Large model enhanced deep learning fitting road network traffic operation risk situation rapid prediction method

Abstract

The invention relates to a large model enhanced deep learning fitting road network traffic operation risk situation rapid prediction method, and belongs to the technical field of traffic risk assessment, and the method comprises the following steps: S1, constructing a multi-scene high-quality road network data set; s2, constructing and training an improved space-time diagram neural network ST-GNN prediction model, and realizing road network state mapping in a period after an accident occurrence moment; s3, expanding a fine tuning data set, introducing a large model, and strengthening the fusion adaptation capability with the ST-GNN prediction model through field fine tuning optimization; and S4, deploying the trained ST-GNN model to an edge computing device, deploying the large model to a cloud end, realizing edge end rapid prediction and cloud end fine judgment and correction, obtaining a dynamic collaborative reasoning result, and displaying risk decision suggestions by using a visual output module.

Description

Technical Field

[0001] This invention belongs to the field of traffic risk assessment technology, and relates to a method for rapid prediction of road network traffic operation risk status using a large model enhanced by deep learning fitting. Background Technology

[0002] Road network congestion and traffic accidents have become core bottlenecks restricting the operational efficiency of urban transportation systems. Related research indicates that congestion on most urban arterial roads is caused by sudden traffic accidents, and effective prediction of road network operational risks can support the formulation of proactive traffic control strategies, significantly reducing secondary congestion or accidents and shortening congestion time. However, current road network operational risk prediction technologies still have many limitations, making it difficult to meet the real-time, precise, and comprehensive needs of traffic emergency management.

[0003] Current road network operation risk prediction technologies can be divided into three generations, each with significant shortcomings: The first generation is a numerical simulation method based on traffic simulation software, using mainstream traffic simulation software such as VISSIM and SUMO as the core. It iteratively calculates by inputting accident scenarios and road network parameters to output the road network status for subsequent periods. The core flaw of this technology is its extremely poor timeliness—a single one-hour road network prediction requires a long computation time, while the golden window for secondary traffic prevention and traffic diversion guidance is limited. The simulation results often lag behind emergency needs, failing to provide timely support for on-site dispatch. At the same time, the simulation software has high hardware requirements, making it difficult to achieve parallel computing on large-scale road networks.

[0004] The second generation is based on traditional machine learning prediction models, including algorithms such as Support Vector Machines (SVM) and Random Forests. These models build prediction models by mining the correlation between historical traffic flow data and accident data. While these models shorten computation time, they have two major limitations: First, they cannot capture the dual characteristics of road network spatial topology and time-series dependencies, such as ignoring the traffic flow transmission effect between intersections and adjacent road segments, leading to large prediction errors. Second, they have poor generalization ability; models trained for specific areas experience a sharp drop in performance when applied across road segments and time periods.

[0005] The third generation is based on deep learning-based prediction methods, such as using Long Short-Term Memory (LSTM) networks and ordinary Graph Neural Networks (GNNs) to process traffic data. While these technologies can initially adapt to the characteristics of spatiotemporal data, they still have significant shortcomings. LSTM models struggle to model the complex spatial topology of road networks and cannot accurately quantify the impact range of accidents on surrounding road segments. Ordinary GNNs use static graph structures and cannot reflect the dynamic adjustment of road segment correlation strength caused by real-time changes in traffic flow, resulting in large errors in congestion diffusion prediction.

[0006] From a technical perspective, existing solutions suffer from the following shortcomings: First, their response time is delayed. The iterative computational nature of simulation software means it cannot meet the emergency response requirements of seconds after an accident. While traditional machine learning and basic deep learning models offer faster speeds, they still require considerable time and cannot cover real-time prediction scenarios across the entire road network. Second, spatiotemporal modeling is insufficient. Modeling methods based on a single time or spatial dimension cannot fully depict the spatial diffusion and temporal evolution of congestion caused by traffic accidents. Third, their generalization ability is weak. Model training relies on labeled data from specific regions, resulting in poor adaptability to changes in road network structure and traffic flow characteristics, and high costs for cross-regional deployment. In summary, existing technologies are insufficient to meet the demands of traffic emergency management for second-level response, accurate quantification, and multi-dimensional output in predicting road network operational status. Summary of the Invention

[0007] In view of this, the purpose of this invention is to address the core problems of delayed response, insufficient spatiotemporal modeling, and weak generalization ability in existing road network operation status prediction technologies, and to propose a large-model-enhanced spatiotemporal graph neural network (ST-GNN) technical solution.

[0008] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting, comprising the following steps: S1: Construct a high-quality road network dataset for multiple scenarios; S2: Construct and train an improved spatiotemporal graph neural network (ST-GNN) prediction model to realize the mapping of road network state within a period after the accident occurs; S3: Expand the fine-tuning dataset, introduce large models and optimize them through domain fine-tuning to enhance the fusion and adaptation capabilities with the ST-GNN prediction model; S4: Deploy the trained ST-GNN model on edge computing devices and deploy the large model in the cloud to achieve rapid prediction at the edge and precise judgment and correction in the cloud, obtain dynamic collaborative reasoning results, and use the visualization output module to display risk decision-making suggestions.

[0009] In one possible implementation, step S1, which involves constructing a high-quality road network dataset for multiple scenarios, includes the following steps: S11: Input the road network and traffic flow baseline simulation data into the traffic simulation software, design diverse disturbance scenarios, including various accident types, accident locations, and different traffic flow baseline time periods, and generate various typical scenario combinations; the road network simulation data includes road grade, number of lanes, intersection signal timing, speed limit information, and special road segment attributes, covering typical road network structures of different city sizes and road network densities; the traffic flow baseline simulation data includes traffic volume, vehicle speed, and density at different time periods, constructing a traffic flow baseline that conforms to actual traffic operation patterns; S12: For each scenario, use traffic simulation software to perform continuous dynamic simulation and collect road network status simulation data.

