CN-121982902-A - Road network traffic running risk situation rapid prediction method by large model reinforced deep learning fitting

CN121982902ACN 121982902 ACN121982902 ACN 121982902ACN-121982902-A

Abstract

The invention relates to a road network traffic running risk situation rapid prediction method of large model reinforcement deep learning fitting, which belongs to the technical field of traffic risk assessment and comprises the following steps of 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 to realize road network state mapping in a period after an accident occurrence moment, S3, expanding a fine adjustment data set, introducing the large model, performing field fine adjustment optimization, strengthening fusion adaptation capability with the ST-GNN prediction model, S4, deploying the trained ST-GNN model on edge computing equipment, deploying the large model on cloud, realizing edge end rapid prediction and cloud fine judgment correction, obtaining dynamic collaborative reasoning results, and displaying risk decision suggestions by using a visual output module.

Inventors

YU SHANCHUAN
JIANG WEIWEI
WEN BINGQI
LIU BIQI
SONG LANG
GOU YI
CHEN ZHEN
ZHOU JIJUN
LI YUANZHE
LI YEMING

Assignees

招商局重庆交通科研设计院有限公司

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. A road network traffic running risk situation rapid prediction method of large model reinforced deep learning fitting is characterized by comprising 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 to realize the mapping of the network state of the road in a period after the occurrence moment of the event; S3, expanding a fine tuning data set, introducing a large model, and strengthening fusion adaptation capability with the ST-GNN prediction model through field fine tuning optimization; And S4, deploying the trained ST-GNN model on edge computing equipment, deploying the large model and the cloud end, realizing quick prediction of the edge end and accurate judgment and correction of the cloud end, obtaining a dynamic collaborative reasoning result, and displaying risk decision suggestions by using a visual output module.
2. The rapid prediction method for road network traffic running risk situation by using the large model reinforced deep learning fit according to claim 1 is characterized in that the constructing of the multi-scene high-quality road network data set in the step S1 comprises the following steps: S11, inputting road network and traffic flow baseline simulation data into traffic simulation software, designing diversified disturbance scenes comprising various accident types, various accident occurrence positions and different traffic flow baseline periods, and generating various typical scene combinations, wherein the road network simulation data comprise road grades, lane numbers, intersection signal timing, speed limit information and special road section attributes, and cover typical road network structures with different city scales and different road network densities; s12, carrying out continuous dynamic simulation by using traffic simulation software aiming at each scene, and collecting road network state simulation data.
3. The rapid prediction method for road network traffic running risk situation by large model reinforcement deep learning fitting according to claim 2 is characterized in that in the improved space-time diagram neural network ST-GNN prediction model, a dynamic weighted road network diagram structure is firstly constructed, a road network is abstracted into a dynamic diagram structure G= (V, E, W), wherein a node V represents a road section or an intersection, an edge E represents a communication relation between road sections and only connects adjacent road sections with traffic flow conduction, an edge weight W is a dynamic weight, and the formula is calculated by a real-time traffic flow transfer coefficient between the two road sections: Wherein, the For the traffic flow from road segment i to road segment j at time t, For a set of adjacent segments for segment i, k represents segment k adjacent to segment i, The traffic flow from road section i to road section k at time t.
4. The road network traffic running risk situation rapid prediction method for large model reinforcement deep learning fitting according to claim 3 is characterized in that the improved space-time diagram neural network ST-GNN prediction model comprises a spatial feature extraction layer, a time dependent modeling layer and an attention fusion module; the spatial feature extraction layer adopts a graph attention network GAT, distributes differential weights for different adjacent nodes through an attention mechanism, and adaptively captures the influence intensity of the accident node on the peripheral road network, wherein the formula is as follows: Wherein, the For the attention weight of road segment j to road segment i, In order to pay attention to the coefficient vector, The feature vectors of the road segments i, j, k respectively, As a matrix of weights, the weight matrix, The new feature vector of the road section i after being fused by the attention mechanism, To activate the function, the spatial feature extraction layer outputs a spatial feature vector of a fixed dimension ; The Bi-LSTM comprises a bidirectional LSTM structure configured