Deep learning-based road traffic flow short-term prediction method and system

Abstract

The invention discloses a road traffic flow short-term prediction method and system based on deep learning, and relates to the technical field of intelligent traffic and deep learning. According to the invention, accurate short-term prediction of the road traffic flow is realized through multi-source data acquisition and preprocessing in combination with a deep learning double-branch architecture, the system can dynamically capture the influence of road network deformation on the traffic flow, and an association rule of road network deformation and traffic flow spatial-temporal characteristics is deeply mined through a deformation-guided attention coupling algorithm, so that the traffic flow prediction accuracy is improved. The weight of the prediction model is dynamically corrected, so that the prediction precision is improved, the real-time performance of the prediction result is ensured, and timely and accurate data support is provided for traffic management and control.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and deep learning technology, specifically to a method and system for short-term prediction of highway traffic flow based on deep learning. Background Technology

[0002] With the acceleration of urbanization and the rapid development of intelligent transportation systems, highway traffic flow management has become a key link in improving urban operational efficiency and ensuring traffic safety. Accurately predicting short-term changes in highway traffic flow is of great significance for traffic control departments to formulate scientific and reasonable traffic management strategies, reduce traffic congestion, and improve road capacity.

[0003] Traditional traffic flow prediction techniques mainly rely on statistical models or simple machine learning algorithms. These methods have significant limitations when dealing with complex and ever-changing traffic environments. On the one hand, statistical models are usually based on the average trend of historical data for prediction, making it difficult to capture the impact of sudden and nonlinear traffic events on traffic flow, such as traffic accidents and severe weather. On the other hand, although simple machine learning algorithms can handle a certain degree of nonlinear relationships, they often fail to achieve ideal prediction results when faced with large-scale, high-dimensional traffic data due to limited feature extraction capabilities and insufficient model complexity. In addition, most traditional methods require pre-setting the correlation rules between road network deformation and traffic flow, making it difficult to adapt to dynamic changes in the road network, especially when dealing with sudden and unconventional road network deformations.

[0004] In view of the problems of insufficient prediction accuracy, poor real-time performance and weak adaptability of traditional traffic flow prediction technology, the present invention proposes a short-term highway traffic flow prediction method and system based on deep learning, which is of particular importance. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for short-term prediction of highway traffic flow based on deep learning. It can achieve accurate short-term prediction of highway traffic flow by collecting and preprocessing multi-source data and combining it with a deep learning dual-branch architecture. The system can dynamically capture the impact of road network deformation on traffic flow and deeply mine the correlation between road network deformation and the spatiotemporal characteristics of traffic flow through a deformation-guided attention coupling algorithm. It can also dynamically adjust the weights of the prediction model, thereby significantly improving the prediction accuracy and real-time performance. At the same time, the system does not require preset association rules and can autonomously adapt to the dynamic changes of the road network. It has strong adaptability and self-learning capabilities, providing more scientific and efficient data support for traffic management.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a short-term prediction method for highway traffic flow based on deep learning, the specific steps of which are as follows: Multi-source data acquisition and preprocessing: real-time traffic flow monitoring data, historical traffic flow data and road network deformation data of the target road segment are acquired. The road network deformation data includes deformation type, deformation range and deformation duration. The data quality is optimized by using a hierarchical multi-source acquisition + graded preprocessing method. Dual-branch architecture construction and feature extraction: A deep learning dual-branch architecture is constructed, namely a road network deformation feature encoding branch and a traffic flow spatiotemporal feature branch. The road network deformation feature encoding branch adopts a three-level serial architecture of preprocessing-extraction-enhancement, and the traffic flow spatiotemporal feature branch adopts a spatiotemporal coupled feature extraction algorithm to extract the core features of road network deformation and the spatiotemporal coupled features of traffic flow, respectively. Dual-branch feature coupling and dynamic weight correction: The features output by the dual branches are deeply coupled using a deformation-guided attention coupling algorithm to mine the correlation between road network deformation features and traffic flow spatiotemporal features. Based on the correlation, the deformation influence weight iterative algorithm is used to dynamically correct the prediction model weights. Short-term forecast output and full-process optimization: The deep learning model based on the corrected weights outputs the short-term traffic flow forecast results of the target road segment in a multi-dimensional hierarchical output manner. It also performs closed-loop verification and iterative optimization of the forecast accuracy, so as to achieve autonomous adaptation to the dynamic changes of the road network without the need to preset the road network deformation and traffic flow association rules. It has broad adaptability and convenient deployment capability, and completes the short-term forecast of highway traffic flow.