[0010] In one possible implementation, the improved spatiotemporal graph neural network (ST-GNN) prediction model first constructs a dynamically weighted road network graph structure, abstracting the road network into a dynamic graph structure G=(V,E,W), where node V represents a road segment or intersection; edge E represents the connectivity between road segments, connecting only adjacent road segments with traffic flow transfer; and edge weight W is a dynamic weight calculated from the real-time traffic flow transfer coefficient between two road segments, as shown in the following formula:

[0011] in, Let t be the time from road segment i to section of road j Traffic flow, For road section i The set of adjacent road segments, k Indicates road segment i Adjacent road sections k , Let t be the time from road segment i to section of road k Traffic flow.

[0012] In one possible implementation, the improved spatiotemporal graph neural network (ST-GNN) prediction model includes a spatial feature extraction layer, a temporal dependency modeling layer, and an attention fusion module. The spatial feature extraction layer employs a graph attention network (GAT), which assigns differentiated weights to different neighboring nodes through an attention mechanism to adaptively capture the impact intensity of accident nodes on the surrounding road network. The formula is as follows:

[0013]

[0014] in, For road section j For the section i Attention weights For attention coefficient vectors, Each is a road segment i , j , k eigenvectors, This is the weight matrix. For road section i The new feature vector after fusion via the attention mechanism As the activation function, the spatial feature extraction layer outputs a spatial feature vector of fixed dimensions. ; The time-dependency modeling layer employs a bidirectional long short-term memory (Bi-LSTM) network to mine the time-series dependencies of traffic flow from both the past to the present and present to the future directions, capturing the periodicity and trend of congestion caused by accidents. The Bi-LSTM consists of two symmetrically configured bidirectional LSTM structures with multiple hidden layers, each containing 256 units. The activation function is tanh. The input consists of spatial features from multiple consecutive time slices, bidirectionally encoding the traffic flow sequence from both the past to the present and present to the future directions, and outputting a fixed-dimensional time feature vector. ; The attention fusion module employs a multilayer perceptron (MLP) to calculate the fusion weights of spatial and temporal features, performing weighted fusion of the two types of features to highlight the different importance of spatial diffusion features in the early stages of an accident and temporal evolution features in the later stages. The final output is a fixed-dimensional fusion feature vector for subsequent traffic flow parameter prediction. The improved spatiotemporal graph neural network (ST-GNN) prediction model uses a multi-task training and optimization strategy to construct a multi-task loss function, simultaneously optimizing three prediction targets: vehicle speed, traffic flow, and congestion range. The multi-task loss function is as follows:

[0015] in, Mean squared error loss, used for vehicle speed v ,flow q The regression prediction, in which , These represent the actual and predicted values ​​of the vehicle speed, respectively. , These represent the actual and predicted traffic flow values, respectively. The cross-entropy loss is used for classification prediction of congestion range r, which is divided into four categories: no congestion, light congestion, moderate congestion, and heavy congestion. , ω1, ω2, and ω3 represent the actual and predicted values ​​of the congestion range, respectively; ω1, ω2, and ω3 are the weight coefficients of each task, determined through experiments, and satisfy ω1 + ω2 + ω3 = 1.

[0016] In one possible implementation, the augmentation and fine-tuning of the dataset described in step S3 includes: Based on the basic simulation data, specialized fine-tuning data is added, and then the final dataset is divided into training, validation, and test sets; the specialized fine-tuning data includes: Spatiotemporal feature annotation data: Extract the fused feature vector output by the ST-GNN model to represent the prediction of future road network change risks; Knowledge data in the transportation sector: Organizing rules and road network topology knowledge in the field of transportation engineering to form a structured knowledge graph; Decision-making scenario case data: Construct a case library, which includes scenario descriptions, decision-making solutions, and effect scores for each decision-making scenario; The basic simulation data is used to train the improved ST-GNN model; the spatiotemporal feature annotation dataset, the transportation domain knowledge dataset, and the decision-making scenario case dataset are used for fine-tuning of the large model domain. In step S3, the large model selected is a general-purpose model with a moderate number of parameters and strong text understanding and logical reasoning capabilities, retaining its core Transformer architecture. A spatiotemporal feature encoding module is added to the input layer of the large model. The spatiotemporal feature encoding module consists of a linear projection layer, a normalized LayerNorm layer, and a ReLU activation function. The input dimension of the linear projection layer is consistent with the dimension of the ST-GNN output fused feature vector. Through linear mapping, the high-dimensional spatiotemporal fused feature vector output by ST-GNN is directly converted into a feature vector with the same dimension as the word embedding of the large model. After normalization and activation, it is sent to the large model Transformer encoder to achieve direct feature input. A multi-task head is added to the output layer, including a regression head, a classification head, and a generation head, corresponding to road network risk scoring, secondary accident risk level, and decision suggestion generation, respectively, to match the decision-making needs of the transportation field.

[0017] In one possible implementation, the domain fine-tuning described in step S3 employs a three-stage fine-tuning optimization, including the following steps: Phase 1: Knowledge Injection Pre-training. Using a structured traffic knowledge graph as input, pre-training is performed using a masked language model and knowledge alignment loss, enabling the large model to master traffic-specific concepts, rules, and road network topology associations. The second stage is feature alignment fine-tuning. Input the spatiotemporal features of ST-GNN and the corresponding label data, freeze the bottom parameters of the large model, and train only the top Transformer layer and the newly added encoding module and task head. The optimization targets are mean squared error loss and cross-entropy loss. During training, a multi-task loss balancing strategy is adopted. First, the mean squared error loss and cross-entropy loss are scaled to the same numerical order of magnitude by min-max normalization. Then, the two types of losses are fused by equal weighting to achieve accurate alignment between the large model and the feature output of ST-GNN. The third stage: fine-tuning of decision-making scenarios. Input decision-making scenario case data and adopt the instruction fine-tuning mode to guide the large model to learn the mapping relationship between state, decision and effect in traffic emergency scenarios. The optimization objective is generative loss and effect score reward. The reward function directly adopts the pre-labeled treatment effect score in the case library and normalizes it to the 0-1 range as the reinforcement learning reward value. The better the solution effect, the higher the reward value. The quantified treatment effect score is used as the reward value. A unified feature space approach is adopted to collaboratively train the three types of data in the same large model across the three stages. The text data is encoded using the same traffic domain-specific word segmenter to ensure a consistent language feature format. The spatiotemporal features output by ST-GNN are converted into vectors with the same word embedding dimension as the large model through a unified spatiotemporal feature encoding module, thus enabling input within the same feature space as the text data.

