A high-precision traffic flow prediction method based on multi-source disturbance features

Abstract

The application provides a high-precision high-speed traffic flow prediction method based on multi-source disturbance characteristics, relates to the field of data prediction, and specifically comprises the following steps: firstly, a multivariate entropy-driven interaction field module maps multi-source disturbance characteristics into a unified energy field, calculates joint information entropy density, constructs a joint interaction field, and processes an original feature sequence; secondly, a collaborative disturbance reconstruction module adopts a learnable mapping matrix and a multi-scale mechanism, extracts dynamic differences of features at different time scales, fuses global disturbance collaborative responses, and generates reinforced disturbance response features through a decoupling network; then, a spatial manifold mapping partition module realizes spatial expression and partition modeling of traffic tension evolution trends; finally, a prediction module combines disturbance amplitude factors and weighted disturbance characteristics, constructs an asymmetric prediction structure, adopts mean square error, introduces a disturbance constraint term, trains the model, and outputs a final traffic flow prediction result through a trained high-speed traffic flow prediction model.

Description

Technical Field

[0001] This invention belongs to the field of data prediction, specifically relating to a high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics. Background Technology

[0002] With the continuous expansion of the expressway network and the sustained increase in traffic density, traffic congestion and frequent emergencies have become important factors restricting road operation efficiency and traffic safety. As a core component of intelligent transportation systems, expressway traffic flow prediction directly affects traffic guidance, signal optimization, and road resource scheduling. However, traditional traffic flow prediction methods mostly rely on historical average traffic flow models, regression fitting, and time series analysis. They are often based on a single vehicle detector and historical traffic flow curves for inference, which has problems such as limited data sources, insufficient response to complex disturbances, and strong prediction lag. Especially in scenarios with drastic fluctuations in traffic conditions, it is difficult to meet the requirements for high-precision, real-time prediction.

[0003] In recent years, with the development of vehicle-to-everything (V2X) networks, road sensors, and meteorological monitoring systems, traffic prediction methods based on deep learning have emerged. Some studies have introduced neural networks to model the temporal dependence and spatial correlation of traffic flow, which can significantly improve prediction performance. Other scholars have combined attention mechanisms with graph neural networks to mine road topology and spatiotemporal correlations. These methods have made some progress in alleviating the problem of traditional models relying on a single data source and enhancing feature representation capabilities.

[0004] However, current high-speed traffic flow prediction still faces several challenges: First, the multi-source heterogeneous disturbance features are inconsistent in terms of spatial and temporal scales, making it difficult for traditional feature fusion methods to accurately reflect real traffic dynamics; second, existing models are insufficient in modeling the asymmetric response mechanism of traffic flow during surges and drops, making it difficult to capture the rapid fluctuation patterns in high-speed scenarios; in addition, prediction models generally lack the ability to adaptively regulate the evolution trend of traffic flow under disturbance backgrounds, making it difficult to achieve high-frequency and high-precision predictions. Therefore, it is urgent to propose new high-speed traffic flow prediction methods to improve the ability to perceive traffic operation status and provide core support for the construction of smart highways and intelligent transportation systems. Summary of the Invention

[0005] This invention provides a high-precision prediction method for high-speed traffic flow based on multi-source disturbance features. It proposes a prediction model for complex multi-source feature data of high-speed traffic flow, which consists of a high-order disturbance encoder module, a dynamic gating feature generator module, a cross-variable interactive modeling module, and a prediction module.

[0006] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps:

[0007] S1. Collect traffic flow data on highways and construct a raw dataset, which includes vehicle speed, weather, static road conditions, and multi-source disturbance characteristics of dynamic events.

[0008] S2. Based on the original dataset, the multi-source perturbation features are mapped into a unified system energy to construct an energy field. The joint information entropy density is calculated by simulating the joint information distribution using the energy field. A joint interaction field is constructed based on the energy field and the joint information entropy density. The original features are then processed to obtain the first dataset.

[0009] S3. Capture the feature perturbation response in the first dataset through the asymmetric perturbation channel, introduce a multi-scale mechanism to perform sliding window residual calculation on the feature perturbation response, perform weighted summation operation, and use a decoupled network to obtain the second dataset.

[0010] S4. Calculate the structural tension tensor based on the tension attenuation coefficient according to the second dataset, construct a deformation response model based on the structural tension tensor, introduce a dynamic scaling factor to calculate the spatial grouping weight, and perform a weighted aggregation operation to obtain the third dataset, which is divided into a training set and a prediction set.

[0011] S5. The training set is mapped to a prediction vector by a compression mechanism of feature perturbation intensity. An asymmetric prediction structure is constructed based on the prediction vector. Mean square error is used, and a perturbation constraint term is introduced to train the model.

[0012] S6. The prediction set is input into the trained high-speed traffic flow prediction model, and the final output is the traffic flow prediction value.