[0007] Furthermore, in the multi-source data acquisition and preprocessing, real-time traffic flow monitoring data, historical traffic flow data, and road network deformation data are acquired and optimized using a layered multi-source acquisition + hierarchical preprocessing approach. The specific implementation steps are as follows: first, real-time traffic flow monitoring data is acquired using roadside loop detectors, high-definition checkpoint cameras, and microwave radar; then, historical traffic flow data for the target road section over the past six months is retrieved from the traffic control cloud database; finally, road network deformation data is acquired using roadside inspection terminals, traffic control instruction issuance platforms, and drone aerial imagery. The road network deformation types explicitly include five categories: road construction occupation, lane closure, emergency control, temporary detours, and road surface defects occupying lanes. The system supplements traffic flow and deformation data from adjacent road segments as auxiliary references. Then, it performs hierarchical preprocessing on the three types of basic data: first, outlier identification and removal; then, data standardization and alignment to unify temporal and spatial granularity; and finally, data completion. For missing data points, interpolation using data from adjacent time periods and adjacent road segments is used for completion. The entire process requires no manual intervention, ensuring the integrity and consistency of the input data. This lays a reliable data foundation for subsequent dual-branch feature extraction, effectively avoiding interference from invalid data on model training, improving the accuracy of subsequent feature extraction, and adapting to real-world scenarios where data quality fluctuates in highway monitoring equipment.

[0008] Furthermore, the road network deformation feature encoding branch in the dual-branch support construction and feature extraction adopts a three-level serial architecture of preprocessing-extraction-enhancement to accurately extract the core features of road network deformation. The specific implementation steps are as follows: First, the preprocessing unit performs feature screening on the road network deformation data, prioritizing the retention of core dimension data strongly correlated with traffic flow changes, and removing irrelevant information such as redundant control instructions and invalid inspection records; then, the deformation feature extraction unit performs deep feature mining on the screened data, first performing spatial gridding encoding on the deformation range, performing temporal segmentation encoding on the deformation duration, and performing category-based one-hot encoding on the deformation type, and then completing the feature mapping of the encoded data through convolutional layers to achieve the initial extraction of deformation features; finally, the feature enhancement unit strengthens the key deformation features, focusing on strengthening the core deformation features that affect traffic flow, and using a residual connection structure to fuse the key features with the initially extracted features, avoiding information loss during the feature extraction process, ensuring that the output core features of road network deformation can accurately reflect the potential impact of road network deformation on traffic flow, and providing high-quality feature support for subsequent dual-branch coupling.

[0009] Furthermore, in the construction and feature extraction of the dual-branch structure, the traffic flow spatiotemporal feature branch employs a spatiotemporal coupled feature extraction algorithm to accurately mine the spatiotemporal coupled features of the traffic flow. This branch is based on a spatiotemporal convolutional network and a gated recurrent unit architecture, and completes feature output through a traffic flow spatiotemporal coupled feature fusion formula, which is as follows: ,in As the core feature of the spatiotemporal coupling of traffic flow, For the spatial topological relationship features of traffic flow, The characteristics of traffic flow temporal evolution, For spatial convolution operations, For timing-gated operations, The weights for spatiotemporal feature fusion are adaptively determined by the spatiotemporal correlation of traffic flow in the target road segment; if the spatial transmissibility of traffic flow is strong, then... The value ranges from 0.6 to 0.8, indicating strong fluctuations in traffic flow timing. The value ranges from 0.2 to 0.4. The spatial feature weight matrix is ​​generated through training on the road segment node topology. The time feature weight matrix is ​​generated by training based on the historical traffic flow time sequence patterns of the past three months. This formula breaks through the limitations of simple spatiotemporal feature splicing in the traditional method and achieves deep fusion of spatiotemporal features. At the same time, it first extracts the spatial features of traffic flow transmission between road segments and traffic flow distribution between lanes through spatial branches, and then extracts the temporal features of traffic flow peaks and troughs and sudden fluctuations through time branches. Finally, it is substituted into the formula to output coupled features, ensuring the comprehensiveness and accuracy of the features.

