network are first extracted, and the weighted motif strength of the nodes is calculated by multiplying the edge weights of the nodes in the motif. To construct the motif matrix W M To ensure the robustness of the features, the motif strength is... Temporal smoothing is performed, incorporating historical information to reduce interference from dynamic changes. Then, the powerful feature aggregation capability of graph neural networks is utilized to optimize motif intensity. And the directed weighted adjacency matrix A t By combining and inputting features into a two-layer graph neural network, a richer and more accurate node embedding Z is generated. t These embeddings not only contain local connectivity information of nodes but also incorporate global motif structural information. Finally, a weighted K-means algorithm is applied within this embedding space, and cluster centers are initialized using the cell centroids from the previous time step, thereby dividing the dynamic urban functional cells to achieve accurate capture and stable tracking of the urban network structure. This application extracts motif features and smooths them over time, uses a graph neural network to aggregate node features to generate embeddings, and then uses a weighted K-means algorithm based on these embeddings to divide the dynamic urban functional cells, achieving in-depth analysis and dynamic tracking of the urban network structure.

[0118] Traditional urban network analysis often relies on edge connections, neglecting higher-order structures (such as loops or feedback patterns), leading to insufficient feature representation, especially in dynamic urban networks. Furthermore, abrupt changes in dynamic urban networks (such as sudden changes in nodes or edges) frequently cause feature instability. This application addresses this issue by extracting motif features and performing temporal smoothing for each static urban network—that is, urban snapshots within each time period—to capture higher-order topology and enhance continuity. Motif features enrich the structural representation of the network, while temporal smoothing reduces the interference of dynamic changes, improving feature robustness and the accuracy of cell segmentation, providing a reliable foundation for interactive prediction.

[0119] As a preferred example of this application, in order to fully explore the high-order interaction features in the transportation network and improve the expressive ability of the node structural roles in the static urban network of each time period, this method selects directed triangle motifs with cyclic representation as structural units. In the motif feature extraction stage, the system first identifies all directed triangle motif instances involved by each node based on the graph structure, and multiplies the edge weights involved in each motif to obtain a weighted motif intensity value reflecting the local interaction intensity. Then, these intensities are summarized by node to construct a weighted motif matrix representing the high-order structural features of the entire network. Considering that nodes and edges may undergo abrupt changes in different time periods in dynamic networks, leading to feature instability, this method performs time smoothing processing on the motif intensity values. That is, by setting a smoothing coefficient, the features of the previous time period are introduced into the current features to improve continuity. Furthermore, the motif matrix is ​​normalized column-wise to unify the feature dimensions and scale. The resulting smooth and normalized motif matrix is ​​used as input for graph neural network feature aggregation, thereby providing a more stable and expressive structural input for subsequent node embedding and functional cell division.

[0120] In step S31, the motif feature extraction is specifically as follows:

[0121] Motif selection: A directed triangle motif is selected to reflect the cyclical interaction patterns in the urban network.

[0122] Motif strength calculation: Calculation Weighted motif strength:

[0123]

[0124] in, It contains nodes The triangular set, with weights multiplied to reflect the interaction strength.

[0125] Temporal smoothing: To avoid abrupt changes, the motif strength is smoothed.

[0126]

[0127] Where α is the first smoothing factor, such as α = 0.2, initially

[0128] Constructing a weighted motif matrix:

[0129]

[0130] Normalization:

[0131]

[0132] smooth:

[0133]

[0134] γ is the second smoothing factor.

[0135] To improve algorithm efficiency, this step uses random walk sampling to calculate the motif, resulting in... and For use by GNN in the next step.

[0136] This application breaks through the limitations of traditional urban network analysis, which relies solely on adjacency relationships, by extracting motif structural features reflecting cyclic interactions and performing temporal smoothing and normalization on them. This greatly enhances the ability of features to express the complex structural roles of nodes, and maintains the continuity and robustness of feature expression even in the face of abrupt changes in nodes or edges in dynamic networks. This effectively improves the stability of subsequent GNN aggregation node embedding, as well as the accuracy and temporal consistency of urban functional cell division. It provides a high-quality structural input foundation for capturing regional linkage patterns in traffic prediction, and helps to improve the overall modeling and prediction capabilities of intelligent transportation systems for complex time-varying structures.

