A traffic accident risk prediction method and system based on multi-model fusion

Abstract

This invention discloses a method and system for predicting traffic accident risks based on multi-model fusion, relating to the field of traffic risk assessment technology. The method involves acquiring traffic accident feature data to construct an original dataset; preprocessing and feature enhancement of the original dataset to transform it into standardized numerical features; training multiple base models based on the standardized numerical features to obtain multiple prediction results; dynamically fusing the multiple prediction results from the multiple base models to obtain a fused prediction probability; and converting the fused prediction probability into a specific risk level according to a preset risk threshold. This invention addresses the problems of insufficient prediction accuracy, class imbalance, and poor data adaptability of traditional single-model predictions by achieving accurate prediction of road traffic accident risks through feature engineering optimization, multi-base model collaborative training, and a dynamic weighted fusion strategy.

Description

Technical Field

[0001] This invention relates to the field of traffic risk assessment technology, and more specifically to a method and system for predicting traffic accident risks based on multi-model fusion. Background Technology

[0002] With the continuous development of urban construction and transportation systems, the operational pressure and safety hazards of road traffic are increasing, making traffic safety an unavoidable social issue. To ensure public travel safety and maintain public transportation order, traffic accident risk prediction technology has become a crucial tool and is receiving increasing attention. How to optimize traffic management and avoid personal injury and property damage caused by traffic accidents using data analysis and intelligent monitoring technologies is a key focus for traffic maintenance personnel. Accurately identifying the probability of traffic accidents on high-risk road sections and during high-risk periods can provide a scientific basis for traffic management departments to formulate targeted prevention and control strategies and optimize traffic infrastructure design, which is of significant practical importance in reducing accident rates and minimizing casualties and property damage.

[0003] Traffic accident risk prediction is essentially a high-dimensional classification problem with unevenly distributed categories. Currently, there are many methods on the market that use single machine learning models for traffic accident risk prediction, but these methods have many shortcomings: The factors that influence the occurrence of accidents are wide-ranging, including road conditions, environmental factors, traffic flow characteristics, and historical accident information. Although a single machine learning model has certain advantages in handling nonlinear relationships, its feature representation ability is still limited when faced with multiple features, and it is easily affected by noise interference. Accident data typically exhibits the characteristics of low frequency of occurrence but severe consequences, belonging to a typical class imbalance distribution. In this case, traditional methods generally cannot simultaneously achieve both prediction accuracy and model generalization ability. While deep learning models excel in feature extraction and representation learning, they often lack stability in scenarios with small sample sizes and class imbalance, and are more prone to overfitting during training.

[0004] Therefore, how to provide a traffic accident risk prediction method that can effectively alleviate class imbalance, improve feature representation capabilities, and integrate the advantages of multiple models is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a traffic accident risk prediction method and system based on multi-model fusion, which solves the problems of class imbalance, poor adaptability, insufficient feature representation and limited generalization ability in traditional methods. By using the complementarity of different models, the fusion model achieves the optimal balance in terms of accuracy, recall and robustness, thereby improving the accuracy and robustness of risk prediction.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A traffic accident risk prediction method based on multi-model fusion includes: S1. Obtain traffic accident feature data and construct the original dataset; S2. Perform data preprocessing and feature enhancement on the original dataset to transform it into standardized numerical features; S3. Train multiple base models based on the standardized numerical features to obtain multiple prediction results; S4. Dynamically fuse the multiple prediction results of the multiple base models to obtain the fused prediction probability; S5. Based on a preset risk threshold, the fusion prediction probability is converted into a specific risk level.

[0007] Optionally, the raw dataset includes detailed records of historical traffic accidents, real-time traffic data, historical and real-time meteorological data, road network and geospatial data, and vehicle and driver data.

[0008] Optionally, S2 specifically includes: S21. Data reading and cleaning: Read the original dataset, delete duplicate and irrelevant columns, and retain feature columns and label columns related to accident risk; S22. Perform deduplication and consistency processing on the cleaned data. Identify and merge duplicate records by defining a unique identifier. At the same time, unify different representations of the same concept into a standard code. S23. Use deletion or statistical methods to fill in missing values ​​in the deduplicated data; S24. Standardize or normalize the data after missing data processing; S25. Balance the category distribution of the data after processing in S24; S26. Perform feature enhancement on the data processed in S25 to obtain standardized numerical features.

