A highway traffic congestion prediction method and system based on deep learning

Abstract

The application belongs to the technical field of expressway, and particularly relates to an expressway traffic congestion prediction method and system based on deep learning. The method comprises the following steps: acquiring traffic flow characteristic information, identifying a traffic bottleneck node based on the traffic flow characteristic information, summarizing reference characteristic values of the traffic bottleneck node to form an overall traffic flow characteristic data set, performing discrete degree analysis on the data set to obtain a characteristic index, extracting a traffic congestion concentrated interval according to the characteristic index, identifying a risk time point in the interval, and outputting a predicted expressway congestion period and a corresponding congestion area according to the risk time point and the traffic congestion concentrated interval. The application can improve the accuracy of congestion root identification by identifying a traffic bottleneck node and quantitatively analyzing the evolution trend of the bottleneck state, and can output a region to be regulated and a regulation period, thereby providing decision support for traffic relief.

Description

Technical Field

[0001] This invention belongs to the field of highway technology, specifically relating to a method and system for predicting highway traffic congestion based on deep learning. Background Technology

[0002] With the rapid development of intelligent transportation systems, modern highways have become an infrastructure supporting the operation of the national economy and society. Today's highways are more like a complex network, consisting of main lines, ramps and branch lines. Against this backdrop, efficient and intelligent management and scheduling of traffic flow in the entire road network to improve traffic capacity and ensure driving safety has become the focus of attention in the field of traffic engineering. To achieve proactive traffic management and control, the prerequisite is to effectively monitor and accurately predict traffic congestion.

[0003] Existing highway traffic congestion management technologies have many shortcomings and bottlenecks in practical applications. At the data acquisition level, traditional solutions heavily rely on a large number of physical sensing devices deployed along the route, such as inductive loop detectors, microwave radar, and video surveillance. These devices are not only costly to purchase, install, and rent, but also require continuous investment of manpower and funds for maintenance and calibration. This severely limits the coverage density and breadth of the monitoring system, making it difficult to achieve comprehensive, blind-spot-free perception of the road network. Furthermore, existing technologies struggle to understand the complex dynamic evolution of traffic flow, making it difficult to accurately predict the occurrence, development, and dissipation of congestion. In particular, they cannot accurately assess the cascading effects of localized congestion on adjacent routes and even the entire road network, resulting in management decisions that fail to achieve truly precise traffic management and proactive flow allocation.

[0004] In view of this, this application provides a method and system for predicting highway traffic congestion based on deep learning. Summary of the Invention

[0005] The purpose of this invention is to provide a highway traffic congestion prediction method based on deep learning, in order to solve the problem that the deployment of a large number of sensing devices in the existing technology not only involves rental costs, but also requires investment of manpower and resources for construction and maintenance.

[0006] The specific technical solution adopted by this invention is as follows: This invention provides a deep learning-based method for predicting highway traffic congestion. When the feature indicators obtained from the discreteness analysis of the overall traffic flow feature dataset of the target highway meet the preset congestion prediction trigger conditions, the following steps are executed: Identify risk points within traffic congestion concentration zones extracted based on characteristic indicators; Based on the risk time points and traffic congestion concentration areas, the predicted highway congestion time periods and their corresponding congestion areas are output. The dispersion analysis of the overall traffic flow feature dataset of the target highway includes: arranging the reference feature values ​​in the overall traffic flow feature dataset in time period order; calculating the difference between adjacent reference feature values ​​in turn; determining a weight factor for each difference value according to a preset weighting rule; and weighting the weight factor and the difference value to obtain a set of feature values ​​to be classified as feature indicators. Identifying risk points within traffic congestion clusters extracted based on feature indicators includes: processing the set of feature values ​​to be classified using clustering methods to output multiple discrete feature scores; identifying fluctuation and abrupt change points in the discrete feature scores where the numerical changes exceed a preset abrupt change threshold, and defining the areas on both sides of the fluctuation and abrupt change points as traffic congestion clusters; and identifying, within the traffic congestion clusters, the time points where the increase in discrete feature scores exceeds a preset score change threshold, and defining these time points as risk points.