[0137] In existing traffic prediction technologies, traditional embedding methods neglect edge weights and directionality, and are sensitive to changes in dynamic networks, leading to unstable embeddings or insufficient expressive power. This application utilizes a Graph Neural Network (GNN) to capture complex structures through a multi-layer aggregation strategy when aggregating node features in a dynamic urban network. It combines motif features and temporal smoothing to adapt to the dynamic urban network. In step S32, by designing a motif-strong GNN and a temporal smoothing mechanism, node embedding integrates motif strength, degree information, and network topology, providing robust feature representations. Temporal smoothing ensures the continuity of the dynamic network and improves the stability of cell partitioning and prediction.

[0138] In this application, when performing GNN feature aggregation, the node features are aggregated using GNN, combined with the directed weighted adjacency matrix A and the weighted motif matrix W. M Generate node embeddings.

[0139] Initial features: calculated for each node Weighted motif strength: for each node Constructing eigenvectors Including out-degree, in-degree, and smooth motif intensity.

[0140] Using two-layer GNN aggregation:

[0141]

[0142] in I is the identity matrix, with self-loops added to preserve the information of the nodes themselves; H (l) It is the node embedding of the l-th layer. These are learnable weights, and σ is the activation function; yes The out-degree matrix is ​​used as a combination matrix to fuse edge weight information and motif information; λ∈[0,1] is a hyperparameter, such as λ=0.3 balancing the directed weighted adjacency matrix A and the smooth motif matrix.

[0143] Embedding a third smoothing factor β: β = 0.2, based on which the node embedding Z is obtained. t The formula is as follows:

[0144]

[0145] In this application, when aggregating node features in a dynamic urban network, a two-layer aggregation strategy is adopted in the graph neural network (GNN) to more comprehensively integrate structural information and interaction weights. In the first aggregation layer, the identity matrix is ​​added to the directed adjacency matrix A to enhance the representation of node features. Simultaneously, a combination matrix is ​​formed by combining the node out-degree matrix to integrate edge weight information and smoothing motif matrix information. An adjustable hyperparameter λ is used to dynamically balance the influence of these two types of structural inputs, enabling the model to have stronger adaptability in different traffic structure environments. In the second aggregation layer, the node embedding Z is further refined. t To improve its representation ability and enhance the nonlinear expression depth of the model, a time smoothing mechanism is introduced in the method to address the potential abrupt changes in node embeddings between different time steps. The node embeddings generated at the current time step are weighted and fused with the results of the previous time step by a preset smoothing factor β, so that the embedding results still have stability and continuity under dynamic changes, thereby providing a more stable embedding basis for the continuous division and interactive modeling of functional cells.

[0146] The cell structure of dynamic urban networks evolves over time. Traditional partitioning methods (such as Louvain or spectral clustering) lack sufficient support for weights and directions, and are prone to generating unstable or fragmented cells in dynamic networks. This application, in the partitioning of dynamic urban functional cells, generates continuous and meaningful cell partitions through weighted K-means and dynamic centroid initialization, simplifying the complex network into a modular structure, facilitating the analysis of inter-regional interactions. Dynamic adjustment ensures that the partitioning adapts to changes in the urban network, improving the accuracy and stability of predictions.

[0147] As a preferred example of this application, in step S33, the weighted K-means method is used to identify dynamic urban functional cells by utilizing the geometric distance and node weights of the embedding space. When initializing the cluster centers, the centroids of urban functional cells at t-1 are referenced, and the optimization objective is as follows:

[0148]

[0149] in, It is a node The weights, combined with the intensity and degree of the smoothed motif, μ t,i It is the functional cell of the city The weighted centroid is calculated using the following formula:

[0150]

[0151] When t=0, randomly initialize k0 (10, the number can be set according to the actual city situation) centroids. If t>=1, use... The center of mass μ t-1,i As initial values, add or delete centroids to match k. t .

[0152] Adjustment of functional cell body number:

[0153]

[0154] Where δ is a preset adjustment parameter that is dynamically adjusted according to the network size, such as δ = 1.0.