[0013] Preferably, the method for constructing the original dataset of high-speed traffic flow involves collecting multi-source information from the highway operating environment. The original dataset includes vehicle speed, weather characteristics, static road condition characteristics, and dynamic event characteristics. Weather characteristics include temperature, wind speed, rainfall, snowfall, humidity, and haze. Static road condition characteristics include road speed limits, number of lanes, road length, and number of entrances and exits. Dynamic event characteristics include information on traffic accidents, road construction, and holidays. Vehicle operation data is collected in real time by vehicle speed monitoring equipment, traffic flow monitoring cameras, and inductive loop sensors installed on key sections of the highway. Weather characteristic data is collected by meteorological monitoring stations along the route and IoT environmental sensors. Temperature and humidity are obtained using automatic weather station sensors with resolutions of 0.1℃ and 0.1%, respectively. Wind speed is collected using ultrasonic anemometers with an accuracy of 0.1 m / s. Rainfall and snowfall are recorded using rain gauges and snow depth sensors with an accuracy of 0.1 m / s. mm; haze concentration was collected by PM2.5 / PM10 laser particulate matter monitor; static road condition characteristics were obtained from the road infrastructure database; dynamic event characteristics were obtained from the automatic traffic accident detection system, construction management platform and holiday calendar database, and finally the original dataset of highway traffic flow was obtained for subsequent module processing.

[0014] Preferably, in step S2, based on the original dataset, the multi-source perturbation features are mapped to a unified system energy to construct an energy field. The joint information entropy density is calculated by simulating the joint information distribution using the energy field. A joint interaction field is constructed based on the energy field and the joint information entropy density, and the original features are processed to obtain the first dataset.

[0015] The multi-source perturbation features are mapped to a unified system energy to construct the energy field. It consists of dynamic energy, environmental energy, static road condition energy, and dynamic event energy. The feature mean is obtained at each moment. The squared Euclidean distance between the current feature vector and the feature mean is accumulated to obtain the overall deviation value. The deviation value is then weighted and scaled using a preset kernel width. The scaled deviation value is then negatively converted to an exponential value and normalized to obtain the probability distribution density value. After taking the negative value of the density and then performing a logarithmic transformation, the joint information entropy density is obtained. Calculate the average joint information entropy density of the samples. and average energy field The joint information entropy density and energy field at each time step are subtracted from the average joint information entropy density and average energy field, respectively, and weighting coefficients are introduced. The differences are fused proportionally to obtain the joint interaction field. For the four types of variables in the original data—vehicle speed, weather, static road conditions, and dynamic events—correction coefficients are assigned to each. These coefficients are then weighted with the joint interaction field according to their corresponding coefficients. Finally, the weighted result is subtracted from the original feature vector to obtain the processed feature sequence. This serves as the first dataset.

[0016] Furthermore, considering the diverse disturbance factors, uneven feature distribution, and drastic changes in traffic conditions in high-speed traffic, this invention proposes a multivariate entropy-driven interaction field module. First, by mapping the multi-source disturbance features of high-speed vehicle speed, weather, static road conditions, and dynamic events to a unified system energy, a system energy field is constructed. Dynamic energy, environmental energy, static road condition energy, and dynamic event energy are calculated separately and fused to express the disturbance intensity. This effectively solves the problem of inconsistency between multi-dimensional heterogeneous disturbance features, improving the physical interpretability of feature fusion and the model's adaptability. Then, a joint information entropy density is introduced to characterize the dense distribution of multi-source disturbance features, and a joint interaction field of the energy field and the joint information entropy density is constructed to dynamically perceive the salience of disturbances. This enables explicit correction and reconstruction of the original features, suppressing inefficient disturbances while retaining key information, thus improving the model's robustness and generalization ability.

[0017] Preferably, in step S3, the feature perturbation response in the first dataset is captured through an asymmetric perturbation channel, a multi-scale mechanism is introduced to perform sliding window residual calculation on the feature perturbation response, a weighted summation operation is performed, and a decoupled network is used to obtain the second dataset.

[0018] The external disturbance feature sequences from the first dataset are input into the asymmetric perturbation channel, and a learnable mapping matrix is ​​used. Perform a linear transformation on the feature sequence, and then compare the transformation result with the bias vector. The values ​​are summed, processed by the hyperbolic tangent function to obtain intermediate activation values, and then used with a learnable mapping matrix. Perform another linear transformation on the activation value to output the characteristic perturbation response. By introducing the multi-scale mechanism, the dynamic differences of each disturbance channel at different time scales are extracted by calculating the sliding window residual for each disturbance path. A cross-scale weighted fusion mechanism is introduced, utilizing learnable weights. The weighted summation of the dynamic differences at different time scales yields the global disturbance cooperative response. The decoupling network retains effective perturbation features, and the decoupling transformation matrix is ​​used. Perform a linear transformation and superimpose the bias vectors. The feature matrix is ​​obtained after processing with the activation function. This serves as the second dataset.