[0010] Furthermore, the feature depth coupling calculation of the dual-branch output in the dual-branch feature coupling and dynamic weight correction is implemented using a deformation-guided attention coupling algorithm. The association pattern is mined through the dual-branch feature coupling weight allocation formula, which is: ,in It is a two-branch fusion feature. The spatiotemporal coupling characteristics of traffic flow As the core feature of road network deformation, The deformation-guided attention weighting coefficient is determined by quantifying the impact of road network deformation. The larger the deformation range, the longer its duration, and the stronger its impact on traffic capacity, the better. The closer the value is to 1, the closer it is to 0.3. This formula breaks through the limitations of traditional attention mechanisms that lack targeted guidance. It takes road network deformation as the core guide to achieve precise coupling of dual-branch features. During the coupling calculation, the road network deformation features are used as the guide. Differentiated attention weights are allocated through this formula, and then the dual-branch features are fused dimension by dimension. It autonomously discovers the inherent correlation between road network deformation and traffic flow spatiotemporal changes without the need to pre-set any correlation rules. It is suitable for diverse road network deformation scenarios and provides core fusion feature support for subsequent weight correction.

[0011] Furthermore, in the dual-branch feature coupling and dynamic weight correction, the dynamic correction of the prediction model weights based on correlation rules is achieved using a deformation influence degree weight iterative algorithm. The dynamic weight update is completed through the prediction model weight correction formula, which is: ,in To correct the weights of the fully connected layers in the prediction model, The original weights before correction, and the initial weights for model pre-training, are generated from a massive amount of highway traffic flow and deformation samples. The learning rate for the weights is adaptively adjusted based on the stability of the fused features; smaller feature fluctuations result in a lower learning rate. Take a value between 0.001 and 0.003; if the characteristic fluctuation is large... Take a value between 0.005 and 0.008. The residual coefficient is the factor representing the effect of deformation on traffic flow. It is calculated from the factor representing the effect of road network deformation on traffic flow. The larger the factor representing the effect, the greater the impact. The larger the value, As a transpose matrix of dual-branch fusion features, this formula breaks through the limitations of traditional static weights and realizes incremental dynamic correction of weights. When correcting weights, the residual coefficient of deformation influence is first calculated based on the fusion features, and then substituted into the formula to complete the iterative update of weights. There is no need to retrain the model with the full amount of data, which not only ensures the adaptability of weights, but also improves the efficiency of weight update, and adapts to the real-time needs of dynamic changes in road network deformation.

[0012] Furthermore, the short-term prediction output and the short-term traffic flow prediction results of the target road segment output by the deep learning model based on the corrected weights in the whole process optimization are implemented in a multi-dimensional hierarchical output manner. The specific implementation steps are as follows: First, the core duration range of the short-term prediction is determined to be 15 minutes to 3 hours, and it is divided into three levels according to the time granularity: 15 minutes, 30 minutes, and 1 hour, to meet the differentiated needs of different traffic control scenarios; then, the basic traffic flow prediction is completed based on the model with corrected weights, and the core prediction data of the target road segment within each time level is output, including the total traffic flow, peak traffic flow value, etc. The system outputs peak traffic flow times, trough traffic flow values, and trough traffic flow times. Next, for road network deformation scenarios of the target road segment, it additionally outputs traffic flow distribution trend prediction data, including the direction of traffic congestion spread, congestion duration, lane traffic flow distribution ratio, and traffic flow transmission trends between adjacent road segments. Finally, all prediction data is standardized and packaged, output according to the traffic control platform's preset format, and includes a prediction result credibility rating. The credibility rating is determined based on the model's current weight fit and the completeness of the input data, providing traffic control personnel with accurate and comprehensive decision-making basis and enhancing the practical application value of the prediction results.

