Expressway traffic control strategy self-evolution method, device and electronic equipment

Abstract

The application provides a highway traffic control strategy self-evolution method and device, electronic equipment and storage medium, belonging to the technical field of intelligent transportation. The method comprises the following steps: obtaining multi-source traffic data of a highway road network and inputting a pre-trained congestion prediction model to obtain congestion prediction results of each road section; generating a scene semantic representation according to the prediction results and a road network topology structure and converting the scene semantic representation into a structured prompt text; inputting the prompt text into a generative large language model to obtain a natural language control strategy output by the generative large language model, analyzing and extracting control parameters to generate a structured traffic control instruction; inputting the structured instruction into a traffic simulation environment for execution, and iteratively optimizing the congestion prediction model and / or strategy generation according to the simulation results by using a reinforcement learning algorithm. The application can improve the intelligent decision-making level and response efficiency of highway traffic control.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to a method, apparatus, and electronic equipment for the self-evolution of highway traffic control strategies. Background Technology

[0002] Highway traffic management technology aims to alleviate traffic congestion and improve traffic efficiency by real-time sensing of road network operation status, accurate prediction of traffic flow evolution trends, and the formulation of dynamic control strategies (such as dynamic speed limits, lane control, and ramp control). Traditional control methods rely on fixed monitoring equipment to collect traffic parameters and make decisions based on statistical analysis and human experience rules. In recent years, traffic prediction methods based on models such as graph neural networks have been able to learn the spatiotemporal dependencies of the road network from historical data and predict future traffic conditions; deep reinforcement learning has also been explored for generating adaptive control strategies, learning optimal decisions in dynamic scenarios through interaction between intelligent agents and the traffic environment.

[0003] In the process of developing this invention, the inventors discovered that existing methods have at least the following drawbacks: First, traffic prediction and control strategy generation are disconnected. The prediction model only outputs traffic state values ​​and fails to transform the prediction results into semantic scene representations that can be directly used by the decision-making system. The formulation of control strategies still requires manual semantic conversion from "data readings" to "strategy judgments," resulting in low decision-making efficiency. Second, existing reinforcement learning methods typically train policy networks from scratch, and the action space is limited to a single type of control action, making it difficult to output multi-type collaborative control instructions. The high-dimensional mixed action space makes training convergence difficult and hinders stable deployment in complex traffic scenarios. Third, existing technologies lack a complete mechanism to automatically convert the natural language policy descriptions output by generative artificial intelligence into executable structured instructions, and there is a lack of effective automated detection and resolution methods for potential spatiotemporal conflicts between multiple policy instructions.

[0004] Therefore, how to improve the intelligence level and response efficiency of highway traffic management from prediction to decision-making is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] In view of this, it is necessary to provide a method, device and electronic equipment for the self-evolution of highway traffic control strategies, so as to solve the technical problems of low level of traffic control intelligence and poor response efficiency caused by the separation of traffic prediction and control strategy generation in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a self-evolving method for highway traffic control strategies, comprising: Multi-source traffic data of the highway network is acquired and input into a pre-trained congestion prediction model to obtain congestion prediction results for each section of the highway. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data and historical congestion pattern data. Based on the congestion prediction results and the road network topology of the highway network, a scene semantic representation of the highway traffic status is generated, and the scene semantic representation is converted into structured prompt text. The structured prompt text is input into the generative large language model to obtain the natural language control strategy output by the generative large language model. The natural language control strategy is then parsed to extract the control parameters in the natural language control strategy. Based on the control parameters, a structured traffic control instruction is generated. The structured traffic control instructions are input into a traffic simulation environment for simulation execution, and the simulation execution results are obtained. Based on the simulation execution results, a reinforcement learning algorithm is used to iteratively optimize the congestion prediction model and / or the generation strategy of the structured traffic control instructions.

[0007] In one possible implementation, the self-evolutionary method for highway traffic control strategies includes the following steps before inputting the multi-source traffic data into a pre-trained congestion prediction model: A pre-trained large language model is used as the teacher network, and a graph convolutional neural network to be trained is used as the student network. Historical traffic data is input into the teacher network and the student network respectively, and a first prediction result output by the teacher network and a second prediction result output by the student network are obtained. Based on the first prediction result and the second prediction result, a multi-objective distillation loss is calculated, wherein the multi-objective distillation loss includes at least two of congestion prediction loss, feature representation loss and inference logic loss. Based on the multi-objective distillation loss, the parameters of the student network are updated by minimizing the inverse KL divergence until the training stopping condition is met, thus obtaining the pre-trained congestion prediction model.

[0008] In one possible implementation, the multi-objective distillation loss includes the congestion prediction loss, the feature representation loss, and the inference logic loss; The step of calculating the multi-objective distillation loss based on the first prediction result and the second prediction result includes: The congestion prediction loss is determined based on the difference between the first prediction result and the second prediction result in the numerical value of congestion status. The feature representation loss is determined based on the difference between the semantic features output by the teacher network and the spatiotemporal features output by the student network. The inference logic loss is determined based on the difference between the inference logic distribution output by the teacher network and the inference logic distribution output by the student network. The congestion prediction loss, the feature representation loss, and the inference logic loss are weighted and summed to obtain the multi-objective distillation loss.

[0009] In one possible implementation, updating the parameters of the student network further includes: Obtain congestion status labels at various spatiotemporal scales from the training sample set; A spatiotemporal pyramid structure is constructed, and features are extracted from the multi-source traffic data in the training samples at multiple preset time granularities and multiple preset spatial granularities to obtain traffic features at multiple spatiotemporal scales. The traffic characteristics at multiple spatiotemporal scales are input into the student network to predict congestion, and congestion prediction results at each spatiotemporal scale are obtained. Based on the consistency between the congestion prediction results at each spatiotemporal scale and the corresponding congestion status labels, the consistency loss is calculated. The parameters of the student network are jointly updated based on the weighted sum of the consistency loss and the multi-objective distillation loss.

[0010] In one possible implementation, generating a scene semantic representation of the highway traffic state based on the congestion prediction result and the highway network topology includes: The highway network is modeled as a directed graph, where nodes represent road segments and edges represent the connections between road segments. Based on the congestion prediction results, feature vectors for each node are determined, and the feature vectors include traffic state information, traffic density information, and average speed information. Feature extraction is performed on the directed graph using a graph convolutional network to obtain high-dimensional features of each node; The association strength between nodes is calculated using a self-attention mechanism, and the scene semantic representation is generated based on the high-dimensional features of each node and the association strength between them.

[0011] In one possible implementation, converting the scene semantic representation into structured prompt text includes: Obtain a preset prompt template, which includes a location placeholder, a congestion level placeholder, a traffic density placeholder, an average speed placeholder, a prediction time placeholder, and a congestion trend placeholder. Extract current location information, current congestion level information, current traffic density information, current average speed information, prediction time information, and congestion trend information from the scene semantic representation; The extracted current location information, current congestion level information, current traffic density information, current average speed information, predicted time information, and congestion trend information are respectively filled into the corresponding placeholders in the prompt template to generate the structured prompt text.