[0019] Furthermore, addressing the challenge of fully representing multidimensional dynamic changes in feature disturbance response extraction from high-speed traffic flow, this invention proposes a collaborative disturbance reconstruction module. First, based on the nonlinear evolution of multidimensional disturbance features across spatiotemporal scales, an asymmetric disturbance channel is designed and transformed channel by channel to obtain disturbance response values, enhancing the model's ability to capture disturbance change trends in different directions. Then, a multi-scale mechanism is introduced to calculate the sliding window residual for each disturbance path, obtaining the dynamic differences between channels and enhancing the model's sensitivity to changes in disturbance amplitude and direction. This effectively captures disturbance instability driven by weather, road conditions, and events under high-speed traffic conditions and establishes a linkage between time scale and disturbance expression. Finally, a cross-scale weighted fusion mechanism is introduced to weightedly fuse dynamic differences at different scales to form a unified global disturbance collaborative response, significantly improving the integrated expression capability of multi-scale disturbance information. Furthermore, a decoupled network is used to extract the dominant components from the disturbance signal, effectively suppressing redundant disturbances and refining features, thus obtaining a second dataset for accurate input to downstream prediction tasks.

[0020] Preferably, in step S4, the structural tension tensor is calculated based on the second dataset using the tension attenuation coefficient, a deformation response model is constructed based on the structural tension tensor, a dynamic scaling factor is introduced to calculate the spatial grouping weights, and a weighted aggregation operation is performed to obtain a third dataset, which is divided into a training set and a prediction set.

[0021] Each feature vector Treating each element as a spatial node in the tension field, we calculate the L1 norm between any two eigenvectors at any two time points and the square of the L2 norm of the sum of the two vectors. We then utilize the eigendimensionality and introduce a tension attenuation coefficient. The influence of the features is exponentially controlled to obtain the structural tension tensor. Based on the structural tension tensor, the attenuation of the tension effect is adjusted according to the reciprocal of the distance, and a positive number is introduced. To prevent numerical anomalies caused by extremely small distances, the deformation response model is constructed to obtain the deformation response vector. Mapping the deformation response vector to For each spatial group, the square of the L2 norm of the deformation response vector and the learnable center vector is calculated. The dynamic scaling factor is then introduced to construct the spatial grouping for each time step. Weights of spatial groups Weighted aggregation is performed based on grouping weights, and the aggregated representation of the output spatial groupings is combined. The third dataset is divided into a training set and a prediction set in a 7:3 ratio.

[0022] Furthermore, addressing the issue of drastic temporal fluctuations and spatial diffusion of disturbance characteristics in high-speed traffic flow, this invention proposes a spatial manifold mapping partitioning module. First, based on the relative positional relationships and tension potential energy coupling between features, a structural tension tensor is constructed, enabling dynamic perception of the attraction and repulsion between feature points and accurately characterizing the propagation intensity and direction of disturbances in local space. Then, a deformation response model is constructed under the drive of the structural tension tensor to simulate vector offset behavior under local tension disturbances. Dynamic clustering and center updates of features are achieved through dynamic scaling factors and weighted fusion mechanisms, thereby enhancing the model's ability to distinguish deformation responses to different types of disturbances and improving its adaptability to complex road conditions and multi-source disturbances.

[0023] Preferably, in step S5, the training set is mapped to a prediction vector using a feature perturbation intensity compression mechanism, an asymmetric prediction structure is constructed based on the prediction vector, and the model is trained by introducing a perturbation constraint term using mean square error.

[0024] Based on the compression mechanism of the aforementioned characteristic perturbation intensity, each aggregated representation is... The perturbation amplitude factor is formed by the learnable center vector, processed by a logarithmic function, and then element-wise multiplied with the original aggregate representation to obtain the weighted perturbation feature. Finally, the predicted vector is obtained by normalizing and weighting all spatial groups. According to the trend difference The positive and negative signs are dynamically selected, and a linearly weighted prediction output is performed to obtain the asymmetric prediction structure. Based on the mean square error, the perturbation suppression strength factor is used. The perturbation constraint term is constructed to measure and optimize the difference between the predicted value and the true label, thereby obtaining a trained high-speed traffic flow prediction model.

[0025] Furthermore, the prediction module first constructs prediction vectors based on the aggregated spatial grouping representation, improving the anti-perturbation ability of the compressed features, reducing prediction noise interference, and enhancing prediction robustness. Then, it constructs a historical direction symmetry judgment structure based on an asymmetric trend judgment strategy, effectively enhancing the model's sensitivity and response strength to sudden traffic flow changes, and further improving the ability to accurately characterize traffic flow trends. Finally, it measures the prediction error based on the mean square error and combines it with perturbation suppression regularization constraints to improve model stability and generalization ability, thereby significantly enhancing the prediction accuracy and dynamic adjustment performance of the high-speed traffic flow prediction model in complex scenarios.

[0026] Preferably, in step S6, the prediction set is input into the trained high-speed traffic flow prediction model, and the final output is the traffic flow prediction value.

