Traffic network scheduling control method and system based on space-time big data, and electronic equipment

Abstract

The embodiment of the invention discloses a traffic network scheduling control method and system based on space-time big data and electronic equipment, and belongs to the technical field of control or regulation systems.The method comprises the steps that a first network state feature tensor representing a local space-time evolution mode is extracted from a space-time grid through a three-dimensional convolutional neural network, and aggregating node information on the dynamic weighting graph by using the graph attention network, generating a second network state feature tensor reflecting the global association and congestion propagation situation of the road network, fusing the two feature tensors, and inputting the fused feature tensors into an attention enhancement sequence model for multi-step prediction to obtain a future multi-period whole network traffic state. And cooperatively inputting the prediction result, the real-time state, the rule and the capacity constraint into the optimization model, generating a comprehensive scheduling instruction set including signal timing, lane control and path induction, and issuing and executing the comprehensive scheduling instruction set. According to the embodiment of the invention, holographic perception, prospective prediction and cooperative regulation and control of the traffic network are realized, and the traffic efficiency and scheduling response capability of the road network are improved.

Description

A traffic network scheduling and control method, system, and electronic equipment based on spatiotemporal big data. Technical Field

[0001] This application relates to a traffic network scheduling and control method, system, and electronic device based on spatiotemporal big data, belonging to the technical field of control or regulation systems. Background Technology

[0002] Modern urban transportation networks are highly complex and dynamically changing systems, and ensuring their efficient and safe operation is crucial for safeguarding socio-economic activities. With accelerating urbanization and the continuous growth of motor vehicle ownership, traffic congestion has become one of the major bottlenecks restricting urban development. To address this challenge, intelligent transportation systems (ITS) have emerged. ITS utilizes information, communication, and control technologies to achieve proactive management and optimized scheduling of traffic flow.

[0003] In existing technologies, traffic network scheduling and control primarily rely on real-time data response systems based on fixed detectors (such as inductive loop detectors and video cameras). These systems typically use individual intersections or short road segments as control units, employing preset schemes or local adjustments to signal timing based on simple real-time traffic thresholds (such as saturation), such as inductive signal control or green wave coordination control. Furthermore, some recent technologies have begun to utilize historical data statistical patterns to predict short-term traffic flow and fine-tune control strategies accordingly. However, these existing technologies are increasingly revealing their inherent limitations when dealing with large-scale road networks and complex traffic situations, specifically in the following aspects:

[0004] First, the perception dimension is limited, and the feature extraction capability is insufficient. Existing technologies mainly rely on sparsely deployed fixed detectors to obtain local point data such as cross-sectional flow and occupancy. This data not only has a large number of spatial coverage blind spots, but more seriously, it is low-dimensional and fragmented, making it difficult to comprehensively and deeply depict the complex nonlinear evolution patterns of traffic flow in time and space, such as the dynamic process of congestion generation, propagation, and dissipation. Although some studies have attempted to introduce floating car (such as GPS navigation systems for taxis and ride-sharing) trajectory data to supplement coverage, how to effectively integrate these two heterogeneous and asynchronous spatiotemporal big data—point cross-sectional data and linear trajectory data—and extract deep features that can simultaneously reflect local micro-evolution and global macro-correlation remains a key problem that existing technologies have not yet adequately solved.

[0005] Second, the modeling methods are limited and struggle to depict global dynamic relationships. Traditional traffic models (such as traffic assignment models and cellular transport models) or simple rule-based controllers typically treat the road network as a static or quasi-static system, with fixed or slowly adjusting model parameters. These models struggle to accurately describe and predict drastic changes in traffic conditions caused by unforeseen events, weather changes, or special activities. More importantly, existing methods often view individual intersections or road segments in isolation, lacking effective means to model the dynamic, long-range, and nonlinear interactions (i.e., congestion propagation) between nodes (intersections, road segments) within the complex road network. This leads to control strategies that are often locally optimal rather than globally optimal, and may even trigger secondary congestion in the road network due to a lack of coordination.

[0006] Third, the prediction and decision-making processes are disconnected, resulting in poor forward-looking and adaptive control. Currently, most systems operate traffic state prediction and control decision-making as two independent modules running in series. The prediction module often uses statistical time series models, which have limited accuracy and duration. Furthermore, the prediction results (mostly traffic flow or speed for the next few minutes) are passed to the decision-making module in a simplified form. The decision-making module then performs calculations based on these potentially inaccurate predictions and simple optimization objectives (such as minimizing current intersection delays). This loose coupling between the prediction and decision-making modules leads to a lack of true forward-looking capabilities in the output. It fails to iteratively generate a coordinated optimal control instruction set (such as a combination of signal timing, dynamic lane function allocation, and path guidance strategies) based on multi-step spatiotemporal situation projections, thus appearing passive and sluggish in the face of dynamic changes.

[0007] Fourth, the optimization objectives and constraints are too simple, making it difficult to achieve system-level optimization. Existing scheduling optimization models often only consider a single objective (such as maximizing capacity or minimizing delay), and the constraints are too simplified (such as ignoring the feedback effect of driver route selection behavior on control effectiveness). They fail to comprehensively consider multiple constraints and objectives, such as dynamic traffic demand (OD matrix), multi-modal traffic flow, complex intersection channelization, variable lanes, the impact of guidance information, and road physical capacity, within a unified optimization framework. This results in the generated scheduling instructions potentially being ineffective in actual execution, or even producing unexpected negative effects.

[0008] In summary, existing traffic network scheduling and control methods and systems have some shortcomings in areas such as deep fusion and deep feature extraction of multi-source heterogeneous spatiotemporal big data, accurate modeling of global dynamic correlations in the road network, and closed-loop coupling of long-term accurate prediction and multi-objective collaborative decision-making. These shortcomings easily lead to problems such as low control accuracy, delayed response, and insufficient coordination, making it difficult to fundamentally alleviate the problem of large-scale traffic congestion in modern urban road networks. Existing technologies can no longer meet people's needs and urgently need to be improved. Summary of the Invention

[0009] The main objective of this application is to provide a traffic network scheduling and control method, system, and electronic device based on spatiotemporal big data, thereby addressing the shortcomings of existing technologies.

[0010] This application adopts the following technical solution: According to one aspect of this application, a traffic network scheduling and control method based on spatiotemporal big data is provided, including: acquiring spatiotemporal big data of a target traffic network, constructing dynamic representation information of the traffic network, mapping the dynamic representation information of the traffic network into a spatiotemporal grid tensor according to geographical location, and extracting the spatiotemporal grid tensor using a three-dimensional convolutional neural network to generate a first network state feature tensor; constructing a graph model based on the traffic network topology, wherein nodes represent intersections or road segments; dynamically calculating the weights of each edge in the graph according to real-time traffic flow speed or density to form a dynamically weighted graph; using a graph attention network, with the dynamic edge weights as the basis for the correlation strength between nodes, aggregating and updating the feature information of nodes in the dynamically weighted graph to generate a second network state feature tensor. The second network state feature tensor is used to reflect the global correlation and congestion propagation trend of the road network. The first and second network state feature tensors are input into the fusion prediction module to form a network spatiotemporal state code. The network spatiotemporal state code is input into an attention-enhanced sequence prediction model. The sequence prediction model has a built-in attention mechanism and outputs the traffic state prediction results of each node in the road network in multiple future time periods through the attention mechanism. The traffic state prediction results include flow rate, speed and congestion index. The traffic state prediction results, the current real-time road network state, the preset scheduling rules and road capacity are input into a network scheduling decision optimization model to generate a comprehensive scheduling control instruction set. The comprehensive scheduling control instruction set is issued to the traffic signal control system to perform real-time regulation of the traffic network.