[0007] In a preferred technical solution, forming the overall traffic flow feature dataset includes: Based on traffic flow characteristic information obtained from the target highway, multiple traffic bottleneck nodes are identified; The reference feature values ​​corresponding to all traffic bottleneck nodes are summarized to form an overall traffic flow feature dataset.

[0008] In a preferred technical solution, traffic flow characteristic information includes congested areas, and the methods for obtaining congested areas include: The route of the target highway is divided into grids, and the vehicle target information associated with each grid is collected; Based on preset congestion rules, determine whether the vehicle target information indicates that vehicles are arranged in a continuous congestion sequence; If the determination is yes, the corresponding grid is marked as congested, and one or more grids in the congested state are identified as congested areas.

[0009] In a preferred technical solution, identifying multiple traffic bottleneck nodes based on traffic flow characteristic information obtained from the target highway includes: For grids within congested areas, extract their center points as nodes to be determined; Determine whether the node to be judged has a reverse blocking effect. If the node to be judged can stop vehicles in at least one direction and the duration of the stop exceeds a preset stop duration threshold, then it is determined to have a reverse blocking effect and is identified as a traffic bottleneck node.

[0010] In a preferred technical solution, determining the reference feature value includes: For each traffic bottleneck node, obtain the area of ​​its associated congested area within multiple preset time periods; The areas within multiple preset time periods are sorted, and the maximum area value is recorded as the reference feature value of the traffic bottleneck node.

[0011] In a preferred embodiment, after outputting the predicted highway congestion periods and their corresponding congestion areas, the method further includes: Determine whether there are any points in the predicted highway congestion period where the change in traffic flow exceeds a preset traffic flow change threshold. If no specific time point exists, the risk time point and the corresponding congested area will be output as the area to be regulated and the regulation period.

[0012] This invention also provides a deep learning-based highway traffic congestion prediction system, comprising the following modules: The traffic flow feature acquisition module is used to acquire traffic flow feature information of the target highway in order to construct an overall traffic flow feature dataset; The congestion risk analysis module is used to perform dispersion analysis on the overall traffic flow feature dataset to obtain feature indicators, and when the feature indicators meet the preset congestion prediction trigger conditions, it identifies the risk time points in the traffic congestion concentration area extracted based on the feature indicators.

[0013] In a preferred embodiment, the technology further includes: The congestion prediction output module is used to output the predicted highway congestion period and its corresponding congestion area in response to the congestion risk analysis module's identification of risk time points. The congestion prediction output module is also used to determine whether there are time points in the predicted highway congestion period where the traffic flow change value exceeds the preset traffic flow change threshold, and if there are no time points, output the risk time point and the corresponding congestion area as the area to be controlled and the control period.

[0014] In a preferred technical solution, traffic flow characteristic information includes congested areas, and the methods for obtaining congested areas include: The route of the target highway is divided into grids, and the vehicle target information associated with each grid is collected; Based on preset congestion rules, determine whether the vehicle target information indicates that vehicles are arranged in a continuous congestion sequence; If the determination is yes, the corresponding grid is marked as congested, and one or more grids in the congested state are identified as congested areas.

[0015] In a preferred technical solution, the identification of risk points includes: The clustering method is used to process the set of feature values ​​to be classified, which are used as feature indicators, and output multiple discrete feature scores. The fluctuation and mutation points in the discrete feature scores that exceed the preset mutation threshold are identified, and the areas on both sides of the fluctuation and mutation points are determined as traffic congestion concentration areas. In addition, within areas of concentrated traffic congestion, the time point at which the increase in discrete feature scores exceeds a preset score change threshold is identified, and this time point is determined as a risk point.

[0016] Beneficial effects This invention identifies traffic bottleneck nodes by determining whether the node to be judged has a reverse blocking effect, even when vehicles are stopped for more than a preset stopping time threshold. It accurately identifies the source of congestion from a physical perspective, effectively distinguishing between slow-moving vehicles and traffic bottleneck nodes that trigger chain reactions. Compared with relying solely on traditional methods, this invention improves the accuracy of congestion source identification and provides a reliable data foundation for subsequent predictive analysis.