[0027] In summary, this invention proposes a high-precision prediction method for high-speed traffic flow based on multi-source disturbance features. The method comprises a multivariate entropy-driven interaction field module, a cooperative disturbance reconstruction module, a spatial manifold mapping partitioning module, and a prediction module. First, the multivariate entropy-driven interaction field module maps multi-source disturbance features of vehicle speed, weather, static road conditions, and dynamic events into a unified energy field, calculates the joint information entropy density, and constructs a joint interaction field. This differentially modulates the original feature sequence, effectively improving the system's unified representation of the impact of multivariate disturbances. Second, the cooperative disturbance reconstruction module employs a learnable mapping matrix and a multi-scale mechanism to extract the dynamic differences of features at different time scales. After fusing the global disturbance cooperative response, it generates enhanced disturbance response features through a decoupling network, enhancing the response to high-frequency changes and asymmetry. The modeling robustness of sudden events is demonstrated. Next, the spatial manifold mapping partitioning module constructs a structural tension tensor and, based on a deformation response model and dynamic scaling mechanism, groups and aggregates it according to spatial heterogeneity to achieve spatial expression and partitioned modeling of the traffic flow tension evolution trend. Then, the prediction module combines disturbance amplitude factors and weighted disturbance features to construct an asymmetric prediction structure and uses mean square error. Disturbance constraint terms are introduced to train the model, and the trained high-speed traffic flow prediction model is used to predict the test set, outputting the final traffic flow prediction result. This invention fully integrates multi-source nonlinear disturbance features to construct a full-process modeling framework from energy density regulation to structural response driving, possessing stronger generalization ability, stability, and sudden adaptability, providing solid support for traffic flow prediction and regulation in complex scenarios in intelligent transportation systems. Attached Figure Description

[0028] Figure 1 This is a step-by-step diagram of a high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics.

[0029] Figure 2 This is a structural diagram of a high-speed traffic flow prediction model.

[0030] Figure 3 This is a structural diagram of a multivariable entropy-driven interaction field module.

[0031] Figure 4 The structural diagram of the collaborative disturbance reconstruction module.

[0032] Figure 5 This is a structural diagram of the spatial manifold mapping partitioning module.

[0033] Figure 6 This is a diagram of the model training process.

[0034] Figure 7 The image shows the fitting effect of the prediction model to evaluate the high-speed traffic flow. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please see Figures 1-7 This invention provides a technical solution: a high-precision prediction method for high-speed traffic flow based on multi-source perturbation features, comprising a multivariate entropy-driven interaction field module, a cooperative perturbation reconstruction module, a spatial manifold mapping partitioning module, and a prediction module. First, the multivariate entropy-driven interaction field module processes the multi-source features. Second, the cooperative perturbation reconstruction module introduces a multi-scale mechanism to calculate the sliding window residual and performs a weighted summation operation to extract enhanced perturbation response features. Next, the spatial manifold mapping partitioning module constructs a structural tension tensor and introduces a deformation response model to complete spatial partitioning and tension aggregation representation. Finally, an asymmetric prediction structure is constructed, and the model is trained based on the mean square error and by introducing perturbation constraint terms to achieve high-precision prediction of traffic flow. Specific steps are as follows: Figure 1 As shown.

[0037] Construct a high-speed traffic flow prediction model, with the following structure: Figure 2 As shown.

[0038] S1. Collect traffic flow data on highways and construct a raw dataset, which includes vehicle speed, weather, static road conditions, and multi-source disturbance characteristics of dynamic events.

[0039] Furthermore, the dataset of this invention contains 1000 traffic flow data points, and the dataset is divided into a training set and a test set in a 7:3 ratio.

[0040] S2. Based on the original dataset, the multi-source perturbation features are mapped into a unified system energy to construct an energy field. The joint information entropy density is calculated by simulating the joint information distribution using the energy field. A joint interaction field is constructed based on the energy field and the joint information entropy density. The original features are then processed to obtain the first dataset.

[0041] Furthermore, the multivariate entropy-driven interaction field module structure is as follows: Figure 3 As shown, time The eigenvectors below are , The dimension of the feature vector is 14. For high-speed vehicles, For weather variables, For static road condition variables, As dynamic event variables, the multi-source disturbance characteristics are first mapped to a unified system energy to construct an energy field. The mathematical model is:

[0042] ;

[0043] In the formula, As kinetic energy, it is calculated by the speed variable of high-speed vehicles. The square of the number divided by 2 is used to obtain the result. For environmental energy, for each weather variable, a normalized squared energy is defined, and the mathematical model is as follows:

[0044] ;

[0045] In the formula, This represents the number of dimensions for the weather variables, with a value of 6, indicating six dimensions: temperature, wind speed, rainfall, snowfall, humidity, and haze. For the first Weather variables in several dimensions This represents the global mean of the weather variable. The global standard deviation, The weighting parameters for weather variables represent the impact of different weather variables on highway traffic flow. The weighting parameter for rainfall is 0.26, the weighting parameter for snowfall is 0.22, the weighting parameter for haze is 0.18, the weighting parameter for wind speed is 0.14, the weighting parameter for humidity is 0.12, and the weighting parameter for temperature is 0.08.