[0013] Furthermore, the specific implementation steps of the short-term prediction output and the closed-loop verification and iterative optimization steps of prediction accuracy in the whole process optimization are as follows: After outputting the short-term traffic flow prediction results, the actual traffic flow monitoring data of the target road segment corresponding to the time period is collected in real time, and the actual data and prediction data are aligned and matched according to the same time granularity and spatial range; then the prediction accuracy is calculated using the relative error method, and the preset error threshold is defined. The prediction threshold is set to ≤3% for 15-minute level, ≤4% for 30-minute level, and ≤5% for 1-hour level. If the prediction error does not exceed the preset threshold, the prediction is deemed qualified and the current model weight is saved; if the prediction error exceeds the preset threshold, the error data is used as the core feedback signal, and the traffic flow features and road network deformation features corresponding to the error are extracted. The sensitivity of the dual-branch feature extraction and the correlation of the dual-branch coupling are optimized in a targeted manner. Then, the model weight is retrained iteratively based on the optimized parameters to complete the model optimization; at the same time, the scenario type of each error exceeding the threshold is recorded to form an error scenario library. Subsequently, the model feature learning is strengthened for similar scenarios to achieve continuous iterative improvement of prediction accuracy and ensure the long-term adaptability of the model.

[0014] Furthermore, the method also includes core technical logic that allows for autonomous adaptation to dynamic changes in the road network without pre-setting rules for associating road network deformation with traffic flow. The specific implementation steps are as follows: During the model training phase, massive amounts of road network deformation data of different types, ranges, and durations, along with corresponding traffic flow data samples, are input. This allows the deep learning dual-branch architecture to autonomously learn the traffic flow variation patterns under different road network deformation scenarios, without requiring manual setting of rules for which deformation corresponds to which traffic flow change. During the model prediction phase, for the real-time input of unknown road network deformation data, the core features are first extracted through the road network deformation feature encoding branch. The model incorporates features that are coupled with the spatiotemporal characteristics of traffic flow, allowing it to autonomously access the association patterns learned during training without the need for additional rule base loading. For sudden and unconventional road network deformations, the model autonomously mines the potential correlation between new deformations and traffic flow through deep coupling of dual-branch features, dynamically adjusting weights for adaptation. Simultaneously, the model possesses self-learning capabilities, supplementing the sample database with new deformation-traffic flow data pairs after each prediction and performing incremental pre-training periodically to continuously improve its adaptability to new road network deformation scenarios, completely eliminating reliance on preset rules and filling the technical gap in traditional models that require manual rule setting.

[0015] On the other hand, a short-term highway traffic flow prediction system based on deep learning includes: a multi-source data acquisition and preprocessing module, a dual-branch feature encoding module, a feature coupling and dynamic weight correction module, and a multi-dimensional prediction output and accuracy optimization module. The multi-source data acquisition and preprocessing module acquires real-time traffic flow monitoring data, historical traffic flow data, and road network deformation data of the target road segment, and optimizes data quality using a layered multi-source acquisition + hierarchical preprocessing approach. The dual-branch feature encoding module: constructs a deep learning dual-branch architecture, including a road network deformation feature encoding submodule and a traffic flow spatiotemporal feature encoding submodule. The road network deformation feature encoding submodule uses a three-level serial architecture of preprocessing-extraction-enhancement to extract the core features of road network deformation, and the traffic flow spatiotemporal feature encoding submodule uses a spatiotemporal coupling feature extraction algorithm to extract the spatiotemporal coupling features of traffic flow. The feature coupling and weight dynamic correction module: adopts deformation-guided attention coupling algorithm to perform deep coupling calculation on the dual-branch output features, explores the correlation between road network deformation and traffic flow spatiotemporal features, and dynamically corrects the prediction model weights based on the correlation. The multi-dimensional prediction output and accuracy optimization module: Based on the deep learning model with corrected weights, it outputs short-term traffic flow prediction results for the target road segment in a multi-dimensional hierarchical manner, and performs closed-loop verification and iterative optimization of prediction accuracy. It achieves autonomous adaptation to dynamic changes in the road network without the need for preset road network deformation and traffic flow association rules, and has broad adaptability and convenient deployment capabilities.

[0016] Compared with existing technologies, this deep learning-based method and system for short-term prediction of highway traffic flow has the following advantages: I. This invention achieves accurate short-term prediction of highway traffic flow by combining multi-source data acquisition and preprocessing with a deep learning dual-branch architecture. The system can dynamically capture the impact of road network deformation on traffic flow and deeply mine the correlation between road network deformation and the spatiotemporal characteristics of traffic flow through a deformation-guided attention coupling algorithm, thereby dynamically correcting the prediction model weights. This process not only improves prediction accuracy but also ensures the real-time nature of the prediction results, providing timely and accurate data support for traffic management.