[0012] In one possible implementation, the natural language control strategy for obtaining the output of the generative large language model includes: The structured prompt text and the thought chain reasoning prompt mark are input into the generative large language model. The thought chain reasoning prompt mark is used to guide the generative large language model to perform step-by-step reasoning in four stages: scenario analysis, trend prediction, strategy generation, and effect evaluation. Obtain the natural language control strategy output by the generative large language model after inference in the four stages; The natural language control strategy is calibrated using a pre-trained traffic knowledge enhancement module.

[0013] In one possible implementation, parsing the natural language control strategy, extracting control parameters from the natural language control strategy, and generating structured traffic control instructions based on the control parameters includes: The natural language control strategy is sequence labeled using a bidirectional long short-term memory network-conditional random field policy parser to identify the operation type field, target region field, parameter value field, and time window field in the natural language control strategy. The control parameters are determined based on the identified operation type field, target area field, parameter value field, and time window field. Based on the control parameters, the natural language control strategy is mapped into structured traffic control instructions, the instruction types of which include at least one of dynamic speed limit instructions, lane control instructions, ramp control instructions, and variable information sign instructions.

[0014] On the other hand, the present invention also provides a self-evolving device for highway traffic control strategies, comprising: The road congestion prediction module is used to acquire multi-source traffic data of the highway network and input the multi-source traffic data into a pre-trained congestion prediction model to obtain congestion prediction results for each road segment of the highway. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data and historical congestion pattern data. The semantic representation generation module is used to generate a scene semantic representation of the highway traffic state based on the congestion prediction results and the road network topology of the highway network, and convert the scene semantic representation into structured prompt text; The traffic control instruction generation module is used to input the structured prompt text into the generative large language model, obtain the natural language control strategy output by the generative large language model, parse the natural language control strategy, extract the control parameters in the natural language control strategy, and generate structured traffic control instructions based on the control parameters. The traffic control optimization iteration module is used to input the structured traffic control instructions into the traffic simulation environment for simulation execution, obtain the simulation execution results, and iteratively optimize the congestion prediction model and / or the generation strategy of the structured traffic control instructions using reinforcement learning algorithms based on the simulation execution results.

[0015] Thirdly, the present invention also provides an electronic device, including a data acquisition interface, a memory, and a processor; The data acquisition interface is used to connect to the roadside sensing device of the highway and receive multi-source traffic data collected by the sensing device. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data and historical congestion pattern data. The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the self-evolution method of highway traffic control strategy described in any of the above implementations.

[0016] The beneficial effects of this invention are as follows: The self-evolutionary method for highway traffic control strategies provided by this invention obtains congestion prediction results for each road segment by acquiring multi-source traffic data and inputting it into a pre-trained congestion prediction model, providing a precise spatiotemporal data foundation for subsequent strategy generation; it generates scene semantic representations based on the prediction results and road network topology and converts them into structured prompt text, upgrading the numerical prediction results to a semantic scene description that can be understood by a large language model, bridging the semantic gap between prediction and decision-making; it inputs the prompt text into a large language model to obtain natural language control strategies, and parses and extracts control parameters to generate structured traffic control instructions, transforming the reasoning ability of the large model into directly executable control actions, overcoming the problem that natural language output cannot be directly parsed by downstream systems; it inputs the structured instructions into a traffic simulation environment for execution, and uses reinforcement learning algorithms to iteratively optimize the congestion prediction model and / or strategy generation based on the simulation results, enabling the prediction model and strategy generation model to be continuously jointly optimized in simulation feedback, effectively improving the intelligent decision-making level and the accuracy of strategy generation for highway traffic control. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic flowchart of an embodiment of the self-evolution method for highway traffic control strategy provided by the present invention; Figure 2 A schematic flowchart of another embodiment of the self-evolution method for highway traffic control strategy provided by the present invention; Figure 3 For the present invention Figure 2 A schematic diagram of an embodiment of S203; Figure 4 For the present invention Figure 1 A schematic diagram of an embodiment of S102; Figure 5 For the present invention Figure 1 A schematic diagram of another embodiment of S102; Figure 6 For the present invention Figure 1 A schematic diagram of an embodiment of S103; Figure 7 For the present invention Figure 1 A schematic diagram of another embodiment of S103; Figure 8 A schematic diagram of an embodiment of the self-evolution device for highway traffic control strategies provided by the present invention; Figure 9 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0019] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0020] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] This invention provides a method, apparatus, and electronic device for the self-evolution of highway traffic control strategies. The technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Figure 1 This is a schematic flowchart of an embodiment of the self-evolution method for highway traffic control strategies provided by the present invention, as shown below. Figure 1 As shown, the self-evolution method for highway traffic control strategies includes: S101. Obtain multi-source traffic data of the highway network and input the multi-source traffic data into a pre-trained congestion prediction model to obtain congestion prediction results for each section of the highway. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data, and historical congestion pattern data.

[0024] Specifically, multi-source traffic data is collected through various sensing devices deployed on the roadside of highways, including but not limited to video image data collected by video surveillance equipment, vehicle trajectory data collected by microwave radar or millimeter-wave radar, meteorological data collected by meteorological sensors, ETC transaction data and path identification data collected by ETC gantry system, and historical congestion pattern data obtained from roadside units or cloud servers.

[0025] Traffic flow data reflects the number of vehicles passing through each road segment per unit time; vehicle speed data reflects the average speed and speed distribution of vehicles on each road segment; meteorological data reflects the impact of weather conditions such as rainfall, fog, and snowfall on traffic conditions; and historical congestion pattern data reflects the periodic congestion patterns of road segments under specific time periods or conditions. The congestion prediction model is a lightweight model pre-trained through knowledge distillation. This model can output the congestion status prediction results of each road segment within a preset future time window based on the input multi-source traffic data, including the predicted congestion level, congestion duration, and congestion propagation range.

[0026] S102. Based on the congestion prediction results and the road network topology of the highway network, generate a scene semantic representation of the highway traffic status, and convert the scene semantic representation into structured prompt text.

[0027] Specifically, the road network topology of a highway network describes the spatial connectivity and directional constraints between road segments. Based on congestion prediction results and the road network topology, the process of generating a scene semantic representation essentially involves structured modeling of highway traffic conditions. This transforms the predicted numerical congestion information into a graph-structured representation containing multi-dimensional semantic information such as location, severity, density, and speed. This representation can be effectively understood and utilized by subsequent processing stages. Converting the scene semantic representation into structured prompt text involves transforming the graph-structured scene semantic information into a text description that conforms to natural language expression habits and contains complete context, so that it can be input into a generative large language model for understanding and reasoning.

[0028] S103. Input the structured prompt text into the generative large language model, obtain the natural language control strategy output by the generative large language model, parse the natural language control strategy, extract the control parameters in the natural language control strategy, and generate structured traffic control instructions based on the control parameters.

[0029] Specifically, the generative large language model is a language model pre-trained on a large-scale text corpus, possessing the ability to understand complex scenarios and perform multi-step reasoning. After inputting structured prompt text into the model, it infers based on the traffic scenario described in the prompt text and outputs a traffic control strategy described in natural language. Since the natural language description contains specific control parameters required for execution, such as the type of control operation, the affected road segment range, specific numerical parameters (e.g., speed limits, ramp signal cycles), and execution time windows, the output natural language strategy needs to be parsed to extract the structured control parameters. Based on the extracted control parameters, a structured traffic control instruction is generated that can be directly recognized by the downstream traffic control execution system. This instruction has a clear instruction type and parameter format.