[0009] Optionally, the multiple base models are TabNet (Table Attention Neural Network) model, LightGBM (Lightweight Gradient Boosting Machine) model, and XGBoost (Extreme Gradient Boosting Tree) model.

[0010] Optionally, the training process of the table attention neural network model includes: Initialize the table attention neural network model by setting the model parameters, output dimension, optimizer type, learning rate, and weights, and output the training log. The early stop strategy is used, with an early stop callback function. The validation set AUC is used as the early stop metric, the metric is maximized, and a preset patience value is set to measure the model's ability to distinguish risks. During the training of the table attention neural network model, the values ​​of maximum training rounds, batch size, and virtual batch size are set, and the training set is used to train the model, while the validation set monitors the training process. After training, save the table attention neural network model and extract the predicted probabilities of the training and validation sets; The fusion feature is constructed by concatenating the predicted probabilities of the table attention neural network model as new features with the original features. Optionally, the training process of the lightweight gradient boosting machine model includes: Parameter grid settings: Configure the hyperparameters of the search grid, including maximum depth, learning rate, number of leaves, subsample ratio, column sampling ratio, and regularization coefficient; The grid search optimization employs a 3-fold hierarchical cross-validation method, using AUC as the scoring metric, to search for the optimal hyperparameters. A lightweight gradient boosting machine model is trained using optimal parameters and fused features. An early stopping strategy is set, the lightweight gradient boosting machine model is saved, and the predicted probabilities of the validation set are obtained. Optionally, the training process of the extreme gradient boosting tree model includes: Parameter grid settings: Configure the hyperparameters of the search grid, including maximum depth, learning rate, subsample ratio, column sampling ratio, and regularization coefficient; The grid search optimization employs a 3-fold hierarchical cross-validation method, using AUC as the scoring metric, to search for the optimal hyperparameters. The extreme gradient boosting tree model is trained using optimal parameters and fused features. An early stopping strategy is set, the extreme gradient boosting tree model is saved, and the predicted probabilities of the validation set are obtained. Optionally, S4 specifically includes: S41. Obtain the predicted probabilities of the lightweight gradient boosting machine model and the extreme gradient boosting tree model based on the fused features, wherein the fused features are formed by concatenating the original features with the predicted probabilities output by the table attention neural network model. S42. Dynamic weight calculation: Calculate the AUC value of each base model on the validation set, and use the ratio of the base model AUC value to the calculated AUC values ​​of the lightweight gradient booster and extreme gradient booster tree on the validation set as the weight of the base model. S43. Calculate the dynamic weights based on the AUC corresponding to the lightweight gradient booster and the extreme gradient booster tree, and perform weighted fusion of the prediction probabilities of the two models to obtain the fused prediction probability.

[0011] A traffic accident risk prediction system based on multi-model fusion includes: The data acquisition module obtains traffic accident characteristic data and constructs the raw dataset; The data preprocessing module performs data preprocessing and feature enhancement on the original dataset, transforming it into standardized numerical features; The model training module trains multiple base models based on the standardized numerical features to obtain multiple prediction results. The model fusion module dynamically fuses the multiple prediction results of the multiple base models to obtain the fused prediction probability; The results output module converts the fusion prediction probability into a specific risk level based on a preset risk threshold.

