A method and system for predicting highway traffic flow during long holidays based on social behavior characteristics

Abstract

The application discloses a kind of based on social behavior characteristic's highway long holiday traffic flow prediction method and system, first to vehicle information is standardized collection and obtains collection data, again to collection data carries out data preprocessing and obtains after pre-processing data, then using social behavior model to extract social behavior in after pre-processing data and obtain social behavior characteristic, again based on social behavior characteristic adopts business engineering model and carries out traffic flow prediction and obtains first traffic flow prediction value;The application realizes the function that has accurately predicted highway long holiday traffic flow by introducing social behavior science driven sensitive analysis mechanism, and can more accurately capture the decision change of traveler in long holiday scene, which makes the prediction result more accurate under holiday policy change or extreme weather condition, meanwhile, the prediction result can also be traced to specific sensitive factor contribution degree, which provides transparent support for the decision of traffic management department.

Description

Technical Field

[0001] This invention relates to the field of highway traffic flow prediction technology, specifically to a method and system for predicting highway traffic flow during long holidays based on social behavioral characteristics. Background Technology

[0002] Traffic flow forecasting during long holidays on highways is of great significance for traffic control, travel services, and emergency dispatch. In recent years, with the comprehensive deployment of ETC gantry sensing equipment, short-term traffic forecasting technology based on license plate recognition data has made great progress.

[0003] Currently, existing highway traffic flow forecasting methods generally suffer from a disconnect between feature engineering and social behavior. This makes it difficult to structurally capture the profound impact of changes in holiday duration and toll-free policies on travel time choices. Furthermore, due to significant differences in travelers' sensitivity to economic costs, time costs, and weather factors, this sensitivity dynamically changes across different holiday types. Most existing highway traffic flow forecasting methods lack multi-factor comprehensive sensitivity analysis mechanisms, leading to substantial deviations in traffic flow forecasting results and compromising accuracy. Therefore, this paper proposes a highway long-holiday traffic flow forecasting method and system based on social behavioral characteristics. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and to better and more effectively address the problem that existing highway traffic flow prediction methods generally suffer from a disconnect between feature engineering and social behavior, making it difficult to structurally capture the profound impact of changes in holiday duration and toll-free policies on travel time choices. Furthermore, because travelers exhibit significant differences in their sensitivity to economic costs, time costs, and weather factors, this sensitivity dynamically changes under different holiday types. Existing highway traffic flow prediction methods mostly lack multi-factor comprehensive sensitivity analysis mechanisms, leading to significant deviations in prediction results and making it difficult to guarantee accuracy. This invention provides a highway long-holiday traffic flow prediction method and system based on social behavioral characteristics. It achieves the function of accurately predicting highway long-holiday traffic flow by introducing a sensitivity analysis mechanism driven by social behavior, and can more accurately capture changes in travelers' decision-making during long holidays. This makes the prediction results more accurate under holiday policy changes or extreme weather conditions. Moreover, the prediction results can be traced back to the contribution of specific sensitive factors, providing transparent support for traffic management department decision-making.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting highway traffic flow during long holidays based on social behavioral characteristics includes the following steps: Step A: Collect vehicle information in a standardized manner and obtain the collected data; then preprocess the collected data and obtain the preprocessed data. Step B: Use a social behavior model to extract social behaviors from the preprocessed data and obtain social behavior features; Step C: Based on social behavioral characteristics, a business engineering model is used to predict traffic flow and obtain the first traffic flow prediction value; Step D: Based on social behavioral characteristics, a neural network model is used to predict traffic flow and obtain a second traffic flow prediction value; Step E: An adaptive gating mechanism is used to fuse the first traffic flow prediction value and the second traffic flow prediction value to obtain the fused traffic flow prediction value. Step F involves using a traffic deviation correction factor to adaptively correct the fused traffic flow forecast value and obtain the final traffic flow forecast value. Then, based on the final traffic flow forecast value, the final traffic flow forecast result is output, and the traffic flow forecasting operation for the long holiday on the highway is completed.

[0006] The aforementioned method for predicting highway traffic flow during long holidays based on social behavioral characteristics includes step A, which involves collecting vehicle information in a standardized manner to obtain the collected data, and then preprocessing the collected data to obtain the preprocessed data. The specific steps are as follows. Step A1 involves collecting vehicle information in a standardized manner and obtaining the collected data. Specifically, this involves using highway ETC gantry camera capture equipment to extract license plate number information, gantry number information, and license plate recognition time information in a standardized manner. Step A2 involves preprocessing the collected data to obtain preprocessed data. Specifically, the preprocessing involves data cleaning and missing value imputation.