[0018] In one possible implementation, during fine-tuning optimization, LoRA (Low-Rank Adaptation) technology is used to train only low-rank matrix parameters, compressing the number of fine-tuning parameters; mixed-precision training and gradient accumulation are introduced to adapt to the differences in computing power between edge devices and the cloud; a traffic-specific word segmenter is constructed to optimize the encoding effect of road network names and traffic parameter-specific expressions. The traffic-specific word segmenter is built based on traffic-specific corpora, adopts BPE sub-word segmentation rules, and expands the traffic-specific vocabulary on the basis of the general word segmenter vocabulary.

[0019] In one possible implementation, after fine-tuning and optimization, the large model, based on the road network state features and structured knowledge input by ST-GNN, scores risks from two core dimensions: congestion spread speed and the probability of secondary accidents, outputting a comprehensive score and dimensional scores. Employing a causal reasoning mechanism, combining road network state features and traffic rule knowledge, it identifies high-risk scenarios such as sudden speed drops and intersection queuing, outputting a list of high-risk road segments and their risk levels at each time point, clarifying the highest congestion level and secondary accident risk for each road segment at each time point. Finally, combining a simulation case library and real-time road network status, it outputs targeted emergency dispatch suggestions.

[0020] In one possible implementation, in step S4, the trained ST-GNN model is deployed on a roadside edge computing device. The device collects accident information and traffic flow data in real time at the moment the accident occurs, completes subsequent road network status prediction and preliminary risk assessment, and outputs basic prediction results. A three-factor trigger mechanism is set up to call the cloud-based large model for fine-tuning when any of the following conditions are met: the ST-GNN prediction confidence is lower than a set threshold; the Scene Complexity Index (SCI) is higher than a set threshold; where SCI comprehensively considers the severity of the accident, traffic flow density, and road network complexity, and the formula is as follows:

[0021] This represents the severity coefficient of the accident. Traffic flow density coefficient The road network complexity coefficient is... , , The weights for each factor are set based on historical experience to satisfy... + + =1; The confidence-weighted fusion method is used to integrate the results from the edge and cloud, as shown in the following formula:

[0022] in, Predict confidence levels for edge devices. Confidence of large cloud models.

[0023] In one possible implementation, in step S4, the real-time simulation verification data collected by the roadside edge computing device is used to dynamically correct the prediction results at fixed intervals, updating the traffic flow status and risk level; a visualization output platform is then constructed to present the results, including: Road network traffic flow heat map: Real-time display of congestion levels on different road sections at different times; Risk score change curve: shows the changing trends of the overall risk score of the road network and the scores of the three dimensions at each time point; Risk probability change curve: shows the evolution of congestion level and secondary accident risk over time in high-risk road sections; Emergency decision-making recommendations: Based on a large model combined with simulation cases and traffic rules, targeted emergency dispatch recommendations are generated.

[0024] In a second aspect, the present invention provides a terminal including a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the step instructions as described in the first aspect of the present invention.

[0025] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform some or all of the steps described in the first aspect of the present invention.

[0026] Fourthly, the present invention provides a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in the first aspect of the present invention. The computer program product may be a software installation package.

[0027] The beneficial effects of this invention are as follows: Through the efficient inference of the edge-end ST-GNN model, the entire process prediction can be completed within seconds after an accident occurs, improving efficiency and fully covering the golden window for emergency decision-making, providing timely support for rapid response. The improved ST-GNN model, through dynamic graph structure and attention fusion technology, effectively reduces the vehicle speed prediction error at different times, significantly improving accuracy compared to traditional GNN models; the accuracy of secondary accident risk location is at a high level, accurately identifying high-risk road sections. A complete output system is constructed, from state prediction to risk assessment, risk location, and finally decision recommendations. Compared to traditional technologies that only output basic data, this system adds multiple types of core decision information, supporting the formulation of refined emergency strategies based on time, road section, and level, such as deploying police forces in advance on high-risk road sections and guiding vehicles to detour via alternative routes. Deep model-level fusion is achieved through large-scale model domain fine-tuning and parametric fusion design. Trained on multi-scenario simulation datasets, the model maintains stable performance across different urban road network structures and traffic flow characteristics, enhancing its adaptability for cross-regional deployment. The ST-GNN model's parameter size is controlled within a reasonable range, supporting edge device deployment and significantly reducing deployment costs compared to cloud-dependent solutions. In simulated urban core area road network scenarios, the emergency dispatch strategy based on this technology can significantly shorten accident congestion dissipation time and improve road network traffic efficiency during peak hours; it effectively reduces the probability of secondary accidents, lowers indirect losses from traffic delays and reduces police force costs. Combined with the advantages of pure simulation data, it possesses significant promotional value.

[0028] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 Flowchart of a method for rapid prediction of road network traffic operation risk status to enhance deep learning fitting for large models; Figure 2 Overall structure diagram of the road network traffic operation risk situation prediction model for enhancing deep learning fitting of large models; Figure 3 This is a comparison chart of simulated and predicted values ​​of road network status changing over time. Detailed Implementation

[0030] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0031] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0032] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0033] Example 1: Please see Figure 1 This invention provides a rapid prediction method for road network traffic operation risk status using large-model enhanced deep learning fitting. It takes spatiotemporal data modeling as the core, large-model fine-tuning enhancement as the key, and dynamic correction as the guarantee, to achieve rapid and accurate prediction of road network operation risk status. Figure 2 The technical architecture of this method and the collaborative interaction between its modules are presented. From top to bottom, they are: multi-scenario dataset construction layer, improved ST-GNN prediction layer, large model domain fine-tuning enhancement layer, and dynamic collaborative inference and result output layer. The core components of each layer, the data flow path, and the model-level fusion logic of ST-GNN and large model from feature input, fusion inference to result output are clearly presented.