by two layers of symmetrical structures, a plurality of hidden layers are arranged, the number of units of each hidden layer is set to be 256, an activation function adopts tanh, spatial characteristics of a plurality of continuous time slices are input, traffic flow sequences are bidirectionally encoded from the past to the present and the present to the future respectively, and a time characteristic vector with fixed dimension is output ; The attention fusion module calculates fusion weights of spatial features and time features by adopting a multi-layer perceptron MLP, performs weighted fusion on the two types of features, highlights different importance of initial spatial diffusion features and later time evolution features of accidents, and finally outputs fusion feature vectors with fixed dimensionality for predicting subsequent traffic flow parameters; The improved space-time diagram neural network ST-GNN prediction model adopts a multi-task training and optimizing strategy to construct a multi-task loss function, and simultaneously optimizes three prediction targets of speed, flow and congestion range, wherein the multi-task loss function is as follows: Wherein, the For regression prediction of vehicle speed v, flow q, as mean square error loss, where , Respectively representing a true value and a predicted value of the vehicle speed, , Respectively representing a true value and a predicted value of the vehicle flow; for cross entropy loss, the method is used for classifying and predicting the congestion range r and is divided into congestion-free, light, moderate and heavy congestion categories, , And omega 1 、ω 2 、ω 3 is a weight coefficient of each task, and is determined through experiments, and omega 1 +ω 2 +ω 3 =1 is satisfied.
5. The rapid prediction method for road network traffic risk situation based on large model reinforcement deep learning fitting according to claim 4, wherein the expanding the fine tuning dataset in step S3 comprises: on the basis of basic simulation data, expanding fine tuning special data, and dividing a finally constructed data set into a training set, a verification set and a test set, wherein the fine tuning special data comprises the following components: Extracting fusion feature vectors output by the ST-GNN model, and representing prediction of future road network change risks; The traffic domain knowledge data is used for arranging traffic engineering domain rules and road network topology knowledge to form a structured knowledge map; The decision scene case data is used for constructing a case library, wherein the case library comprises scene descriptions, decision schemes and effect scores of all decision scenes; the space-time characteristic labeling dataset, the traffic field knowledge dataset and the decision scene case dataset are used for fine adjustment in the field of large models; the method comprises the steps of S3, selecting a general large model with moderate parameter quantity and strong text understanding and logic reasoning capability, reserving a core transducer architecture, adding a time-space feature coding module in an input layer of the large model, wherein the time-space feature coding module consists of a layer of linear projection layer, a layer of normalization LayerNorm and a ReLU activation function, keeping the input dimension of the linear projection layer consistent with the dimension of an ST-GNN output fusion feature vector, directly converting the high-dimensional time-space fusion feature vector output by the ST-GNN into the feature vector embedded in the same dimension with a large model word through linear mapping, and sending the feature vector into a large model transducer encoder after normalization and activation to realize direct feature input, and adding a multitasking head comprising a regression head, a classification head and a generation head corresponding to road network risk score, secondary accident risk level and decision suggestion generation respectively to match the decision requirement of the traffic field.
6. The rapid prediction method for road network traffic running risk situation by using large model reinforcement deep learning fitting according to claim 5, wherein the domain fine tuning in step S3 adopts three-stage fine tuning optimization, and comprises the following steps: the first stage, knowledge injection pre-training, namely pre-training by taking a structured traffic knowledge graph as input and adopting a mask language model and knowledge alignment loss, so that a large model can master the special concept, rules and road network topology associated knowledge in the traffic field; The second stage, fine tuning the characteristic alignment, inputting ST-GNN space-time characteristics and corresponding label data, freezing parameters of a bottom layer of a large model, training a top layer transducer layer, a newly added coding module and a task head, optimizing the target to mean square error loss and cross entropy loss, adopting a multi-task loss balancing strategy in training, firstly carrying out minimum-maximum normalization scaling on the mean square error loss and the cross entropy loss to the same numerical magnitude, and then fusing the two types of losses in an equal weight weighting mode to realize the