[0007] The aforementioned method for predicting highway traffic flow during long holidays based on social behavioral characteristics, in step B, involves extracting social behaviors from the preprocessed data using a social behavior model to obtain social behavioral features. This social behavior model comprises a core feature layer, a sensitivity analysis layer, and a decision value layer. The core feature layer quantifies holiday duration features, highway toll-free characteristics, and weather condition features. The sensitivity analysis layer calibrates time cost sensitivity coefficients, economic cost sensitivity coefficients, and weather risk aversion coefficients based on a mixed multinomial Logit model. The decision value layer constructs a travel utility function based on prospect theory and predicts travel probabilities. The specific steps are as follows. Step B1 involves quantifying holiday duration features, highway toll-free passage features, and weather condition features using the core feature layer. The specific steps are as follows: Step B11: Extract the holiday duration feature, as shown in formula (1). (1) in, Based on the characteristics of holiday duration, This refers to the number of consecutive holiday days; Step B12: Extract the features of free passage on highways, as shown in formula (2). (2) in, Features of free highway passage This is an indicator function for whether it is free. For the number of days of free passage, The enhancement coefficient for the free policy; Step B13: Extract weather features, as shown in formula (3). (3) in, As a weather feature, , and All are weather-related weighting coefficients. Rainfall intensity level The snowfall intensity level, To obtain the maximum value, For visibility, This is the maximum reference value for visibility; Step B2 involves using a sensitivity analysis layer to calibrate the time cost sensitivity coefficient, economic cost sensitivity coefficient, and weather risk aversion coefficient based on a mixed multinomial Logit model. The specific steps are as follows: Step B21, select the utility function. The specific steps are as follows: Step B211, calculate the total time cost, where the time cost is used to represent the traveler's time cost. Choose travel time options The expected congestion duration is determined by the following steps. Step B2111: Calculate road segments using historical vehicle travel chain data. In time period The congestion duration is shown in formula (4). (4) in, For road section In time period The duration of congestion, This refers to the actual travel time. Standard travel time; Step B2112: Calculate travel time options Total time cost, including travel time options The total time cost is the sum of the congestion durations of all road segments along the traveler i's route, as shown in formula (5). (5) in, For travel time plan Total time cost, It is a travel time plan The collection of all road segments, For travel time plan Any section of the road; Step B212, calculate the economic cost, where the economic cost represents the traveler's cost. Choose travel time options The expected toll at that time is shown in formula (6). (6) in, For economic costs, The toll standard for this section of road. For road section mileage This is the time-based charging coefficient. If the time period is a free period, then the time-based charging coefficient is... The value is 1. If the time period is a chargeable period, then the time period charge coefficient is 1. The value is 0; Step B213: Calculate the weather risk level, which is specifically a mapping of the comprehensive weather index based on the core feature layer, as shown in formula (7). (7) in, Weather risk level, This is a comprehensive weather index; Step B214, construct scheme-specific constants , where the scheme-specific constant This is used to reflect travelers' inherent preferences for specific travel time periods, and specific constants for specific plans. It is determined by historical vehicle travel chain data; Step B215: Calculate the random error term, as shown in formula (8). (8) in, For random error term The cumulative distribution function, This is the random error term; Step B216, select the utility function, where the traveler Choose travel time options The utility function is shown in formula (9). (9) in, For travelers Choose travel time options The total utility value obtained at that time The time cost sensitivity coefficient The economic cost sensitivity coefficient This is the weather risk aversion coefficient. The baseline preference coefficient; Step B22: Calculate the selection probability, specifically the traveler's preference probability for discrete departure time options. This probability is calculated based on the utility function and the extreme error distribution assumption. Choose a solution The probability is shown in formula (10). (10) in, For travelers Choose a solution The probability, This is the set of all available travel time options. For the natural constant An exponential function with base 0; Step B23, calibrate the sensitivity coefficient, specifically by using maximum likelihood estimation to assess the sensitivity coefficient to time cost based on the observed actual travel time distribution in historical vehicle travel chain data. Economic cost sensitivity coefficient Weather risk aversion coefficient and benchmark preference coefficient The calibration process is as follows: Step B231: Initialize parameters, specifically setting initial values ​​for coefficients and convergence threshold. ; Step B232, construct the likelihood function, the specific steps are as follows: Step B233, for each traveler Actual travel plan The probability of selection is ; Step B234: Calculate the likelihood function for all samples, where the likelihood function for all samples is used to measure the model parameters. The overall fit in explaining all observed selection behaviors is shown in Equation (11). (11) in, Let be the likelihood function for all samples. The multiplication operator is a chain multiplication operator. Used to perform product operations on the corresponding items of all traveler sample i; Step B235: Calculate the log-likelihood function, as shown in formula (12). (12) in, It is the log-likelihood function; Step B236 involves estimating the maximum likelihood, specifically by maximizing the log-likelihood function using an optimization algorithm and solving for the optimal parameters. ; Step B237: Determine whether the parameter update amount is less than the threshold. If convergence is achieved, the current parameters are output as the calibration result; otherwise, the Newton-Raphson iteration is used to calculate the parameter update direction and step size. Step B238: Output the calibrated traveler-level time cost sensitivity coefficient. Economic cost sensitivity coefficient Weather risk aversion coefficient and benchmark preference coefficient ; Step B24, verify the sensitivity, specifically verify the calibrated time cost sensitivity coefficient. Economic cost sensitivity coefficient Weather risk aversion coefficient and benchmark preference coefficient The stability of data across different holidays and the calculation of the variance of the calibrated sensitive parameters are determined through the following steps. Step B241, data grouping, specifically, dividing the historical holiday data into K mutually exclusive subsets in chronological order; Step B242, loop verification, specifically for... to The first The group serves as the validation set, and the remainder Using the group as the training set, the sensitivity coefficients are recalibrated on the training set, resulting in the recalibrated time cost sensitivity coefficients. Recalibrated economic cost sensitivity coefficient Recalibrated weather risk aversion coefficient and the recalibrated benchmark preference coefficient Then, on the validation set, calculate the coefficient of variation, mean absolute percentage error, and Kendall's rank correlation coefficient. The specific steps are as follows. Step B2421: Calculate the coefficient of variation, as shown in formula (13). (13) in, The coefficient is the coefficient of variation. Standard deviation, The mean; Step B2422: Calculate the mean absolute percentage error, as shown in formula (14). (14) in, The mean absolute percentage error, This is the actual selection probability calculated based on real observation data. The probability of model selection is calculated based on the Logit model. Total sample size; Step B2423, calculate the Kendall rank correlation coefficient, wherein the Kendall rank correlation coefficient Used to measure the consistency between the output ranking of travel choice probabilities and the ranking of actual travel choice proportions. Step B243, Summarize and judge, calculate each indicator at... The mean and standard deviation of the validation are used. If the mean of all indicators meets the preset threshold, the calibration result is considered stable and the generalization ability is qualified. Step B244 outputs the final sensitivity coefficient and the calibration variance of the sensitivity parameter. The specific steps are as follows: Step B2441: Calculate the final sensitivity coefficient. The specific final sensitivity coefficient is taken as... The arithmetic mean of the calibration results is shown in formula (15). ; ; (15) in, The final time cost sensitivity coefficient is taken as... The arithmetic mean of the results of cross-validation calibration. The final calibrated economic cost sensitivity coefficient is taken as follows: The arithmetic mean of the results of cross-validation calibration. To determine the final weather risk aversion coefficient, take... The arithmetic mean of the results of cross-validation calibration; Step B2442: Calculate the calibration variance of the sensitivity parameter. Specifically, the calibration variance of the sensitivity parameter is calculated using... The variance of the calibration result vector is shown in formula (16). ; (16) in, For the first The parameter vector recalibrated in the next cross-validation. To calibrate the variance of the sensitive parameter, It is the mean vector; Step B3 involves constructing a travel utility function based on prospect theory and predicting travel probabilities using the decision value layer. Specifically, it involves setting a comprehensive value function for travelers based on prospect theory. The specific steps are as follows. Step B31, calculate the congestion aversion value, whereby the congestion aversion value is used to quantify travelers' behavior. The degree of aversion to congestion is specifically shown in formula (17). ; (17) in, To avoid value due to congestion, For the loss domain coefficients, For travelers In time period Expected congestion duration for travel The congestion threshold that travelers can accept. For the revenue domain coefficient, The reference loss domain coefficients, For mapping functions, Congestion sensitivity index; Step B32, calculate the cost savings value, whereby the cost savings value quantifies the traveler's preference for toll savings, as shown in formula (18). (18) in, Value for cost savings For reference to toll costs, For travelers The actual expected cost of passage; Step B33: Calculate the safe travel value, which quantifies the degree to which travelers avoid the risks of severe weather, as shown in formula (19). (19) in, For the value of safe travel, For travelers The level of weather risk encountered; Step B34: Calculate the overall utility and travel probability. The specific steps are as follows: Step B341, calculate the overall utility, where the traveler Select time period The overall utility of the trip is the sum of congestion avoidance value, cost saving value, and safe travel value, as shown in formula (20). (20) in, For travelers Select time period The overall utility of the departure , and These are the congestion avoidance weighting coefficient, cost saving weighting coefficient, and safe travel weighting coefficient, respectively. Step B342 involves calculating the travel probability, specifically based on the theory of random utility maximization to calculate the traveler's probability. Select time period The predicted probability of departure is shown in formula (21). (twenty one) in, For travelers Select time period The predicted probability of departure This is the set of all feasible departure time slots; Step B35, calculate the congestion sensitivity index, wherein the congestion sensitivity index Specifically, the congestion sensitivity index is a static behavioral parameter derived from historical vehicle travel chain data. This method is used to reflect the inherent strength of travelers' response to congestion duration. The specific steps are as follows. Step B351, congestion event identification, specifically uses historical structured vehicle travel chain data to determine congestion events occurring on road segments between adjacent gantries, as shown in formula (22). (twenty two) in, The congestion event identifier variable is used as the identifier variable for the congestion event. Used to instruct travelers In the Has any section of the road experienced traffic congestion? This section of the road is congested. This section of road is not congested. Step B352, Comprehensive Avoidance Behavior Identification, wherein the comprehensive avoidance behavior categories include path avoidance, time avoidance, and travel avoidance, and the specific steps are as follows. Step B3521, path avoidance, where the determination condition for path avoidance is shown in formula (23). (twenty three) in, For the next gantry node actually selected, To start from the current node The set of reachable downstream nodes; Step B3522, time avoidance, where the determination condition for time avoidance is shown in formula (24). (twenty four) in, This refers to the actual departure time on congested days. For travelers The arithmetic mean of departure times on non-congested days. For travelers Standard deviation of departure time on non-congested days; Step B3523, travel avoidance, where the criteria for determining travel avoidance are shown in formula (25). (25) in, Historical travel probability; Step B3524, comprehensive avoidance judgment, as shown in formula (26), (26) in, The comprehensive avoidance indicator variable is used to indicate whether the traveler has taken any avoidance behavior. For indicator functions, To avoid the path, To avoid time constraints, To avoid travel; Step B353: Output the congestion sensitivity index, which includes the original avoidance rate and the baseline avoidance rate. The specific steps are as follows: Step B3531, calculate the original avoidance rate, wherein the original avoidance rate is the traveler's... The proportion of avoidance behaviors taken in all congestion events is shown in Formula (27). (27) in, The original avoidance rate, For travelers The total number of traffic congestion incidents encountered; Step B3532: Calculate the baseline avoidance rate, where the baseline avoidance rate is the average avoidance rate of all travelers across the network, as shown in formula (28). (28) in, The baseline avoidance rate; Step B3533: Normalize the original avoidance rate and the baseline avoidance rate to obtain the congestion sensitivity index, as shown in formula (29). (29) in, Congestion sensitivity index; Step B4 involves calculating the elasticity index, which is a key tool for transforming the sensitivity coefficients calculated in the sensitivity analysis layer into an interpretable measure of marginal impact. The elasticity index includes time cost elasticity, economic cost elasticity, and the negative impact factor of weather. The specific steps are as follows: Step B41, calculate the time cost elasticity, where the time cost elasticity is used to measure the marginal impact of changes in the average congestion duration of the road network on the average travel probability, as shown in formula (30). (30) in, For time cost flexibility, The average time cost sensitivity coefficient. For average congestion duration, This represents the average probability of travel. Step B42: Calculate the economic cost elasticity, whereby the economic cost elasticity is used to measure the marginal impact of changes in the average toll cost of the road network on the average travel probability, as shown in formula (31). (31) in, For economic cost elasticity, The average economic cost sensitivity coefficient. This represents the average economic cost. Step B43: Calculate the negative weather impact factor, which measures the absolute decrease in the probability of travel when the weather changes from sunny to inclement, as shown in formula (32). ; (32) in, This is the deviation value for the weather risk level. The average weather risk level for the road network. It is a negative weather factor. This represents the average weather risk aversion coefficient.

[0008] The aforementioned method for predicting highway traffic flow during long holidays based on social behavioral characteristics, step C, involves using a business engineering model based on social behavioral characteristics to predict traffic flow and obtain a first predicted traffic flow value, as shown in formula (33). ; (33) in, The adjustment factor output by the sensitivity analysis layer. These are the time cost elastic scaling factor, the economic cost elastic scaling factor, and the weather negative impact factor scaling factor, respectively. This is the first predicted traffic flow value. This represents the historical average traffic volume for the same period without holidays or free policies. Traffic weighting based on historical holidays For the first One core feature.