[0012] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a traffic accident risk prediction method and system based on multi-model fusion, which has the following beneficial effects: 1. Multi-model synergy advantages: It integrates the advantages of three models: TabNet, LightGBM, and XGBoost. TabNet extracts high-dimensional features, LightGBM captures non-linear interactions, and XGBoost ensures robustness. Information complementarity between models is achieved through feature fusion. 2. Dynamic weighted fusion: Based on the model performance AUC, weights are dynamically allocated, which better reflects the differences in model advantages and improves the fusion effect compared to equal-weighted fusion. 3. Adaptive threshold optimization: Dynamically selects the optimal classification threshold based on the characteristics of data distribution, improving the model's prediction performance on imbalanced data, especially suitable for scenarios with few high-risk samples in traffic accident risk prediction. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0014] Figure 1 Flowchart of the traffic accident risk prediction method provided by the present invention; Figure 2 A comparison chart of precision, recall, and F1 score of the fusion model under different thresholds provided by this invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] This invention discloses a traffic accident risk prediction method based on multi-model fusion, such as... Figure 1 As shown, it includes: S1. Obtain traffic accident feature data and construct the original dataset; S2. Perform data preprocessing and feature enhancement on the original dataset to transform it into standardized numerical features; S3. Train multiple base models based on the standardized numerical features to obtain multiple prediction results; Specifically, three base models, TabNet, LightGBM and XGBoost, are trained respectively to fully leverage the advantages of different models in high-dimensional feature extraction, non-linear interaction capture and robust prediction. S4. Dynamically fuse the multiple prediction results of the multiple base models to obtain the fused prediction probability; S5. Based on a preset risk threshold, the fusion prediction probability is converted into a specific risk level.

[0017] Specifically, such as Figure 2 As shown, each threshold is iterated through, and the optimal threshold (which balances precision and recall and maximizes the F1 score) is selected. Samples with a fusion prediction probability greater than the optimal threshold are classified as high-risk (label 1), and those without are classified as low-risk (label 0). The fusion prediction probability is classified using a preset risk threshold to obtain the final traffic accident risk prediction result, providing a decision-making basis for upper-level applications.

[0018] In a specific embodiment, S1 acquires data. To achieve comprehensive and accurate prediction of traffic accident risks, it is first necessary to construct a rich and multi-dimensional dataset. The specific steps are as follows: S11. Obtain detailed records of traffic accidents that have occurred in the past few years. Each record should include the type of accident, the severity of the accident, the types of vehicles involved, weather conditions, road conditions, lighting conditions, etc. S12. Obtain real-time traffic data, including: average vehicle speed, traffic flow, vehicle density, lane occupancy, etc. These data can dynamically reflect the degree of road congestion and the stability of traffic flow, and are key dynamic features for predicting the probability of accidents. S13. Obtain historical and real-time meteorological data, including temperature, rainfall or snowfall, visibility, wind speed, and whether the road surface is icy. S14. Obtain road network and geospatial data, including road geometry, road class, number of lanes, speed limit, curvature, and slope. S15. Obtain vehicle and driver data: instantaneous speed, acceleration, braking force, steering wheel angle, distance to the vehicle in front, etc., and use them as important inputs for risk prediction.

[0019] In a specific embodiment, data preprocessing and feature enhancement in S2 are mainly used to thoroughly clean the collected data. The specific steps are as follows: S21. Data reading and cleaning: Read the original traffic accident dataset, delete duplicate and irrelevant columns, and retain feature columns and label columns related to accident risk; S22. Data deduplication and consistency processing: Duplicate records are identified and merged by defining a unique identifier. Simultaneously, different representations of the same concept are standardized using a unified encoding. S23. Handling Missing Values: Missing values ​​are prevalent in the data and require appropriate methods for processing. The choice of processing method depends on the proportion of missing values ​​and the nature of the features. This invention uses two methods, including: The deletion method can be used to directly delete features or samples with a very high missing rate if their information content is insufficient to support effective imputation. Statistical imputation methods can be used for numerical features, such as the mean, median, or mode. For categorical features, the mode or a specific category can be used. S24. Data standardization and normalization: The units of measurement for different features may vary greatly. In order to eliminate the influence of the units of measurement on model training, it is necessary to standardize or normalize the data. S241. Standardization transforms the data into a normal distribution with a mean of 0 and a standard deviation of 1. This method is suitable for data distributions that are approximately normal. x Represents the original data points. μ Represents the mean of the original features. σ The standard deviation of the original feature is represented by the normal distribution expression:

[0020] S242. Normalization: This method shrinks the data to the interval [0,1] or [-1,1]. It is suitable for situations where the data distribution is unknown or not normally distributed, and is particularly important for models such as neural networks. x Represents the original data points. The minimum value of the feature column is represented by max(x), and the maximum value of the feature column is represented by max(x). The scaling expression is as follows:

[0021] S25. To balance the class distribution of the training set, SMOTE oversampling is used to synthesize new high-risk samples, thereby balancing the class distribution of the training set. S26. A new feature engineering module (feature enhancement) has been added, including physical meaning cross features, target encoding features, and statistical aggregation features; In a specific embodiment, S3 trains the base models respectively, and the specific steps are as follows: S31. The TabNet model is used as the deep feature extraction model as follows: S311. Model initialization, setting model parameters as follows: , , , , The output dimension is 2, and the optimizer uses... Learning rate Weight decay ,set up Output training logs, where the relevant parameters are specifically represented as follows: and , representing the dimensions of feature transformation and attention transformation in the decision-making step, respectively. These two parameters together determine the model's capacity and complexity; This indicates the number of decision-making steps. More steps mean the model can perform deeper feature interaction reasoning, but it also increases computational cost and the risk of overfitting. gamma is a parameter that controls the sparsity of feature selection. A higher gamma value encourages the model to select only a few key features at each step; This is used to control how the features are masked at each step. The activation generation mask has the characteristic of strong sparsity; Verbose is used to control the level of detail in log output. This allows TabNet to print logs after each training epoch, making it easier to observe the results. Sparse regularization coefficients are used to further control the complexity and interpretability of the model; Optimizer: The selection of the optimizer and its parameters such as learning rate and weight decay; S312. Early stopping strategy: adopts an early stopping callback function, uses the validation set AUC as the early stopping metric, maximizes the metric, and sets the patience value to 10 to avoid model overfitting. S313. Model training: Set the maximum number of training rounds to 30, batch size to 512, virtual batch size to 128, train the model using the training set, monitor the training process using the validation set, and after training is complete, save the model and extract the predicted probabilities of the training set and validation set. S314. After training is complete, save the model and extract the predicted probabilities of the training set and validation set. S315. Construction of fused features: The predicted probabilities of TabNet are used as new features and concatenated with the original features to construct fused features, which are then used for training the LightGBM and XGBoost models. S32. The LightGBM model is used as follows: S321. Parameter grid settings: Set the hyperparameters for the search grid, including maximum depth [4,6], learning rate [0.05,0.1], number of estimators

[100] , number of leaves [15,31], subsample ratio [0.7,0.8], column sampling ratio [0.7,0.8], L1 regularization coefficient [0.1,0.5], and L2 regularization coefficient [0.1,0.5]. S322. Grid search optimization, using 3-fold hierarchical cross-validation, with AUC as the scoring metric, searches for the optimal hyperparameters. The optimal parameters are: =0.7、 =0.1、 =6、 =100、 =31、 =0.1、 =0.1、 =0.7; where the relevant parameters are specifically expressed as follows: This indicates that random sampling is performed on the feature column; The learning rate controls the step size of weight updates in each iteration. A smaller learning rate usually requires more iterations to achieve better performance. Controlling the complexity of each tree is one of the most important parameters affecting model performance; it is related to... Related, but The settings are more flexible; Limiting the maximum depth of a tree is another way to control model complexity; , indicating how many decision trees the model needs to build; , represents the L1 regularization coefficient, which makes the weights of many unimportant features directly become 0, thus preventing the model from overfitting during training. It will only reduce and smooth out the feature weights as a whole, but will not set them to zero; S323, Model training: Train the model using optimal parameters and set an early stopping strategy. Save the model and obtain the predicted probabilities of the validation set; S33. The XGBoost model is used as follows: S331. Parameter grid settings: Set the hyperparameters for the search grid, including maximum depth [3,5], learning rate [0.05,0.1], number of estimators