[0017] Second, this invention does not require pre-setting association rules between road network deformation and traffic flow, and can autonomously adapt to dynamic changes in the road network. During the model training phase, by inputting massive amounts of road network deformation data of different types, ranges, and durations, along with corresponding traffic flow data samples, the deep learning dual-branch architecture autonomously learns the changing patterns of traffic flow under different road network deformation scenarios. During the prediction phase, the system can autonomously call upon the association rules learned during the training phase to make predictions for unknown road network deformation data input in real time, and dynamically adjust weights for sudden and unconventional road network deformations. In addition, the system also has self-learning capabilities. After each prediction, it adds new deformation-traffic flow data pairs to the sample library and performs incremental pre-training periodically to continuously improve its adaptability to new road network deformation scenarios. This feature greatly enhances the model's adaptability and long-term value.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 The flowchart shows a method for short-term prediction of highway traffic flow based on deep learning. Figure 2 A flowchart outlining the specific steps involved in data collection for a deep learning-based method for short-term highway traffic flow prediction. Figure 3 This is a flowchart of a deep learning-based short-term highway traffic flow prediction system. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structure, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] Example 1

[0023] This embodiment follows the process of a deep learning-based method for short-term prediction of highway traffic flow (e.g.) Figure 1 (As shown) To conduct predictions, the specific steps are as follows: Three types of core data are acquired using a layered, multi-source acquisition method. The acquisition process is as follows: Figure 2 As shown: Real-time traffic flow monitoring data for the main road is collected via roadside loop detectors, high-definition checkpoint cameras, and microwave radar; historical traffic flow data for the past 6 months is retrieved from the traffic management cloud database; the specific scope of road occupancy during construction is recorded by roadside inspection terminals, and the construction progress is reported in real time by these terminals; traffic control requirements during construction are obtained from the traffic control instruction issuance platform; and drone aerial imagery confirms the road occupancy boundary. Simultaneously, traffic flow data and road conditions of two adjacent parallel branch roads are collected as auxiliary references. The three types of data are preprocessed in a tiered manner: first, abnormal traffic flow data caused by equipment failure or extreme weather is removed using an outlier identification algorithm; then, real-time monitoring data, historical data, and road network deformation data are standardized and aligned to unify the temporal and spatial granularity of the data; finally, for data gaps caused by equipment offline in certain time periods, interpolation is performed using data from adjacent time periods and adjacent road sections to ensure data integrity.

[0024] Constructing a dual-branch architecture for deep learning, such as Figure 3 As shown, this architecture corresponds to the dual-branch feature encoding module in the system, which performs feature extraction separately. The road network deformation feature encoding branch adopts a three-level serial architecture of preprocessing-extraction-enhancement: First, the preprocessing unit filters out the core data directly related to the impact of traffic flow; then, the deformation feature extraction unit performs deep feature mining, spatially gridding the scope of construction occupation, temporally segmenting the duration of construction, and class-specific one-hot encoding the deformation type. Subsequently, the convolutional layer completes the feature mapping of the encoded data, realizing the initial extraction of deformation features; finally, the feature enhancement unit strengthens the key deformation features, and the key features are fused with the initially extracted features through the residual connection structure to obtain the core features of road network deformation. The traffic flow spatiotemporal feature branch is based on a spatiotemporal convolutional network and a gated recurrent unit, and adopts a spatiotemporal coupled feature extraction algorithm, the formula of which is: ,in As the core feature of the spatiotemporal coupling of traffic flow, For the spatial topological relationship features of traffic flow, The characteristics of traffic flow temporal evolution, For spatial convolution operations, For timing-gated operations, For spatiotemporal feature fusion weights, The spatial feature weight matrix is... As the time feature weight matrix, combined with the spatial topology and temporal evolution of the road segment, the spatiotemporal coupling features of traffic flow are extracted using a spatiotemporal coupling feature fusion formula. This accurately captures the spatial distribution differences and temporal variation patterns of traffic flow. The formula is as follows: ,in It is a two-branch fusion feature. The spatiotemporal coupling characteristics of traffic flow As the core feature of road network deformation, The deformation guides the attention weighting coefficient.