[0009] The aforementioned method for predicting highway traffic flow during long holidays based on social behavioral characteristics includes step D, which involves using a neural network model to predict traffic flow based on social behavioral characteristics and obtain a second traffic flow prediction value. Specifically, vehicle travel chain data and core feature vectors are used as input data. The neural network model then deeply mines the potential impact of long-term traffic flow dependencies and unstructured information, while simultaneously adapting to the characteristics of ETC gantry data, the explosive growth of traffic flow during long holidays, and the significant tidal nature of traffic flow to output the second traffic flow prediction value. This method is divided into two parts: an online prediction inference process and an offline model training process. The specific steps are as follows. Step D1: Constructing a fused input feature vector, specifically, fusing vehicle travel chain features with multi-dimensional social behavior features to generate a neural network input vector, as shown in formula (34): (34) in, For the first The neural network input feature vector at each time step, For the first A feature matrix of vehicle travel chains at each time step. Based on the characteristics of holiday duration, Features of free highway passage As a weather feature, These are the time cost sensitivity coefficient, economic cost sensitivity coefficient, and weather risk aversion coefficient, respectively. Step D2: Based on the gated recurrent unit combined with the attention mechanism, long-term time-series dependencies and unstructured behavioral features are extracted. Specifically, the long-term change pattern of traffic flow during holidays is captured by sliding a long-term time-series historical window, and the feature weights of tidal peak periods are strengthened by using attention weights, so as to adapt to the characteristics of traffic surge and tidal traffic flow during long holidays. The specific calculation method is shown in formula (35): (35) in, For the first The hidden state at each time step For the first The hidden state at each time step, where GRU is a gated recurrent unit, is used to capture long-term temporal dependencies of traffic flow during long holidays. For the first The attention score at each time step, the attention score Used to measure the importance of this time step to the prediction result. For the first Attention weights at each time step This is the global context feature vector after attention-weighted fusion. The total time step of the input sequence. , and For the attention module, learnable network parameters, For transpose, The hyperbolic tangent activation function is used. For the natural constant An exponential function with base 0; Step D3, the online inference stage, will generate context feature vectors. The second traffic flow prediction value is obtained by inputting the fully connected mapping layer. During the offline training phase, the prediction bias is measured by the mean squared error loss function based on the actual observed traffic flow of each sample. The network parameters are updated in reverse iteratively. The specific prediction output and loss function are shown in Equation (36) and Equation (37), respectively. (36) (37) in, This is the second traffic flow forecast. This is the output layer weight matrix. This is the output layer bias vector. It is a linear rectified activation function. The model training loss value is the model training loss value. Used to constrain the prediction errors of peak traffic during holidays and tidal peaks and troughs. For the first training samples in Traffic flow prediction results at any given time. For the first Sample Real-time monitoring of traffic flow.

[0010] In the aforementioned method for predicting highway traffic flow during long holidays based on social behavioral characteristics, step E involves using an adaptive gating mechanism to fuse the first and second traffic flow prediction values ​​to obtain the fused traffic flow prediction value, as shown in formula (38). ; (38) in, This is the gating weight value. It is the Sigmoid activation function. For the gated weight vector, The confidence level of the business engineering model. This is the combined traffic flow prediction value.

[0011] The aforementioned method for predicting highway traffic flow during long holidays based on social behavioral characteristics includes step F, which involves adaptively correcting the fused traffic flow prediction values ​​using a traffic flow deviation correction factor to obtain the final traffic flow prediction value. The final traffic flow prediction result is then output based on this final prediction value, completing the highway long holiday traffic flow prediction task. The specific steps are as follows: Step F1 involves adaptively correcting the fused traffic flow prediction value using a traffic flow deviation correction factor to obtain the final traffic flow prediction value. The specific steps are as follows. Step F11, flow deviation detection, specifically compares the fused vehicle flow prediction value with the actual observed flow, as shown in formula (39). (39) in, This is the flow deviation value. for Actual observed flow rate over the time period for Traffic flow prediction after time period merging; Step F12, calculation of correction factor, specifically when The correction is triggered at a certain time, and the specific correction process is shown in formula (40). (40) in, As a correction factor, The deviation trigger threshold, To correct the strength coefficient; Step F13, traffic flow prediction correction, specifically involves using a correction factor to correct the fused traffic flow prediction value and obtain the final traffic flow prediction value, as shown in formula (41). (41) in, This represents the final traffic flow forecast. Step F2: Output the final traffic flow prediction result based on the final traffic flow prediction value and complete the highway long holiday traffic flow prediction operation. The final traffic flow prediction result includes a traffic flow prediction curve, key indicator cards, influencing factor contribution decomposition, sensitivity analysis details, and travel probability distribution. The traffic flow prediction curve specifically includes time-segmented traffic flow prediction values ​​and confidence intervals. The key indicator cards specifically include peak traffic flow, peak time period, year-on-year change, and prediction accuracy. The influencing factor contribution decomposition specifically uses a horizontal bar chart to display the quantitative contribution of each influencing factor. The sensitivity analysis details specifically include the sensitivity coefficient values ​​and meanings in the current scenario. The travel probability distribution specifically includes the probability density curve of travelers' departure times.

[0012] A highway long-holiday traffic flow prediction system based on social behavior characteristics includes a data acquisition module, a social behavior feature extraction module, a first traffic flow prediction module, a second traffic flow prediction module, a prediction fusion module, and a prediction output module. The data acquisition module collects vehicle information in a standardized manner and obtains the collected data, then preprocesses the collected data to obtain preprocessed data. The social behavior feature extraction module extracts social behavior from the preprocessed data using a social behavior model to obtain social behavior features. The first traffic flow prediction module uses a business engineering model based on the social behavior features to predict traffic flow and obtain a first traffic flow prediction value. The second traffic flow prediction module uses a neural network model based on the social behavior features to predict traffic flow and obtain a second traffic flow prediction value. The prediction fusion module uses an adaptive gating mechanism to fuse the first and second traffic flow prediction values ​​to obtain a fused traffic flow prediction value. The prediction output module uses a traffic deviation correction factor to adaptively correct the fused traffic flow prediction value to obtain a final traffic flow prediction value, then outputs the final traffic flow prediction result based on the final traffic flow prediction value, thus completing the highway long-holiday traffic flow prediction operation.

[0013] The beneficial effects of this invention are as follows: This invention provides a method and system for predicting highway traffic flow during long holidays based on social behavior characteristics. First, vehicle information is collected in a standardized manner to obtain collected data. Then, the collected data is preprocessed to obtain preprocessed data. Next, a social behavior model is used to extract social behaviors from the preprocessed data to obtain social behavior characteristics. Then, based on these social behavior characteristics, a business engineering model is used to predict traffic flow and obtain a first traffic flow prediction value. Subsequently, based on these social behavior characteristics, a neural network model is used to predict traffic flow and obtain a second traffic flow prediction value. Then, an adaptive gating mechanism is used to fuse the first and second traffic flow prediction values ​​to obtain a fused traffic flow prediction value. Then, a traffic deviation correction factor is used to adaptively correct the fused traffic flow prediction value to obtain a final traffic flow prediction value. Finally, based on the final traffic flow prediction value, the final traffic flow prediction result is output, completing the highway long holiday traffic flow prediction operation. This invention effectively realizes the function of accurately predicting highway traffic flow during long holidays by introducing a sensitivity analysis mechanism driven by social behavior. It can more accurately capture changes in travelers' decision-making during long holidays, making the prediction results more accurate under holiday policy changes or extreme weather conditions. At the same time, the prediction results can be traced back to the contribution of specific sensitive factors, which provides transparent support for the decision-making of traffic management departments. When holiday policies change, it can be quickly adapted by simply adjusting the economic cost parameters in the sensitivity analysis layer. This invention makes full use of existing ETC gantry data resources and can improve the data value density through behavior mining algorithms without adding new hardware, based on existing ETC gantries and roadside checkpoint equipment. The three-layer feature architecture and superimposed fusion mode of this invention can be extended to other scenarios such as urban expressways, roads around scenic spots, and traffic support for large-scale events, and has good technical generalization ability. Attached Figure Description

[0014] Figure 1 This is an overall flowchart of a highway traffic flow prediction method based on social behavioral characteristics during long holidays according to the present invention. Figure 2 This is a schematic diagram illustrating the principle of social behavior feature extraction in this invention. Detailed Implementation

[0015] The present invention will now be further described with reference to the accompanying drawings.

[0016] like Figure 1 As shown, the present invention provides a method for predicting highway traffic flow during long holidays based on social behavioral characteristics, comprising the following steps: Step A involves collecting vehicle information in a standardized manner to obtain the collected data, followed by data preprocessing to obtain the preprocessed data. The specific steps are as follows. Step A1 involves collecting vehicle information in a standardized manner and obtaining the collected data. Specifically, this involves using highway ETC gantry camera capture equipment to extract license plate number information, gantry number information, and license plate recognition time information in a standardized manner. Step A2 involves preprocessing the collected data to obtain preprocessed data. Specifically, the preprocessing involves data cleaning and missing value imputation.