[100] , subsample ratio [0.7,0.8], column sampling ratio [0.7,0.8], L1 regularization coefficient [0.1,1], and L2 regularization coefficient [0.1,1]. S332, Grid Search Optimization: A 3-fold hierarchical cross-validation method is used, with AUC as the scoring metric, to search for the optimal hyperparameters. The optimal parameters are... =0.7、 =0.1、 =5、 =100、 =0.1、 =1、 =0.8; where the relevant parameters are specifically expressed as follows: This controls the depth of the tree; This indicates how many decision trees to use; Eta, the learning rate, controls the step size for updating weights in each iteration. A smaller learning rate usually requires more iterations to achieve better performance. and These correspond to LightGBM respectively. and Used for random sampling of rows and columns; alpha and lambda: are the coefficients of L1 and L2 regularization terms, respectively, and are the core parameters for controlling model complexity and preventing overfitting; This determines the sum of the sample weights of the smallest leaf node, which is used to control overfitting. S333, Model training, setting early stopping strategy ( =15), save the model and obtain the predicted probabilities of the validation set; S4. The prediction results are dynamically fused as follows: S41. Prediction Probability Acquisition: After the TabNet model completes training, its output prediction probabilities are used as new features, concatenated with the original features to form fused features, and then input into the LightGBM and XGBoost models respectively for training to obtain the prediction probabilities output by the LightGBM and XGBoost models. The prediction probabilities output by the TabNet model are only used for fused feature construction and do not participate in the final prediction result fusion. S42. Model performance evaluation: Hierarchical K-fold cross-validation is used to evaluate the generalization performance of the LightGBM model and the XGBoost model respectively. The AUC values ​​of the two models on each fold validation set are calculated, and the average AUC is used as the model performance evaluation index. S43. Dynamic weight calculation: Based on the average AUC value of each model obtained from cross-validation, this invention designs a dynamic weight calculation mechanism. Let the average values ​​of LightGBM and XGBoost be respectively... and The expression for calculating dynamic weights is:

[0022] in, It is the first i The weights of each model are determined by a normalized linear weighting method that ensures the sum of all weights is 1. Furthermore, the weight of each model is proportional to its performance on the validation set (AUC value). The better the performance of a model, the higher its weight and the greater its contribution to the final prediction result. S44. Weighted Fusion: For a new sample to be predicted, firstly, a pre-trained TabNet model is used to generate a prediction probability. This prediction probability is then concatenated with the original features to form a fused feature. This fused feature is then input into both the LightGBM and XGBoost models to obtain corresponding prediction probabilities. Finally, the prediction probabilities from the LightGBM and XGBoost models are weighted and summed according to dynamic weights to obtain the final fused prediction probability. The calculation formula is as follows: as follows:

[0023] This final The traffic accident risk probability output by the method of this invention can be converted into a specific risk level, such as high risk or low risk, based on a preset risk threshold, and corresponding warning information can be output. It can be concluded that the feature engineering of this invention can effectively improve feature discrimination, the multi-model fusion strategy can give full play to the advantages of each model, and the dynamic weight and adaptive threshold optimization further improve the prediction accuracy. This method has extremely high performance in traffic accident risk prediction and can be effectively applied to scenarios such as traffic risk prevention and control and road safety optimization.

[0024] A traffic accident risk prediction system based on multi-model fusion includes: The data acquisition module obtains traffic accident characteristic data and constructs the raw dataset; The data preprocessing module performs data preprocessing and feature enhancement on the original dataset, transforming it into standardized numerical features; The model training module trains multiple base models based on the standardized numerical features to obtain multiple prediction results. The model fusion module dynamically fuses the multiple prediction results of the multiple base models to obtain the fused prediction probability; The results output module converts the fusion prediction probability into a specific risk level based on a preset risk threshold.

[0025] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0026] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A traffic accident risk prediction method based on multi-model fusion, characterized in that, include: S1. Obtain traffic accident feature data and construct the original dataset; S2. Perform data preprocessing and feature enhancement on the original dataset to transform it into standardized numerical features; S3. Train multiple base models based on the standardized numerical features to obtain multiple prediction results; S4. Dynamically fuse the multiple prediction results of the multiple base models to obtain the fused prediction probability; S5. Based on a preset risk threshold, the fusion prediction probability is converted into a specific risk level.

2. The traffic accident risk prediction method based on multi-model fusion according to claim 1, characterized in that, The original dataset includes detailed records of historical traffic accidents, real-time traffic data, historical and real-time meteorological data, road network and geospatial data, and vehicle and driver data.