[0030] S104. Input the structured traffic control instructions into the traffic simulation environment for simulation execution, obtain the simulation execution results, and use reinforcement learning algorithms to iteratively optimize the congestion prediction model and / or the generation strategy of the structured traffic control instructions based on the simulation execution results.

[0031] Specifically, the traffic simulation environment is a digital simulation system for highway traffic flow. It can receive structured traffic control instructions and simulate their execution within a simulated road network, outputting the changes in traffic state after execution as the simulation result. The simulation result reflects the effectiveness of the generated control strategy in the virtual environment, including its impact on traffic efficiency and safety. The reinforcement learning algorithm, through continuous interaction between the agent and the simulation environment, continuously adjusts the model parameters based on the reward signals provided by the simulation results, enabling the congestion prediction model and / or strategy generation model to gradually evolve towards better outcomes. Specifically, when the traffic state represented by the simulation result is better than the pre-execution state, the reinforcement learning algorithm provides a positive reward, reinforcing the model's tendency to output such a strategy; conversely, it provides a negative reward, suppressing the model's tendency to output such a strategy. Through this iterative optimization process, the congestion prediction model can output more accurate prediction results, and the strategy generation model can generate control strategies more suitable for the current traffic scenario.

[0032] In this embodiment, by fusing multi-source traffic data for congestion prediction, a comprehensive perception of highway traffic conditions is achieved. By converting the prediction results into scene semantic representations and structured prompt text, effective integration between the predicted information and a generative large language model is realized. By parsing natural language policies to generate structured instructions, the transformation link from semantic reasoning to executable control commands is established. Through the introduction of a simulation environment and closed-loop iterative optimization using reinforcement learning, the prediction model and policy generation model can continuously evolve through continuous feedback. These technical means collectively improve the intelligent decision-making level and the accuracy of policy generation in highway traffic management.

[0033] In some embodiments of the present invention, such as Figure 2 As shown, before inputting multi-source traffic data into the pre-trained congestion prediction model, this self-evolutionary method for highway traffic control strategies includes a training step for the congestion prediction model, as detailed below: S201. Obtain a pre-trained large language model as the teacher network, and obtain a graph convolutional neural network to be trained as the student network.

[0034] Specifically, the teacher network is a pre-trained general-purpose large language model, which has been pre-trained on a large-scale text corpus and possesses the ability to understand traffic scenarios and perform semantic reasoning. In this embodiment, the teacher network is preferably a DeepSeek large model. The student network is a graph convolutional neural network to be trained. A graph convolutional neural network is a neural network architecture that can directly perform convolution operations on graph-structured data. Its core idea is to extend convolution operations from Euclidean space to graph space, and update the feature representation of the current node by aggregating the information of neighboring nodes in the neighborhood of each node in the graph, thereby effectively capturing the spatial dependencies in the road network topology.

[0035] S202. Input the historical traffic data into the teacher network and the student network respectively, and obtain the first prediction result output by the teacher network and the second prediction result output by the student network.

[0036] Specifically, historical traffic data includes traffic flow data, vehicle speed data, weather data, and congestion status records for historical time periods. This data serves as training samples, inputted into both the teacher network and the student network. The teacher network, leveraging its powerful semantic understanding capabilities, performs a comprehensive analysis of the historical traffic data and outputs a first prediction result. The student network, using graph convolution operations, extracts and models the spatiotemporal features from the historical traffic data and outputs a second prediction result. Both the first and second prediction results represent predictions of traffic congestion status within the same historical time period.

[0037] S203. Based on the first prediction result and the second prediction result, calculate the multi-objective distillation loss, which includes at least two of the following: congestion prediction loss, feature representation loss, and inference logic loss.

[0038] Specifically, knowledge distillation is a method for transferring knowledge contained in a large, complex model (teacher network) to the training of a small, efficient model (student network).

[0039] In this embodiment, by using the first prediction result output by the teacher network as a soft objective, the student network is guided to learn the prediction distribution and inference logic of the large model. This allows the lightweight student network to significantly reduce inference latency and computational resource consumption while maintaining high prediction accuracy. Multi-objective distillation loss is a loss function that measures the prediction difference between the student network and the teacher network from multiple dimensions, and it includes at least two of congestion prediction loss, feature representation loss, and inference logic loss.

[0040] Among them, congestion prediction loss measures the deviation between the two networks in predicting specific congestion values, feature representation loss measures the distribution difference between the two networks in the intermediate layer feature representation, and inference logic loss measures the logical consistency between the two networks in the decision-making and inference process.

[0041] S204. Based on the multi-objective distillation loss, the parameters of the student network are updated by minimizing the inverse KL divergence until the training stopping condition is met, thus obtaining the pre-trained congestion prediction model.

[0042] Specifically, KL divergence is an asymmetric measure used to measure the difference between two probability distributions.

[0043] In this embodiment, minimizing the inverse KL divergence is used as the optimization objective, aiming to make the output distribution of the student network as close as possible to the output distribution of the teacher network. The optimization direction of the inverse KL divergence tends to make the student network cover the main modes of the teacher network, that is, to make the student network prioritize learning the most likely modes predicted by the teacher network and ignore secondary modes. This helps the student network focus on learning the most important traffic modes and avoids spreading the model capacity on sparse secondary modes. Training stopping conditions include, but are not limited to, reaching a preset maximum number of training epochs, the multi-objective distillation loss value converging to below a preset threshold, or the prediction accuracy on the validation set no longer improving.

[0044] This embodiment trains the congestion prediction model using knowledge distillation, and utilizes a large language model as a teacher network to guide the lightweight graph convolutional neural network in learning, so that the trained congestion prediction model has both prediction accuracy and computational efficiency. The multi-objective distillation loss constrains the learning direction of the student network from multiple dimensions such as numerical prediction, feature representation and reasoning logic, ensuring the comprehensiveness of knowledge transfer. The optimization strategy of minimizing the inverse KL divergence makes the student network focus on the main traffic modes, which is conducive to improving the prediction stability of the model in mainstream traffic scenarios.

[0045] In some embodiments of the present invention, such as Figure 3 As shown, step S203 calculates the multi-objective distillation loss based on the first and second prediction results, specifically including: S301. Determine the congestion prediction loss based on the difference between the first prediction result and the second prediction result in the numerical value of the congestion status.

[0046] Specifically, congestion state numerical values ​​refer to numerical indicators used to quantify the degree of traffic congestion, such as congestion index, average speed, and traffic density. Congestion prediction loss measures the deviation between the teacher network and the student network in predicting congestion state numerical values, and can be calculated using regression loss functions such as mean squared error or mean absolute error. This loss term ensures that the student network approximates the teacher network in terms of numerical prediction accuracy.

[0047] S302. Determine the feature representation loss based on the difference between the semantic features of the teacher's network output and the spatiotemporal features of the student's network output.