[0034] First, a high-quality road network dataset covering multiple scenarios is constructed. This dataset is the core foundation for model training. This method utilizes simulation generation and data augmentation to construct a high-quality dataset covering basic spatiotemporal data, domain knowledge data, and decision-making scenario data, addressing the problems of scarce real-world road network operation data and high annotation costs. Details are as follows: Basic Data Input and Scenario Design: SUMO or VISSIM traffic simulation software is selected as the core data generation carrier. Two types of basic simulation data are input: First, high-precision road network simulation data, including road grade, number of lanes, intersection signal timing, speed limit information, and special road segment attributes such as tunnels / bridges, covering typical road network structures of different city sizes and road network densities; second, multi-time period traffic flow baseline simulation data, collecting core parameters such as traffic volume, speed, and density at different times such as peak, off-peak, and low-peak periods to construct a traffic flow baseline that conforms to actual traffic operation patterns. Diverse disturbance scenarios are designed, employing multi-dimensional orthogonal experimental design to ensure complete scenario coverage. These diverse disturbance scenarios include combinations of minor, moderate, and severe accident types (rear-end collisions, collisions, obstructions, etc.), accident locations such as ordinary straight road sections, intersection entrances and exits, and road network bottlenecks, as well as multi-dimensional parameter combinations of different traffic flow baseline time periods such as morning peak, evening peak, off-peak, and low-peak periods, generating multiple typical scenario combinations covering different city sizes, road network densities, and traffic operation states.

[0035] Dynamic simulation and data slicing: For each scenario, a traffic accident is triggered at time t0 (the time of the accident) and the congestion is set to last for 1 hour. SUMO software is then started for continuous dynamic simulation. A slicing period of 5 minutes is used to synchronously collect data from t1 to t2. 12 Simulation data of road network status at 12 time points.

[0036] An improved spatiotemporal graph neural network (ST-GNN) prediction model was constructed and trained, tested, and validated using a slice dataset obtained from dynamic simulation using SUMO software. The improved ST-GNN model realizes the prediction from accident information at time t0 to time t1-t. 12 Accurate mapping of road network state at any time, model optimization includes graph structure construction, model architecture design and training strategies.

[0037] The graph structure construction involves abstracting the road network into a dynamic graph structure G=(V,E,W), where nodes V represent road segments or intersections; edges E represent the connectivity between road segments, connecting only adjacent road segments with traffic flow transfer; and edge weights W are dynamic weights calculated from the real-time traffic flow transfer coefficient between two road segments, as shown in the following formula:

[0038] in, Let t be the time from road segment i to section of road j Traffic flow, For road section i The set of adjacent road segments, k Indicates road segment i Adjacent road sections k , Let t be the time from road segment i to section of road k Traffic flow. Dynamic weights can reflect the adjustment of road segment correlation strength caused by changes in traffic flow in real time, breaking through the limitations of traditional static graph structures.

[0039] The model architecture adopts a three-stage structure of spatial feature extraction, temporal dependency modeling, and attention fusion output to achieve in-depth mining of spatiotemporal features, including a spatial feature extraction layer, a temporal dependency modeling layer, and an attention fusion module.

[0040] The spatial feature extraction layer employs a graph attention network (GAT), which assigns differentiated weights to different neighboring nodes through an attention mechanism to adaptively capture the impact intensity of accident nodes on the surrounding road network. The formula is as follows:

[0041]

[0042] in, For road section j For the section i Attention weights For attention coefficient vectors, Each is a road segment i , j , k eigenvectors, This is the weight matrix. For road section i The new feature vector after fusion via the attention mechanism This is the activation function. This layer outputs a fixed-dimensional spatial feature vector. .

[0043] The time-dependency modeling layer employs a bidirectional long short-term memory (Bi-LSTM) network to mine the time-series dependencies of traffic flow from both the past to the present and the present to the future, capturing the periodicity and trend of congestion caused by accidents. The Bi-LSTM uses multiple hidden layers, taking spatial features from consecutive time slices as input and outputting a fixed-dimensional temporal feature vector. The Bi-LSTM network is configured as a two-layer bidirectional LSTM structure with 256 hidden layer units per layer. The forward and backward LSTM units are configured symmetrically, and the activation function is tanh. The traffic flow sequence is bidirectionally encoded from the past to the present and from the present to the future.

[0044] The attention fusion module uses a multilayer perceptron (MLP) to calculate the fusion weights of spatial and temporal features, and performs weighted fusion of the two types of features to highlight the different importance of spatial diffusion features in the early stage of an accident and temporal evolution features in the later stage. Finally, it outputs a fixed-dimensional fusion feature vector for subsequent traffic flow parameter prediction.

[0045] The multi-task training and optimization strategy specifically includes constructing a multi-task loss function and simultaneously optimizing three prediction targets: vehicle speed, traffic flow, and congestion range, thereby improving the overall performance of the model.

[0046] in, Mean squared error loss, used for vehicle speed v ,flow q The regression prediction, in which , These represent the actual and predicted values ​​of the vehicle speed, respectively. , These represent the actual and predicted traffic flow values, respectively. Cross-entropy loss is used for classification prediction of congestion range r, categorizing it into no congestion, mild congestion, moderate congestion, and severe congestion. , Let ω1, ω2, and ω3 represent the actual and predicted values ​​of the congestion range, respectively; ω1, ω2, and ω3 are the weight coefficients of each task, satisfying ω1 + ω2 + ω3 = 1. In the experiment, combinations with 0.1 intervals were compared, and the parameter combination ω1 = 0.4, ω2 = 0.3, and ω3 = 0.3 was determined under the objective of achieving the best prediction accuracy. An adaptive optimizer was used, combined with a learning rate scheduling strategy and an early stopping mechanism for model training. Training was stopped when the validation set loss did not decrease for several consecutive rounds, and the optimal model was saved.