[0011] According to at least one specific embodiment of the present application, the step of acquiring spatiotemporal big data of the target traffic network and constructing dynamic representation information of the traffic network further includes: acquiring multi-source heterogeneous raw traffic flow data in real time based on fixed sensing devices and floating car trajectory data sources in the target traffic network; preprocessing the raw traffic flow data and formatting it under a preset geographic coordinate system and a unified time reference to obtain standardized spatiotemporal sequence data; using a pre-divided geographic regular grid, associating the floating car trajectory with the traffic network through a map matching algorithm, and assigning attribute allocation for trajectory points / segments based on the spatial relationship between the traffic network and the geographic regular grid; and... The detector data is located to its geographically regular grid. For each geographically regular grid, all allocated data are aggregated and calculated within each preset time slice to generate a five-dimensional spatiotemporal tensor. The dimensions of the five-dimensional spatiotemporal tensor correspond to: time step sequence, grid row index, grid column index, and traffic feature channel, respectively. The constructed five-dimensional spatiotemporal tensor is input into a pre-trained three-dimensional convolutional neural network model. Feature extraction is performed by the three-dimensional convolutional kernel that slides synchronously in the spatial and temporal dimensions of the three-dimensional convolutional neural network model, and the first network state feature tensor is output. The first network state feature tensor is used to characterize the spatiotemporal evolution pattern of traffic status in a local area.

[0012] According to at least one specific embodiment of the present application, the preprocessing of the original traffic flow data specifically involves spatiotemporal alignment, outlier removal, and missing value imputation of the original traffic flow data; the traffic feature channel includes traffic flow obtained by aggregating data from fixed detectors and spatial average speed calculated from floating car trajectory data; during feature extraction using three-dimensional convolutional kernels, the three-dimensional convolutional neural network model is downsampled via pooling layers to capture nonlinear spatiotemporal dependencies across grids and time periods layer by layer.

[0013] According to at least one specific embodiment of the present application, the step of constructing a graph model based on the traffic network topology, wherein nodes represent intersections or road segments; dynamically calculating the weights of each edge in the graph based on real-time traffic flow speed or density to form a dynamically weighted graph; using a graph attention network, with the dynamic edge weights as the basis for the strength of association between nodes, aggregating and updating the feature information of nodes in the dynamically weighted graph to generate a second network state feature tensor, further includes: constructing a basic skeleton of the graph model based on the topology of the target traffic network, wherein the node set in the basic skeleton includes all signal-controlled intersections and key points of road segments determined according to preset objective rules, and the edge set in the node set is used to connect nodes with direct passage relationships; Based on real-time acquired traffic flow speed or density data, the weight of each edge in the graph is calculated. The weight is calculated based on the real-time travel time or its reciprocal, and the weight update frequency is consistent with the traffic data collection frequency or scheduling decision cycle. The initial feature vectors and weighted adjacency matrices of each node in the dynamically weighted graph are input into a graph attention network. The initial feature vectors are composed of the real-time state and static attributes of the corresponding nodes. The element values ​​of the weighted adjacency matrix are generated by the dynamic weights of the corresponding edges. Through the multi-head graph attention layer in the graph attention network, the node features are iteratively aggregated so that each node obtains a corresponding updated feature vector. The updated feature vectors of all nodes together constitute the second network state feature tensor.

[0014] According to at least one specific implementation of the embodiments of this application, the preset objective rules include: road grade, number of lanes or historical traffic threshold, and the key points of the road segment include the road segment endpoints or the road segment center point; the initial feature vector is composed of the real-time state and static attributes of the corresponding node, specifically: the initial feature vector is composed of the real-time state and static attributes of the intersection or road segment.

[0015] According to at least one specific embodiment of the present application, the step of inputting the first network state feature tensor and the second network state feature tensor into a fusion prediction module to form a network spatiotemporal state code, and inputting the network spatiotemporal state code into an attention-enhanced sequence prediction model, wherein the sequence prediction model has a built-in attention mechanism and outputs the traffic state prediction results of each node of the road network in multiple future time periods through the attention mechanism, further includes: inputting the first and second network state feature tensors into a fusion prediction module, integrating the input tensors through the fusion prediction module to generate a unified network spatiotemporal state code; inputting the unified network spatiotemporal state code into an attention-enhanced sequence prediction model, wherein the sequence prediction model obtains the long-range spatiotemporal dependencies across nodes and across time in the unified network spatiotemporal state code through its built-in multi-head self-attention mechanism; and decoding the unified network spatiotemporal state code through a decoder in the sequence prediction model to obtain multi-channel prediction results corresponding to each node of the road network in multiple consecutive future time periods, wherein the multi-channel prediction results correspond to the predicted traffic flow, predicted speed, and predicted congestion index of the future traffic state, respectively.

[0016] According to at least one specific embodiment of the present application, the step of inputting the traffic state prediction result, the current real-time road network state, the preset scheduling rules, and the road capacity into a network scheduling decision optimization model to generate a comprehensive scheduling control instruction set, and then sending the comprehensive scheduling control instruction set to the traffic signal control system to perform real-time regulation of the traffic network, further includes: constructing a network scheduling decision optimization model based on the traffic state prediction result, the current real-time road network state, the dynamic OD matrix, and the road capacity data; the network scheduling decision optimization model has a built-in model predictive control framework, in which the optimal control instruction sequence for several future cycles is solved with the current state as the initial point; generating a comprehensive scheduling control instruction set based on the immediate control instructions output by the network scheduling decision optimization model, the comprehensive scheduling control instruction set including dynamic signal timing schemes, variable lane control instructions, and route driving suggestions, and then sending the comprehensive scheduling control instruction set to the traffic signal control system in real time to perform corresponding traffic regulation operations.

[0017] According to at least one specific embodiment of the present application, the dynamic OD matrix is ​​specifically the vehicle's intention to move between regions based on historical and real-time floating car trajectories; the step of solving for the optimal control command sequence for several future cycles with the current state as the initial point specifically involves: taking the current state as the initial point, and based on the traffic state prediction results, iteratively solving for the optimal control command sequence for future cycles under the conditions of satisfying signal, lane, and capacity constraints; the integrated scheduling control command set includes: signal timing schemes, variable lane control commands, and route suggestion combinations; the traffic signal control system includes electronic traffic signs, vehicle-mounted traffic platforms, or mobile terminal traffic platforms.

[0018] According to another aspect of the embodiments of this application, a traffic network scheduling and control system based on spatiotemporal big data is provided to implement the aforementioned traffic network scheduling and control method based on spatiotemporal big data. The system includes: a first network state feature tensor generation module, which acquires spatiotemporal big data of a target traffic network, constructs dynamic representation information of the traffic network, maps the dynamic representation information of the traffic network to a spatiotemporal grid tensor according to geographical location, and extracts the spatiotemporal grid tensor using a three-dimensional convolutional neural network to generate a first network state feature tensor; and a second network state feature tensor generation module, which constructs a graph model based on the traffic network topology, where nodes represent intersections or road segments; dynamically calculates the weights of each edge in the graph according to real-time traffic flow speed or density to form a dynamically weighted graph; and uses a graph attention network, with the dynamic edge weights as the basis for the strength of association between nodes, aggregates and updates the feature information of nodes in the dynamically weighted graph to generate a second network. The system comprises a state feature tensor, the second network state feature tensor being used to reflect the global correlation and congestion propagation trend of the road network; a traffic state prediction result output module, which inputs the first and second network state feature tensors into the fusion prediction module to form a network spatiotemporal state code, and inputs the network spatiotemporal state code into an attention-enhanced sequence prediction model, which has a built-in attention mechanism to output the traffic state prediction results of each node in the road network for multiple future time periods, including flow rate, speed, and congestion index; and a real-time traffic network control module, which inputs the traffic state prediction results, the current real-time road network state, preset scheduling rules, and road capacity into a network scheduling decision optimization model to generate a comprehensive scheduling control instruction set, which is then sent to the traffic signal control system to perform real-time control of the traffic network.

[0019] According to another aspect of the embodiments of this application, an electronic device is provided, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method.