[0048] Specifically, semantic features are deep feature representations extracted by the teacher network in its intermediate layers, incorporating semantic understanding of the traffic scene. Spatiotemporal features are feature representations extracted by the student network in its intermediate layers, integrating road network spatial structure information and time-series change information. Feature representation loss measures the difference in distribution between the two feature representations in the feature space, typically calculated using mean squared error or cosine similarity loss. This loss term motivates the student network to learn the teacher network's semantic understanding of the traffic scene.

[0049] S303. Determine the inference logic loss based on the difference between the inference logic distribution of the teacher's network output and the inference logic distribution of the student's network output.

[0050] Specifically, the inference logic distribution refers to the logic layer output of the network before the final prediction result, reflecting the network's confidence distribution across different prediction categories. The inference logic loss measures the difference in logic layer output distribution between the teacher and student networks and can be calculated using KL divergence. This loss term encourages the student network to learn the teacher network's reasoning process and decision-making logic, rather than simply imitating the final prediction result.

[0051] S304. The congestion prediction loss, feature representation loss, and inference logic loss are weighted and summed to obtain the multi-objective distillation loss.

[0052] Specifically, the multi-objective distillation loss is obtained by multiplying each loss term by its corresponding weight coefficient and then summing the results. The weight coefficients are used to adjust the contribution ratio of each loss term to the total loss and can be preset or dynamically adjusted according to the specific application scenario and training effect. By minimizing this multi-objective distillation loss, the student network simultaneously approximates the teacher network in terms of numerical prediction accuracy, feature semantic representation, and inference logic.

[0053] This embodiment decomposes the knowledge distillation process into three dimensions of loss terms: numerical prediction, feature representation, and reasoning logic, and calculates them separately. This achieves refined constraints on the student network, which is beneficial for improving the knowledge transfer effect of the student network at different levels.

[0054] In some embodiments of the present invention, multi-scale spatiotemporal consistency learning is introduced as an additional training constraint during the parameter update process of the student network. Specifically, congestion status labels at various spatiotemporal scales are obtained from the training sample set. The training sample set contains historical traffic data and actual congestion statuses within corresponding time periods, either manually labeled or recorded by the system. These labels cover different temporal and spatial granularities. A spatiotemporal pyramid structure is constructed, and feature extraction is performed on the multi-source traffic data in the training samples at multiple preset temporal and spatial granularities to obtain traffic features at multiple spatiotemporal scales.

[0055] The preset time granularity includes, but is not limited to, 5 minutes, 15 minutes, and 30 minutes, while the preset spatial granularity includes, but is not limited to, road segment level, region level, and road network level. The spatiotemporal pyramid structure captures the evolution patterns of traffic flow at different spatiotemporal scales by performing hierarchical sampling and feature aggregation of traffic data across different time spans and spatial ranges.

[0056] Furthermore, traffic features at multiple spatiotemporal scales are input into the student network for congestion prediction, yielding congestion prediction results at each scale. The student network outputs a corresponding congestion prediction result for each spatiotemporal scale's input features. A consistency loss is calculated based on the consistency between the congestion prediction results at each scale and the corresponding congestion state label. The consistency loss measures the degree of agreement between the student network's predictions at different spatiotemporal scales and the actual congestion state, and can be calculated using mean squared error or cross-entropy loss. The parameters of the student network are jointly updated based on the weighted sum of the consistency loss and the multi-objective distillation loss.

[0057] Specifically, the total loss function is the sum of the consistency loss and the multi-objective distillation loss multiplied by their respective weight coefficients. This total loss is used as the optimization objective for backpropagation and parameter updates of the student network. This joint update method ensures that while the student network fits the output of the teacher network, the prediction results also conform to the evolution of traffic flow at different spatiotemporal scales at the physical level.

[0058] This embodiment introduces multi-scale spatiotemporal consistency learning, which ensures that the congestion prediction model maintains consistency and coherence in its prediction results at different temporal and spatial granularities. This avoids situations where the model is accurate in prediction at one scale but contradicts or violates the objective laws of traffic flow at other scales, which is beneficial to improving the model's generalization ability and prediction reliability in complex traffic scenarios.

[0059] In some embodiments of the present invention, such as Figure 4 As shown, step S102 generates a scene semantic representation of the highway traffic state based on the congestion prediction results and the highway network topology, specifically including: S401. Model the highway network as a directed graph, where nodes represent road segments and edges represent the connections between road segments.

[0060] Specifically, a directed graph is a graph structure whose edges have directionality, representing the directional constraints of traffic flow. For example, on a highway, vehicles can only travel in the direction of travel and cannot travel in the opposite direction. Each node corresponds to a segment of the highway, and the directed edges between nodes represent the connectivity between adjacent segments. The direction of the edges indicates the permitted direction of traffic flow. In this way, the highway network is transformed into graph structure data that can be directly processed by graph neural networks.

[0061] S402. Determine the feature vector of each node based on the congestion prediction results. The feature vector includes traffic state information, traffic density information and average speed information.

[0062] Specifically, for each node in the directed graph, an initial feature vector is constructed based on the predicted output of the road segment corresponding to that node in the congestion prediction model. The feature vector contains at least traffic state information (such as congestion level or congestion index), traffic density information, and average speed information. This information together constitutes a numerical description of the traffic state of the road segment corresponding to that node.

[0063] S403. Feature extraction is performed on the directed graph using a graph convolutional network to obtain high-dimensional features of each node.

[0064] Specifically, the directed graph obtained from the modeling and the initial feature vectors of each node are input into a graph convolutional network. The graph convolutional network performs convolution operations on the graph structure, aggregating the features of each node with those of its neighboring nodes, and transforming the features using a learnable weight matrix to generate a high-dimensional feature representation of that node. After multiple layers of graph convolution operations, the high-dimensional features of each node integrate its own information with information from its multi-hop neighborhood, enabling it to characterize the traffic state of that node within the overall road network context.

[0065] S404. Calculate the association strength between nodes through a self-attention mechanism, and generate a scene semantic representation based on the high-dimensional features of each node and the association strength between nodes.

[0066] Specifically, self-attention is a mechanism used to compute the interdependencies between elements in a sequence or graph. In the context of directed graphs, self-attention quantifies the degree of mutual influence between traffic states of different road segments by calculating the attention weights between any two nodes. By aggregating the high-dimensional features of each node with the corresponding inter-node correlation strengths, a scene semantic representation incorporating global contextual information is obtained. This scene semantic representation is presented in the form of graph-structured data, containing information on the road network topology, traffic state information of each node, and spatiotemporal dependencies between nodes.

[0067] This embodiment models the highway network as a directed graph, uses graph convolutional networks to extract high-dimensional features of nodes in the network, and captures the correlation strength between nodes with the help of a self-attention mechanism. This makes the generated scene semantic representation retain the spatial structure features of the road network, while incorporating traffic status information and the mutual influence between nodes, providing a complete and information-rich scene description for subsequent prompt text conversion.

[0068] In some embodiments of the present invention, such as Figure 5 As shown, step S102 converts the scene semantic representation into structured prompt text, specifically including: S501. Obtain a preset prompt template, which includes a location placeholder, a congestion level placeholder, a traffic density placeholder, an average speed placeholder, a prediction time placeholder, and a congestion trend placeholder.