[0155] In the process of dynamic urban functional cell identification and adjustment, this application aims to improve the temporal coherence and structural rationality of the segmentation results. First, after the node embedding is generated, a comprehensive weight is constructed for each node by combining the smooth motif intensity, out-degree and in-degree information. This weight not only reflects the centrality of the node in the structure, but also the degree of its interaction in the motif. Then, a weighted K-means algorithm is used to segment urban functional cells in the embedding space. At t=0, the system randomly initializes several centroids. In subsequent time steps t≥1, the centroids of urban functional cells identified in the previous time step are preferentially inherited as the initial cluster centers. Centroids are automatically added or deleted according to the current network structure changes to dynamically adapt to actual changes. This application significantly improves the accuracy and robustness of functional cell segmentation by constructing a weighted similarity index through the fusion of multi-source structural features when segmenting dynamic urban functional cells. Especially in the context of frequent changes in node connections and drastic fluctuations in traffic flow in dynamic urban networks, the weighted K-means clustering based on dual information of node embedding and structural weights can more effectively reflect the structural consistency within urban areas. At the same time, the clustering initialization based on historical centroids and the identification of functional cell state changes ensure the continuity and stability of the segmentation results in the time dimension, thereby providing a well-structured and controllable analytical basis for subsequent traffic flow prediction, congestion analysis, and regional scheduling strategies.

[0156] As a preferred example of this application, the following steps are included when predicting changes in the intensity of intercellular feedback between functional cells in subsequent time segments:

[0157] S41: Extract cell-level features, including urban functional cell-body embeddings. Motif intensity and urban functional cell in / out influx / outflow, constructing mutual feedback features of urban functional cell pairs. Based on historical interaction feature sequences, the fundamental feedback intensity of urban functional cell pairs at future time step t+h is predicted using a time series model.

[0158] S42: Combining the dynamic states of urban functional cells, including birth, expansion, shrinkage, and death, the predicted value of interaction intensity is adjusted through weighted similarity matching.

[0159] In the mutual feedback intensity prediction process, this application first constructs a multi-dimensional feature vector that integrates cell embedding location, motif interaction intensity, and in-degree and out-degree activity to fully characterize the structural performance and connectivity activity of urban functional cells. Then, the feature vectors are combined in pairs to form mutual feedback features representing regional interaction relationships, and the input is organized into the LSTM model according to the time series. The model is used to perform deep learning on the potential temporal dependencies in the historical sequence to predict the mutual feedback intensity in future time steps. On this basis, the system detects the birth, expansion, shrinkage, or death status of urban functional cells in real time, and dynamically corrects the LSTM output by weighted similarity matching with historical cells, thereby forming the final mutual feedback prediction result consistent with the actual network evolution.

[0160] In intelligent traffic prediction methods, traditional time-series prediction methods focus on a single time series, ignoring network structure and regional interactions, resulting in limited prediction accuracy, especially in dynamic urban networks. Cellular-level interaction prediction requires the fusion of topological and temporal information. This application achieves accurate prediction through cell embedding and LSTM models. Mutual feedback strength prediction provides insights into the future state of the network, supporting the optimization and management of dynamic networks; the cell-level perspective simplifies the analysis of complex networks, improving computational efficiency and prediction interpretability.

[0161] As a preferred example of this application, step S41 includes:

[0162] S411: Cell-level feature extraction:

[0163] Obtaining node embedding Z t Based on this, each cell body is generated. The embedding formula is as follows:

[0164]

[0165] S412: Constructing basic interfeedback features of the cell body;

[0166] First, for each cell body Constructing eigenvectors The formula is as follows:

[0167]

[0168] in, It is the sum of the motif intensities of all nodes within the cell body. These are the sum of the out-degree and in-degree of nodes within the cell body, respectively.