[0046] Furthermore, in the mathematical model of the energy field For static road condition energy, highway road conditions are relatively fixed, but they still affect traffic flow. The mathematical model is as follows:

[0047] ;

[0048] In the formula, This represents the number of dimensions for static road condition variables; a value of 4 indicates that these dimensions include speed limit, number of lanes, road length, and number of entrances / exits. The weighted parameters for road conditions are as follows: number of lanes: 0.3; speed limit: 0.26; number of entrances / exits: 0.24; road length: 0.2. For the first Static road condition variables in several dimensions To map highway traffic condition variables to functions of energy, where the speed limit variable is an energy function. The mathematical model is as follows:

[0049] ;

[0050] In the formula, To adjust the parameter, the value is set to 0.3 to prevent the current speed limit from being applied. When the minimum speed limit is The case where it is 0, This is the minimum speed limit for the entire highway. The energy function of the maximum speed limit and the number of lanes for the entire highway section. The energy function for road length is calculated by adding 1 to the number of highway lanes and then taking the reciprocal. The energy function for the number of entrances and exits is calculated by adding 1 to the road length and then taking the reciprocal. It is calculated by adding 1 to the number of entrances and exits and then taking the reciprocal.

[0051] Furthermore, in the mathematical model of the energy field For dynamic event energy, the mathematical model is:

[0052] ;

[0053] In the formula, This represents the number of dimensions for the dynamic event variable; a value of 3 indicates three dimensions: accident, construction, and holiday. The energy weights for different events are as follows: accidents have a weight of 0.37, holidays have a weight of 0.33, and construction has a weight of 0.3. A value of 1 indicates that the event has occurred, and a value of 0 indicates that the event has not occurred. For the first Dynamic event variables in multiple dimensions.

[0054] Furthermore, by simulating the probability distribution of joint features, and based on the principles of information entropy and thermodynamic distribution, a joint information entropy density is constructed. The mathematical model is:

[0055] ;

[0056] In the formula, Let be the probability density value, and the mathematical model is:

[0057] ;

[0058] In the formula, These are normalization parameters used to ensure... At time t, the sum of all density values ​​equals 1, thus forming an effective joint perturbation characteristic distribution. The mathematical model is as follows:

[0059] ;

[0060] In the formula, This represents the number of samples participating in the normalization calculation at any given time. Its value depends on the actual number of samples at that time. This represents the kernel width, with a value of 1.5. For time The j-th eigenvector below, For a moment The characteristic mean, The squared Euclidean distance between the current feature vector and the global center is used; the larger the distance, the higher the degree of feature perturbation at time point. Then, a joint interaction field is constructed based on the principle of cross-entropy gradient. The mathematical model used to reflect the spatial propagation direction and intensity of the information perturbation gradient and the dynamic perturbation gradient is as follows:

[0061] ;

[0062] In the formula, For time points The average joint information entropy density of all samples, Time point The average energy field of all samples, This is the weighting coefficient for the coupling of energy and information, with a value of 0.25.

[0063] Furthermore, by utilizing the joint interaction field to process the features of the original data, the mathematical model is as follows:

[0064] ;

[0065] In the formula, The vector represents the correction coefficients for the variables. The correction coefficient for vehicle speed is 0.23, for weather is 0.1, for static road conditions is 0.2, and for dynamic events is 0.18. This results in the processed features. As the first dataset.

[0066] S3. Capture the feature perturbation response in the first dataset through the asymmetric perturbation channel, introduce a multi-scale mechanism to perform sliding window residual calculation on the feature perturbation response, perform weighted summation operation, and use a decoupled network to obtain the second dataset.

[0067] Furthermore, in step S3, the first dataset is input into the collaborative perturbation reconstruction module for deep reconstruction, thereby extracting stable perturbation structure features and providing enhanced representation for high-precision prediction of high-speed traffic flow. The process is as follows: Figure 4 As shown, the specific steps are as follows:

[0068] To address the nonlinear and diverse trajectory behavior exhibited by multidimensional perturbation features on the time axis, a design was developed. There are three asymmetric perturbation channels, each capturing a characteristic perturbation response through a nonlinear path mapping function. The mathematical model is as follows:

[0069] ;

[0070] In the formula, This indicates the asymmetric perturbation channel index. and This represents the learnable mapping matrix corresponding to the channel. This represents the bias vector. For the first The path at time The characteristic perturbation response.

[0071] Furthermore, in this example, The value is set to 3 to characterize the weather, static road conditions and dynamic event disturbance mechanisms in the high-speed traffic flow data, ensuring that the model has the ability to model multi-channel disturbance behavior. Initialization using uniform distribution, i.e. According to the backpropagation algorithm, updates are performed via gradient descent. The error of each layer is propagated back through the chain rule, the gradient is calculated, and the update is performed in each training iteration. The update formula is: , Let the learning rate be set to... , The cross-entropy loss function; Normal initialization is used, that is The updated formula is as follows Through gradient updates, and It will be gradually optimized during training to accurately map the perturbation path, thereby improving the model's ability to learn different perturbation channels; bias vector To avoid bias in the initial stage, the network is optimized from scratch during training, initially set to a zero vector, and subsequently updated during backpropagation. The update formula is as follows: .