[0047] Fine-tuning Dataset Expansion: Based on the basic simulation data, three types of dedicated fine-tuning data are added: spatiotemporal feature-annotated data, traffic domain knowledge data, and decision-making scenario case data. Of the four datasets mentioned above, the basic simulation dataset is used for training the improved ST-GNN model, while the spatiotemporal feature-annotated dataset, traffic domain knowledge dataset, and decision-making scenario case dataset are used for fine-tuning the larger model. The spatiotemporal feature-annotated data extracts the fused feature vectors output by the ST-GNN model, representing predictions of future road network change risks. The traffic domain knowledge data organizes rules in the field of traffic engineering, including, in this embodiment, accident handling priorities set by relevant regulations, congestion diffusion thresholds in traffic flow theory, and road network topology knowledge, forming a structured knowledge graph. The decision-making scenario case data constructs a case library containing scenario descriptions, decision schemes, and effect scores. The final dataset is divided into training, validation, and test sets according to conventional proportions, using simulation data throughout to ensure data accessibility and consistency.

[0048] This method introduces a large model and upgrades it from general reasoning to traffic domain-specific reasoning through domain fine-tuning, enhancing its integration and adaptation capabilities with ST-GNN. The core of this method includes three parts: model selection and modification, phased fine-tuning process, and fine-tuning optimization techniques.

[0049] The basic model is a general-purpose large model with a moderate number of parameters and strong text understanding and logical reasoning capabilities, retaining its core Transformer architecture. In this embodiment, the number of parameters is selected as 10B-30B. A spatiotemporal feature encoding module is added to the input layer of the large model. This module consists of a linear projection layer, a normalization LayerNorm layer, and a ReLU activation function. The input dimension of the linear projection layer is consistent with the dimension of the ST-GNN output fused feature vector. Through linear mapping, the high-dimensional spatiotemporal fused feature vector output by ST-GNN is directly converted into a feature vector with the same dimension as the word embedding of the large model. After normalization and activation, it is fed into the large model Transformer encoder to realize direct feature input and avoid information loss in natural language conversion. A multi-task head is added to the output layer, corresponding to road network risk scoring (i.e., regression head), secondary accident risk level (i.e., classification head), and decision suggestion generation (i.e., generation head), which accurately matches the decision-making needs in the transportation field.

[0050] A lightweight adaptation and domain knowledge injection fine-tuning mode is adopted. The fine-tuning optimization technique uses LoRA (Low-Rank Adaptation) technology, which trains only low-rank matrix parameters, compressing the number of fine-tuning parameters to 1%-5% of the original model, significantly reducing computational power consumption and training costs. It combines mixed precision training and gradient accumulation to adapt to the differences in computational power, and uses a traffic-specific word segmenter to optimize the domain data encoding effect. This word segmenter is built on corpora in the fields of traffic engineering terminology, road network names, traffic flow parameters, accident types, congestion levels, and emergency dispatch instructions. It adopts the BPE (Byte Pair Encoding) sub-word segmentation rule and expands the vocabulary of general word segmenters with traffic-specific vocabulary, such as dedicated tokens for road network segments, intersections, signal timing, traffic control, and accident levels, to avoid the incorrect segmentation of traffic professional expressions and avoid the high computational power and high time cost of full-scale fine-tuning. This fine-tuning mode is divided into three progressive implementation stages.The first stage is knowledge injection pre-training. It uses a structured traffic knowledge graph, formed by organizing rules and road network topology knowledge from the traffic engineering domain, as input. A masked language model (MLM) is used to randomly mask traffic-specific vocabulary and road network topology-related statements, allowing the model to perform prediction and reconstruction. Simultaneously, a knowledge alignment loss is introduced. By calculating the cosine distance loss between the knowledge representation output by the large model and the structured knowledge embedding vectors in the traffic knowledge graph, the large model accurately grasps core knowledge such as traffic-specific concepts, accident handling priorities, congestion diffusion thresholds, and road network topology relationships, completing the initial adaptation of the general-purpose large model to the traffic domain. The second stage is feature alignment fine-tuning. It inputs the spatiotemporal fusion feature vectors of the road network output by the ST-GNN model, along with the corresponding real road network risk scores and congestion level labels. The parameters of the bottom 70% of the Transformer base layers of the large model are strictly frozen, and only the top Transformer is trained. The layer, along with the newly added spatiotemporal feature encoding module and multi-task head, optimizes the regression prediction effect of road network risk score with mean squared error loss and optimizes the classification prediction effect of secondary accident risk level and congestion range with cross-entropy loss. During training, a multi-task loss balancing strategy is adopted. First, the mean squared error loss and cross-entropy loss are scaled to the same numerical order of magnitude using minimum-maximum normalization. Then, the two types of losses are fused using an equal-weighted method to avoid single-task loss dominating gradient updates and ensure stable convergence of multi-task collaboration. Iterative training in batches achieves precise alignment between the large model feature representation and the spatiotemporal feature output of ST-GNN, ensuring the matching and correlation of the feature dimensions of the two. The third stage is fine-tuning of the decision scenario, with input including accident scenario description, emergency dispatch decision plan, and quantitative score of handling effect. By utilizing emergency decision-making scenario case library data, a fine-tuning mode is adopted to transform case data into instruction pairs. Specifically, emergency decision-making suggestions are obtained through road network status and risk characteristics. The large model is trained by understanding and generating instructions, with the optimization objective being generative loss and effect score reward. The reward function directly uses the pre-labeled handling effect scores in the case library, normalizing them to the 0-1 range as reinforcement learning reward values. The better the solution effect, the higher the reward value. This quantified handling effect score is used as the reward value, and a reinforcement learning strategy is used to reward the model that generates high-quality decision-making suggestions. This guides the large model to deeply learn the mapping relationship between road network status, risk characteristics, decision-making schemes, and handling effects in traffic emergency scenarios, significantly improving the relevance, practicality, and implementability of the model's output emergency decision-making suggestions.To address the compatibility issue of different types of data within the same model, this invention employs a unified feature space approach to enable collaborative training of three types of data within the same large model: First, all text data are encoded using the same traffic domain-specific word segmenter to ensure a consistent language feature format; Second, the spatiotemporal features output by ST-GNN are converted into vectors with the same word embedding dimension as the large model through a unified spatiotemporal feature encoding module, thus enabling input within the same feature space as the text data.