[0169] Subsequently, the mutual feedback features between cell body pairs are constructed. The formula is as follows:

[0170]

[0171] in This indicates that at time t, the cell body... The basic mutual feedback strength between them is given by the following formula:

[0172]

[0173] in, It is the edge Weight at time t;

[0174] S413: Timing modeling;

[0175] LSTM is used to model the time series of urban cell-level feedback intensity, with historical feedback features as input. With a hidden layer dimension of 128 and τ = 5, the LSTM model is constructed as follows:

[0176] h t ,c t =LSTM(X) i,j ,h t-1 ,c t-1 )

[0177]

[0178] Where h t ,c t These are the hidden states and cell states of the LSTM; W out It is the weight of the output layer, b out For the output layer bias, the loss function is:

[0179]

[0180] N represents the number of city cell pairs. This represents the basic mutual feedback strength of the urban functional cell pairs at that moment. This represents the predicted value of the interaction strength.

[0181] In the cell-level mutual feedback prediction process, this application first constructs a cell embedding vector by aggregating the motif strength, out-degree, and in-degree information of all nodes within each functional cell based on the node embedding obtained in the previous stage. The motif strength is used to characterize the structural tightness between nodes within the cell, and the out-degree and in-degree reflect the output and input connectivity of the cell, respectively. Then, the edge weights between any two cells are extracted from the urban network graph as the basis for their interaction strength. The mutual feedback feature vectors of cell pairs are constructed comprehensively, and the mutual feedback feature sequence between each pair of cells is constructed as a temporal input. Finally, an LSTM model is used to learn and model these sequences. Through training, the inherent laws of the evolution of mutual feedback features over time are captured, and the changes in the mutual feedback strength between cell pairs in the future are predicted. This enables temporal perception and quantitative expression of regional traffic evolution trends, thereby achieving a forward-looking analysis and judgment of regional interaction patterns in dynamic traffic networks.

[0182] As a preferred example of this application, in step S42, the weighted similarity between dynamic urban functional cells is calculated to identify... and The cell state and the cell feedback intensity of the city are:

[0183]

[0184] Dynamic urban functional cell birth: if With all of but For the functional cells of a newly emerging city;

[0185] Dynamic urban functional cell death: if With all of but die;

[0186] Dynamic urban functional cell body persistence: if and satisfy and but yes The continuation;

[0187] Dynamic urban functional cell expansion: if and but merge

[0188] Dynamic urban functional cell shrinkage: if and The city's functional cells have shrunk;

[0189] Dynamic urban cell birth: if For newly formed cells, the predicted mutual feedback intensity is initialized based on the average mutual feedback intensity at the current time step:

[0190]

[0191] Dynamic urban cell death: if They die at time t+h, meaning there are no matching cell bodies.

[0192] Dynamic urban cell expansion and contraction: If and Matching, adjusting according to size changes:

[0193]

[0194] in This represents the basic mutual feedback strength of the urban functional cell pairs at that moment.

[0195] This represents the predicted value of the mutual feedback intensity.

[0196] The cellular structure of dynamic urban networks changes over time (e.g., new nodes are added or cells disappear). Traditional prediction methods ignore these evolutions, leading to unreasonable or inaccurate prediction results. In the example of this application, to adapt to the changes brought about by the structural evolution in urban transportation networks and to maintain the rationality and continuity of the mutual feedback prediction results, this application introduces a dynamic urban functional cell state identification and mutual feedback intensity adjustment mechanism into the prediction process. First, by comparing the structural characteristics and distribution of urban functional cells in consecutive time steps, a weighted similarity between each pair of cells is calculated. This similarity comprehensively considers indicators such as node composition overlap, node embedding feature distance, and overall scale similarity. If a cell's similarity with all other cells is below a set threshold in the current time step, it is identified as a newly emerging urban functional cell; if it has no matching similarity in subsequent time steps... If a cell is identified as dead, it is considered dead. If the current cell has a high similarity to a cell from a previous time period and their node sizes are basically the same, it is considered a persistent state. If the similarity is high but the size increases significantly, it is considered an expansion. If the number of nodes decreases, it is considered a shrinkage. After the state identification is completed, the mutual feedback strength between cells is adjusted according to different states. The mutual feedback relationship of newly created cells is initialized with the average value of the current overall mutual feedback strength. The mutual feedback strength of dead cells is cleared to zero to reflect their exit from the network. The mutual feedback strength of expansion or shrinkage states is linearly adjusted according to the change ratio of cell size, thereby ensuring a reasonable transition of the network mutual feedback structure in time.