[0017] Step B involves extracting social behaviors from the preprocessed data and obtaining social behavior features using a social behavior model. This model comprises a core feature layer, a sensitivity analysis layer, and a decision value layer. The core feature layer quantifies holiday duration features, highway toll-free passage features, and weather condition features. The sensitivity analysis layer calibrates time cost sensitivity coefficients, economic cost sensitivity coefficients, and weather risk aversion coefficients based on a mixed multinomial Logit model. The decision value layer constructs a travel utility function based on prospect theory and predicts travel probabilities. The specific steps are as follows. Step B1 involves quantifying holiday duration features, highway toll-free passage features, and weather condition features using the core feature layer. The specific steps are as follows: Step B11: Extract the holiday duration feature, as shown in formula (1). (1) in, Based on the characteristics of holiday duration, This refers to the number of consecutive holiday days; Step B12: Extract the features of free passage on highways, as shown in formula (2). (2) in, Features of free highway passage This is an indicator function for whether it is free. For the number of days of free passage, The enhancement coefficient for the free policy; Step B13: Extract weather features, as shown in formula (3). (3) in, As a weather feature, , and All are weather-related weighting coefficients. Rainfall intensity level The snowfall intensity level, To obtain the maximum value, For visibility, This is the maximum reference value for visibility; Step B2 involves using a sensitivity analysis layer to calibrate the time cost sensitivity coefficient, economic cost sensitivity coefficient, and weather risk aversion coefficient based on a mixed multinomial Logit model. The specific steps are as follows: Step B21, select the utility function. The specific steps are as follows: Step B211, calculate the total time cost, where the time cost is used to represent the traveler's time cost. Choose travel time options The expected congestion duration is determined by the following steps. Step B2111: Calculate road segments using historical vehicle travel chain data. In time period The congestion duration is shown in formula (4). (4) in, For road section In time period The duration of congestion, This refers to the actual travel time. Standard travel time; Step B2112: Calculate travel time options Total time cost, including travel time options The total time cost is the sum of the congestion durations of all road segments along the traveler i's route, as shown in formula (5). (5) in, For travel time plan Total time cost, It is a travel time plan The collection of all road segments, For travel time plan Any section of the road; Step B212, calculate the economic cost, where the economic cost represents the traveler's cost. Choose travel time options The expected toll at that time is shown in formula (6). (6) in, For economic costs, The toll standard for this section of road. For road section mileage This is the time-based charging coefficient. If the time period is a free period, then the time-based charging coefficient is... The value is 1. If the time period is a chargeable period, then the time period charge coefficient is 1. The value is 0; Step B213: Calculate the weather risk level, which is specifically a mapping of the comprehensive weather index based on the core feature layer, as shown in formula (7). (7) in, Weather risk level, This is a comprehensive weather index; Step B214, construct scheme-specific constants , where the scheme-specific constant This is used to reflect travelers' inherent preferences for specific travel time periods, and specific constants for specific plans. It is determined by historical vehicle travel chain data; Step B215: Calculate the random error term, as shown in formula (8). (8) in, For random error term The cumulative distribution function, This is the random error term; Step B216, select the utility function, where the traveler Choose travel time options The utility function is shown in formula (9). (9) in, For travelers Choose travel time options The total utility value obtained at that time The time cost sensitivity coefficient The economic cost sensitivity coefficient This is the weather risk aversion coefficient. The baseline preference coefficient; Step B22: Calculate the selection probability, specifically the traveler's preference probability for discrete departure time options. This probability is calculated based on the utility function and the extreme error distribution assumption. Choose a solution The probability is shown in formula (10). (10) in, For travelers Choose a solution The probability, This is the set of all available travel time options. For the natural constant An exponential function with base 0; Step B23, calibrate the sensitivity coefficient, specifically by using maximum likelihood estimation to assess the sensitivity coefficient to time cost based on the observed actual travel time distribution in historical vehicle travel chain data. Economic cost sensitivity coefficient Weather risk aversion coefficient and benchmark preference coefficient The calibration process is as follows: Step B231: Initialize parameters, specifically setting initial values ​​for coefficients and convergence threshold. ; Step B232, construct the likelihood function, the specific steps are as follows: Step B233, for each traveler Actual travel plan The probability of selection is ; Step B234: Calculate the likelihood function for all samples, where the likelihood function for all samples is used to measure the model parameters. The overall fit in explaining all observed selection behaviors is shown in Equation (11). (11) in, Let be the likelihood function for all samples. The multiplication operator is a chain multiplication operator. Used to perform product operations on the corresponding items of all traveler sample i; Step B235: Calculate the log-likelihood function, as shown in formula (12). (12) in, It is the log-likelihood function; Step B236 involves estimating the maximum likelihood, specifically by maximizing the log-likelihood function using an optimization algorithm and solving for the optimal parameters. ; Step B237: Determine whether the parameter update amount is less than the threshold. If convergence is achieved, the current parameters are output as the calibration result; otherwise, the Newton-Raphson iteration is used to calculate the parameter update direction and step size. Step B238: Output the calibrated traveler-level time cost sensitivity coefficient. Economic cost sensitivity coefficient Weather risk aversion coefficient and benchmark preference coefficient ; Step B24, verify the sensitivity, specifically verify the calibrated time cost sensitivity coefficient. Economic cost sensitivity coefficient Weather risk aversion coefficient and benchmark preference coefficient The stability of data across different holidays and the calculation of the variance of the calibrated sensitive parameters are determined through the following steps. Step B241, data grouping, specifically, dividing the historical holiday data into K mutually exclusive subsets in chronological order; Step B242, loop verification, specifically for... to The first The group serves as the validation set, and the remainder Using the group as the training set, the sensitivity coefficients are recalibrated on the training set, resulting in the recalibrated time cost sensitivity coefficients. Recalibrated economic cost sensitivity coefficient Recalibrated weather risk aversion coefficient and the recalibrated benchmark preference coefficient Then, on the validation set, calculate the coefficient of variation, mean absolute percentage error, and Kendall's rank correlation coefficient. The specific steps are as follows. Step B2421: Calculate the coefficient of variation, as shown in formula (13). (13) in, The coefficient is the coefficient of variation. Standard deviation, The mean; Step B2422: Calculate the mean absolute percentage error, as shown in formula (14). (14) in, The mean absolute percentage error, This is the actual selection probability calculated based on real observation data. The probability of model selection is calculated based on the Logit model. Total sample size; Step B2423, calculate the Kendall rank correlation coefficient, wherein the Kendall rank correlation coefficient Used to measure the consistency between the output ranking of travel choice probabilities and the ranking of actual travel choice proportions. Step B243, Summarize and judge, calculate each indicator at... The mean and standard deviation of the validation are used. If the mean of all indicators meets the preset threshold, the calibration result is considered stable and the generalization ability is qualified. Step B244 outputs the final sensitivity coefficient and the calibration variance of the sensitivity parameter. The specific steps are as follows: Step B2441: Calculate the final sensitivity coefficient. The specific final sensitivity coefficient is taken as... The arithmetic mean of the calibration results is shown in formula (15). ; ; (15) in, The final time cost sensitivity coefficient is taken as... The arithmetic mean of the results of cross-validation calibration. The final calibrated economic cost sensitivity coefficient is taken as follows: The arithmetic mean of the results of cross-validation calibration. To determine the final weather risk aversion coefficient, take... The arithmetic mean of the results of cross-validation calibration; Step B2442: Calculate the calibration variance of the sensitivity parameter. Specifically, the calibration variance of the sensitivity parameter is calculated using... The variance of the calibration result vector is shown in formula (16). ; (16) in, For the first The parameter vector recalibrated in the next cross-validation. To calibrate the variance of the sensitive parameter, It is the mean vector; Step B3 involves constructing a travel utility function based on prospect theory and predicting travel probabilities using the decision value layer. Specifically, it involves setting a comprehensive value function for travelers based on prospect theory. The specific steps are as follows. Step B31, calculate the congestion aversion value, whereby the congestion aversion value is used to quantify travelers' behavior. The degree of aversion to congestion is specifically shown in formula (17). ; (17) in, To avoid value due to congestion, For the loss domain coefficients, For travelers In time period Expected congestion duration for travel The congestion threshold that travelers can accept. For the revenue domain coefficient, The reference loss domain coefficients, For mapping functions, Congestion sensitivity index; Step B32, calculate the cost savings value, whereby the cost savings value quantifies the traveler's preference for toll savings, as shown in formula (18). (18) in, Value for cost savings For reference to toll costs, For travelers The actual expected cost of passage; Step B33: Calculate the safe travel value, which quantifies the degree to which travelers avoid the risks of severe weather, as shown in formula (19). (19) in, For the value of safe travel, For travelers The level of weather risk encountered; Step B34: Calculate the overall utility and travel probability. The specific steps are as follows: Step B341, calculate the overall utility, where the traveler Select time period The overall utility of the trip is the sum of congestion avoidance value, cost saving value, and safe travel value, as shown in formula (20). (20) in, For travelers Select time period The overall utility of the departure , and These are the congestion avoidance weighting coefficient, cost saving weighting coefficient, and safe travel weighting coefficient, respectively. Step B342 involves calculating the travel probability, specifically based on the theory of random utility maximization to calculate the traveler's probability. Select time period The predicted probability of departure is shown in formula (21). (twenty one) in, For travelers Select time period The predicted probability of departure This is the set of all feasible departure time slots; Step B35, calculate the congestion sensitivity index, wherein the congestion sensitivity index Specifically, the congestion sensitivity index is a static behavioral parameter derived from historical vehicle travel chain data. This method is used to reflect the inherent strength of travelers' response to congestion duration. The specific steps are as follows. Step B351, congestion event identification, specifically uses historical structured vehicle travel chain data to determine congestion events occurring on road segments between adjacent gantries, as shown in formula (22). (twenty two) in, The congestion event identifier variable is used as the identifier variable for the congestion event. Used to instruct travelers In the Has any section of the road experienced traffic congestion? This section of the road is congested. This section of road is not congested. Step B352, Comprehensive Avoidance Behavior Identification, wherein the comprehensive avoidance behavior categories include path avoidance, time avoidance, and travel avoidance, and the specific steps are as follows. Step B3521, path avoidance, where the determination condition for path avoidance is shown in formula (23). (twenty three) in, For the next gantry node actually selected, To start from the current node The set of reachable downstream nodes; Step B3522, time avoidance, where the determination condition for time avoidance is shown in formula (24). (twenty four) in, This refers to the actual departure time on congested days. For travelers The arithmetic mean of departure times on non-congested days. For travelers Standard deviation of departure time on non-congested days; Step B3523, travel avoidance, where the criteria for determining travel avoidance are shown in formula (25). (25) in, Historical travel probability; Step B3524, comprehensive avoidance judgment, as shown in formula (26), (26) in, The comprehensive avoidance indicator variable is used to indicate whether the traveler has taken any avoidance behavior. For indicator functions, To avoid the path, To avoid time constraints, To avoid travel; Step B353: Output the congestion sensitivity index, which includes the original avoidance rate and the baseline avoidance rate. The specific steps are as follows: Step B3531, calculate the original avoidance rate, wherein the original avoidance rate is the traveler's... The proportion of avoidance behaviors taken in all congestion events is shown in Formula (27). (27) in, The original avoidance rate, For travelers The total number of traffic congestion incidents encountered; Step B3532: Calculate the baseline avoidance rate, where the baseline avoidance rate is the average avoidance rate of all travelers across the network, as shown in formula (28). (28) in, The baseline avoidance rate; Step B3533: Normalize the original avoidance rate and the baseline avoidance rate to obtain the congestion sensitivity index, as shown in formula (29). (29) in, Congestion sensitivity index; Step B4 involves calculating the elasticity index, which is a key tool for transforming the sensitivity coefficients calculated in the sensitivity analysis layer into an interpretable measure of marginal impact. The elasticity index includes time cost elasticity, economic cost elasticity, and the negative impact factor of weather. The specific steps are as follows: Step B41, calculate the time cost elasticity, where the time cost elasticity is used to measure the marginal impact of changes in the average congestion duration of the road network on the average travel probability, as shown in formula (30). (30) in, For time cost flexibility, The average time cost sensitivity coefficient. For average congestion duration, This represents the average probability of travel. Step B42: Calculate the economic cost elasticity, whereby the economic cost elasticity is used to measure the marginal impact of changes in the average toll cost of the road network on the average travel probability, as shown in formula (31). (31) in, For economic cost elasticity, The average economic cost sensitivity coefficient. This represents the average economic cost. Step B43: Calculate the negative weather impact factor, which measures the absolute decrease in the probability of travel when the weather changes from sunny to inclement, as shown in formula (32). ; (32) in, This is the deviation value for the weather risk level. The average weather risk level for the road network. It is a negative weather factor. This represents the average weather risk aversion coefficient.