[0069] Specifically, the prompt template is a pre-designed natural language text framework with pre-defined grammatical structures and placeholders. It is used to fill fixed positions with dynamic information extracted from the scene semantic representation to form a complete and standardized structured prompt text. Placeholders are reserved fields in the prompt template to be filled, and each placeholder corresponds to a type of scene semantic information.

[0070] S502. Extract current location information, current congestion level information, current traffic density information, current average speed information, prediction time information, and congestion trend information from the scene semantic representation.

[0071] Specifically, the scene semantic representation contains complete traffic status information for each road segment in the form of graph-structured data. Information items corresponding to each placeholder in the prompt template are extracted from this scene semantic representation. Among them, congestion trend information is determined by comparing the current congestion status with the historical congestion status or the congestion status of adjacent road segments, and is presented as a trend description such as "intensifying", "easing" or "maintaining".

[0072] S503. Fill the extracted current location information, current congestion level information, current traffic density information, current average speed information, predicted time information, and congestion trend information into the corresponding placeholders in the prompt template to generate structured prompt text.

[0073] Specifically, each information item is filled into the corresponding position in the prompt template according to the correspondence of placeholders, replacing the placeholders to form a complete structured prompt text. This prompt text is presented in the grammatical structure of natural language and contains complete scene parameter information, which can be accurately understood and processed by subsequent generative large language models.

[0074] This embodiment generates prompt text by using a preset prompt template and combining it with structured information extracted from the scene semantic representation. This ensures that the input of the generative large language model has a unified format specification and a complete scene description, which is beneficial to improving the accuracy and stability of the model's output strategy.

[0075] In some embodiments of the present invention, such as Figure 6 As shown, step S103, obtaining the natural language control strategy output by the generative large language model, specifically includes: S601. Input structured prompt text and thought chain reasoning prompts into the generative large language model. The thought chain reasoning prompts are used to guide the generative large language model to perform step-by-step reasoning in four stages: scenario analysis, trend prediction, strategy generation, and effect evaluation.

[0076] Specifically, the thought chain reasoning prompt mark is an instructional mark embedded in the input prompt, used to guide the large language model to think step by step according to the preset reasoning chain, rather than directly giving a conclusion. In this embodiment, by inserting the reasoning prompt mark "Let's analyze step by step" into the API call, the generative large language model is guided to reason sequentially in four stages: scenario analysis, trend prediction, strategy generation, and effect evaluation. The scenario analysis stage is used to identify the congestion level, spatial distribution, and causes of the current traffic scenario; the trend prediction stage is used to predict the subsequent evolution trend based on the congestion propagation law; the strategy generation stage is used to design intervention measures for key bottleneck sections; and the effect evaluation stage is used to analyze the expected impact of the generated strategy on traffic flow.

[0077] S602. Obtain the natural language control strategy output by the generative large language model after reasoning in four stages.

[0078] Specifically, the generative large language model completes the reasoning process in four stages based on the input structured prompt text and thought chain reasoning prompts, ultimately outputting a control strategy described in natural language. This natural language control strategy includes a judgment of the current scenario, a description of the control measures to be taken, the scope of the measures, and an explanation of the expected effects.

[0079] S603. The natural language control strategy is calibrated through a pre-trained traffic knowledge enhancement module.

[0080] Specifically, the traffic knowledge enhancement module is an auxiliary model pre-trained based on knowledge from the traffic engineering field. It is used to review and correct the natural language strategies output by the generative large language model. This module can identify parts of the strategy description that do not conform to traffic engineering principles, such as suggested speed limits exceeding reasonable limits or control measures contradicting traffic flow theory, and calibrate them to make the output strategies more in line with engineering practice.

[0081] This embodiment guides the large language model through multi-stage step-by-step reasoning by using thought chain reasoning prompts, which helps to fully leverage the reasoning capabilities of the large language model and generate logically complete and well-considered control strategies. The output strategy is calibrated by the traffic knowledge enhancement module to ensure the rationality and feasibility of the generated strategy at the traffic engineering level.

[0082] In some embodiments of the present invention, such as Figure 7 As shown, step S103 involves parsing the natural language control strategy, extracting the control parameters from the natural language control strategy, and generating structured traffic control instructions based on the control parameters. Specifically, this includes: S701. Sequence labeling of natural language control policies is performed using a bidirectional long short-term memory network-conditional random field policy parser to identify the operation type field, target region field, parameter value field, and time window field in the natural language control policy.

[0083] Specifically, Bidirectional Long Short-Term Memory (BiLSTM) is a variant of recurrent neural network capable of capturing contextual dependencies from both positive and negative directions in sequence data, while Conditional Random Field (CRF) is a structured prediction model for sequence labeling that can predict labels at each position based on the global dependencies of the label sequence. The BiLSTM-CRF policy parser, formed by combining these two technologies, can perform sequence labeling on natural language text, that is, label each word or phrase in the text with its corresponding semantic label. In this embodiment, the semantic labels include operation type labels, target region labels, parameter value labels, and time window labels. The operation type field describes the type of control action, such as speed limit, lane control, ramp control, or information dissemination; the target region field describes the scope of the control action; the parameter value field describes the specific numerical parameters of the control action; and the time window field describes the execution period of the control action.

[0084] S702. Determine the control parameters based on the identified operation type field, target area field, parameter value field, and time window field.

[0085] Specifically, the fields identified by sequence labeling are combined and formatted to form a structured set of control parameters. This set of control parameters contains complete parameter information required to execute a traffic control instruction, including what to do, where to do it, to what extent, and when to execute it.

[0086] S703. Based on the control parameters, the natural language control strategy is mapped into structured traffic control instructions. The instruction types of the structured traffic control instructions include at least one of dynamic speed limit instructions, lane control instructions, ramp control instructions, and variable information sign instructions.

[0087] Specifically, the instruction type is determined based on the operation type in the control parameters, and the various fields of the instruction are filled in according to fields such as target area, parameter value, and time window to generate a structured traffic control instruction with a standardized format. Dynamic speed limit instructions are used to set or adjust the maximum permissible speed on specific sections of highways; lane control instructions are used to control the opening and closing status of lanes, such as closing congested lanes or opening emergency lanes; ramp control instructions are used to control the signal cycle and merging rate of ramps merging into the main line; and variable message sign instructions are used to control the display content of roadside variable message signs, providing guidance information to drivers. For example, for the natural language strategy "It is recommended to implement a speed limit of 80 km / h on the G4 highway section from K100 to K120," a dynamic speed limit instruction can be generated after parsing and mapping. In this instruction, the operation type is dynamic speed limit, the target area is the K100 to K120 section, and the parameter value is 80 km / h.

[0088] It should be understood that the above instruction types are merely illustrative examples, and the structured traffic control instructions in this embodiment are not limited to the above types. More instruction types can be expanded according to actual control needs.