[0072] Furthermore, to enhance the model's ability to learn from external disturbance signals, a multi-scale mechanism is introduced based on the theoretical foundation of trend residual modeling. By calculating the sliding window residual for each disturbance path, the dynamic differences of each disturbance channel at different time scales are extracted. The mathematical model for capturing the changing patterns of disturbance features over time is as follows:

[0073] ;

[0074] In the formula, It is used to capture disturbance features at different scales, including weather, static road conditions, and dynamic events, helping models extract multi-scale disturbance differences, enhancing their response to disturbance signals at different scales, and improving the model's generalization performance. For the first The characteristic perturbation response of the path at time j.

[0075] Furthermore, to fully utilize the information in multi-scale residuals, this example introduces a cross-scale weighted fusion mechanism based on the information weighting theory. Through learnable weights, dynamic differences at different scales are weighted and synthesized to form a unified global disturbance cooperative response. This is used for subsequent structural decoupling, and the mathematical model is as follows:

[0076] ;

[0077] In the formula, Let be the learnable weight, representing the th Path and the The fusion coefficients at each scale are weighted to account for the responses to perturbations at different scales, ensuring that the model can dynamically adjust its learning and fusion methods based on information from different perturbation paths and scales. The initial value is set to 0.3, and subsequent values ​​are... The values ​​can be dynamically selected between them.

[0078] Furthermore, in obtaining a global coordinated response to disturbances... Then, a decoupled network is used for processing to separate the dominant component and noise component in the perturbation signal, retaining the effective perturbation characteristics. The mathematical model is as follows:

[0079] ;

[0080] In the formula, This is the decoupling transformation matrix, used to adjust the global disturbance response. Initialization using uniform distribution, i.e. The updated formula is: , Let the learning rate be set to... , The cross-entropy loss function is used, and regularization is employed to avoid overfitting, ensuring the weight matrix is ​​not too large and maintaining model complexity. The formula is: , Here is the regularization coefficient, with a value of ; The bias vector is initially set to an all-zero vector and is subsequently updated during backpropagation using the following formula: Indicates time The decoupled disturbance signal, through adaptive decoupling transformation, removes noise disturbances while retaining the main disturbance factors affecting high-speed traffic flow, resulting in more refined features and stronger predictive ability. The final output feature matrix is ​​as follows: This serves as the second dataset for use as input in subsequent modules.

[0081] S4. Calculate the structural tension tensor based on the tension attenuation coefficient according to the second dataset, construct a deformation response model based on the structural tension tensor, introduce a dynamic scaling factor to calculate the spatial grouping weight, and perform a weighted aggregation operation to obtain the third dataset, which is divided into a training set and a prediction set.

[0082] Furthermore, in step S4, the second dataset... The input, the spatial manifold mapping partitioning module performs spatial grouping processing on the features, the process is as follows: Figure 5 As shown, the specific steps are as follows:

[0083] Highway traffic flow data exhibits significant disturbance coupling, time drift, and strong nonlinear evolution characteristics. This example, based on tension mechanics, decomposes each feature vector... Treating these as spatial nodes in a tension field, the attraction and repulsion relationships between features are simulated, driving the feature points to generate deformation responses in the manifold, thereby completing a spatial clustering-style dynamic grouping operation. First, for any two time points... Define the structural tension tensor The mathematical model for simulating the tension relationship between cooperative perturbation and external signal fusion in the characteristic space is as follows:

[0084] ;

[0085] In the formula, This is the tension attenuation coefficient, used to control the effect of tension over long distances, and its value is 0.7. This indicates a measurement of relative positional relationships. It represents the coupling amount of the overall strength of the features of two nodes, enhancing the perception of the total energy of the features.

[0086] Furthermore, based on the structural tension tensor, a deformation response model is constructed to simulate the vector offset behavior under local tension perturbation. The mathematical model is as follows:

[0087] ;

[0088] In the formula, Indicates time The deformation response vector, To prevent positive numbers introduced by division by zero, the value is set to The above design embodies a kind of "tension manifold attraction", which drives the eigenvector to produce a directional shift under tension guidance.

[0089] Furthermore, the deformation response vector Mapped to Each space is grouped into _ , and a dynamic scaling factor is introduced to construct the _ ... Weights of spatial groups The mathematical model is:

[0090] ;

[0091] In the formula, For the first Learnable center vectors of a spatial group The mathematical model for the dynamic scaling factor is:

[0092] ;

[0093] In the formula, The preset number of partition groups, with a value of 500. For the first Learnable center vectors for a set of spaces.