[0025] A deformation-guided attention coupling algorithm is used to perform deep coupling calculations on the features output from the two branches mentioned above, such as... Figure 3 As shown, the correlation between road network deformation features and traffic flow spatiotemporal features is mined using a dual-branch feature coupling weight allocation formula. Based on the mined correlation, a deformation influence weight iterative algorithm is used to dynamically correct the prediction model weights. The dynamic update of the weights is achieved through the prediction model weight correction formula, which is as follows: ,in To correct the weights of the fully connected layers in the prediction model, To correct the original weights, The learning rate for weight iteration. The residual coefficient is affected by deformation. The model is a dual-branch fusion feature transpose matrix. It focuses on increasing the feature weights corresponding to peak construction periods and core road occupation areas, while reducing the weights of irrelevant features during periods without construction impact, to ensure that the model can accurately adapt to traffic flow patterns under construction road occupation scenarios.

[0026] Based on a deep learning model with corrected weights, the prediction results are output using a multi-dimensional, hierarchical output method. The core duration range for short-term predictions is defined as 15 minutes to 3 hours, divided into three levels based on time granularity: 15-minute, 30-minute, and 1-hour levels. For 15-minute level predictions, the total traffic volume, peak traffic volume and its occurrence time, and valley traffic volume and its occurrence time are output. For 30-minute and 1-hour level predictions, in addition to outputting the total traffic volume, peak and valley traffic volumes and their corresponding time periods, additional traffic distribution trend prediction data is output, including the direction of traffic congestion spread, congestion duration, traffic distribution ratio of each lane, and traffic transmission trends between adjacent road segments. All prediction data is standardized and packaged according to the traffic control platform's preset format, and includes a prediction result reliability rating. Simultaneously, a closed-loop verification of prediction accuracy is performed. After each prediction period ends, real-time monitoring data of traffic flow for the corresponding time period is collected for that road segment. The actual data and the predicted data are aligned and matched with the same time granularity and spatial range. The prediction accuracy is calculated using the relative error method, with preset error thresholds of ≤3% for 15-minute intervals, ≤4% for 30-minute intervals, and ≤5% for 1-hour intervals. If the prediction error does not exceed the threshold, the prediction is considered qualified, and the current model weights are saved. If the prediction error at a certain level exceeds the threshold, the error data is used as the core feedback signal to extract the traffic flow features and road network deformation features corresponding to that time period. The sensitivity of the dual-branch feature extraction and the correlation of the dual-branch coupling are optimized in a targeted manner. Then, the model weights are iteratively trained based on the optimized parameters to complete the model optimization. Throughout the process, the model does not require manual pre-setting of rules for traffic flow changes corresponding to road occupancy during construction. Through the massive construction scenario samples learned during the training phase, it autonomously adapts to the dynamic changes in the road network during construction. Even if the construction progress is temporarily adjusted, the model can autonomously explore the potential correlation between new deformations and traffic flow through the deep coupling of dual-branch features and dynamically adjust the weights accordingly.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A short-term highway traffic flow prediction method based on deep learning, characterized in that, The specific steps of this method are as follows: Multi-source data acquisition and preprocessing: real-time traffic flow monitoring data, historical traffic flow data and road network deformation data of the target road segment are acquired. The road network deformation data includes deformation type, deformation range and deformation duration. The data quality is optimized by using a hierarchical multi-source acquisition + graded preprocessing method. Dual-branch architecture construction and feature extraction: A deep learning dual-branch architecture is constructed, namely a road network deformation feature encoding branch and a traffic flow spatiotemporal feature branch. The road network deformation feature encoding branch adopts a three-level serial architecture of preprocessing-extraction-enhancement, and the traffic flow spatiotemporal feature branch adopts a spatiotemporal coupled feature extraction algorithm to extract the core features of road network deformation and the spatiotemporal coupled features of traffic flow, respectively. Dual-branch feature coupling and dynamic weight correction: The features output by the dual branches are deeply coupled using a deformation-guided attention coupling algorithm to mine the correlation between road network deformation features and traffic flow spatiotemporal features. Based on the correlation, the deformation influence weight iterative algorithm is used to dynamically correct the prediction model weights. Short-term forecast output and full-process optimization: The deep learning model based on the corrected weights outputs the short-term traffic flow forecast results of the target road segment in a multi-dimensional hierarchical output manner, and performs closed-loop verification and iterative optimization of the forecast accuracy to complete the short-term forecast of highway traffic flow.