[0051] Hybrid precision training employs a computational mode combining FP16 and FP32. FP16 is used for core computations to reduce memory usage, while FP32 is retained for critical loss calculations and gradient updates, with a gradient scaling mechanism added to prevent gradient vanishing or overflow. The cloud can flexibly switch to higher precision to fully utilize computing power. Gradient accumulation uses dynamically adjustable step sizes, accumulating multiple batches of gradients before updating parameters, ensuring stable and efficient training across the board. Simultaneously, a dedicated word segmenter for the transportation domain is built, optimizing the description of traffic parameters and road network topology based on traffic engineering specifications, road network datasets, and maintenance business texts, thereby improving the accuracy of the large model's understanding of domain data.

[0052] After fine-tuning, the large model has the following functions: Quantitative assessment of road network traffic operation risks: based on t1-t of ST-GNN input 12 Based on real-time road network status characteristics and structured knowledge, risk scoring is performed from two core dimensions: congestion spread speed and probability of secondary accidents, outputting comprehensive score and dimension score; Intelligent reasoning for traffic operation risks: Employing a causal reasoning mechanism, combining road network state characteristics with traffic rule knowledge, it identifies high-risk scenarios such as sudden speed drops and intersection queues, and outputs t1-t. 12 A list of high-risk road sections and their risk levels at different times, clearly indicating the highest level of congestion and secondary accident risk at which times and on which road sections; Precise decision-making suggestions are generated: By combining the simulation case library with the real-time road network status, targeted emergency dispatch suggestions are output, such as police deployment on high-risk road sections, alternative detour route planning, and intersection signal timing adjustment.

[0053] This invention utilizes a dynamic collaborative reasoning and result output framework to design a collaborative reasoning mechanism that enables rapid prediction at the edge and precise correction in the cloud, balancing prediction efficiency and accuracy. Simultaneously, a visualization output module is constructed to meet the practical application needs of traffic management departments. The trained ST-GNN model is deployed on roadside edge computing devices. In this embodiment, an intelligent traffic camera terminal is used. The device collects accident information and traffic flow data in real time at time t0. The accident information includes type, location, and severity, completing the data from t1 to t2. 12The system predicts road network conditions and conducts preliminary risk assessments at any given time, outputting basic prediction results. A three-factor trigger mechanism is implemented to automatically invoke a cloud-based large-scale model for fine-tuning when any of the following conditions are met: the ST-GNN prediction confidence level is lower than a set threshold; or the Scene Complexity Index (SCI) is higher than a set threshold. The SCI comprehensively considers accident severity, traffic flow density, and road network complexity, and is calculated using the following formula:

[0054] The severity coefficient of the accident is calculated as follows: (full lane blockage > multiple lane occupancy > single lane occupancy). Traffic flow density coefficient (peak > off-peak > low-peak). The road network complexity coefficient (core area > main roads > suburban roads). , , The weights of each factor satisfy the following conditions: + + =1, based on historical traffic emergency response cases, the optimal value is selected as 1. , , Users can actively adjust the settings within a reasonable range based on different city sizes, road network levels, control requirements, and time period characteristics. , , All use a level assignment within the range of 0-1. Assign values ​​of 0.3 for single-lane occupancy, 0.7 for multi-lane occupancy, and 1.0 for full-lane blockage; Values ​​are assigned based on traffic flow density: 0.2 for low peak, 0.5 for off-peak, and 1.0 for high peak. Assign values ​​based on the characteristics of the road network: 0.2 for suburban roads, 0.6 for main roads, and 1.0 for the core area road network.

[0055] The confidence-weighted fusion method is used to integrate the results from the edge and cloud, as shown in the following formula:

[0056] in, For the final prediction result, Predict confidence levels for edge devices. Confidence of large cloud-based models. For edge prediction results, This is a cloud-based prediction result.

[0057] Real-time simulation verification data or simulated real monitoring data collected by roadside equipment are used to dynamically correct prediction results at fixed intervals, updating traffic flow status and risk levels. A visualization output platform is constructed to present the results in the following form: Road network traffic flow heat map: real-time display of t1-t2. 12 The congestion level of each road segment at any given time is defined as follows: red for severe, orange for moderate, yellow for mild, and green for normal; the risk score change curve shows the time interval from t1 to t2. 12 The model displays the changing trends of the overall risk score of the road network at any time and the scores of the three dimensions; the risk probability change curve shows the evolution of congestion and secondary accident risk of high-risk road sections over time; emergency decision-making suggestions are provided by combining the large model with simulation cases and traffic rules to output targeted emergency dispatch suggestions.

[0058] like Figure 3 As shown, the horizontal axis represents the change of road network status over time. The figure uses different gray levels to represent the results of the real road network status simulated by SUMO simulation software and the prediction results of this model. The numerical fit between the two at each time point can be compared intuitively, clearly demonstrating the accurate fitting ability of the model of this invention to the time evolution law of road network status, as well as the prediction accuracy advantage compared with traditional methods.

[0059] Example 2: Consistent with Embodiment 1, the system includes a processor, an input device, an output device, and a memory, which are interconnected. The memory stores a computer program comprising program instructions, and the processor is configured to invoke these instructions. The program includes instructions for performing the following steps: S1: Construct a high-quality road network dataset for multiple scenarios; S2: Construct and train an improved spatiotemporal graph neural network (ST-GNN) prediction model to realize the mapping of road network state within a period after the accident occurs; S3: Introducing a large model and optimizing it through domain fine-tuning enhances its fusion and adaptation capabilities with the ST-GNN prediction model; S4: Deploy the trained ST-GNN model on edge computing devices and deploy the large model in the cloud to achieve rapid prediction at the edge and precise judgment and correction in the cloud, obtain dynamic collaborative reasoning results, and use the visualization output module to display risk decision-making suggestions.

[0060] This embodiment primarily describes the solution from the perspective of the method-side execution process. It is understood that, in order to achieve the aforementioned functions, the terminal includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the examples described in the embodiments provided herein, this patent can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this patent.