[0020] The beneficial technical effects of the embodiments of this application are as follows: The traffic network scheduling and control method based on spatiotemporal big data provided in this application adopts a collaborative computing approach of dual-path feature perception, fusion prediction, and closed-loop decision-making. By generating a first network state feature tensor, a three-dimensional convolutional neural network is used to extract local spatiotemporal evolution features from a spatiotemporal grid, and then a second network state feature tensor is generated. A graph attention network is used to aggregate node information on a dynamically weighted graph to capture the global congestion propagation trend. Then, the two types of features are fused and input into an attention-enhanced sequence model for multi-step accurate prediction. Finally, the prediction results are collaboratively input into an optimization model with real-time status, rules, and capacity constraints to generate a comprehensive scheduling command. This enables comprehensive perception, forward prediction, and collaborative regulation of the traffic network, achieving the technical effects of improving the overall traffic efficiency of the road network, suppressing congestion spread, and enhancing the agility of scheduling response. It solves the technical problems of low regulation accuracy, delayed response, and insufficient coordination caused by the separation of local and global features and the disconnect between prediction and decision-making links in traditional methods. Attached Figure Description

[0021] To more clearly illustrate the specific implementation methods of the embodiments of this application or the technical solutions in the prior art, the drawings used in the description of the specific implementation methods or the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 is a flowchart of steps S1 to S4 of an embodiment of this application.

[0023] Figure 2 is a flowchart of the optimized technical solution method for steps S11 to S14 in the embodiment of this application.

[0024] Figure 3 is a flowchart of the optimized technical solution method for steps S21 to S24 in the embodiment of this application.

[0025] Figure 4 is a flowchart of the optimized technical solution method for steps S31 to S33 in the embodiments of this application.

[0026] Figure 5 is a flowchart of the optimized technical solution method for steps S41 to S43 in the embodiments of this application.

[0027] Figure 6 is a system architecture diagram of an embodiment of this application.

[0028] Figure 7 is a schematic diagram of the electronic device. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the embodiments of this application, and not all embodiments. Based on the specific implementation methods in the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the embodiments of this application.

[0030] The embodiments of this application can be implemented in various application scenarios, such as routine traffic management during morning and evening rush hours in cities, vehicle scheduling on road networks in central business districts (CBDs), tidal traffic management, traffic control on accident-prone road sections, and so on.

[0031] This application scenario takes a road network covering approximately 10 square kilometers of the target CBD as an example. By integrating traffic flow and occupancy data from fixed detectors (inductive loops and video checkpoints) with real-time GPS trajectory data from approximately 500 taxis and ride-hailing vehicles within the area, a dynamic representation of the traffic network is constructed. The core computing cycle is initiated approximately one hour before and after the morning and evening rush hours each day.

[0032] Path 1, Local Spatiotemporal Features: The system divides the road network into 50m*50m geographical grids, using 5-minute time slices. Historical data (traffic flow, average speed) from the past 60 minutes is vectorized. The mapping dimensions of the spatiotemporal grid tensor are: 12 time slices, 200 rows of grid, 200 columns of grid, and 2 feature channels. This spatiotemporal grid tensor is input into a pre-trained 3D convolutional neural network. The 3D convolutional neural network uses three layers of 3D convolution and pooling operations to progressively abstract features: the first layer identifies the start and stop fluctuations of traffic flow at each entrance lane of a single intersection; the second layer captures the green / red wave phenomenon formed by convoys between adjacent intersections (the green / red wave phenomenon refers to whether vehicles can pass smoothly at multiple signal-controlled intersections); and the third layer outlines the spatiotemporal contours of congestion generation and dissipation within a small area (such as a street block). Finally, it outputs the first network state feature tensor, which describes the real-time evolution patterns of multiple micro-"hotspots" within the area.

[0033] Path Two, Global Relationship Features: The system synchronously constructs a road network map model, with nodes representing the center points of 156 signal-controlled intersections and 87 key arterial road segments within the region. Based on floating car GPS data (which can be provided by taxis, ride-hailing services, or carpooling), the average travel time of each road segment (edge) is calculated in real time, and its reciprocal is used as the dynamic edge weight (traffic efficiency), forming a dynamically weighted graph. This dynamically weighted graph is then input into the graph attention network. The GAT layer of the graph attention network calculates the attention coefficient based on the dynamic edge weight: for example, when the traffic efficiency of an arterial road drops sharply due to an accident (edge ​​weight increases), the upstream and downstream intersection nodes directly connected to it will pay more attention to this high-weight edge during feature aggregation, thus enabling congestion information to be quickly perceived and aggregated into the features of these nodes. After propagation through two layers of GAT, the features of a node on the periphery of the CBD can already contain the impact intensity information propagated from the core congestion point through two road segments. The final generated second network state feature tensor clearly reveals the global situation and pressure transmission path of congestion spreading from the core accident point along the arterial road to the outward radial roads.

[0034] The two feature tensors mentioned above are fed into the fusion prediction module. The fusion prediction module adopts a cross-attention mechanism, which enables local features (such as abnormal left-turn traffic flow at a key intersection) to query the most relevant parts of the global features (such as the pressure situation of upstream related road sections) and perform weighted fusion to generate a comprehensive and unified network spatiotemporal state code.

[0035] The encoding was then fed into a sequence prediction model based on the Transformer architecture. The model used its multi-head self-attention mechanism to analyze the encoded information: one part of the head focused on time dependence, identifying that the current congestion pattern was similar to the historical pattern in which traffic flow began to gather 45 minutes ago; another part of the head focused on spatial dependence, discovering that several key diversion nodes would play a decisive role in the future road network state. Based on this, the model made rolling predictions of the overall network traffic state (flow, speed, congestion index) for six time slices with 5-minute intervals within the next 30 minutes. The final prediction results showed that without intervention, congestion would spread to two key cross-regional bridges in 20 minutes, causing regional network paralysis.

[0036] Based on this judgment, the system immediately starts the network scheduling decision optimization model. The network scheduling decision optimization model is based on the model predictive control framework. It combines the prediction results of the above steps, the current real-time status, the preset scheduling rules such as the minimum green light time of traffic lights, the physical capacity of each road, and the dynamic OD matrix (travel demand distribution) of the current time period obtained from historical data into a rolling time domain optimization problem. The optimization objective is to minimize the total travel delay of the entire network within the next 30 minutes.

[0037] After the model is solved, a set of coordinated integrated scheduling and control instructions is generated: Instruction set 1, signal timing scheme: Dynamic traffic interception is carried out at 12 intersections upstream of the accident point, the green light time of relevant phases in their signal cycle is reduced step by step, and at the same time, the green light time of downstream intersections is increased to form a relief green wave.

[0038] Instruction set 2, Variable lane control instruction: The middle section of a main road leading to the congestion direction, which is currently a two-way traffic lane, will be adjusted to a tidal flow lane via lane signals, temporarily changing to in-way traffic for the next 30 minutes, providing an additional channel for traffic flow to dissipate.

[0039] Command set 3, route guidance suggestion combination: Through roadside variable message signs and navigation APP, three levels of guidance information are issued to vehicles about to enter the area: Level 1 guidance suggests that distant vehicles detour via the outer ring; Level 2 guidance suggests that nearby vehicles use the newly opened tidal flow lane; Level 3 guidance prompts the real-time travel time of alternative routes.

[0040] Multiple instruction sets, transmitted via fiber optic networks and 4G / 5G communication, can be rapidly distributed to relevant traffic signals and variable message signs at multiple intersections, and pushed to partner map service platforms. Once executed, the system continuously monitors real-time traffic flow changes, repeating the feature extraction and instruction generation process at fixed intervals to achieve dynamic adjustments. Applying this embodiment, the system successfully confined predicted severe congestion to the core area in this application scenario, preventing a network-wide shutdown. Average vehicle speeds during peak hours in the region improved, and the peak congestion index decreased significantly.

[0041] As shown in Figure 1, this application discloses a traffic network scheduling and control method based on spatiotemporal big data, including: Step S1, acquiring spatiotemporal big data of the target traffic network, constructing dynamic representation information of the traffic network, mapping the dynamic representation information of the traffic network into a spatiotemporal grid tensor according to geographical location, and extracting the spatiotemporal grid tensor using a three-dimensional convolutional neural network to generate a first network state feature tensor. The first network state feature tensor in step S1 is mainly used to capture the spatiotemporal evolution pattern of local areas in the road network.