2. The method for short-term prediction of highway traffic flow based on deep learning according to claim 1, characterized in that, The multi-source data acquisition and preprocessing process employs a layered multi-source acquisition + hierarchical preprocessing approach to acquire and optimize real-time traffic flow monitoring data, historical traffic flow data, and road network deformation data. The specific implementation steps are as follows: First, real-time traffic flow monitoring data is acquired using roadside loop detectors, high-definition checkpoint cameras, and microwave radar. Historical traffic flow data for the target road segment over the past six months is retrieved from the traffic control cloud database. Road network deformation data is acquired using roadside inspection terminals, traffic control instruction issuance platforms, and drone aerial imagery. The road network deformation types explicitly include five categories: road construction occupation, lane closure, emergency control, temporary detours, and road surface defects. Traffic flow and deformation data from adjacent road segments are also supplemented as auxiliary references. Next, the three types of basic data undergo hierarchical preprocessing. First, outlier identification and removal are performed. Then, data standardization and alignment are performed to unify the temporal and spatial granularities of the data. Finally, data completion is performed, with interpolation of adjacent time periods and adjacent road segments from the same source used to fill in missing data points.

3. The short-term highway traffic flow prediction method based on deep learning according to claim 1, characterized in that, The road network deformation feature encoding branch in the dual-branch support construction and feature extraction adopts a three-level serial architecture of preprocessing-extraction-enhancement to accurately extract the core features of road network deformation. The specific implementation steps are as follows: First, the road network deformation data is filtered for features through the preprocessing unit; then, the deformation feature extraction unit performs deep feature mining on the filtered data, firstly, spatial gridding encoding is performed on the deformation range, temporal segmentation encoding is performed on the deformation duration, and category-based one-hot encoding is performed on the deformation type; then, the feature mapping of the encoded data is completed through the convolutional layer to achieve the initial extraction of deformation features; finally, the key deformation features are enhanced through the feature enhancement unit, focusing on enhancing the core deformation features that affect traffic flow, and the key features are fused with the initially extracted features through the residual connection structure.

4. The method for short-term prediction of highway traffic flow based on deep learning according to claim 1, characterized in that, In the construction and feature extraction of the dual-branch support, the traffic flow spatiotemporal feature branch adopts a spatiotemporal coupled feature extraction algorithm to accurately mine the spatiotemporal coupled features of the traffic flow. This branch is based on a spatiotemporal convolutional network and a gated recurrent unit architecture, and completes feature output through a traffic flow spatiotemporal coupled feature fusion formula, which is as follows: ,in As the core feature of the spatiotemporal coupling of traffic flow, For the spatial topological relationship features of traffic flow, The characteristics of traffic flow time sequence evolution, For spatial convolution operations, For timing-gated operations, For spatiotemporal feature fusion weights, The spatial feature weight matrix, This is the time feature weight matrix.

5. The method for short-term prediction of highway traffic flow based on deep learning according to claim 1, characterized in that, The feature depth coupling calculation of the bi-branch output in the bi-branch feature coupling and dynamic weight correction is implemented using a deformation-guided attention coupling algorithm. The association pattern is mined through the bi-branch feature coupling weight allocation formula, which is: ,in It is a two-branch fusion feature. The spatiotemporal coupling characteristics of traffic flow As the core feature of road network deformation, The deformation-guided attention weighting coefficient.

6. The short-term highway traffic flow prediction method based on deep learning according to claim 1, characterized in that, In the bi-branch feature coupling and dynamic weight correction, the dynamic correction of the prediction model weights based on correlation rules is achieved using a deformation influence degree weight iterative algorithm. The dynamic weight update is accomplished through the prediction model weight correction formula, which is: ,in To correct the weights of the fully connected layers in the prediction model, To correct the original weights, The learning rate for weight iteration. The residual coefficient is affected by deformation. This is the transpose matrix of the bi-branch fusion feature.

7. The method for short-term prediction of highway traffic flow based on deep learning according to claim 1, characterized in that, The short-term prediction output and the short-term traffic flow prediction results of the target road segment output by the deep learning model based on the corrected weights in the whole process optimization are achieved through a multi-dimensional hierarchical output method. The specific implementation steps are as follows: First, the core duration range of the short-term prediction is determined to be 15 minutes to 3 hours, divided into three levels according to the time granularity: 15 minutes, 30 minutes, and 1 hour, to meet the differentiated needs of different traffic control scenarios; then, the basic traffic flow prediction is completed based on the model with corrected weights, and the core prediction data of the target road segment within each time level is output, including the total traffic flow, peak traffic flow value, peak traffic flow time, valley traffic flow value, and valley traffic flow time; next, for the road network deformation scenario of the target road segment, additional traffic flow distribution trend prediction data is output, including the direction of traffic congestion spread, congestion duration, lane traffic flow distribution ratio, and traffic flow transmission trend of adjacent road segments; finally, all prediction data are standardized and packaged, output according to the preset format of the traffic control platform, and a prediction result credibility rating is attached. The credibility rating is determined based on the current weight adaptation of the model and the completeness of the input data, providing traffic control personnel with accurate and comprehensive decision-making basis.