[0061] This embodiment can divide the terminal into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this patent embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0062] Example 3: This embodiment provides a computer-readable storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the large model-enhanced deep learning fitting methods for rapid prediction of road network traffic operation risk status as described in the above method embodiments.

[0063] Example 4: This embodiment provides a computer program product, including a computer program and a non-transitory computer-readable storage medium storing the computer program. The computer program causes a computer to perform some or all of the steps of any of the large model-enhanced deep learning fitting methods for rapid prediction of road network traffic operation risk status as described in the above method embodiments.

[0064] It should be noted that, for the sake of simplicity, the aforementioned method embodiments are all described as a series of actions. However, those skilled in the art should understand that this patent is not limited to the described order of actions, as some steps can be performed in other orders or simultaneously according to this patent. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this patent.

[0065] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0066] In the several embodiments provided in this patent, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0067] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0068] Furthermore, the functional units in the various embodiments of the application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software program module.

[0069] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0070] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0071] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0072] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0073] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0074] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0075] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0076] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for rapid prediction of road network traffic operation risk status using large-scale model-enhanced deep learning fitting, characterized by: Includes the following steps: S1: Construct a high-quality road network dataset for multiple scenarios; S2: Construct and train an improved spatiotemporal graph neural network (ST-GNN) prediction model to realize the mapping of road network state within a period after the accident occurs; S3: Expand the fine-tuning dataset, introduce large models and optimize them through domain fine-tuning to enhance the fusion and adaptation capabilities with the ST-GNN prediction model; S4: Deploy the trained ST-GNN model on edge computing devices and deploy the large model in the cloud to achieve rapid prediction at the edge and precise judgment and correction in the cloud, obtain dynamic collaborative reasoning results, and use the visualization output module to display risk decision-making suggestions.

2. The method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting as described in claim 1, characterized in that: Step S1, which involves constructing a high-quality road network dataset for multiple scenarios, includes the following steps: S11: Input the road network and traffic flow baseline simulation data into the traffic simulation software, design diverse disturbance scenarios, including various accident types, accident locations, and different traffic flow baseline time periods, and generate various typical scenario combinations; the road network simulation data includes road grade, number of lanes, intersection signal timing, speed limit information, and special road segment attributes, covering typical road network structures of different city sizes and road network densities; the traffic flow baseline simulation data includes traffic volume, vehicle speed, and density at different time periods, constructing a traffic flow baseline that conforms to actual traffic operation patterns; S12: For each scenario, use traffic simulation software to perform continuous dynamic simulation and collect road network status simulation data.

3. The method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting as described in claim 2, characterized in that: In the improved spatiotemporal graph neural network (ST-GNN) prediction model, a dynamically weighted road network structure is first constructed. The road network is abstracted into a dynamic graph structure G=(V,E,W), where node V represents a road segment or intersection; edge E represents the connectivity between road segments, connecting only adjacent road segments where traffic flow transfer exists; and edge weight W is a dynamic weight calculated from the real-time traffic flow transfer coefficient between two road segments, as shown in the following formula: in, Let t be the time from road segment i to section of road j Traffic flow, For road section i The set of adjacent road segments, k Indicates road segment i Adjacent road sections k , Let t be the time from road segment i to section of road k Traffic flow.

4. The method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting as described in claim 3, characterized in that: The improved spatiotemporal graph neural network (ST-GNN) prediction model includes a spatial feature extraction layer, a temporal dependency modeling layer, and an attention fusion module. The spatial feature extraction layer employs a graph attention network (GAT), which assigns differentiated weights to different neighboring nodes through an attention mechanism to adaptively capture the impact intensity of accident nodes on the surrounding road network. The formula is as follows: in, For road section j For the section i Attention weights For attention coefficient vectors, Each is a road segment i , j , k eigenvectors, This is the weight matrix. For road section i The new feature vector after fusion via the attention mechanism As the activation function, the spatial feature extraction layer outputs a spatial feature vector of fixed dimensions. ; The time-dependency modeling layer employs a bidirectional long short-term memory (Bi-LSTM) network to mine the time-series dependencies of traffic flow from both the past to the present and present to the future directions, capturing the periodicity and trend of congestion caused by accidents. The Bi-LSTM consists of two symmetrically configured bidirectional LSTM structures with multiple hidden layers, each containing 256 units. The activation function is tanh. The input consists of spatial features from multiple consecutive time slices, bidirectionally encoding the traffic flow sequence from both the past to the present and present to the future directions, and outputting a fixed-dimensional time feature vector. ; The attention fusion module uses a multilayer perceptron (MLP) to calculate the fusion weights of spatial and temporal features, and performs weighted fusion of the two types of features to highlight the different importance of spatial diffusion features in the early stage of an accident and temporal evolution features in the later stage. Finally, it outputs a fixed-dimensional fusion feature vector for subsequent traffic flow parameter prediction. The improved spatiotemporal graph neural network (ST-GNN) prediction model employs a multi-task training and optimization strategy to construct a multi-task loss function, simultaneously optimizing three prediction targets: vehicle speed, traffic flow, and congestion range. The multi-task loss function is as follows: in, Mean squared error loss, used for vehicle speed v ,flow q The regression prediction, in which , These represent the actual and predicted values ​​of the vehicle speed, respectively. , These represent the actual and predicted traffic flow values, respectively. The cross-entropy loss is used for classification prediction of congestion range r, which is divided into four categories: no congestion, light congestion, moderate congestion, and heavy congestion. , ω1, ω2, and ω3 represent the actual and predicted values ​​of the congestion range, respectively; ω1, ω2, and ω3 are the weight coefficients of each task, determined through experiments, and satisfy ω1 + ω2 + ω3 = 1.