[0042] Step S2: Construct a graph model based on the traffic network topology, where nodes represent intersections or road segments. Dynamically calculate the weights of each edge in the graph based on real-time traffic flow speed or density to form a dynamically weighted graph. Using a graph attention network, with the dynamic edge weights as the basis for the strength of association between nodes, aggregate and update the feature information of the nodes in the dynamically weighted graph to generate a second network state feature tensor. This second network state feature tensor reflects the global association and congestion propagation trend of the road network.

[0043] Step S3: Input the first network state feature tensor and the second network state feature tensor into the fusion prediction module to form a network spatiotemporal state code. Input the network spatiotemporal state code into an attention-enhanced sequence prediction model. The sequence prediction model has a built-in attention mechanism and outputs the traffic state prediction results of each node of the road network in multiple future time periods through the attention mechanism. The traffic state prediction results include flow rate, speed and congestion index.

[0044] Step S4: Input the traffic state prediction results, the current real-time road network state, the preset scheduling rules, and the road capacity into a network scheduling decision optimization model to generate a comprehensive scheduling control instruction set. Then, send the comprehensive scheduling control instruction set to the traffic signal control system to perform real-time regulation of the traffic network.

[0045] The optimization solutions provided in steps S1 to S4, through a collaborative computing framework, achieve a systematic understanding and forward-looking regulation of the traffic system from micro to macro levels, and from the current situation to future evolution. Specifically, step S1 utilizes a three-dimensional convolutional neural network to process the spatiotemporal grid tensor. Its convolutional kernel slides synchronously in both time and geographic dimensions, capturing the continuous nonlinear evolution patterns of traffic flow parameters (such as speed and density) within local areas like intersections and road segments over short time periods (e.g., convoy start-stop, short-term congestion formation and dissipation). Step S2 abstracts the road network into a dynamically weighted graph and processes it using a graph attention network. This allows the graph attention network to dynamically adjust the attention weights for information aggregation between nodes based on real-time travel costs (e.g., travel time), reflecting the global topological structure of the traffic network. The information in this global topological structure includes, but is not limited to, the impact of congestion, the path and intensity of traffic pressure propagating dynamically along network connections, and so on. Steps S1 and S2 respectively extract two complementary core dimensions of the traffic network: state evolution and structural correlation, providing a data foundation for subsequent fusion and overcoming the problem of the one-sidedness of a single feature dimension.

[0046] Step S3 utilizes mechanisms such as cross-attention to enable deep interaction between features representing local patterns and features representing global correlations. Through the graph model, it learns that when global features indicate high-pressure propagation on a main road (output of step S2), greater attention should be paid to the details of local traffic surges at key upstream diversion nodes (output of step S1). Step S3, based on the outputs of steps S1 and S2, forms a unified network spatiotemporal state code, preventing the separation of local and global information. This code is then processed by a sequence prediction model based on an attention mechanism, enabling the network spatiotemporal state code to adaptively identify historical moments and network nodes that directly impact future predictions—the role of graph attention in step S2 is demonstrated here. This achieves efficient modeling of complex spatiotemporal long-range dependencies and outputs accurate multi-step future network state predictions.

[0047] Finally, step S4 receives the multi-step prediction results from S3, integrates the current real-time state, road physical capacity, and management rules, and iteratively solves the problem in the network scheduling decision optimization model. When calculating the current optimal control command, the feedback impact of these commands on the predicted future state (from step S3) is simultaneously simulated, ensuring that the commands are not only effective in the present but also influence future traffic conditions. Therefore, the integrated scheduling control command set generated in this step is consistent in both time and space. After the commands are issued and executed, the system state changes periodically, triggering round after round of perception-decision loops, repeating the process from step S1 to step S4, forming a complete closed-loop feedback.

[0048] In summary, the optimization solutions provided in steps S1 to S4 significantly improve the overall traffic efficiency and stability of the road network, actively suppress and quickly dissipate local congestion to prevent its global spread, and greatly enhance the responsiveness and multi-means coordination capabilities of the traffic control system in response to complex dynamic changes. They systematically solve problems such as the disconnect between local and global cognition, the separation between prediction and decision-making, as well as low control accuracy, delayed response, and malfunctioning control effects.

[0049] Definition: Dynamic OD matrix: Describes dynamic data on the intention of vehicles to move between different traffic zones, and represents the predicted or real-time travel volume between each origin (O) and destination (D) over time in matrix form.

[0050] Graph Attention Networks: Neural networks that process graph data by aggregating neighbor information through calculating attention weights (rather than fixed weights) between nodes, making them suitable for analyzing dynamically changing traffic networks.

[0051] Dynamically weighted graph: The traffic network is abstracted as a graph, where nodes are intersections / road segments, and the weights (such as travel time) on the connecting edges are dynamically updated according to real-time traffic conditions (such as congestion).

[0052] Dynamic edge weights: In a dynamically weighted graph, each edge is associated with a weight value that changes over time. Dynamic edge weights quantify the real-time state of a connection (such as travel cost) and are the core manifestation of network dynamism.

[0053] Network spatiotemporal state coding: a comprehensive and structured data representation that integrates the spatial state (the situation at each point) and temporal evolution (the trend of change) of the transportation network at a specific moment and in a short historical period.

[0054] Fusion prediction module: A neural network component used to integrate features from different sources or types (such as local and global features) to generate a more comprehensive and consistent feature representation for use in subsequent prediction steps.

[0055] Network scheduling decision optimization models are general or specialized mathematical models or algorithmic frameworks that can comprehensively predict, consider real-time conditions and constraints, and calculate the optimal traffic control instructions (such as signal timing) for a future period of time through iterative updates (rolling optimization).

[0056] Spatiotemporal grid tensor: A three-dimensional data structure formed by dividing traffic data into grids according to geographic space and stacking them in the time dimension. Two dimensions represent spatial location and one dimension represents time series. It is used to structure and store and express the spatiotemporal changes of traffic status.

[0057] Graph model: used to abstractly represent the mathematical structure of transportation networks. Nodes in a graph model represent entities in the transportation network (such as intersections or road segments), and edges represent the connections between entities. The topology of the road network is described by nodes, edges, and their attributes.

[0058] The fusion prediction module integrates features from different feature extraction paths (such as raster features extracted by CNN and graph features extracted by GAT). The fusion prediction module uses features to form a unified spatiotemporal state representation for subsequent prediction through feature concatenation, weighting, or cross-attention.

[0059] Sequence prediction models: Predictive models used to process time series data. Sequence prediction models can predict future states in multiple steps based on historical state sequences. In the transportation field, they are often used to predict future traffic flow, speed and other indicators. Commonly used models include recurrent neural networks, temporal convolutional networks or Transformers.

[0060] Attention mechanism: A computational mechanism that simulates human attention allocation. The attention mechanism dynamically evaluates the importance of different parts of the input data and assigns higher weights to important parts, thereby improving the model's ability to capture key information.

[0061] As shown in Figure 2, preferably, in step S1, the acquisition of spatiotemporal big data of the target traffic network and the construction of dynamic representation information of the traffic network further includes: step S11, based on the fixed sensing devices and floating car trajectory data sources in the target traffic network, acquiring multi-source heterogeneous traffic flow raw data in real time, preprocessing the traffic flow raw data, and formatting it under a preset geographic coordinate system and a unified time reference to obtain standardized spatiotemporal sequence data.

[0062] Step S12: Using a pre-divided geographic rule grid, the floating car trajectory is associated with the traffic network through a map matching algorithm, and the attributes of trajectory points / segments are assigned according to the spatial relationship between the traffic network and the geographic rule grid.

[0063] Step S13: Locate the fixed detector data to the geographic rule grid where it is located. For each geographic rule grid, aggregate and calculate all the allocated data in each preset time slice to generate a five-dimensional spatiotemporal tensor, so that the dimensions of the five-dimensional spatiotemporal tensor correspond to: time step sequence, grid row index, grid column index, and traffic feature channel.

[0064] Step S14: Input the constructed five-dimensional spatiotemporal tensor into the pre-trained three-dimensional convolutional neural network model, and extract features through the three-dimensional convolutional kernel that slides synchronously in the spatial and temporal dimensions of the three-dimensional convolutional neural network model, and output the first network state feature tensor. The first network state feature tensor is used to characterize the spatiotemporal evolution pattern of traffic status in a local area.