[0089] Furthermore, optionally, when mapping natural language control strategies to structured traffic control instructions based on control parameters, an encoder-decoder architecture can be used for semantic-to-parameter conversion. Specifically, the encoder encodes the natural language strategy text to obtain a strategy text vector, and the decoder or parameter matching module matches the target parameter set from the candidate parameter set based on the strategy text vector, generating structured traffic control instructions based on the target parameter set. This approach enables accurate mapping between natural language descriptions and structured parameters.

[0090] This embodiment uses a BiLSTM-CRF policy parser to perform sequence annotation on natural language control policies, which can effectively identify the structured parameters hidden in the policy text and transform free-format natural language descriptions into standardized control parameters. Based on parameter mapping, structured instructions are generated, enabling the control policies to be accurately parsed and automatically executed by downstream execution systems, overcoming the interface barriers between natural language output and unstructured execution systems.

[0091] In some embodiments of the present invention, when mapping natural language control strategies into structured traffic control instructions according to control parameters, a dual encoder architecture is used for semantic-to-parameter mapping.

[0092] Specifically, the natural language traffic control strategy is encoded using a BERT encoder to obtain a strategy text vector. BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model based on the Transformer architecture, capable of bidirectional encoding of input text to extract vector representations rich in contextual semantic information. Simultaneously, a parametric encoder encodes each candidate parameter set to obtain candidate parameter vectors. The candidate parameter set consists of a pre-defined combination of various structured control parameters, each corresponding to a possible control action and its parameter configuration, such as different speed values ​​for dynamic speed limits or different signal cycles for ramp control. The parametric encoder can use the same or different encoding network structure as the BERT encoder to encode the candidate parameter set into a fixed-dimensional vector representation. The similarity between the strategy text vector and each candidate parameter vector is calculated, using metrics such as cosine similarity or dot product similarity. Based on the calculated similarity, a target parameter set is determined from each candidate parameter set; for example, the candidate parameter set with the highest similarity is selected as the target parameter set, and structured traffic control instructions are generated based on this target parameter set. By using the above method, the control strategies expressed in natural language are mapped into structured control parameters through semantic similarity matching, which improves the accuracy and flexibility of the mapping.

[0093] This embodiment constructs a dual encoder architecture using a BERT encoder and a parameter encoder, mapping the natural language strategy and candidate parameter set to the same vector space, and using similarity calculation to achieve matching of the optimal parameter set, effectively improving the conversion accuracy of natural language strategy to structured control instructions.

[0094] In some embodiments of the present invention, after generating structured traffic control instructions, a step of conflict detection and resolution of multiple instructions is also included.

[0095] Specifically, a dependency matrix is ​​constructed between strategies. This matrix is ​​two-dimensional, and each element represents the dependency type between any two structured traffic control instructions. Dependency types include mutually exclusive, complementary, and independent types. A mutually exclusive type indicates that two instructions cannot be executed simultaneously under the same spatiotemporal conditions. For example, two different speed limit instructions may exist on the same road segment, or lane closure instructions and lane opening instructions may coexist on the same road segment and at the same time. A complementary type indicates that two instructions can cooperate and be executed in a coordinated manner. For example, an upstream speed limit instruction and a downstream lane control instruction may work together to alleviate the same congestion event. An independent type indicates that there is no obvious mutual influence between two instructions.

[0096] A conflict detection graph is constructed based on the inter-strategy dependency matrix. The conflict detection graph is a graph structure with structured traffic control instructions as nodes and mutually exclusive relationships between instructions as edges. A minimum conflict subgraph search is performed on the conflict detection graph to identify sets of structured traffic control instructions with mutual exclusion relationships. Specifically, the minimum conflict subgraph search refers to finding the subgraph with the most mutually exclusive edges in the conflict detection graph; the instruction set corresponding to this subgraph is the set of instructions with execution conflicts. Conflict resolution processing is performed on the identified sets of structured traffic control instructions with mutual exclusion relationships, including deleting some instructions or adjusting the parameters of some instructions, so that the final set of executed structured traffic control instructions remains consistent in time and space. For example, when two speed limit instructions, 80 km / h and 100 km / h, are detected on the same road segment, one is retained and the other is deleted based on the instruction's generation timestamp or confidence score; when a lane closure instruction and a lane diversion instruction are detected pointing to the same lane, the target area of ​​one instruction is adjusted to stagger the execution range.

[0097] This embodiment constructs a dependency matrix between strategies and a conflict detection graph to automatically detect conflicts in multiple structured traffic control instructions. By using minimum conflict subgraph search and conflict resolution, it eliminates spatiotemporal conflicts between multiple instructions, ensuring the consistency and executability of the final set of executed instructions.

[0098] In some embodiments of the present invention, based on the simulation results, reinforcement learning algorithms are used to iteratively optimize the generation strategy of congestion prediction models and / or structured traffic control instructions, specifically including the following steps.

[0099] First, an Actor-Critic reinforcement learning framework is constructed, where the Actor network employs a graph attention network and the Critic network employs a graph convolutional network. The Actor-Critic framework is a hybrid architecture combining policy-based and value-based reinforcement learning methods. The Actor network is responsible for selecting actions based on the current state, and the Critic network is responsible for evaluating the value of the actions selected by the Actor network. Graph attention networks are a variant of graph neural networks that assign different weights to different neighbors through an attention mechanism when aggregating neighbor node information, adaptively focusing on neighboring nodes that have a greater impact on the current node. Graph convolutional networks extract spatial features of nodes by performing convolution operations on the graph structure. The current traffic state of the highway network is used as the state space input into the Actor-Critic reinforcement learning framework. The state space is represented by a graph structure, where each node contains flow characteristics, speed characteristics, occupancy characteristics, and control state characteristics, where the control state characteristics reflect whether the current node already has control instructions being executed.

[0100] Next, the control actions corresponding to the structured traffic control instructions are used as the action space. The action space includes at least one control dimension, namely, dynamic speed limit, ramp signal cycle, and lane opening / closing state. Specifically, the dynamic speed limit ranges from 30 to 120 km / h, the ramp signal cycle ranges from 5 to 30 seconds, and the lane opening / closing state includes three states: open, closed, and variable guidance. Since the action space includes both discrete actions (such as lane opening / closing state) and continuous actions (such as speed limit and signal cycle), the action space in this embodiment is a multi-dimensional discrete-continuous hybrid space.

[0101] Finally, the reward function value is calculated based on the simulation results. This reward function value is used to evaluate traffic efficiency and safety. Traffic efficiency can be measured by indicators such as average road segment speed, traffic volume, or total travel time, while safety can be measured by indicators such as vehicle speed variance, emergency braking frequency, or accident risk index. Based on the reward function value, the parameters of the Actor network and Critic network are updated using the Proximal Policy Optimization (PPO) algorithm. The PPO algorithm is a reinforcement learning algorithm based on policy gradients. It introduces a pruning mechanism during policy update to limit the step size of each update, thus avoiding training instability caused by excessively large update increments.

[0102] Specifically, in each iteration, the Actor network selects a control action based on the current traffic state. The simulation environment executes this action and returns a new traffic state and reward function value. The Critic network evaluates the expected reward of the current state. The advantage function is calculated based on the difference between the actual reward function value and the expected reward evaluated by the Critic network. The pruning loss of the PPO algorithm is calculated based on the advantage function, and the parameters of both the Actor and Critic networks are updated by minimizing this loss. Through this process, the Actor network gradually learns a strategy that can select the optimal control action under different traffic states, and the Critic network gradually learns a more accurate state value assessment.