[0197] This application introduces a weighted similarity recognition mechanism and a feedback intensity state adjustment strategy, which enables accurate judgment and efficient response to the state of functional cells even in the context of frequent changes in the dynamic urban network structure. This effectively avoids the problem of prediction distortion caused by the appearance of new areas or the disappearance of old areas. At the same time, by adjusting the state matching of the feedback intensity, it ensures that the prediction results are highly consistent with the actual state of the network, which significantly improves the adaptability and prediction stability of the intelligent transportation system to the evolution of urban traffic structure.

[0198] As a preferred example of this application, this application also discloses an intelligent traffic prediction and control system based on dynamic functional cell coupling, comprising:

[0199] The data acquisition module is used to collect vehicle trajectory data and construct an urban static traffic network using the taxi operating area as the basic geographical unit. The nodes of the static traffic network represent geographical locations, and the edges represent the traffic interaction relationships between nodes and have weighted directions.

[0200] The dynamic network construction module is used to construct a multi-time-lapse sequence of urban network snapshots from a static network according to a given time interval Δ, thereby forming a dynamic urban network.

[0201] The feature extraction and embedding module includes a motif extraction unit and a graph neural network (GNN) unit. The motif extraction unit is used to identify the directed motif structure in the traffic network at each time step and calculate the intensity of node participation in the motif. The graph neural network (GNN) unit is used to fuse the directed weighted adjacency matrix and the weighted motif matrix, and generate node embeddings by aggregating node features.

[0202] The dynamic urban cell identification module is used to divide urban functional cells based on node embedding results and combined with time smoothing strategy, using the weighted K-means method, and to identify and update the cell status (birth, death, expansion, shrinkage).

[0203] The time-series prediction module extracts cell-level features and constructs mutual feedback features of urban functional cell pairs. These features are used to model the cell mutual feedback sequence based on a long short-term memory network (LSTM) and predict the traffic mutual feedback intensity between cell pairs in the future time period.

[0204] The strategy output module is used to generate traffic scheduling suggestions based on the prediction results to support traffic management tasks such as signal timing optimization, route navigation, congestion warning and infrastructure planning.

[0205] The intelligent traffic prediction and control system based on dynamic functional cell coupling described in this application constructs static and dynamic traffic networks through multi-source trajectory data. It inputs motifs reflecting higher-order structures and basic adjacency information into a graph neural network to generate node embeddings. Then, it combines time smoothing and weighted clustering to dynamically divide functional cells and identify their evolutionary states. By aggregating cell-level embeddings, motif strength, and in-degree, it constructs a mutual feedback feature sequence and performs time-series prediction using LSTM. Finally, it transforms the predicted mutual feedback strength results into traffic management strategies such as signal timing, navigation paths, and congestion warnings. Specifically, the intelligent traffic prediction and control system based on dynamic functional cell coupling described in this application first collects high-precision trajectory data of taxis and shared vehicles with time and speed information through an urban traffic management platform and third-party travel interfaces during the data acquisition phase. Taxi service areas are then abstracted into network nodes as geographical units, and inter-regional vehicle flow paths are converted into weighted directed edges to construct a static traffic network. Subsequently, in the dynamic network construction phase, a series of network snapshots reflecting urban traffic conditions at different time periods are generated at preset time intervals (given time intervals) and combined into a dynamic graph. A motif extraction unit identifies directed motifs reflecting regional cyclical interactions in each snapshot and calculates node values. The motif matrix is ​​constructed by weighting the motif intensity of the points. Then, the adjacency matrix and the motif matrix are fused by the GNN unit to generate node embeddings through multi-level aggregation. Next, in the dynamic urban cell identification module, time-smooth embedding is used in combination with node weights to dynamically divide functional cells using the weighted K-means algorithm and track their birth, expansion, shrinkage and extinction states. Then, in the time series prediction stage, the mutual feedback feature sequence composed of embedding, motif intensity and in-degree is calculated by pairwise combination of functional cells and input into the LSTM model to predict the future mutual feedback intensity. Finally, the strategy output module provides executable scheduling schemes for traffic light timing, path guidance, congestion warning and regional planning based on the prediction results.