accurate alignment of the large model and ST-GNN characteristic output; The third stage, fine tuning decision scene, inputting decision scene case data, adopting instruction fine tuning mode to guide large model to learn the mapping relation of state, decision and effect in traffic emergency scene, optimizing target to generate loss and effect scoring rewards, adopting the pre-marked treatment effect scoring in case library directly by rewarding function, normalizing to 0-1 interval as reinforcement learning rewards value, using quantized treatment effect scoring as rewards value as higher as scheme effect is better rewards value; And the space-time characteristics output by the ST-GNN are converted into vectors with the same embedding dimension as the large model words by a unified space-time characteristic coding module, so that the input of the text data in the same characteristic space is realized.
7. The road network traffic running risk situation rapid prediction method based on the large model reinforcement deep learning fitting of claim 6 is characterized in that a LoRA technology is adopted in fine tuning optimization, only low-rank matrix parameters are trained, the fine tuning parameters are compressed, mixed precision training and gradient accumulation are introduced, the calculation power difference between an edge end and a cloud end is adapted, a traffic domain exclusive word segmentation device is constructed, the encoding effect of road network names and traffic parameter exclusive expression is optimized, the traffic domain exclusive word segmentation device is constructed based on traffic domain corpus, BPE sub word segmentation rules are adopted, and traffic exclusive words are expanded on the basis of a universal word segmentation device vocabulary.
8. The rapid prediction method for road network traffic running risk situation by means of large model reinforcement deep learning fitting according to claim 7 is characterized in that after fine adjustment and optimization, the large model carries out risk scoring from two core dimensions of congestion diffusion speed and secondary accident occurrence probability based on road network state characteristics and structured knowledge input by ST-GNN, outputs comprehensive score and dimension score, adopts a causal reasoning mechanism, combines road network state characteristics and traffic rule knowledge, identifies vehicle speed dip, and intersections queue high risk scenes, outputs a high risk road section list and risk level at each moment, confirms the highest congestion degree and secondary accident risk of each road section at each moment, and combines a simulation case library and real-time road network state, and outputs targeted emergency scheduling suggestions.
9. The rapid prediction method for road network traffic running risk situation by using the large model enhanced deep learning fitting according to claim 8 is characterized in that in step S4, a trained ST-GNN model is deployed on road side edge computing equipment, the equipment collects accident information and traffic flow data at the moment of accident occurrence in real time, road network state prediction and preliminary risk assessment at the subsequent moment are completed, and a basic prediction result is output; Setting a three-factor trigger mechanism, and calling a cloud large model to accurately judge when any condition that the ST-GNN prediction confidence is lower than a set threshold value and the scene complexity index SCI is higher than the set threshold value is met, wherein the SCI comprehensively considers the accident severity, the traffic flow density and the road network complexity, and the formula is as follows: As a factor of the severity of the accident, Is the density coefficient of traffic flow As a coefficient of complexity of the road network, 、、 For each factor weight, according to the historical experience setting, satisfy + + =1; And integrating the edge end and the cloud end result by adopting a confidence weighting fusion method, wherein the formula is as follows: Wherein, the The confidence level is predicted for the edge-side, And the confidence is the confidence of the cloud large model.
10. The rapid prediction method for road network traffic running risk situation by using the large model reinforcement deep learning fitting according to claim 9 is characterized in that in the step S4, real-time simulation verification data collected by road side edge computing equipment dynamically corrects a prediction result according to a fixed period to update a traffic flow state and a risk level, and a visual output platform presentation result is constructed, wherein the method comprises the following steps: displaying the congestion degree of each road section at each moment in real time; the risk score change curve is used for displaying the change trend of the road network risk comprehensive score and the three-dimensional score at each moment; The risk probability change curve shows the evolution of the congestion degree of the high-risk road section and the secondary accident risk along with time; And (3) emergency decision suggestion, namely combining the simulation cases and the traffic rules by a large model, and outputting targeted emergency scheduling suggestion.