[0103] This embodiment constructs an Actor-Critic reinforcement learning framework, using a graph attention network for the Actor network to adaptively focus on important neighboring nodes, and a graph convolutional network for the Critic network to extract road network spatial structure features, enabling the reinforcement learning agent to fully utilize road network topology information; parameter updates are performed through a near-end policy optimization algorithm, ensuring training stability and convergence efficiency.

[0104] In some embodiments of the present invention, the above-mentioned reward function value is a multi-objective reward function value, that is, a comprehensive reward value that considers multiple optimization objectives simultaneously. Specifically, traffic efficiency evaluation values ​​and safety evaluation values ​​are determined based on the simulation execution results. The traffic efficiency evaluation value is used to quantitatively characterize the degree of improvement in traffic efficiency by the control strategy, and can be determined based on a comparison of the average speed of the road segment, traffic flow, or total travel time after the simulation execution with that before the execution. The safety evaluation value is used to quantitatively characterize the degree of impact of the control strategy on driving safety, and can be determined based on a comparison of the vehicle speed variance, number of emergency braking events, or collision risk index after the simulation execution with that before the execution.

[0105] This invention employs a multi-objective optimization algorithm to search for the optimal weight combination of traffic efficiency and safety assessment values ​​on the Pareto front. The Pareto front refers to the set of all Pareto optimal solutions in a multi-objective optimization problem. A Pareto optimal solution is one that cannot further improve any other objective without compromising at least one objective. This embodiment uses NSGA-II (Nondominated Sorting Genetic Algorithm II) as the multi-objective optimization algorithm, searching for the Pareto optimal weight combination in the weight space through genetic evolution.

[0106] The traffic efficiency and safety assessment values ​​are weighted and summed based on the optimal weight combination obtained from the search to obtain the multi-objective reward function value. This embodiment uses the NSGA-II multi-objective optimization algorithm to search for the optimal weight combination of efficiency and safety on the Pareto front, enabling the reinforcement learning agent to achieve a dynamic balance between efficiency and safety. It adaptively adjusts the optimization focus according to different traffic scenarios, avoiding the problem of insufficient adaptability of fixed weight combinations in diverse scenarios.

[0107] To better implement the self-evolution method for highway traffic control strategies in this invention embodiment, based on the self-evolution method for highway traffic control strategies, correspondingly, as follows: Figure 8 As shown, this embodiment of the invention also provides a self-evolving device for highway traffic control strategies. The self-evolving device 800 for highway traffic control strategies includes: The road segment congestion prediction module 801 is used to acquire multi-source traffic data of the highway network and input the multi-source traffic data into a pre-trained congestion prediction model to obtain congestion prediction results for each road segment of the highway. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data and historical congestion pattern data. The semantic representation generation module 802 is used to generate a scene semantic representation of the highway traffic state based on the congestion prediction results and the road network topology of the highway network, and convert the scene semantic representation into structured prompt text. The traffic control instruction generation module 803 is used to input the structured prompt text into the generative large language model, obtain the natural language control strategy output by the generative large language model, parse the natural language control strategy, extract the control parameters in the natural language control strategy, and generate structured traffic control instructions based on the control parameters. The traffic control optimization iteration module 804 is used to input the structured traffic control instructions into the traffic simulation environment for simulation execution, obtain the simulation execution results, and iteratively optimize the congestion prediction model and / or the generation strategy of the structured traffic control instructions using a reinforcement learning algorithm based on the simulation execution results.

[0108] The highway traffic control strategy self-evolution device 800 provided in the above embodiments can realize the technical solutions described in the above highway traffic control strategy self-evolution method embodiments. The specific implementation principles of each module or unit can be found in the corresponding content in the above highway traffic control strategy self-evolution method embodiments, which will not be repeated here.

[0109] like Figure 9 As shown, the present invention also provides an electronic device 900. The electronic device 900 includes a processor 901, a memory 902, a display 903, and a data acquisition interface 904. Figure 9 Only some components of the electronic device 900 are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0110] In some embodiments, processor 901 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 902 or process data, such as the self-evolution method for highway traffic control strategy in this invention.

[0111] In some embodiments, processor 901 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 901 may be local or remote. In some embodiments, processor 901 may be implemented on a cloud platform. In one embodiment, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-cloud, etc., or any combination thereof.

[0112] In some embodiments, memory 902 may be an internal storage unit of electronic device 900, such as a hard disk or memory of electronic device 900. In other embodiments, memory 902 may also be an external storage device of electronic device 900, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 900.

[0113] Furthermore, the memory 902 may include both internal storage units of the electronic device 900 and external storage devices. The memory 902 is used to store application software and various types of data installed on the electronic device 900.

[0114] In some embodiments, display 903 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. Display 903 is used to display information from electronic device 900 and to display a visual user interface.

[0115] The data acquisition interface 904 is used to connect to the roadside sensing device of the highway and receive multi-source traffic data collected by the sensing device. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data and historical congestion pattern data.

[0116] In one embodiment, the data acquisition interface 904 includes at least one of the following interfaces: a first interface for connecting to a video surveillance device and receiving video image data; a second interface for connecting to a microwave radar or millimeter-wave radar and receiving radar detection data; a third interface for connecting to a meteorological sensor and receiving meteorological data; a fourth interface for connecting to an ETC gantry system and receiving ETC transaction data and route identification data; and a fifth interface for connecting to a communication module and receiving data from a roadside unit or a cloud server.

[0117] In one embodiment, when processor 901 executes the highway traffic control strategy self-evolution program in memory 902, the following steps can be implemented: Multi-source traffic data of the highway network is acquired and input into a pre-trained congestion prediction model to obtain congestion prediction results for each section of the highway. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data and historical congestion pattern data. Based on the congestion prediction results and the road network topology of the highway network, a scene semantic representation of the highway traffic status is generated, and the scene semantic representation is converted into structured prompt text. The structured prompt text is input into the generative large language model to obtain the natural language control strategy output by the generative large language model. The natural language control strategy is then parsed to extract the control parameters in the natural language control strategy. Based on the control parameters, a structured traffic control instruction is generated. The structured traffic control instructions are input into a traffic simulation environment for simulation execution, and the simulation execution results are obtained. Based on the simulation execution results, a reinforcement learning algorithm is used to iteratively optimize the congestion prediction model and / or the generation strategy of the structured traffic control instructions.

[0118] It should be understood that when the processor 901 executes the highway traffic control strategy self-evolution program in the memory 902, in addition to the functions mentioned above, it can also perform other functions, as detailed in the description of the corresponding method embodiments above.

[0119] Furthermore, the embodiments of the present invention do not specifically limit the type of the electronic device 900 mentioned. The electronic device 900 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, the electronic device 900 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0120] Accordingly, this application also provides a computer-readable storage medium for storing computer-readable programs or instructions. When the programs or instructions are executed by a processor, they can implement the steps or functions of the self-evolution method for highway traffic control strategies provided in the above-described method embodiments.