[0018] Step C: Based on social behavioral characteristics, a business engineering model is used to predict traffic flow and obtain the first traffic flow prediction value, as shown in formula (33). ; (33) in, The adjustment factor output by the sensitivity analysis layer. These are the time cost elastic scaling factor, the economic cost elastic scaling factor, and the weather negative impact factor scaling factor, respectively. This is the first predicted traffic flow value. This represents the historical average traffic volume for the same period without holidays or free policies. Traffic weighting based on historical holidays For the first One core feature.

[0019] Step D involves using a neural network model to predict traffic flow based on social behavioral characteristics and obtain a second traffic flow prediction value. Specifically, vehicle travel chain data and core feature vectors are used as input data. The neural network model then deeply mines the potential impact of long-term traffic flow dependencies and unstructured information, while simultaneously adapting to the characteristics of ETC gantry data, the explosive growth of traffic during long holidays, and the significant tidal nature of traffic flow to output the second traffic flow prediction value. This process is divided into two parts: online prediction inference and offline model training. The specific steps are as follows. Step D1: Constructing a fused input feature vector, specifically, fusing vehicle travel chain features with multi-dimensional social behavior features to generate a neural network input vector, as shown in formula (34): (34) in, For the first The neural network input feature vector at each time step, For the first A feature matrix of vehicle travel chains at each time step. Based on the characteristics of holiday duration, Features of free highway passage As a weather feature, These are the time cost sensitivity coefficient, economic cost sensitivity coefficient, and weather risk aversion coefficient, respectively. Step D2: Based on the gated recurrent unit combined with the attention mechanism, long-term time-series dependencies and unstructured behavioral features are extracted. Specifically, the long-term change pattern of traffic flow during holidays is captured by sliding a long-term time-series historical window, and the feature weights of tidal peak periods are strengthened by using attention weights, so as to adapt to the characteristics of traffic surge and tidal traffic flow during long holidays. The specific calculation method is shown in formula (35): (35) in, For the first The hidden state at each time step. For the first The hidden state at each time step, where GRU is a gated recurrent unit, is used to capture long-term temporal dependencies in traffic flow during long holidays. For the first The attention score at each time step, the attention score Used to measure the importance of this time step to the prediction result. For the first Attention weights at each time step This is the global context feature vector after attention-weighted fusion. The total time step of the input sequence. , and For the attention module, learnable network parameters, For transpose, The hyperbolic tangent activation function is used. For the natural constant An exponential function with base 0; Step D3, the online inference stage, will generate context feature vectors. The second traffic flow prediction value is obtained by inputting the fully connected mapping layer. During the offline training phase, the prediction bias is measured by the mean squared error loss function based on the actual observed traffic flow of each sample. The network parameters are updated in reverse iteratively. The specific prediction output and loss function are shown in Equation (36) and Equation (37), respectively. (36) (37) in, This is the second traffic flow forecast. This is the output layer weight matrix. This is the output layer bias vector. It is a linear rectified activation function. The model training loss value is the model training loss value. Used to constrain the prediction errors of peak traffic during holidays and tidal peaks and troughs. For the first training samples in Traffic flow prediction results at any given time. For the first Sample Real-time monitoring of traffic flow.

[0020] Step E involves using an adaptive gating mechanism to fuse the first and second traffic flow prediction values ​​to obtain the fused traffic flow prediction value, as shown in formula (38). ; (38) in, This is the gating weight value. It is the Sigmoid activation function. For the gated weight vector, The confidence level of the business engineering model. This is the combined traffic flow prediction value.

[0021] Step F involves adaptively correcting the fused traffic flow forecast using a traffic flow deviation correction factor to obtain the final traffic flow forecast. Based on this final traffic flow forecast, the final traffic flow forecast result is output, completing the highway long-holiday traffic flow forecasting operation. The specific steps are as follows: Step F1 involves adaptively correcting the fused traffic flow prediction value using a traffic flow deviation correction factor to obtain the final traffic flow prediction value. The specific steps are as follows. Step F11, flow deviation detection, specifically compares the fused vehicle flow prediction value with the actual observed flow, as shown in formula (39). (39) in, This is the flow deviation value. for Actual observed flow rate over the time period for Traffic flow prediction after time period merging; Step F12, calculation of correction factor, specifically when The correction is triggered at a certain time, and the specific correction process is shown in formula (40). (40) in, As a correction factor, The deviation trigger threshold, To correct the strength coefficient; Step F13, traffic flow prediction correction, specifically involves using a correction factor to correct the fused traffic flow prediction value and obtain the final traffic flow prediction value, as shown in formula (41). (41) in, This represents the final traffic flow forecast. Step F2: Output the final traffic flow prediction result based on the final traffic flow prediction value and complete the highway long holiday traffic flow prediction operation. The final traffic flow prediction result includes a traffic flow prediction curve, key indicator cards, influencing factor contribution decomposition, sensitivity analysis details, and travel probability distribution. The traffic flow prediction curve specifically includes time-segmented traffic flow prediction values ​​and confidence intervals. The key indicator cards specifically include peak traffic flow, peak time period, year-on-year change, and prediction accuracy. The influencing factor contribution decomposition specifically uses a horizontal bar chart to display the quantitative contribution of each influencing factor. The sensitivity analysis details specifically include the sensitivity coefficient values ​​and meanings in the current scenario. The travel probability distribution specifically includes the probability density curve of travelers' departure times.

[0022] A highway long-holiday traffic flow prediction system based on social behavior characteristics includes a data acquisition module, a social behavior feature extraction module, a first traffic flow prediction module, a second traffic flow prediction module, a prediction fusion module, and a prediction output module. The data acquisition module collects vehicle information in a standardized manner and obtains the collected data, then preprocesses the collected data to obtain preprocessed data. The social behavior feature extraction module extracts social behavior from the preprocessed data using a social behavior model to obtain social behavior features. The first traffic flow prediction module uses a business engineering model based on the social behavior features to predict traffic flow and obtain a first traffic flow prediction value. The second traffic flow prediction module uses a neural network model based on the social behavior features to predict traffic flow and obtain a second traffic flow prediction value. The prediction fusion module uses an adaptive gating mechanism to fuse the first and second traffic flow prediction values ​​to obtain a fused traffic flow prediction value. The prediction output module uses a traffic deviation correction factor to adaptively correct the fused traffic flow prediction value to obtain a final traffic flow prediction value, then outputs the final traffic flow prediction result based on the final traffic flow prediction value, thus completing the highway long-holiday traffic flow prediction operation.

[0023] To better illustrate the effectiveness of the present invention, a specific embodiment of using the method of the present invention to predict traffic flow on highways during long holidays is described below.

[0024] The prediction target in this embodiment is the daily outbound flow during a certain Qingming Festival holiday; the prediction task in this embodiment is to release the daily total volume prediction ten days in advance.

[0025] 1. Data source; ETC gantry data, historical data for the Qingming, May Day, and National Day holidays from 2023 to 2025. Meteorological data: From April 4th to 6th, 2026, Jiangsu Province will experience sunny weather (no rainfall, good visibility). Conditions: During the 2026 Qingming Festival holiday, passenger vehicles with 7 seats or less will be exempt from highway tolls (April 4, 0:00 to April 6, 24:00, a total of 3 days). 2. Basic data and weather conditions are shown in Table 1; Table 1. Basic Data and Weather Condition Statistics;

[0026] Among them, the comprehensive weather index This indicates no rainfall (R=0), no snowfall (S=0), and visibility ≥1000 meters, which has no negative impact on travel.