8. The method for short-term prediction of highway traffic flow based on deep learning according to claim 1, characterized in that, The specific implementation steps of the short-term prediction output and the closed-loop verification and iterative optimization steps of prediction accuracy in the whole process optimization are as follows: After outputting the short-term traffic flow prediction results, the actual traffic flow monitoring data of the target road segment corresponding to the time period is collected in real time, and the actual data and prediction data are aligned and matched according to the same time granularity and spatial range; then the prediction accuracy is calculated using the relative error method, and the preset error threshold is defined. The prediction threshold is set to ≤3% for 15-minute level, ≤4% for 30-minute level, and ≤5% for 1-hour level. If the prediction error does not exceed the preset threshold, the prediction is deemed qualified and the current model weight is saved; if the prediction error exceeds the preset threshold, the error data is used as the core feedback signal, and the traffic flow features and road network deformation features corresponding to the error are extracted. The sensitivity of the dual-branch feature extraction and the correlation of the dual-branch coupling are optimized in a targeted manner, and the model weights are retrained iteratively based on the optimized parameters to complete the model optimization.

9. The method for short-term prediction of highway traffic flow based on deep learning according to claim 1, characterized in that, The method also includes core technical logic that allows for autonomous adaptation to dynamic changes in the road network without pre-setting rules for the association between road network deformation and traffic flow. The specific implementation steps are as follows: During the model training phase, massive amounts of road network deformation data of different types, ranges, and durations, along with corresponding traffic flow data samples, are input. This allows the deep learning dual-branch architecture to autonomously learn the traffic flow change patterns under different road network deformation scenarios, without requiring manual setting of rules for which deformation corresponds to which traffic flow change. During the model prediction phase, for real-time input of unknown road network deformation data, the core features are first extracted through the road network deformation feature encoding branch, and then coupled with the spatiotemporal features of traffic flow. The model autonomously calls the association rules learned during the training phase, without needing to load an additional rule base. For sudden, unconventional road network deformations, the model autonomously mines the potential association between new deformations and traffic flow through deep coupling of dual-branch features, dynamically adjusting the weights for adaptation.

10. A deep learning-based short-term highway traffic flow prediction system, applicable to the deep learning-based short-term highway traffic flow prediction method described in any one of claims 1-9, characterized in that, The system includes: a multi-source data acquisition and preprocessing module, a dual-branch feature encoding module, a feature coupling and dynamic weight correction module, and a multi-dimensional prediction output and accuracy optimization module. The multi-source data acquisition and preprocessing module acquires real-time traffic flow monitoring data, historical traffic flow data, and road network deformation data of the target road segment, and optimizes data quality using a layered multi-source acquisition + hierarchical preprocessing approach. The dual-branch feature encoding module: constructs a deep learning dual-branch architecture, including a road network deformation feature encoding submodule and a traffic flow spatiotemporal feature encoding submodule. The road network deformation feature encoding submodule uses a three-level serial architecture of preprocessing-extraction-enhancement to extract the core features of road network deformation, and the traffic flow spatiotemporal feature encoding submodule uses a spatiotemporal coupling feature extraction algorithm to extract the spatiotemporal coupling features of traffic flow. The feature coupling and weight dynamic correction module: adopts deformation-guided attention coupling algorithm to perform deep coupling calculation on the dual-branch output features, explores the correlation between road network deformation and traffic flow spatiotemporal features, and dynamically corrects the prediction model weights based on the correlation. The multi-dimensional prediction output and accuracy optimization module: Based on the deep learning model with corrected weights, it outputs the short-term prediction results of traffic flow on the target road segment in a multi-dimensional hierarchical manner, and performs closed-loop verification and iterative optimization of prediction accuracy.

Images