[0206] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented using computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, FLASH, floppy disk, magnetic disk or optical disk, hard disk, and cloud storage. The computer processor can be a central processing unit (CPU), microprocessor (MPU), digital signal processor (DSP), or field-programmable gate array (FPGA), meaning that the steps of the intelligent traffic prediction and control method based on dynamic functional cell coupling can be implemented using the aforementioned processor.

[0207] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for intelligent traffic prediction and control based on dynamic functional cell coupling, characterized in that, include: Acquire urban vehicle trajectory data, and construct an initial static urban network with taxi areas as the geographical unit. Network nodes represent urban geographical locations, and network edges represent traffic interaction relationships between nodes with direction and weight. The static urban network is transformed into a dynamic urban network composed of multiple time-segment snapshots of the urban network, in order to characterize the time-varying structure of the urban transportation network. In each time-segment network snapshot, motif features are extracted and a feature matrix is ​​constructed. A GNN model is used to aggregate the dynamic embedding of geographic locations to segment urban functional cells, and the dynamic changes of cells are identified by combining the cell state of the previous time step. The following steps are included in the division of urban functional cells; S31: For directed weighted dynamic city networks Based on the specified directed motif type, extract motif features for each time step t and generate weighted motif intensity. and motif matrix And integrate historical information for time smoothing; S32: The directed weighted adjacency matrix and smooth motif matrix Input graph neural networks perform feature aggregation to generate node embeddings. And time smoothing is used to make the embedding continuous; S33: Node-based embedding The weighted K-means algorithm was used to divide the dynamic urban functional cells. And perform cluster initialization using the cell centroids of the previous time step; Based on the dynamic urban functional cell structure, we construct the mutual feedback feature vector between functional cells and form a mutual feedback time series. We then use the LSTM model to model and predict the changes in the mutual feedback intensity between functional cells in subsequent time segments. When predicting changes in the intensity of functional cell-cell feedback in subsequent time segments, the following steps are included: S41: Extract cell-level features, including urban functional cell-body embeddings. Based on motif intensity and the out- and in-degree of urban functional cells, the mutual feedback characteristics of urban functional cell pairs are constructed. Based on historical interaction feature sequences, the fundamental feedback intensity of urban functional cell pairs at future time step t+h is predicted using a time series model. ; S42: Combining the dynamic states of urban functional cells, including birth, expansion, shrinkage, and death, the predicted value of interaction intensity is adjusted through weighted similarity matching. ; Provide support for urban transportation planning based on the forecast results, including at least one of the following: signal timing, route navigation, traffic warning, and regional dispatch recommendations.

2. The intelligent traffic prediction and control method based on dynamic functional cell coupling according to claim 1, characterized in that, In the process of constructing dynamic urban networks, according to a given time interval Divide a day into a set of discrete time intervals. A city network snapshot is generated at each time interval, representing the dynamic city network as a set of city snapshots across multiple time intervals. ,in, This represents the static city network at the t-th time interval. , Let represent the set of city network nodes and the set of directed edges respectively at the t-th time interval. Each edge in Set a weight , This represents the nodes in the city network during the t-th time interval. and The interaction strength between them is determined by normalization, which limits the edge weights to the range [0,1].

3. The intelligent traffic prediction and control method based on dynamic functional cell coupling according to claim 1, characterized in that, In step S31, motif feature extraction includes: S311: Computation Node Number of motifs at time t and weighted motif intensity The formula is as follows: Where A is The directed weighted adjacency matrix, , This represents the nodes in the city network during the t-th time interval. and The strength of interaction between them This represents the nodes in the city network during the t-th time interval. and The strength of interaction between them This represents the nodes in the city network during the t-th time interval. and The intensity of interaction between them; It contains nodes A set of triangles, It is a node exist The product of the weights of all edges in the middle, For participating motif instances; S312: Employs time-smoothed weighted motif intensity Obtain smooth motif strength , ,in It is the first smoothing factor, initially ; S313: Construct a weighted motif matrix Represents a node The weighted connection strength in a directed motif is calculated using the following formula: After normalization and smoothing, the smoothed motif matrix is ​​obtained: This is the second smoothing factor.