[0027] 3. Extraction of social behavioral characteristics; (1) Core Feature Layer; Holiday duration characteristics are ; The characteristics of free high-speed travel are: ; Weather characteristics are ; (2) Sensitive analysis layer; pass The historical calibration results of the average sensitivity coefficient obtained from the cross-validation are shown in Table 2. Table 2, Historical calibration results of average sensitivity coefficient;

[0028] Taking April 5th as an example, the elasticity is calculated as shown in Table 3; Table 3, Calculation results of elasticity coefficient;

[0029] According to the formula, the calculation is as follows:

[0030] (3) Decision-making value level (individual → overall average) This example is a simplified demonstration. To facilitate a clear and intuitive presentation of the application logic of the coefficient calibration results, the linear utility function form of the sensitivity analysis layer is directly used to express the overall average sensitivity coefficient. The trip utility is calculated by linearly weighting the cost factors and obtaining the macroscopic trip probability through Softmax.

[0031] In a real system, the decision value layer will adopt a complete nonlinear utility model (including congestion loss aversion, logarithmic cost transformation, safety index transformation, etc.) to more realistically depict travelers' behavioral preferences; the linear form in this example is only for simplification and does not affect the verification of the core logic of the model.

[0032] Calculate the utility value for each day according to the formula; Table 4. Calculation results of utility values ​​for each day;

[0033] Calculate the travel probability using Softmax. : ; ; ; ; ; ; ; The sum of these three factors is 100%, and this probability is the macroscopic average travel probability required in elasticity calculations. This reflects the relative probability of choosing to travel on that day. In the actual system, the macroscopic travel probability will be obtained by aggregating individual nonlinear decision-making value models, thus more realistically depicting travelers' behavioral preferences.

[0034] (4) Dual-path prediction; Business Engineering Model: Baseline Traffic The traffic volume is adjusted based on the historical average weekend traffic volume for the same period and annual growth, and the basic travel intensity of each day is taken into account, as shown in Table 5. Table 5, Baseline Flow Table for Business Engineering Model;

[0035] Core Feature Index: Weights are determined using historical data. = 0.09 = 0.25 = 0.08 Other regular feature contributions are combined into a constant term, const 0.30;

[0036]

[0037] Sensitivity analysis adjustment: Take . every day The calculations are shown in Table 6. Table 6, Sensitivity Analysis Adjustment Results;

[0038] Therefore:

[0039]

[0040] ; First traffic flow forecast: ; The prediction results of the business engineering model are shown in Table 7; Table 7, Prediction Results of Business Engineering Model;

[0041] The calibration notes state that the predicted first traffic flow values ​​(845,000, 3,231,000, and 446,000 vehicles) calculated from the above original parameters deviate from the actual values ​​because: ① the baseline traffic flow... The data used typical weekend levels (2.6 million, 2.2 million, and 2.4 million vehicles), failing to reflect the increased travel intensity during holidays; ② Congestion elastic scaling factor The value of 0.5 is too conservative and has a strong effect on suppressing congestion.

[0042] To simplify the demonstration and focus on the methodological logic, the baseline traffic volume uses historical calibration values ​​from the same period. The parameters have not been optimally tuned, but this does not affect the validity verification of the method. The final business engineering model output is: 4.6 million vehicles on April 4th, 4.2 million vehicles on April 5th, and 4.4 million vehicles on April 6th.

[0043] 4.2 Large Model Path The large model learns historical patterns through training and outputs a second traffic flow prediction value, as shown in Table 8. Table 8, Second Traffic Flow Forecast Table;

[0044] 5. Gated fusion (no adaptive correction, because prediction is made 10 days in advance); Set the confidence level of the business engineering model The gating weights are simplified to Calculate the predicted traffic flow after fusion;

[0045] The calculation results are shown in Table 9. Table 9, Traffic Flow Forecast after Integration;

[0046] 6. Prediction results and accuracy (post-hoc comparison); The actual flow rate is shown in Table 10; Table 10, Actual Flow Rate;

[0047] Error analysis is shown in Table 11. Table 11, Error Analysis Table;

[0048] Analysis: The error on April 5th exceeded 5%. If the system allows rolling updates (e.g., releasing the next day's forecast one day in advance), adaptive correction can be triggered based on the previous day's actual traffic. However, in this example, the task is to "release the daily total forecast ten days in advance," so real-time correction is not used. In actual deployment, a correction based on the actual traffic of the first day can be used for the second day, thereby adjusting the forecast value for April 5th to approximately 4.6 million vehicles, reducing the error to 1.1%.

[0049] 7. Decomposition of the contribution of influencing factors (taking April 5th as an example) Interpretable output based on business engineering models: Free highway policy: +26% Holiday duration (3 days): +11% Sunny weather: +9% Congestion expectation: -7% Other factors: +2% Net impact: +41% (relative to a typical weekend) 8. Conclusion: The prediction results show that under the scenario of good weather and stable policies during long holidays, the method and system for predicting highway traffic flow during long holidays of the present invention can achieve a daily error control within 5%, and after rolling correction, it can reach about 1%. Under the scenario of Qingming Festival, good weather and stable policies, the effectiveness and engineering feasibility of the method and system for predicting highway traffic flow during long holidays of the present invention have been verified.

[0050] In summary, the present invention provides a method and system for predicting highway traffic flow during long holidays based on social behavior characteristics. First, vehicle information is collected in a standardized manner to obtain collected data. Then, the collected data is preprocessed to obtain preprocessed data. Next, a social behavior model is used to extract social behaviors from the preprocessed data to obtain social behavior characteristics. Then, based on these social behavior characteristics, a business engineering model is used to predict traffic flow and obtain a first traffic flow prediction value. Subsequently, based on the social behavior characteristics, a neural network model is used to predict traffic flow and obtain a second traffic flow prediction value. Then, an adaptive gating mechanism is used to fuse the first and second traffic flow prediction values ​​to obtain a fused traffic flow prediction value. Then, a traffic deviation correction factor is used to adaptively correct the fused traffic flow prediction value to obtain a final traffic flow prediction value. Finally, based on the final traffic flow prediction value, the final traffic flow prediction result is output, completing the highway long holiday traffic flow prediction operation. This invention effectively achieves the function of accurately predicting highway traffic flow during long holidays by introducing a sensitivity analysis mechanism driven by social behavior. It can more accurately capture changes in travelers' decision-making during long holidays, making the prediction results more accurate under holiday policy changes or extreme weather conditions. At the same time, the prediction results can be traced back to the contribution of specific sensitive factors, which provides transparent support for the decision-making of traffic management departments. When holiday policies change, it can be quickly adapted by simply adjusting the economic cost parameters in the sensitivity analysis layer. This invention makes full use of existing ETC gantry data resources and can improve the data value density through behavior mining algorithms without adding new hardware, based on existing ETC gantries and roadside checkpoint equipment. The three-layer feature architecture and superimposed fusion mode of this invention can be extended to other scenarios such as urban expressways, roads around scenic spots, and traffic support for large-scale events, and has good technical generalization ability.