[0065] For example, the preprocessing of the original traffic flow data specifically involves spatiotemporal alignment, outlier removal, and missing value imputation of the original traffic flow data.

[0066] The traffic characteristic channel includes traffic flow obtained by aggregating data from fixed detectors, and spatial average speed calculated from floating car trajectory data.

[0067] In the process of feature extraction using three-dimensional convolutional kernels, the three-dimensional convolutional neural network model is downsampled through pooling layers to capture nonlinear spatiotemporal dependencies across grids and time periods layer by layer.

[0068] Definition: Fixed detector data: Traffic flow data collected by sensors installed at fixed locations on the road (such as geomagnetic coils, microwave radar, and cameras). Fixed detector data typically includes flow rate, speed, time occupancy, etc., and is an important data source for traffic condition perception and prediction.

[0069] The optimization techniques provided in steps S11 to S14 construct a standardized spatiotemporal data cube and utilize a three-dimensional convolutional neural network for deep mining to achieve a structured representation of the dynamic evolution of traffic conditions in local areas. Step S11 is the foundational step, addressing issues such as inconsistent original data sources, varying frequencies, and limited accuracy through spatiotemporal alignment, outlier removal, and missing value imputation, thus providing standardized spatiotemporal sequence data for subsequent steps. Step S12 uses a map matching algorithm to associate discrete, linear floating car trajectories with the actual road network and assigns attributes based on the spatial relationship between the grid and the road network, ensuring that trajectory data (such as speed) is correctly assigned to the spatial units it influences. Step S13 directly locates point-like fixed detector data (such as traffic flow) to its corresponding grid. Through the coordinated efforts of steps S12 and S13, the two different geometric forms and physical meanings of the original data—points (detectors) and lines (trajectories)—are uniformly transformed and merged into a contribution value for the opposite side (grid unit), providing a foundation for constructing a unified data structure.

[0070] It is worth noting that step S13 is a pivotal step that connects the preceding and following steps. Step S13 encapsulates the data provided in steps S11 and S12 into a five-dimensional spatiotemporal tensor. The dimensional design of this five-dimensional spatiotemporal tensor (time step, row, column, feature channel) is deeply adapted to the industry standard format of subsequent neural network processing. It can integrate key dimensions such as time evolution, two-dimensional spatial distribution, and multiple traffic features (such as flow rate and spatial average speed) into a mathematical object, forming a data cube containing rich spatiotemporal correlation information. By quantizing and encapsulating the data, the original data is transformed into a feature vector that can be understood by the machine.

[0071] Step S14 inputs the structured tensor generated in step S13 into the pre-trained 3D convolutional neural network. The 3D convolutional neural network utilizes its 3D convolutional kernels to slide across two-dimensional space (grid rows and columns) to capture local spatial patterns in adjacent areas (such as the impact of congestion at one intersection on adjacent road segments), and simultaneously slides across the time dimension to directly capture the state evolution of the same area over consecutive time slices (such as the increase in queue length over time). Through the spatiotemporal synchronous convolutional operation in step S14, the goal of capturing spatiotemporal evolution patterns is achieved. Furthermore, through the downsampling operation of the pooling layers in the network, and through alternating cooperation with the convolutional layers, the perception area is gradually expanded, enabling the model to capture nonlinear spatiotemporal dependencies across grids and time periods, from micro to macro and from short-term to long-term. In this application, the nonlinear spatiotemporal dependency is the macroscopic pattern of morning rush hour congestion radiating outwards from the core area, identified by the 3D convolutional neural network model.

[0072] As shown in Figure 3, preferably, in step S2, the graph model is constructed based on the traffic network topology, where nodes represent intersections or road segments; the weights of each edge in the graph are dynamically calculated according to the real-time traffic flow speed or density to form a dynamically weighted graph; using a graph attention network, the dynamic edge weights are used as the basis for the strength of the association between nodes to aggregate and update the feature information of the nodes in the dynamically weighted graph to generate a second network state feature tensor. This further includes: step S21, constructing the basic skeleton of the graph model based on the topology of the target traffic network. The node set in the basic skeleton includes all signal-controlled intersections and key points of road segments determined according to preset objective rules. The edge set in the node set is used to connect nodes with direct passage relationships.

[0073] Step S22: Based on the real-time acquired traffic flow speed or density data, calculate the weight of each edge in the graph. The weight is calculated based on the real-time travel time or its reciprocal, and the weight update frequency is consistent with the traffic data collection frequency or scheduling decision cycle.

[0074] Step S23: Input the initial feature vectors and weighted adjacency matrices of each node in the dynamically weighted graph into a graph attention network. The initial feature vectors are composed of the real-time state and static attributes of the corresponding nodes.

[0075] In step S24, the element values ​​of the weighted adjacency matrix are generated by the dynamic weights of the corresponding edges. The node features are iteratively aggregated through the multi-head graph attention layer in the graph attention network, so that each node obtains a corresponding updated feature vector. The updated feature vectors of all nodes together constitute the second network state feature tensor.

[0076] As an example, in step S21, the preset objective rules include: road grade, number of lanes, or historical traffic flow threshold, and the key points of the road segment include the road segment endpoints or the road segment center point. The initial feature vector is composed of the real-time state and static attributes of the corresponding node, specifically: the initial feature vector is composed of the real-time state and static attributes of the intersection or road segment.

[0077] The optimization solutions provided in steps S21 to S24 integrate the static topology of the traffic network with its dynamic real-time flow, modeling the road network as a system that accurately reflects dynamic relationships such as congestion propagation. Specifically: Step S21 selects nodes based on preset objective rules (such as road grade and number of lanes), establishing the static basic skeleton of the graph model. This static basic skeleton does not include all road details but focuses primarily on key points at intersections and road segments that have a controlling or critical impact on global traffic flow, ensuring the model's computational efficiency and macroscopic representativeness. Step S22 injects dynamic elements into the static basic skeleton by dynamically calculating the weight of each edge based on real-time traffic flow data (such as speed) using real-time travel time or its reciprocal. Step S22, in conjunction with Step S21, uses edge weights to characterize the instantaneous passage cost and efficiency of connections between nodes, giving the abstract graph connections practical physical meaning. The weight update frequency is synchronized with the system decision cycle, ensuring that the graph model can sense and quantify the dynamic changes in traffic conditions within the road network in real time. Through the coordinated action of steps S21 and S22, the static road network topology map is upgraded into a dynamically weighted map whose edge weights fluctuate in real time according to congestion conditions, laying a data foundation for subsequent steps that includes both static structure and dynamic interactive information.

[0078] Step S23 encodes the real-time status (such as current traffic flow and speed) of a node and its static attributes (such as number of lanes and turning restrictions) into an initial feature vector at the node level. This allows each node to integrate its real-time status and static attributes before entering complex graph computation. Static attributes define the prerequisites for a node, while the real-time status reflects its current health status. The combination of the two ensures the completeness of the node's features.

[0079] Step S24 receives the dynamic edge weights (forming a weighted adjacency matrix) from S22 and the initial node features from step S23. The role of the graph attention network (especially the multi-head graph attention layer) in step S24 is to adaptively allocate aggregation weights based not only on whether the adjacency matrix is ​​connected, but more importantly, on the tightness of the connections represented by the dynamic edge weights, reflecting the travel costs. For example, when an edge experiences a significant increase in travel costs due to congestion (the edge weight increases), the attention mechanism will give more attention or adjust the propagation of information along this edge. Through iterative aggregation across multiple attention layers, the updated feature vector obtained by each node is no longer its initial feature, but rather a second network state feature tensor composed of globally weighted information from dynamic traffic conditions. Therefore, it can transcend local perspectives and characterize the global correlation, propagation path, and influence intensity of states such as congestion pressure and traffic efficiency in the road network.