[0094] Furthermore, before model training, the deformation response vectors of all high-speed traffic flow data samples are first processed. Clustering initialization is performed to obtain the initial center vector for each spatial group. The mathematical model is as follows:

[0095] ;

[0096] To improve stability and cluster consistency, this example introduces an exponential moving average mechanism to dynamically update the center vector without gradients. The mathematical model is as follows: ;

[0097] In the formula, This is a smoothing factor used to control the update magnitude, initially set to 0.8, and subsequently... The values ​​are dynamically selected between them. For the first The weighted average of all deformation response vectors in each spatial group, through the above design, enables the model to adjust its center over time, thereby adapting to different disturbance modes.

[0098] Furthermore, weighted aggregation is performed according to the weights to generate an aggregated representation for each spatial group. The mathematical model is:

[0099] ;

[0100] The final output is the combination of all space groups. As the third dataset, the third dataset is divided into training and test sets in a 7:3 ratio.

[0101] S5. The training set is mapped to a prediction vector using a compression mechanism based on the feature perturbation intensity. An asymmetric prediction structure is constructed based on the prediction vector. Mean square error is used, and a perturbation constraint term is introduced to train the model.

[0102] Furthermore, the training set was input into the high-speed traffic flow prediction model. The model uses the PyTorch deep learning framework and is accelerated using an NVIDIA V100 32GB GPU. During training, the batch size was set to 128. In the training phase, the training set was first mapped to a prediction vector based on a compression mechanism based on the feature perturbation intensity. To capture high-speed traffic flow change patterns, the mathematical model is as follows:

[0103] ;

[0104] In the formula, For the first The aggregated representation of spatial groups reduces small perturbation interference by implicitly fusing perturbation perception and feature compression.

[0105] Furthermore, an asymmetric prediction structure is constructed based on historical traffic flow trends to improve the response to drastic fluctuations in traffic flow, forming a precise prediction mechanism for traffic surges and drops in high-speed scenarios. The mathematical model is as follows:

[0106] ;

[0107] In the formula, This is a forecast of highway traffic volume. The difference between the current traffic volume and the previous traffic volume is used as a basis for trend judgment. The trainable weight vector representing the upward trend branch is initialized using a normal distribution, i.e. The updated formula is as follows , Let the learning rate be set to... , Let cross-entropy be the loss function. The trainable weight vector representing the downward trend branch is initialized using a uniform distribution, i.e. The updated formula is as follows , This is the bias term, initially set as a vector of all zeros, and subsequently updated during backpropagation. The update formula is: , For the bias term, the initial settings and update formulas are the same as those for the bias term. Consistent.

[0108] Furthermore, to further improve the model's prediction accuracy, this example uses mean squared error and adds a perturbation constraint term to enhance stability, measuring and optimizing the difference between the predicted values ​​and the true labels. The mathematical model is as follows:

[0109] ;

[0110] In the formula, For a moment Traffic flow forecast, To reflect real highway traffic volume, The perturbation suppression strength factor, 0.08, controls the trade-off between the two parts of the loss function. During training, the model minimizes this loss function using the Adam optimizer, iteratively updating network parameters to gradually approach the globally optimal prediction structure. The introduction of the perturbation suppression term effectively curbs the perturbation shift caused by abnormal data or sudden events, improving the model's prediction stability under highly volatile and uncertain road conditions. The training process is as follows: Figure 6 As shown in the figure, the number of training rounds is 70. It can be seen from the figure that as the number of training rounds increases, the loss value gradually decreases from a relatively high level in the initial stage and tends to stabilize in the later stage, indicating that the model converges and stabilizes during the training process. Finally, a well-trained high-speed traffic flow prediction model is obtained.

[0111] S6. The prediction set is input into the trained high-speed traffic flow prediction model, and the final output is the traffic flow prediction value.

[0112] Furthermore, the high-speed traffic flow prediction model achieves the following fitting effect diagram: Figure 7 As shown in the figure, the horizontal axis represents time, the vertical axis represents traffic flow value, the solid dotted line represents the actual traffic flow value, and the dashed cross line represents the model prediction value. It can be seen from the figure that the trend of the predicted value and the actual highway traffic flow value are roughly similar. The experimental results show that the highway traffic flow prediction model can effectively capture the trend of highway traffic flow data and can predict the highway traffic flow value well.

Claims

1. A high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics, characterized in that: Collect traffic flow data on highways to construct a raw dataset, which includes vehicle speed, weather, static road conditions, and multi-source disturbance characteristics of dynamic events; Based on the original dataset, the multi-source perturbation features are mapped into a unified system energy to construct an energy field. The joint information entropy density is calculated by simulating the joint information distribution using the energy field. A joint interaction field is constructed based on the energy field and the joint information entropy density. The original features are then processed to obtain the first dataset. The first dataset is obtained by capturing the feature perturbation response in the first dataset through an asymmetric perturbation channel, introducing a multi-scale mechanism to perform sliding window residual calculation on the feature perturbation response, performing a weighted summation operation, and using a decoupled network. Based on the second dataset, the structural tension tensor is calculated using the tension attenuation coefficient. A deformation response model is constructed based on the structural tension tensor. A dynamic scaling factor is introduced to calculate the spatial grouping weights. A weighted aggregation operation is performed to obtain the third dataset, which is divided into a training set and a prediction set. The training set is mapped to a prediction vector by a compression mechanism based on the feature perturbation intensity. An asymmetric prediction structure is constructed based on the prediction vector. The model is trained by using mean square error and introducing a perturbation constraint term. The prediction set is input into the trained high-speed traffic flow prediction model, and the final output is the traffic flow prediction value.