3. The traffic accident risk prediction method based on multi-model fusion according to claim 1, characterized in that, S2 specifically includes: S21. Data reading and cleaning: Read the original dataset, delete duplicate and irrelevant columns, and retain feature columns and label columns related to accident risk; S22. Perform deduplication and consistency processing on the cleaned data. Identify and merge duplicate records by defining a unique identifier. At the same time, unify different representations of the same concept into a standard code. S23. Use deletion or statistical methods to fill in missing values ​​in the deduplicated data; S24. Standardize or normalize the data after missing data processing; S25. Balance the category distribution of the data after processing in S24; S26. Perform feature enhancement on the data processed in S25 to obtain standardized numerical features.

4. The traffic accident risk prediction method based on multi-model fusion according to claim 1, characterized in that, The base models are a table attention neural network model, a lightweight gradient booster model, and an extreme gradient booster tree model.

5. The traffic accident risk prediction method based on multi-model fusion according to claim 4, characterized in that, The training process of the table attention neural network model includes: Initialize the table attention neural network model by setting the model parameters, output dimension, optimizer type, learning rate, and weights, and output the training log. The early stop strategy is used, with an early stop callback function. The validation set AUC is used as the early stop metric, the metric is maximized, and a preset patience value is set to measure the model's ability to distinguish risks. During the training of the table attention neural network model, the values ​​of maximum training rounds, batch size, and virtual batch size are set, and the training set is used to train the model, while the validation set monitors the training process. After training, save the table attention neural network model and extract the predicted probabilities of the training and validation sets; The fusion feature is constructed by taking the predicted probability of the table attention neural network model as a new feature and concatenating it with the original features.

6. The traffic accident risk prediction method based on multi-model fusion according to claim 5, characterized in that, The training process of the lightweight gradient booster model includes: Parameter grid settings: Configure the hyperparameters of the search grid, including maximum depth, learning rate, number of leaves, subsample ratio, column sampling ratio, and regularization coefficient; The grid search optimization employs a 3-fold hierarchical cross-validation method, using AUC as the scoring metric, to search for the optimal hyperparameters. A lightweight gradient booster model is trained using optimal parameters and fused features. An early stopping strategy is set, the lightweight gradient booster model is saved, and the prediction probabilities of the validation set are obtained.

7. The traffic accident risk prediction method based on multi-model fusion according to claim 5, characterized in that, The training process of the extreme gradient boosting tree model includes: Parameter grid settings: Configure the hyperparameters of the search grid, including maximum depth, learning rate, subsample ratio, column sampling ratio, and regularization coefficient; The grid search optimization employs a 3-fold hierarchical cross-validation method, using AUC as the scoring metric, to search for the optimal hyperparameters. The extreme gradient boosting tree model is trained using optimal parameters and fused features. An early stopping strategy is set, the extreme gradient boosting tree model is saved, and the predicted probabilities of the validation set are obtained.

8. The traffic accident risk prediction method based on multi-model fusion according to claim 5, characterized in that, S4 specifically includes: S41. Obtain the predicted probabilities of the lightweight gradient boosting machine model and the extreme gradient boosting tree model based on the fused features, wherein the fused features are formed by concatenating the original features with the predicted probabilities output by the table attention neural network model. S42. Dynamic weight calculation: Calculate the AUC value of each base model on the validation set, and use the ratio of the base model AUC value to the calculated AUC values ​​of the lightweight gradient booster and extreme gradient booster tree on the validation set as the weight of the base model. S43. Calculate the dynamic weights based on the AUC corresponding to the lightweight gradient booster and the extreme gradient booster tree, and perform weighted fusion of the prediction probabilities of the two models to obtain the fused prediction probability.

9. A traffic accident risk prediction system based on multi-model fusion, characterized in that, The traffic accident risk prediction method based on multi-model fusion according to any one of claims 1-8 includes: The data acquisition module obtains traffic accident characteristic data and constructs the raw dataset; The data preprocessing module performs data preprocessing and feature enhancement on the original dataset, transforming it into standardized numerical features; The model training module trains multiple base models based on the standardized numerical features to obtain multiple prediction results. The model fusion module dynamically fuses the multiple prediction results of the multiple base models to obtain the fused prediction probability; The results output module converts the fusion prediction probability into a specific risk level based on a preset risk threshold.

Images