[0080] In summary, the optimization solutions provided in steps S21 to S24 achieve unified quantitative modeling of static topology and dynamic flow patterns of the traffic network. They can adaptively capture and quantify the dynamic correlation strength in the road network caused by changes in real-time traffic conditions. They also generate structured high-order features that can comprehensively and deeply reflect the global operation status of the road network, especially the congestion propagation and interaction relationships, providing a global perspective for macro-traffic situation assessment and collaborative control.

[0081] As shown in Figure 4, preferably, in step S3, the first network state feature tensor and the second network state feature tensor are input into the fusion prediction module to form a network spatiotemporal state code. The network spatiotemporal state code is then input into an attention-enhanced sequence prediction model. This sequence prediction model has a built-in attention mechanism, which outputs traffic state prediction results for each node in the road network over multiple future time periods. Further, step S31 involves inputting the first and second network state feature tensors into a fusion prediction module. The fusion prediction module integrates the input tensors to generate a unified network spatiotemporal state code. In this step, the fusion prediction module uses one or more of the following mechanisms to fuse the input tensors: feature concatenation, weighted summation, or cross-attention.

[0082] Step S32 involves inputting the unified network spatiotemporal state code into an attention-enhanced sequence prediction model. This model, through its built-in multi-head self-attention mechanism, obtains the long-range spatiotemporal dependencies across nodes and across time in the unified network spatiotemporal state code. In this step, the sequence prediction model employs an encoder-decoder architecture.

[0083] Step S33: The unified network spatiotemporal state code is decoded by the decoder in the sequence prediction model to obtain the multi-channel prediction results corresponding to each node of the road network in multiple consecutive time periods in the future. The multi-channel prediction results correspond to the predicted traffic flow, predicted speed and predicted congestion index of the future traffic state, respectively.

[0084] The optimization techniques provided in steps S31 to S33 capture complex spatiotemporal dependencies by fusing local and global features, achieving high-precision multi-indicator prediction of future road network traffic conditions. Step S31 serves to complement and unify information, forming a fusion hub that takes heterogeneous information representing local spatiotemporal evolution patterns (first network state feature tensor) and global correlation propagation trends (second network state feature tensor) as input. The fusion prediction module integrates features using feature concatenation, weighted summation, or cross-attention mechanisms, learning the complementary relationship between the two types of features. For example, through the cross-attention mechanism, the model can use global features to query which local areas have the most critical detailed changes under the current global congestion propagation trend, and accordingly perform targeted weighting and fusion of local features. Step S31 fuses the details of the feature tensor with the associated nodes on the map into a complete network spatiotemporal state code, eliminating the separation between local and global information at the prediction entry point and providing high-quality integrated input for subsequent modeling.

[0085] Step S32 involves inputting the unified encoding output from S31 into an attention-enhanced sequence prediction model based on an encoder-decoder architecture. The model incorporates a multi-head self-attention mechanism, which allows multiple attention heads to focus on different aspects of the encoding dependencies in different subspaces. For example, one head can focus on uncovering long-term patterns across time (such as the duration of morning rush hour congestion), while another head focuses on identifying key cross-space associations (such as the decisive impact of congestion on a major upstream traffic artery on the future speed of a downstream area). Applying the multi-head self-attention mechanism in step S32 enables the sequence model to adaptively capture the long-range spatiotemporal dependencies hidden in the unified network's spatiotemporal state encoding. Compared to traditional models like Recurrent Neural Networks (RNNs), which struggle to effectively handle such non-local, bidirectional complex dependencies and cannot transform the fused static encoding into a deep understanding of dynamic patterns, this approach is crucial.

[0086] Step S33 transforms the mined dependencies into specific future states. Utilizing the spatiotemporal dependencies understood by the encoder in step S32, step S33 uses the decoder to progressively generate prediction sequences for multiple consecutive future time periods in an autoregressive or parallel manner. When generating the state of each future moment and each node, the decoder dynamically references all contextual information provided by the encoder (i.e., the dependencies mined in step S32) and the already generated partial future sequences. The final output is a multi-channel prediction result (predicted traffic flow, speed, and congestion index). Step S33, through its collaborative calculation with steps S31 and S32, ensures that the predictions for each channel are not independent but jointly derived under the constraints of a unified spatiotemporal dependency model. This ensures the consistency between the prediction results and actual traffic conditions. The output results, such as predicted high traffic flow and low speed, and high congestion index, are logically self-consistent.

[0087] As shown in Figure 5, preferably, in step S4, the step of inputting the traffic state prediction results, the current real-time road network status, the preset scheduling rules, and the road capacity into a network scheduling decision optimization model to generate a comprehensive scheduling control instruction set, and then issuing the comprehensive scheduling control instruction set to the traffic signal control system to perform real-time regulation of the traffic network, further includes: step S41, constructing a network scheduling decision optimization model based on the traffic state prediction results, the current real-time road network status, the dynamic OD matrix, and the road capacity data. For example, the traffic state prediction results can reveal the future multi-period road network operation status, the current real-time road network status, the dynamic OD matrix and path selection preferences generated by mining historical and real-time floating car trajectory data, preset scheduling rules, and road capacity data, etc.

[0088] Step S42: The network scheduling decision optimization model has a built-in model predictive control framework. In the model predictive control framework, the optimal control command sequence for the next few cycles is solved with the current state as the initial point.

[0089] Step S43: Based on the real-time control instructions output by the network scheduling decision optimization model, generate a comprehensive scheduling control instruction set, which includes dynamic signal timing schemes, variable lane control instructions, and route driving suggestions. Send the comprehensive scheduling control instruction set to the traffic signal control system in real time to execute the corresponding traffic control operations.

[0090] As an example, in step S41, the dynamic OD matrix specifically refers to: Step S411, the vehicle's intention to move between regions is derived based on historical and real-time floating car trajectories. The vehicle's intention to move between regions refers to a pair of key-value pairs, such as a numerical pair from the starting point to the destination.

[0091] Step S412, the step of solving the optimal control command sequence for several future cycles with the current state as the initial point, specifically involves: taking the current state as the initial point, and based on the traffic state prediction results, iteratively solving the optimal control command sequence for future cycles under the conditions of satisfying signal, lane, and capacity constraints.

[0092] Step S413, the integrated dispatch control instruction set includes: traffic light timing scheme, variable lane control instruction, and route suggestion combination.

[0093] Step S414, the traffic signal control system includes electronic traffic signs, vehicle-mounted traffic platforms, or mobile terminal traffic platforms.

[0094] The optimization solutions provided in steps S41 to S43 combine traffic state prediction with dynamic travel demand, generating collaborative control instructions within the optimization framework. This achieves a shift from passively responding to traffic conditions to actively guiding traffic regulation. Specifically, step S41 integrates four types of information with different spatiotemporal scales and properties: traffic state prediction results (providing a view of the system state evolution over multiple future time periods), the current real-time road network state (providing an immediate starting point for decision-making), the dynamic OD matrix (representing the dynamic travel demand structure over future time periods, such as the travel destinations of vehicles), and road capacity data (defining the physical hard constraints of system operation). In particular, the introduction of the dynamic OD matrix (as described in step S411, derived from the start-end pairs mined by floating car trajectories) enables the model to not only know the current traffic conditions but also how vehicles intend to travel. This allows the model to proactively consider the potential impact of control measures on driver route selection during the optimization process, avoiding deviations in traffic control effects caused by ignoring feedback from the current traffic state.

[0095] Step S42 employs a model predictive control framework as the model's solution engine. Starting with the current problem defined in step S41, within a forward-extending finite time window (several future cycles), signal timing, lane function, etc., are used as decision variables. With the goal of minimizing total system delay, and under constraints such as road capacity, an optimal forward-looking control command sequence (e.g., a signal scheme at fixed intervals within a certain future cycle) is iteratively updated and solved. The key feature of step S42 is iteration and feedback. The model executes only the first set of commands corresponding to the current moment in the sequence. At the next decision moment, the system acquires the latest actual state (integrating the effects of the previous control round and new random disturbances), and immediately uses this new state as the starting point to re-predict and optimize, initiating a new round of solution. This closed-loop control, capable of iterative updates within a finite time, gives the system strong anti-interference and adaptive capabilities, allowing for dynamic strategy adjustments based on actual conditions.