5. The method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting as described in claim 4, characterized in that: The expanded fine-tuning dataset mentioned in step S3 includes: Based on the basic simulation data, specialized fine-tuning data is added, and then the final dataset is divided into training, validation, and test sets; the specialized fine-tuning data includes: Spatiotemporal feature annotation data: Extract the fused feature vector output by the ST-GNN model to represent the prediction of future road network change risks; Knowledge data in the transportation sector: Organizing rules and road network topology knowledge in the field of transportation engineering to form a structured knowledge graph; Decision-making scenario case data: Construct a case library, which includes scenario descriptions, decision-making solutions, and effect scores for each decision-making scenario; The basic simulation data is used to train the improved ST-GNN model; the spatiotemporal feature annotation dataset, the transportation domain knowledge dataset, and the decision-making scenario case dataset are used for fine-tuning of the large model domain. In step S3, the large model selected is a general-purpose model with a moderate number of parameters and strong text understanding and logical reasoning capabilities, retaining its core Transformer architecture. A spatiotemporal feature encoding module is added to the input layer of the large model. This module consists of a linear projection layer, a normalized LayerNorm layer, and a ReLU activation function. The input dimension of the linear projection layer is consistent with the dimension of the ST-GNN output fused feature vector. Through linear mapping, the high-dimensional spatiotemporal fused feature vector output by ST-GNN is directly converted into a feature vector with the same dimension as the word embedding of the large model. After normalization and activation, it is fed into the large model Transformer encoder to achieve direct feature input. A multi-task head is added to the output layer, including a regression head, a classification head, and a generation head, corresponding to road network risk scoring, secondary accident risk level, and decision suggestion generation, respectively, to match the decision-making needs of the transportation field.

6. The method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting according to claim 5, characterized in that: The domain fine-tuning described in step S3 employs a three-stage fine-tuning optimization, including the following steps: Phase 1: Knowledge Injection Pre-training. Using a structured traffic knowledge graph as input, pre-training is performed using a masked language model and knowledge alignment loss, enabling the large model to master traffic-specific concepts, rules, and road network topology associations. The second stage is feature alignment fine-tuning. Input the spatiotemporal features of ST-GNN and the corresponding label data, freeze the bottom parameters of the large model, and train only the top Transformer layer and the newly added encoding module and task head. The optimization targets are mean squared error loss and cross-entropy loss. During training, a multi-task loss balancing strategy is adopted. First, the mean squared error loss and cross-entropy loss are scaled to the same numerical order of magnitude by min-max normalization. Then, the two types of losses are fused by equal weighting to achieve accurate alignment between the large model and the feature output of ST-GNN. The third stage: fine-tuning of decision-making scenarios. Input decision-making scenario case data and adopt the instruction fine-tuning mode to guide the large model to learn the mapping relationship between state, decision and effect in traffic emergency scenarios. The optimization objective is generative loss and effect score reward. The reward function directly adopts the pre-labeled treatment effect score in the case library and normalizes it to the 0-1 range as the reinforcement learning reward value. The better the solution effect, the higher the reward value. The quantified treatment effect score is used as the reward value. A unified feature space approach is adopted to collaboratively train the three types of data in the same large model across the three stages. The text data is encoded using the same traffic domain-specific word segmenter to ensure a consistent language feature format. The spatiotemporal features output by ST-GNN are converted into vectors with the same word embedding dimension as the large model through a unified spatiotemporal feature encoding module, thus enabling input within the same feature space as the text data.

7. The method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting as described in claim 6, characterized in that: During fine-tuning and optimization, LoRA technology is used to train only low-rank matrix parameters, thus compressing the number of fine-tuning parameters. Mixed precision training and gradient accumulation are introduced to adapt to the differences in computing power between edge devices and the cloud. A traffic-specific word segmenter is constructed to optimize the encoding effect of road network names and traffic parameter-specific expressions. The traffic-specific word segmenter is built based on traffic-specific corpora and adopts the BPE sub-word segmentation rule, expanding the traffic-specific vocabulary on the basis of the general word segmenter vocabulary.

8. The method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting according to claim 7, characterized in that: After fine-tuning and optimization, the large model, based on the road network state features and structured knowledge input by ST-GNN, scores risks from two core dimensions: congestion spread speed and probability of secondary accidents, and outputs a comprehensive score and dimension scores. It adopts a causal reasoning mechanism, combining road network state features and traffic rule knowledge, to identify high-risk scenarios such as sudden speed drops and intersection queuing, and outputs a list of high-risk road segments and risk levels at each time, clarifying the degree of congestion and the highest risk of secondary accidents for each road segment at each time. By combining simulation case libraries with real-time road network conditions, targeted emergency dispatch suggestions are generated.

9. The method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting as described in claim 8, characterized in that: In step S4, the trained ST-GNN model is deployed on the roadside edge computing device. The device collects accident information and traffic flow data in real time at the moment the accident occurs, completes the prediction of road network status and preliminary risk assessment at subsequent times, and outputs the basic prediction results. A three-factor trigger mechanism is set up to call the cloud-based large model for fine-tuning when any of the following conditions are met: the ST-GNN prediction confidence is lower than a set threshold; the Scene Complexity Index (SCI) is higher than a set threshold; where SCI comprehensively considers the severity of the accident, traffic flow density, and road network complexity, and the formula is as follows: This represents the severity coefficient of the accident. Traffic flow density coefficient The road network complexity coefficient is... , , The weights for each factor are set based on historical experience to satisfy... + + =1; The confidence-weighted fusion method is used to integrate the results from the edge and cloud, as shown in the following formula: in, Predict confidence levels for edge devices. Confidence of large cloud models.

10. The method for rapid prediction of road network traffic operation risk status using large-model enhanced deep learning fitting according to claim 9, characterized in that: In step S4, the real-time simulation verification data collected by the roadside edge computing device is used to dynamically correct the prediction results at fixed intervals, and update the traffic flow status and risk level. Build a visualization output platform to present the results, including: Road network traffic flow heat map: Real-time display of congestion levels on different road sections at different times; Risk score change curve: shows the changing trends of the overall risk score of the road network and the scores of the three dimensions at each time point; Risk probability change curve: shows the evolution of congestion level and secondary accident risk over time in high-risk road sections; Emergency decision-making recommendations: Based on a large model combined with simulation cases and traffic rules, targeted emergency dispatch recommendations are generated.

Images