[0121] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0122] The above provides a detailed description of the self-evolution method, apparatus, and electronic equipment for highway traffic control strategies provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A self-evolutionary method for highway traffic control strategies, characterized in that, include: Multi-source traffic data of the highway network is acquired and input into a pre-trained congestion prediction model to obtain congestion prediction results for each section of the highway. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data and historical congestion pattern data. Based on the congestion prediction results and the road network topology of the highway network, a scene semantic representation of the highway traffic status is generated, and the scene semantic representation is converted into structured prompt text. The structured prompt text is input into the generative large language model to obtain the natural language control strategy output by the generative large language model. The natural language control strategy is then parsed to extract the control parameters in the natural language control strategy. Based on the control parameters, a structured traffic control instruction is generated. The structured traffic control instructions are input into a traffic simulation environment for simulation execution, and the simulation execution results are obtained. Based on the simulation execution results, a reinforcement learning algorithm is used to iteratively optimize the congestion prediction model and / or the generation strategy of the structured traffic control instructions.

2. The self-evolution method for highway traffic control strategy according to claim 1, characterized in that, Before inputting the multi-source traffic data into the pre-trained congestion prediction model, the self-evolution method for highway traffic control strategies includes: A pre-trained large language model is used as the teacher network, and a graph convolutional neural network to be trained is used as the student network. Historical traffic data is input into the teacher network and the student network respectively, and a first prediction result output by the teacher network and a second prediction result output by the student network are obtained. Based on the first prediction result and the second prediction result, a multi-objective distillation loss is calculated, wherein the multi-objective distillation loss includes at least two of congestion prediction loss, feature representation loss and inference logic loss. Based on the multi-objective distillation loss, the parameters of the student network are updated by minimizing the inverse KL divergence until the training stopping condition is met, thus obtaining the pre-trained congestion prediction model.

3. The self-evolution method for highway traffic control strategy according to claim 2, characterized in that, The multi-objective distillation loss includes the congestion prediction loss, the feature representation loss, and the inference logic loss; The step of calculating the multi-objective distillation loss based on the first prediction result and the second prediction result includes: The congestion prediction loss is determined based on the difference between the first prediction result and the second prediction result in the numerical value of congestion status. The feature representation loss is determined based on the difference between the semantic features output by the teacher network and the spatiotemporal features output by the student network. The inference logic loss is determined based on the difference between the inference logic distribution output by the teacher network and the inference logic distribution output by the student network. The congestion prediction loss, the feature representation loss, and the inference logic loss are weighted and summed to obtain the multi-objective distillation loss.

4. The self-evolution method for highway traffic control strategy according to claim 2, characterized in that, The parameter update of the student network also includes: Obtain congestion status labels at various spatiotemporal scales from the training sample set; A spatiotemporal pyramid structure is constructed, and features are extracted from the multi-source traffic data in the training samples at multiple preset time granularities and multiple preset spatial granularities to obtain traffic features at multiple spatiotemporal scales. The traffic characteristics at multiple spatiotemporal scales are input into the student network to predict congestion, and congestion prediction results at each spatiotemporal scale are obtained. Based on the consistency between the congestion prediction results at each spatiotemporal scale and the corresponding congestion status labels, the consistency loss is calculated. The parameters of the student network are jointly updated based on the weighted sum of the consistency loss and the multi-objective distillation loss.

5. The self-evolution method for highway traffic control strategy according to claim 1, characterized in that, The step of generating a scene semantic representation of highway traffic status based on the congestion prediction results and the highway network topology includes: The highway network is modeled as a directed graph, where nodes represent road segments and edges represent the connections between road segments. Based on the congestion prediction results, feature vectors for each node are determined, and the feature vectors include traffic state information, traffic density information, and average speed information. Feature extraction is performed on the directed graph using a graph convolutional network to obtain high-dimensional features of each node; The association strength between nodes is calculated using a self-attention mechanism, and the scene semantic representation is generated based on the high-dimensional features of each node and the association strength between them.

6. The self-evolution method for highway traffic control strategy according to claim 1, characterized in that, The step of converting the scene semantic representation into structured prompt text includes: Obtain a preset prompt template, which includes a location placeholder, a congestion level placeholder, a traffic density placeholder, an average speed placeholder, a prediction time placeholder, and a congestion trend placeholder. Extract current location information, current congestion level information, current traffic density information, current average speed information, prediction time information, and congestion trend information from the scene semantic representation; The extracted current location information, current congestion level information, current traffic density information, current average speed information, predicted time information, and congestion trend information are respectively filled into the corresponding placeholders in the prompt template to generate the structured prompt text.

7. The self-evolution method for highway traffic control strategy according to claim 1, characterized in that, The natural language control strategy for obtaining the output of the generative large language model includes: The structured prompt text and the thought chain reasoning prompt mark are input into the generative large language model. The thought chain reasoning prompt mark is used to guide the generative large language model to perform step-by-step reasoning in four stages: scenario analysis, trend prediction, strategy generation, and effect evaluation. Obtain the natural language control strategy output by the generative large language model after inference in the four stages; The natural language control strategy is calibrated using a pre-trained traffic knowledge enhancement module.

8. The method according to claim 1, characterized in that, The step of parsing the natural language control strategy, extracting the control parameters from the natural language control strategy, and generating structured traffic control instructions based on the control parameters includes: The natural language control strategy is sequence labeled using a bidirectional long short-term memory network-conditional random field policy parser to identify the operation type field, target region field, parameter value field, and time window field in the natural language control strategy. The control parameters are determined based on the identified operation type field, target area field, parameter value field, and time window field. Based on the control parameters, the natural language control strategy is mapped into structured traffic control instructions, the instruction types of which include at least one of dynamic speed limit instructions, lane control instructions, ramp control instructions, and variable information sign instructions.

9. A self-evolving device for highway traffic control strategies, characterized in that, include: The road congestion prediction module is used to acquire multi-source traffic data of the highway network and input the multi-source traffic data into a pre-trained congestion prediction model to obtain congestion prediction results for each road segment of the highway. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data and historical congestion pattern data. The semantic representation generation module is used to generate a scene semantic representation of the highway traffic state based on the congestion prediction results and the road network topology of the highway network, and convert the scene semantic representation into structured prompt text; The traffic control instruction generation module is used to input the structured prompt text into the generative large language model, obtain the natural language control strategy output by the generative large language model, parse the natural language control strategy, extract the control parameters in the natural language control strategy, and generate structured traffic control instructions based on the control parameters. The traffic control optimization iteration module is used to input the structured traffic control instructions into the traffic simulation environment for simulation execution, obtain the simulation execution results, and iteratively optimize the congestion prediction model and / or the generation strategy of the structured traffic control instructions using reinforcement learning algorithms based on the simulation execution results.

10. An electronic device, characterized in that, Includes data acquisition interface, memory, and processor; The data acquisition interface is used to connect to the roadside sensing device of the highway and receive multi-source traffic data collected by the sensing device. The multi-source traffic data includes traffic flow data, vehicle speed data, meteorological data and historical congestion pattern data. The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the self-evolution method of highway traffic control strategy according to any one of claims 1 to 8.

Images