[0096] Step S43 transforms the abstract optimization solution (real-time control command) output in S42 into an executable set of comprehensive scheduling and control commands. This set includes not only dynamic signal timing schemes (time resource allocation), but also variable lane control commands (spatial resource reorganization) and route suggestions (information guidance to influence demand distribution). As shown in step S413, the various commands in the command set are not generated in isolation, but are jointly solved within the network scheduling decision optimization model provided in step S42, ensuring the consistency and complementarity of the command set in spatiotemporal logic (for example, while closing a lane, signal timing and route guidance are simultaneously adjusted to divert traffic). Subsequently, the commands generated in step S413 are transmitted in real-time to the traffic signal control system via the communication network, achieving coordinated guidance and real-time control of traffic flow in three dimensions: time, space, and route selection.

[0097] For the method steps disclosed in the above embodiments, the method steps are described as a series of actions for the purpose of simplicity. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily necessary for the embodiments of this application.

[0098] As shown in Figure 6, this application embodiment also provides a traffic network scheduling and control system based on spatiotemporal big data, used to implement the traffic network scheduling and control method based on spatiotemporal big data described in any specific implementation of this application embodiment, including: a first network state feature tensor generation module, which acquires spatiotemporal big data of the target traffic network, constructs dynamic representation information of the traffic network, maps the dynamic representation information of the traffic network into a spatiotemporal grid tensor according to the geographical location, and extracts the spatiotemporal grid tensor using a three-dimensional convolutional neural network to generate a first network state feature tensor.

[0099] The second network state feature tensor generation module constructs a graph model based on the traffic network topology, where nodes represent intersections or road segments. It dynamically calculates the weights of each edge in the graph according to the real-time traffic flow speed or density, forming a dynamically weighted graph. Using a graph attention network, the module aggregates and updates the feature information of the nodes in the dynamically weighted graph, using the dynamic edge weights as the basis for the strength of the association between nodes, to generate the second network state feature tensor. The second network state feature tensor is used to reflect the global association and congestion propagation trend of the road network.

[0100] The traffic state prediction output module inputs the first network state feature tensor and the second network state feature tensor into the fusion prediction module to form a network spatiotemporal state code. The network spatiotemporal state code is then input into an attention-enhanced sequence prediction model. The sequence prediction model has a built-in attention mechanism and outputs the traffic state prediction results of each node in the road network for multiple future time periods through the attention mechanism. The traffic state prediction results include flow rate, speed, and congestion index.

[0101] The real-time traffic network control module inputs the traffic state prediction results, the current real-time road network status, the preset scheduling rules, and the road capacity into a network scheduling decision optimization model, generates a comprehensive scheduling control instruction set, and sends the comprehensive scheduling control instruction set to the traffic signal control system to perform real-time control of the traffic network.

[0102] The implementation methods of the system described above are merely illustrative. For example, the various functional modules, units, or subsystems within the system may or may not be physically separate, or they may or may not be physical units; that is, they may be located in the same place or distributed across multiple different systems and their subsystems or modules. Those skilled in the art can select some or all of the functional modules, units, or subsystems to achieve the objectives of the embodiments of this application according to actual needs. Those skilled in the art can understand and implement the above-described situations without any creative effort.

[0103] As shown in Figure 7, this application embodiment, in addition to providing a traffic network scheduling and control method and system based on spatiotemporal big data, also provides a corresponding electronic device: an electronic device, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method.

[0104] Explanation of reference numerals in the attached drawings: Electronic device 500, External device 514, Processor 516, Bus 518, Network adapter 520, I / O interface 522, Display device 524, Memory 528, RAM 530, Cache 532, Storage system 534, Program / Utility 540, Program module 542.

[0105] The electronic device shown in Figure 7 is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this application. Typically, this electronic device could be a device within an electronic product based on the call file transfer verification method described above. The electronic device 500 is manifested as a general-purpose computing device. Components of the electronic device 500 may include, but are not limited to: one or more processing units or processors 516, a memory 528, and a bus 518 connecting different system components (including the memory 528 and the processor 516). The bus 518 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus. The electronic device 500 typically includes various computer system readable media. These media can be any available media accessible to the electronic device 500, including volatile and non-volatile media, removable and non-removable media. Memory 528 may include computer system readable media in the form of volatile memory, such as RAM 530 and / or cache 532. Electronic device 500 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 534 may be used to read and write non-removable, non-volatile magnetic media (not shown in the figure, commonly referred to as "hard disk drives"). Although not shown in the figure, storage system 534 may provide disk drives for reading and writing to removable non-volatile disks (e.g., floppy disks, portable hard disks, hot-swappable storage media) and optical disk drives for reading and writing to removable non-volatile optical disks (e.g., CD-ROMs, DVD-ROMs, or other optical media). In these cases, each drive may be connected to bus 518 via one or more data media interfaces. Memory 528 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of various specific embodiments of the present application. A program / utility 540 having a set (at least one) of program modules 542 may be stored, for example, in memory 528. Such program modules 542 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 542 typically perform the functions and / or methods described in the embodiments of this application.Electronic device 500 can also communicate with one or more external devices 514 (e.g., keyboard, pointing device, display device 524, etc.), and with one or more devices that enable a user to interact with electronic device 500, and / or with any device that enables electronic device 500 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed through I / O interface 522. Furthermore, electronic device 500 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 520. Network adapter 520 communicates with other modules of electronic device 500 via bus 518. It should be understood that, although not shown in the figures, those skilled in the art can use other hardware and / or software modules in conjunction with electronic device 500, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems. Processor 516 executes various functional applications and data processing by running programs stored in memory 528, such as implementing the methods provided in any one or more embodiments of this application.

[0106] In the description of the embodiments of this application, the reference to terms such as "an embodiment," "example," "specific example," etc., means that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the embodiments of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0107] Furthermore, the technical solutions of the various implementation methods in this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the embodiments of this application.

[0108] All features disclosed in the embodiments of this application, or all steps in the disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps. Any feature disclosed in the specification of the embodiments of this application, unless specifically stated otherwise, may be replaced by other equivalent or similar alternative features. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features. Throughout the specification, the same reference numerals indicate the same elements.

[0109] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification of embodiments (including the corresponding claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification of embodiments (including the corresponding claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of this application, and are not intended to limit them. Although the embodiments of this application have been described in detail with reference to the foregoing specific embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing specific embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the specific embodiments of this application.

Claims

1. A traffic network scheduling and control method based on spatiotemporal big data, characterized in that, include: The process involves acquiring spatiotemporal big data of the target traffic network, constructing dynamic representation information of the traffic network, mapping this dynamic representation information to spatiotemporal grid tensors according to geographical location, and extracting the spatiotemporal grid tensors using a 3D convolutional neural network to generate a first network state feature tensor. A graph model is constructed based on the traffic network topology, where nodes represent intersections or road segments. The weights of each edge in the graph are dynamically calculated based on real-time traffic flow speed or density to form a dynamically weighted graph. A graph attention network is used, with the dynamic edge weights serving as the basis for the strength of association between nodes, to aggregate and update the feature information of nodes in the dynamically weighted graph, generating a second network state feature tensor. This second network state feature tensor reflects the global association and congestion propagation trend of the road network. The first and second network state feature tensors are input into the fusion prediction module to form a network spatiotemporal state code. The network spatiotemporal state code is then input into an attention-enhanced sequence prediction model. The sequence prediction model has a built-in attention mechanism, which outputs traffic state prediction results for each node of the road network in multiple future time periods. The traffic state prediction results include flow rate, speed, and congestion index. The traffic state prediction results, the current real-time road network state, preset scheduling rules, and road capacity are then input into a network scheduling decision optimization model to generate a comprehensive scheduling control instruction set. The comprehensive scheduling control instruction set is then sent to the traffic signal control system to perform real-time regulation of the traffic network.