Description

Road network traffic running risk situation rapid prediction method by large model reinforced deep learning fitting Technical Field The invention belongs to the technical field of traffic risk assessment, and relates to a road network traffic running risk situation rapid prediction method by using a large model to enhance deep learning fitting. Background Road network congestion and traffic accidents have become core bottlenecks that limit the operating efficiency of urban traffic systems. Related researches show that most urban arterial road congestion is caused by sudden traffic accidents, and effective road network running risk situation prediction can support the formulation of an active traffic management and control strategy, so that secondary congestion or secondary accidents are obviously reduced, and road network congestion time is shortened. However, the current road network operation situation prediction technology still has a plurality of bottlenecks, and the real-time, accurate and three-dimensional requirements of traffic emergency management and control are difficult to meet. The current road network running risk situation prediction technology can be divided into three generations of technology routes, and each generation of technology has prominent short boards, wherein the first generation is a numerical simulation method based on traffic simulation software, takes the main traffic simulation software such as VISSIM, SUMO and the like as a core, carries out iterative computation by inputting accident scenes and road network parameters, and outputs the road network state in the subsequent period. The technology has the core defects that timeliness is extremely poor, long calculation time is required for single road network prediction for 1 hour, the golden window period for secondary traffic prevention and traffic flow-free induction is limited, simulation results often lag behind emergency requirements, timely support cannot be provided for on-site scheduling, and meanwhile, simulation software has high requirements on hardware configuration and is difficult to realize parallel calculation of a large-scale road network. The second generation is a prediction model based on traditional machine learning, and the prediction model is constructed by mining association relations between historical traffic flow data and accident data through algorithms such as Support Vector Machines (SVM), random forests and the like. The model shortens the calculation time, but has two major limitations, namely that firstly, the dual characteristics of road network space topological association and time sequence dependence cannot be captured, such as larger prediction error caused by neglecting traffic flow guiding effect of an intersection and an adjacent road section, and secondly, the generalization capability is poor, and the performance of the model trained for a specific area is suddenly reduced when the model is applied across the road section and the time period. The third generation is a prediction method based on basic deep learning, such as processing traffic data by adopting a long and short term memory network LSTM and a common graph neural network GNN. Although the technology can be preliminarily matched with the time-space data characteristic, obvious defects still exist, the LSTM model is difficult to model a complex space topological structure of a road network, the influence range of accidents on peripheral road sections cannot be accurately quantized, the ordinary GNN adopts a static diagram structure, the dynamic adjustment of road section association strength caused by the real-time change of traffic flow cannot be reflected, and the error is large in congestion diffusion prediction. From technical essence analysis, the existing scheme has the defects that firstly response time is delayed, the iterative computation characteristic of simulation software determines that the simulation software cannot meet the emergency requirement of second-level response after an accident occurs, the traditional machine learning and basic deep learning model is accelerated but still needs a certain time, a full road network real-time prediction scene is difficult to cover, secondly space-time modeling is insufficient, a modeling method with a single time dimension or a space dimension cannot completely describe congestion space diffusion and time evolution rules caused by traffic accidents, thirdly generalization capability is weak, model training depends on labeling data of a specific area, adaptability is poor when road network structures and traffic flow characteristics change, and trans-regional deployment cost is high. In summary, the prior art is difficult to meet the requirements of traffic emergency management and control on second-level response, accurate quantification and multidimensional output of road network operation situation prediction. Disclosure of Invention In view of the above, the invent