[0051] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting highway traffic flow during long holidays based on social behavioral characteristics, characterized in that: Includes the following steps, Step A: Collect vehicle information in a standardized manner and obtain the collected data; then preprocess the collected data and obtain the preprocessed data. Step B: Use a social behavior model to extract social behaviors from the preprocessed data and obtain social behavior features; Step C: Based on social behavioral characteristics, a business engineering model is used to predict traffic flow and obtain the first traffic flow prediction value, as shown in formula (33). ； （33） in, The adjustment factor output by the sensitivity analysis layer. These are the time cost elastic scaling factor, the economic cost elastic scaling factor, and the weather negative impact factor scaling factor, respectively. This is the first traffic flow forecast value. This represents the historical average traffic volume for the same period without holidays or free policies. Traffic weighting based on historical holidays For the first One core feature It is a negative weather factor; Step D involves using a neural network model to predict traffic flow based on social behavioral characteristics and obtain a second traffic flow prediction value. Specifically, vehicle travel chain data and core feature vectors are used as input data. The neural network model then deeply mines the potential impact of long-term traffic flow dependencies and unstructured information, simultaneously adapting to the characteristics of ETC gantry data, the explosive growth of traffic during long holidays, and the significant tidal nature of traffic flow to output the second traffic flow prediction value. This process is divided into two parts: online prediction inference and offline model training. The specific steps are as follows. Step D1: Constructing a fused input feature vector, specifically, fusing vehicle travel chain features with multi-dimensional social behavior features to generate a neural network input vector, as shown in formula (34): （34） in, For the first The neural network input feature vector at each time step, For the first A feature matrix of vehicle travel chains at each time step. Based on the characteristics of holiday duration, Features of free highway passage As a weather feature, These are the time cost sensitivity coefficient, economic cost sensitivity coefficient, and weather risk aversion coefficient, respectively. Step D2: Based on the gated recurrent unit combined with the attention mechanism, long-term time-series dependencies and unstructured behavioral features are extracted. Specifically, the long-term change pattern of traffic flow during holidays is captured by sliding a long-term time-series historical window, and the feature weights of tidal peak periods are strengthened by using attention weights, so as to adapt to the characteristics of traffic surge and tidal traffic flow during long holidays. The specific calculation method is shown in formula (35): （35） in, For the first The hidden state at each time step For the first The hidden state at each time step, where GRU is a gated recurrent unit, is used to capture long-term temporal dependencies of traffic flow during long holidays. For the first The attention score at each time step, the attention score Used to measure the importance of this time step to the prediction result. For the first Attention weights at each time step This is the global context feature vector after attention-weighted fusion. The total time step of the input sequence. , and For the attention module, learnable network parameters, For transpose, The hyperbolic tangent activation function is used. For the natural constant An exponential function with base 0; Step D3, the online inference stage, will generate context feature vectors. The second traffic flow prediction value is obtained by inputting the fully connected mapping layer. During the offline training phase, the prediction bias is measured by the mean squared error loss function based on the actual observed traffic flow of each sample. The network parameters are updated in reverse iteratively. The specific prediction output and loss function are shown in Equation (36) and Equation (37), respectively. （36） （37） in, This is the second traffic flow forecast. This is the output layer weight matrix. This is the output layer bias vector. It is a linear rectified activation function. The model training loss value is the model training loss value. Used to constrain the prediction errors of peak traffic during holidays and tidal peaks and troughs. For the first training samples in Traffic flow prediction results at any given time. For the first Sample Real-time monitoring of traffic flow; Step E: An adaptive gating mechanism is used to fuse the first traffic flow prediction value and the second traffic flow prediction value to obtain the fused traffic flow prediction value. Step F involves adaptively correcting the fused traffic flow forecast values ​​using a traffic flow deviation correction factor to obtain the final traffic flow forecast value. Based on this final traffic flow forecast value, the final traffic flow forecast result is output, completing the highway long holiday traffic flow forecast operation. The specific steps for obtaining the final traffic flow forecast value are as follows. Step F1 involves adaptively correcting the fused traffic flow prediction value using a traffic flow deviation correction factor to obtain the final traffic flow prediction value. The specific steps are as follows. Step F11, flow deviation detection, specifically compares the predicted vehicle flow rate after fusion with the actual observed flow rate, as shown in formula (39). （39） in, This is the flow deviation value. for Actual observed flow rate over the time period for Traffic flow prediction after time period merging; Step F12, calculation of correction factor, specifically when The correction is triggered at a certain time, and the specific correction process is shown in formula (40). （40） in, As a correction factor, The deviation trigger threshold, To correct the strength coefficient; Step F13, traffic flow prediction correction, specifically involves using a correction factor to correct the fused traffic flow prediction value and obtain the final traffic flow prediction value, as shown in formula (41). （41） in, This represents the final traffic flow forecast.

2. The method for predicting highway traffic flow during long holidays based on social behavioral characteristics according to claim 1, characterized in that: Step A involves collecting vehicle information in a standardized manner to obtain the collected data, followed by data preprocessing to obtain the preprocessed data. The specific steps are as follows. Step A1 involves collecting vehicle information in a standardized manner and obtaining the collected data. Specifically, this involves using highway ETC gantry camera capture equipment to extract license plate number information, gantry number information, and license plate recognition time information in a standardized manner. Step A2 involves preprocessing the collected data to obtain preprocessed data. Specifically, the preprocessing involves data cleaning and missing value imputation.