2. The high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics according to claim 1, characterized in that, The original dataset includes vehicle speed, weather, static road conditions, and dynamic event multi-source disturbance features. Weather features include temperature, wind speed, rainfall, snowfall, humidity, and haze. Static road condition features include road speed limits, number of lanes, road length, and number of entrances and exits. Dynamic event features include traffic accidents, road construction, and holiday information, thus obtaining the original dataset.

3. The high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics according to claim 2, characterized in that, The multi-source perturbation features are mapped to a unified system energy to construct the energy field. It consists of dynamic energy, environmental energy, static road condition energy, and dynamic event energy. The feature mean is obtained at each moment. The squared Euclidean distance between the current feature vector and the feature mean is accumulated to obtain the overall deviation value. The deviation value is then weighted and scaled using a preset kernel width. The scaled deviation value is then negatively converted to an exponential value and normalized to obtain the probability distribution density value. After taking the negative value of the density and then performing a logarithmic transformation, the joint information entropy density is obtained. .

4. The high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics according to claim 3, characterized in that, Calculate the average joint information entropy density of the samples and average energy field The joint information entropy density and energy field at each time step are subtracted from the average joint information entropy density and average energy field, respectively, and weighting coefficients are introduced. The differences are fused proportionally to obtain the joint interaction field. For the four types of variables in the original data—vehicle speed, weather, static road conditions, and dynamic events—correction coefficients are assigned to each. These coefficients are then weighted with the joint interaction field according to their corresponding coefficients. Finally, the weighted result is subtracted from the original feature vector to obtain the processed feature sequence. This serves as the first dataset.

5. The high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics according to claim 4, characterized in that, The external disturbance feature sequences from the first dataset are input into the asymmetric perturbation channel, and a learnable mapping matrix is ​​used. Perform a linear transformation on the feature sequence, and then compare the transformation result with the bias vector. The values ​​are summed, processed by the hyperbolic tangent function to obtain intermediate activation values, and then used with a learnable mapping matrix. Perform another linear transformation on the activation value to output the characteristic perturbation response. By introducing the multi-scale mechanism, the dynamic differences of each disturbance channel at different time scales are extracted by calculating the sliding window residual for each disturbance path. .

6. The high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics according to claim 5, characterized in that, Introducing a cross-scale weighted fusion mechanism, utilizing learnable weights The weighted summation of the dynamic differences at different time scales yields the global disturbance cooperative response. The decoupling network retains effective perturbation features, and the decoupling transformation matrix is ​​used. Perform a linear transformation and superimpose the bias vectors. The feature matrix is ​​obtained after processing with the activation function. This serves as the second dataset.

7. The high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics according to claim 6, characterized in that, Each feature vector Treating each element as a spatial node in the tension field, we calculate the L1 norm between any two eigenvectors at any two time points and the square of the L2 norm of the sum of the two vectors. We then utilize the eigendimensionality and introduce a tension attenuation coefficient. The influence of the features is exponentially controlled to obtain the structural tension tensor. Based on the structural tension tensor, the attenuation of the tension effect is adjusted according to the reciprocal of the distance, and a positive number is introduced. To prevent numerical anomalies caused by extremely small distances, the deformation response model is constructed to obtain the deformation response vector. .

8. The high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics according to claim 7, characterized in that, Map the deformation response vector to For each spatial group, the square of the L2 norm of the deformation response vector and the learnable center vector is calculated. The dynamic scaling factor is then introduced to construct the spatial grouping for each time step. Weights of each spatial group Weighted aggregation is performed based on grouping weights, and the aggregated representation of the output spatial groupings is combined. The third dataset is divided into a training set and a prediction set in a 7:3 ratio.

9. The high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics according to claim 8, characterized in that, Based on the compression mechanism of the aforementioned characteristic perturbation intensity, each aggregated representation is... The perturbation amplitude factor is formed by the learnable center vector, processed by a logarithmic function, and then element-wise multiplied with the original aggregate representation to obtain the weighted perturbation feature. Finally, the predicted vector is obtained by normalizing and weighting all spatial groups. According to the trend difference The positive and negative signs are dynamically selected, and the weighted branches are used to perform linear weighted prediction output to obtain the asymmetric prediction structure.

10. The high-precision prediction method for high-speed traffic flow based on multi-source disturbance characteristics according to claim 9, characterized in that, Based on the mean square error, the perturbation suppression strength factor is used. The perturbation constraint term is constructed to measure and optimize the difference between the predicted value and the true label, thereby obtaining a trained high-speed traffic flow prediction model, and finally outputting the traffic flow prediction value.

Images