4. The intelligent traffic prediction and control method based on dynamic functional cell coupling according to claim 3, characterized in that, In step S32, the aggregation formula of GNN is: in, It is a directed weighted adjacency matrix at time t. , It is the identity matrix. For the first Layer node embedding matrix, For learnable weights, It is an activation function. These are hyperparameters used to balance the contributions of the adjacency matrix and the motif matrix. ; Based on this, node embedding is obtained. The formula is as follows: in It is the third smoothing factor.

5. The intelligent traffic prediction and control method based on dynamic functional cell coupling according to claim 4, characterized in that, In step S33, the weighted K-means optimization objective is: in, It is a node The weight, It is the functional cell of the city The weighted centroid, The number of urban functional cells; The node weight The calculation formula is: It is a node The intensity of the smooth motif.

6. The intelligent traffic prediction and control method based on dynamic functional cell coupling according to claim 1, characterized in that, Step S41 includes: S411: Cell-level feature extraction: In obtaining node embeddings Based on this, each cell body is generated. The embedding formula is as follows: S412: Constructing basic interfeedback features of the cell body; First, for each cell body Constructing eigenvectors The formula is as follows: in, It is the sum of the motif intensities of all nodes within the cell body. These are the sum of the out-degree and in-degree of nodes within the cell body, respectively. Subsequently, the mutual feedback features between cell body pairs are constructed. The formula is as follows: in This indicates that at time t, from the cell body to cell body The basic mutual feedback strength between them This indicates that at time t, from the cell body to cell body The basic mutual feedback strength between them The formula is as follows: in, It is an edge ( The weight at time t; S413: Timing modeling; LSTM is used to model the time series of urban cell-level feedback intensity, with historical feedback features as input. The LSTM model is constructed as follows: in These are the hidden states and cell states of the LSTM; These are the weights of the output layer. For the output layer bias, the loss function is: Where N represents the number of urban cell pairs. Let be the basic mutual feedback intensity of the urban functional cell pairs at time t+h. The value is the predicted value of the LSTM mutual feedback intensity at time t+h.

7. The intelligent traffic prediction and control method based on dynamic functional cell coupling according to claim 6, characterized in that, In step S42, the weighted similarity between dynamic urban functional cells is calculated to identify... and The cell state and the cell feedback strength of the city are: like With all of ,but Initialize the new cell body feedback strength for the newly formed urban functional cells: like With all of ,but die, ; like and satisfy ,and ,but yes The continuation; like ,and ,but merge The city's functional cells expand; like ,and The city's functional cells have shrunk; when and Matching, adjusting according to size changes: in This represents the basic mutual feedback strength of the urban functional cell pairs at that moment. This represents the predicted value of the mutual feedback intensity.

8. An intelligent traffic prediction and control system based on dynamic functional cell coupling, characterized in that, The method for implementing the intelligent traffic prediction and control method based on dynamic functional cell coupling as described in any one of claims 1 to 7 includes: The data acquisition module is used to collect vehicle trajectory data and construct urban static traffic networks; The dynamic network construction module is used to construct a dynamic urban network from a static network by creating a sequence of multi-time snapshots of the urban network. The feature extraction and embedding module includes a motif extraction unit for identifying directed motif structures and calculating node participation strength, and a graph neural network (GNN) unit for fusing adjacency matrices and motif matrices to generate node embeddings. The dynamic urban cell identification module is used to divide urban functional cells based on node embedding results and combined with time smoothing strategy, using the weighted K-means method, and to identify and update the cell status. The time-series prediction module extracts cell-level features and constructs feedback features of urban functional cell pairs. These features are used to model cell feedback sequences based on long short-term memory networks and predict the traffic feedback intensity between cell pairs in future time periods. The strategy output module is used to generate traffic scheduling suggestions based on the prediction results to support any traffic management task, including signal timing optimization, route navigation, congestion warning, and infrastructure planning.

Images