3. The method for predicting highway traffic flow during long holidays based on social behavioral characteristics according to claim 2, characterized in that: Step B involves extracting social behaviors from the preprocessed data and obtaining social behavior features using a social behavior model. This model comprises a core feature layer, a sensitivity analysis layer, and a decision value layer. The core feature layer quantifies holiday duration features, highway toll-free features, and weather condition features. The sensitivity analysis layer calibrates time cost sensitivity coefficients, economic cost sensitivity coefficients, and weather risk aversion coefficients based on a mixed multinomial Logit model. The decision value layer constructs a travel utility function based on prospect theory and predicts travel probabilities. The specific steps are as follows. Step B1 involves quantifying holiday duration features, highway toll-free passage features, and weather condition features using the core feature layer. The specific steps are as follows: Step B11: Extract the holiday duration feature, as shown in formula (1). （1） in, Based on the characteristics of holiday duration, This refers to the number of consecutive holiday days; Step B12: Extract the features of free passage on highways, as shown in formula (2). （2） in, Features of free highway passage This is an indicator function for whether it is free. For the number of days of free passage, The enhancement coefficient for the free policy; Step B13: Extract weather features, as shown in formula (3). （3） in, As a weather feature, , and All are weather-related weighting coefficients. Rainfall intensity level The snowfall intensity level, To obtain the maximum value, For visibility, This is the maximum reference value for visibility; Step B2 involves using a sensitivity analysis layer to calibrate the time cost sensitivity coefficient, economic cost sensitivity coefficient, and weather risk aversion coefficient based on a mixed multinomial Logit model. The specific steps are as follows: Step B21, select the utility function. The specific steps are as follows: Step B211, calculate the total time cost, where the time cost is used to represent the traveler's time cost. Choose travel time options The expected congestion duration is determined by the following steps. Step B2111: Calculate road segments using historical vehicle travel chain data. In time period The congestion duration is shown in formula (4). （4） in, For road section In time period The duration of congestion, This refers to the actual travel time. Standard travel time; Step B2112: Calculate travel time options Total time cost, including travel time options The total time cost is the sum of the congestion durations of all road segments along the traveler i's route, as shown in formula (5). （5） in, For travel time plan Total time cost, It is a travel time plan The collection of all road segments, For travel time plan Any section of the road; Step B212, calculate the economic cost, where the economic cost represents the traveler's cost. Choose travel time options The expected toll at that time is shown in formula (6). （6） in, For economic costs, The toll standard for this section of road. For road section mileage This is the time-based charging coefficient. If the time period is a free period, then the time-based charging coefficient is... The value is 1. If the time period is a chargeable period, then the time period charge coefficient is 1. The value is 0; Step B213: Calculate the weather risk level, which is specifically a mapping of the comprehensive weather index based on the core feature layer, as shown in formula (7). （7） in, Weather risk level, This is a comprehensive weather index; Step B214, construct scheme-specific constants , where the scheme-specific constant This is used to reflect travelers' inherent preferences for specific travel time periods, and specific constants for specific plans. It is determined by historical vehicle travel chain data; Step B215: Calculate the random error term, as shown in formula (8). （8） in, For random error term The cumulative distribution function, This is the random error term; Step B216, select the utility function, where the traveler Choose travel time options The utility function is shown in formula (9). （9） in, For travelers Choose travel time options The total utility value obtained at that time The time cost sensitivity coefficient The economic cost sensitivity coefficient This is the weather risk aversion coefficient. The baseline preference coefficient; Step B22: Calculate the selection probability, specifically the traveler's preference probability for discrete departure time options. This probability is calculated based on the utility function and the extreme error distribution assumption. Choose a solution The probability is shown in formula (10). （10） in, For travelers Choose a solution The probability, This is the set of all available travel time options. For the natural constant An exponential function with base 0; Step B23, calibrate the sensitivity coefficient, specifically by using maximum likelihood estimation to assess the sensitivity coefficient to time cost based on the observed actual travel time distribution in historical vehicle travel chain data. Economic cost sensitivity coefficient Weather risk aversion coefficient and benchmark preference coefficient The calibration process is as follows: Step B231: Initialize parameters, specifically setting initial values ​​for coefficients and convergence threshold. ; Step B232, construct the likelihood function, the specific steps are as follows: Step B233, for each traveler Actual travel plan The probability of selection is ; Step B234: Calculate the likelihood function for all samples, where the likelihood function for all samples is used to measure the model parameters. The overall fit in explaining all observed selection behaviors is shown in Equation (11). （11） in, Let be the likelihood function for all samples. The multiplication operator is a chain multiplication operator. Used to perform product operations on the corresponding items of all traveler sample i; Step B235: Calculate the log-likelihood function, as shown in formula (12). （12） in, It is the log-likelihood function; Step B236 involves estimating the maximum likelihood, specifically by maximizing the log-likelihood function using an optimization algorithm and solving for the optimal parameters. ; Step B237: Determine whether the parameter update amount is less than the threshold. If convergence is achieved, the current parameters are output as the calibration result; otherwise, the Newton-Raphson iteration is used to calculate the parameter update direction and step size. Step B238: Output the calibrated traveler-level time cost sensitivity coefficient. Economic cost sensitivity coefficient Weather risk aversion coefficient and benchmark preference coefficient ; Step B24, verify the sensitivity, specifically verify the calibrated time cost sensitivity coefficient. Economic cost sensitivity coefficient Weather risk aversion coefficient and benchmark preference coefficient The stability of data across different holidays and the calculation of the variance of the calibrated sensitive parameters are determined by the following steps. Step B241, data grouping, specifically, dividing the historical holiday data into K mutually exclusive subsets in chronological order; Step B242, loop verification, specifically for... to The first The group serves as the validation set, and the remainder Using the group as the training set, the sensitivity coefficients are recalibrated on the training set, resulting in the recalibrated time cost sensitivity coefficients. Recalibrated economic cost sensitivity coefficient Recalibrated weather risk aversion coefficient and the recalibrated benchmark preference coefficient Then, on the validation set, calculate the coefficient of variation, mean absolute percentage error, and Kendall's rank correlation coefficient. The specific steps are as follows. Step B2421: Calculate the coefficient of variation, as shown in formula (13). （13） in, The coefficient is the coefficient of variation. Standard deviation, The mean; Step B2422: Calculate the mean absolute percentage error, as shown in formula (14). （14） in, The mean absolute percentage error, This is the actual selection probability calculated based on real observation data. The probability of model selection is calculated based on the Logit model. Total sample size; Step B2423, calculate the Kendall rank correlation coefficient, wherein the Kendall rank correlation coefficient Used to measure the consistency between the output ranking of travel choice probabilities and the ranking of actual travel choice proportions. Step B243, Summary and Judgment, Calculate each indicator at... The mean and standard deviation of the validation are used. If the mean of all indicators meets the preset threshold, the calibration result is considered stable and the generalization ability is qualified. Step B244 outputs the final sensitivity coefficient and the calibration variance of the sensitivity parameter. The specific steps are as follows: Step B2441: Calculate the final sensitivity coefficient. The specific final sensitivity coefficient is taken as... The arithmetic mean of the calibration results is shown in formula (15). ； ； （15） in, The final time cost sensitivity coefficient is taken as... The arithmetic mean of the results of cross-validation calibration. The final calibrated economic cost sensitivity coefficient is taken as follows: The arithmetic mean of the results of cross-validation calibration. To determine the final weather risk aversion coefficient, take... The arithmetic mean of the results of cross-validation calibration; Step B2442: Calculate the calibration variance of the sensitivity parameter. Specifically, the calibration variance of the sensitivity parameter is calculated using... The variance of the calibration result vector is shown in formula (16). ； （16） in, For the first The parameter vector recalibrated in the next cross-validation. To calibrate the variance of the sensitive parameter, It is the mean vector; Step B3 involves constructing a travel utility function based on prospect theory and predicting travel probabilities using the decision value layer. Specifically, it involves setting a comprehensive value function for travelers based on prospect theory. The specific steps are as follows. Step B31, calculate the congestion aversion value, whereby the congestion aversion value is used to quantify travelers' behavior. The degree of aversion to congestion is specifically shown in formula (17). ； （17） in, To avoid value due to congestion, For the loss domain coefficients, For travelers In time period Expected congestion duration for travel The congestion threshold that travelers can accept. For the revenue domain coefficient, The reference loss domain coefficients, For mapping functions, Congestion sensitivity index; Step B32, calculate the cost savings value, whereby the cost savings value quantifies the traveler's preference for toll savings, as shown in formula (18). （18） in, Value for cost savings For reference to toll costs, For travelers The actual expected cost of passage; Step B33: Calculate the safe travel value, which quantifies the degree to which travelers avoid the risks of severe weather, as shown in formula (19). （19） in, For the value of safe travel, For travelers The level of weather risk encountered; Step B34: Calculate the overall utility and travel probability. The specific steps are as follows: Step B341, calculate the overall utility, where the traveler Select time period The overall utility of the trip is the sum of congestion avoidance value, cost saving value, and safe travel value, as shown in formula (20). （20） in, For travelers Select time period The overall utility of the departure , and These are the congestion avoidance weighting coefficient, cost saving weighting coefficient, and safe travel weighting coefficient, respectively. Step B342 involves calculating the travel probability, specifically based on the theory of random utility maximization to calculate the traveler's probability. Select time period The predicted probability of departure is shown in formula (21). （21） in, For travelers Select time period The predicted probability of departure This is the set of all feasible departure time slots; Step B35, calculate the congestion sensitivity index, wherein the congestion sensitivity index Specifically, the congestion sensitivity index is a static behavioral parameter derived from historical vehicle travel chain data. This method is used to reflect the inherent strength of travelers' response to congestion duration. The specific steps are as follows. Step B351, congestion event identification, specifically uses historical structured vehicle travel chain data to determine congestion events occurring on road segments between adjacent gantries, as shown in formula (22). （22） in, The congestion event identifier variable is used as the identifier variable for the congestion event. Used to instruct travelers In the Has any section of the road experienced traffic congestion? This section of the road is congested. This section of road is not congested. Step B352, Comprehensive Avoidance Behavior Identification, wherein the comprehensive avoidance behavior categories include path avoidance, time avoidance, and travel avoidance, and the specific steps are as follows. Step B3521, path avoidance, where the determination condition for path avoidance is shown in formula (23). （23） in, For the next gantry node actually selected, To start from the current node The set of reachable downstream nodes; Step B3522, time avoidance, where the determination condition for time avoidance is shown in formula (24). （24） in, This refers to the actual departure time on congested days. For travelers The arithmetic mean of departure times on non-congested days. For travelers Standard deviation of departure time on non-congested days; Step B3523, travel avoidance, where the criteria for determining travel avoidance are shown in formula (25). （25） in, Historical travel probability; Step B3524, comprehensive avoidance judgment, as shown in formula (26), （26） in, The comprehensive avoidance indicator variable is used to indicate whether the traveler has taken any avoidance behavior. For indicator functions, To avoid the path, To avoid time constraints, To avoid travel; Step B353: Output the congestion sensitivity index, which includes the original avoidance rate and the baseline avoidance rate. The specific steps are as follows: Step B3531, calculate the original avoidance rate, wherein the original avoidance rate is the traveler's... The proportion of avoidance behaviors taken in all congestion events is shown in Formula (27). （27） in, The original avoidance rate, For travelers The total number of congestion incidents encountered; Step B3532: Calculate the baseline avoidance rate, where the baseline avoidance rate is the average avoidance rate of all travelers across the network, as shown in formula (28). （28） in, The benchmark avoidance rate; Step B3533: Normalize the original avoidance rate and the baseline avoidance rate to obtain the congestion sensitivity index, as shown in formula (29). （29） in, Congestion sensitivity index; Step B4 involves calculating the elasticity index, which is a key tool for transforming the sensitivity coefficients calculated in the sensitivity analysis layer into an interpretable measure of marginal impact. The elasticity index includes time cost elasticity, economic cost elasticity, and the negative impact factor of weather. The specific steps are as follows: Step B41, calculate the time cost elasticity, where the time cost elasticity is used to measure the marginal impact of changes in the average congestion duration of the road network on the average travel probability, as shown in formula (30). （30） in, For time cost flexibility, The average time cost sensitivity coefficient. For average congestion duration, This represents the average probability of travel. Step B42: Calculate the economic cost elasticity, whereby the economic cost elasticity is used to measure the marginal impact of changes in the average toll cost of the road network on the average travel probability, as shown in formula (31). （31） in, For economic cost elasticity, The average economic cost sensitivity coefficient. This represents the average economic cost. Step B43: Calculate the negative weather impact factor, which measures the absolute decrease in the probability of travel when the weather changes from sunny to inclement, as shown in formula (32). ； （32） in, This is the deviation value for the weather risk level. The average weather risk level for the road network. It is a negative weather factor. This represents the average weather risk aversion coefficient.

4. The method for predicting highway traffic flow during long holidays based on social behavioral characteristics according to claim 3, characterized in that: Step E involves using an adaptive gating mechanism to fuse the first and second traffic flow prediction values ​​to obtain the fused traffic flow prediction value, as shown in formula (38). ； （38） in, This is the gating weight value. It is the Sigmoid activation function. For the gated weight vector, The confidence level of the business engineering model. This is the combined traffic flow prediction value.

5. The method for predicting highway traffic flow during long holidays based on social behavioral characteristics according to claim 4, characterized in that: Step F involves adaptively correcting the fused traffic flow forecast values ​​using a traffic flow deviation correction factor to obtain the final traffic flow forecast value. Based on this final traffic flow forecast value, the final traffic flow forecast result is output, completing the highway long holiday traffic flow forecasting operation. The specific steps for outputting the final traffic flow forecast result are as follows. Step F2: Output the final traffic flow prediction result based on the final traffic flow prediction value and complete the highway long holiday traffic flow prediction operation. The final traffic flow prediction result includes a traffic flow prediction curve, key indicator cards, influencing factor contribution decomposition, sensitivity analysis details, and travel probability distribution. The traffic flow prediction curve specifically includes time-segmented traffic flow prediction values ​​and confidence intervals. The key indicator cards specifically include peak traffic flow, peak time period, year-on-year change, and prediction accuracy. The influencing factor contribution decomposition specifically uses a horizontal bar chart to display the quantitative contribution of each influencing factor. The sensitivity analysis details specifically include the sensitivity coefficient values ​​and meanings in the current scenario. The travel probability distribution specifically includes the probability density curve of travelers' departure times.

6. A highway long holiday traffic flow prediction system based on social behavioral characteristics, wherein the specific prediction process of the highway long holiday traffic flow prediction system is based on the highway long holiday traffic flow prediction method according to any one of claims 1-5, characterized in that: It includes a data acquisition module, a social behavior feature extraction module, a first traffic flow prediction module, a second traffic flow prediction module, a prediction fusion module, and a prediction output module. The data acquisition module is used to collect vehicle information in a standardized manner and obtain the collected data, and then preprocess the collected data to obtain the preprocessed data. The social behavior feature extraction module is used to extract social behavior from the preprocessed data using a social behavior model and obtain social behavior features. The first traffic flow prediction module is used to predict traffic flow based on social behavioral characteristics using a business engineering model and obtain a first traffic flow prediction value. The second traffic flow prediction module is used to predict traffic flow based on social behavioral characteristics using a neural network model and obtain a second traffic flow prediction value; The prediction fusion module is used to fuse the first traffic flow prediction value and the second traffic flow prediction value using an adaptive gating mechanism to obtain the fused traffic flow prediction value. The prediction output module is used to adaptively correct the fused traffic flow prediction value using a traffic flow deviation correction factor and obtain the final traffic flow prediction value. Then, based on the final traffic flow prediction value, it outputs the final traffic flow prediction result and completes the traffic flow prediction operation for highway long holidays.

Images