2. The traffic network scheduling and control method based on spatiotemporal big data according to claim 1, characterized in that, The acquisition of spatiotemporal big data of the target traffic network and the construction of dynamic representation information of the traffic network further include: acquiring multi-source heterogeneous raw traffic flow data in real time based on fixed sensing devices and floating car trajectory data sources in the target traffic network; preprocessing the raw traffic flow data and formatting it under a preset geographic coordinate system and a unified time reference to obtain standardized spatiotemporal sequence data; using a pre-divided geographic regular grid, associating the floating car trajectory with the traffic network through a map matching algorithm, and assigning attributes to trajectory points / segments according to the spatial relationship between the traffic network and the geographic regular grid; and locating the fixed detector data to its... Within each preset time slice of the geographic regular grid, all allocated data for each geographic regular grid are aggregated and calculated to generate a five-dimensional spatiotemporal tensor. The dimensions of the five-dimensional spatiotemporal tensor correspond to: time step sequence, grid row index, grid column index, and traffic feature channel, respectively. The constructed five-dimensional spatiotemporal tensor is input into a pre-trained three-dimensional convolutional neural network model. Feature extraction is performed by the three-dimensional convolutional kernels that slide synchronously in the spatial and temporal dimensions of the three-dimensional convolutional neural network model, and the first network state feature tensor is output. The first network state feature tensor is used to characterize the spatiotemporal evolution pattern of traffic status in a local area.

3. The traffic network scheduling and control method based on spatiotemporal big data according to claim 2, characterized in that, The preprocessing of the raw traffic flow data specifically involves spatiotemporal alignment, outlier removal, and missing value imputation. The traffic feature channels include traffic flow obtained by aggregating data from fixed detectors and spatial average speed calculated from floating car trajectory data. In the process of feature extraction using three-dimensional convolutional kernels, the three-dimensional convolutional neural network model is downsampled through pooling layers to capture nonlinear spatiotemporal dependencies across grids and time periods layer by layer.

4. The traffic network scheduling and control method based on spatiotemporal big data according to claim 1, characterized in that, The method involves constructing a graph model based on the traffic network topology, where nodes represent intersections or road segments. The weights of each edge in the graph are dynamically calculated based on real-time traffic flow speed or density, forming a dynamically weighted graph. A graph attention network is used, with the dynamic edge weights serving as the basis for the strength of association between nodes, to aggregate and update the feature information of nodes in the dynamically weighted graph, generating a second network state feature tensor. Further, the method includes: constructing the basic skeleton of the graph model based on the topology of the target traffic network. The node set in the basic skeleton includes all signal-controlled intersections and key points of road segments determined according to preset objective rules. The edge set in the node set is used to connect nodes with direct passage relationships. Based on real-time acquired traffic flow speed... The graph uses degree or density data to calculate the weight of each edge. The weight is calculated based on the real-time travel time or its reciprocal, and the weight update frequency is consistent with the traffic data collection frequency or scheduling decision cycle. The initial feature vectors and weighted adjacency matrices of each node in the dynamically weighted graph are input into a graph attention network. The initial feature vectors are composed of the real-time state and static attributes of the corresponding nodes. The element values ​​of the weighted adjacency matrix are generated by the dynamic weights of the corresponding edges. The node features are iteratively aggregated through the multi-head graph attention layer in the graph attention network, so that each node obtains a corresponding updated feature vector. The updated feature vectors of all nodes together constitute the second network state feature tensor.

5. The traffic network scheduling and control method based on spatiotemporal big data according to claim 4, characterized in that, The preset objective rules include: road grade, number of lanes or historical traffic flow threshold; the key points of the road segment include the road segment endpoints or the road segment center point; the initial feature vector is composed of the real-time status and static attributes of the corresponding node, specifically: the initial feature vector is composed of the real-time status and static attributes of the intersection or road segment.

6. The traffic network scheduling and control method based on spatiotemporal big data according to claim 1, characterized in that, The process of inputting the first and second network state feature tensors into a fusion prediction module to form a network spatiotemporal state code, and then inputting the network spatiotemporal state code into an attention-enhanced sequence prediction model, which has a built-in attention mechanism, to output traffic state prediction results for each node of the road network in multiple future time periods, further includes: inputting the first and second network state feature tensors into a fusion prediction module to integrate the input tensors and generate a unified network spatiotemporal state code; inputting the unified network spatiotemporal state code into an attention-enhanced sequence prediction model, which uses its built-in multi-head self-attention mechanism to obtain the long-range spatiotemporal dependencies across nodes and across time in the unified network spatiotemporal state code; and decoding the unified network spatiotemporal state code through a decoder in the sequence prediction model to obtain multi-channel prediction results for each node of the road network in multiple consecutive future time periods, wherein the multi-channel prediction results correspond to the predicted traffic flow, predicted speed, and predicted congestion index of the future traffic state, respectively.

7. The traffic network scheduling and control method based on spatiotemporal big data according to claim 1, characterized in that, The process of inputting the traffic state prediction results, the current real-time road network status, preset scheduling rules, and road capacity into a network scheduling decision optimization model to generate a comprehensive scheduling control instruction set, and then issuing the comprehensive scheduling control instruction set to the traffic signal control system to perform real-time regulation of the traffic network, further includes: constructing a network scheduling decision optimization model based on the traffic state prediction results, the current real-time road network status, the dynamic OD matrix, and road capacity data; the network scheduling decision optimization model has a built-in model predictive control framework, in which the optimal control instruction sequence for several future cycles is solved with the current state as the initial point; generating a comprehensive scheduling control instruction set based on the immediate control instructions output by the network scheduling decision optimization model, the comprehensive scheduling control instruction set including dynamic signal timing schemes, variable lane control instructions, and route driving suggestions, and issuing the comprehensive scheduling control instruction set to the traffic signal control system in real time to execute corresponding traffic regulation operations.

8. The traffic network scheduling and control method based on spatiotemporal big data according to claim 7, characterized in that, The dynamic OD matrix is ​​specifically the vehicle's intention to move between regions based on historical and real-time floating car trajectories; the step of solving for the optimal control command sequence for several future cycles with the current state as the initial point is specifically: with the current state as the initial point, based on the traffic state prediction results, and under the condition of satisfying signal, lane and capacity constraints, iteratively solving for the optimal control command sequence for future cycles. The integrated dispatch and control instruction set includes: traffic light timing scheme, variable lane control instruction, and route suggestion combination; the traffic signal control system includes electronic traffic signs, vehicle-mounted traffic platform, or mobile terminal traffic platform.

9. A traffic network scheduling and control system based on spatiotemporal big data, used to implement the traffic network scheduling and control method based on spatiotemporal big data as described in any one of claims 1 to 8, characterized in that, include: The first network state feature tensor generation module acquires spatiotemporal big data of the target traffic network, constructs dynamic representation information of the traffic network, maps the dynamic representation information of the traffic network into a spatiotemporal grid tensor according to geographical location, and extracts the spatiotemporal grid tensor using a three-dimensional convolutional neural network to generate the first network state feature tensor. The second network state feature tensor generation module constructs a graph model based on the traffic network topology, where nodes represent intersections or road segments. It dynamically calculates the weights of each edge in the graph based on real-time traffic flow speed or density, forming a dynamically weighted graph. Using a graph attention network, with the dynamic edge weights as the basis for the strength of association between nodes, it aggregates and updates the feature information of nodes in the dynamically weighted graph, generating the second network state feature tensor. This second network state feature tensor reflects the global association and congestion propagation trend of the road network. The traffic state prediction result output module inputs the first and second network state feature tensors into the fusion prediction module to form the network state prediction result. The system employs an empty-state encoding method, which inputs the network spatiotemporal state encoding into an attention-enhanced sequence prediction model. This model incorporates an attention mechanism to output traffic state predictions for each node in the road network over multiple future time periods. These predictions include traffic flow, speed, and congestion index. A real-time traffic network control module inputs the traffic state predictions, the current real-time road network status, preset scheduling rules, and road capacity into a network scheduling decision optimization model. This model generates a comprehensive scheduling control instruction set, which is then sent to the traffic signal control system to execute real-time control of the traffic network.

10. An electronic device, characterized in that, include: The system includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; the memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the method according to any one of claims 1 to 8.

Images