[0017] This invention summarizes the reference characteristic values ​​of each traffic bottleneck node and obtains characteristic indicators describing the trend of traffic flow changes through dispersion analysis. Based on these characteristic indicators, this invention can extract the concentrated traffic congestion intervals and further identify risk points. By quantifying the bottleneck state and analyzing the dynamic evolution trend, it achieves a refined and multi-level characterization of the congestion process. It not only predicts the occurrence of congestion but also defines the congestion periods and risk points on highways, improving the accuracy and granularity of the prediction results in the time dimension.

[0018] After outputting the prediction results, this invention further determines whether the traffic flow changes within the corresponding time period are stable. If stable, it outputs the area to be regulated and the regulation period, directly linking the prediction results with actual control needs. This extends the process from prediction to decision-making, enabling the screening of stable congestion events and providing clear areas to be regulated and regulation periods. This provides direct and operable decision support for traffic management departments, improving the pertinence and timeliness of traffic diversion measures. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0020] Example 1 Please refer to Figure 1 This embodiment discloses a deep learning-based method for predicting highway traffic congestion, including the following steps: S1. First, traffic flow characteristic information is obtained to establish an accurate and quantitative data foundation for subsequent congestion analysis. The target highway route is processed by gridding, for example, dividing each kilometer segment into multiple 100-meter by 100-meter grid units, assigning a unique identifier to each grid unit, and collecting and aggregating the equipment information in each grid unit and its associated vehicle target information through sensing devices such as cameras, microwave radar, and geomagnetic coils installed on the roadside. The device information includes the sensor's identification (ID) and geographic coordinates, while the vehicle target information includes the instantaneous speed of the vehicle when passing through the grid cell, the vehicle type classification, and license plate recognition data. Based on the collected data, according to the preset congestion rules, it is determined whether the vehicle target information reflects that the vehicles have formed a continuous, low-speed driving or stationary queue. Preset congestion rules refer to identifying situations where traffic flow is interrupted. For example, within a single grid cell, if the average vehicle speed is detected to be below 20 kilometers per hour for three consecutive minutes and the average following distance between vehicles is less than 5 meters, the corresponding grid cell is marked as congested. All grid cells with congested status are spatially aggregated to form a congested area.

[0021] S2. Within the defined congested area, accurately identify traffic bottleneck nodes. For the geometric center point of the grid cell within the congested area, use the extracted center point as the node to be determined. Verify whether the node to be determined has a reverse blocking effect and determine whether the node to be determined is a traffic bottleneck node. Reverse blocking effect is a key indicator for measuring whether a local congestion point has evolved into an upstream traffic flow bottleneck. The judgment condition is: the node to be judged can cause vehicles to stop continuously in the upstream direction of the traffic flow, and the duration of the stop exceeds the preset stop duration threshold. The preset stop duration threshold is set to 120 seconds to filter out brief stops caused by traffic light changes or temporary yielding behavior, ensuring that the identified stop status is a continuous congestion caused by internal traffic flow conflicts. If the node to be judged is confirmed to have a reverse blocking effect, that is, its negative impact on upstream traffic flow is confirmed to be significant, then the node to be judged is identified as a traffic bottleneck node, and the preset area surrounding the node, such as a circular area with a radius of 500 meters centered on the traffic bottleneck node or other nine adjacent grid units, is marked as the bottleneck area. At the same time, the specific preset time period when the traffic bottleneck node is first confirmed is recorded as its occurrence time period.

[0022] S3. To quantify the impact of traffic bottleneck nodes, the reference feature values ​​are summarized and the dataset is constructed. The specific processing method is as follows: for the identified traffic bottleneck nodes, every 5 minutes is taken as an observation period within a continuous preset time period, and the area of ​​the corresponding bottleneck area is continuously tracked and calculated. The area is a direct quantitative representation of its impact range. The larger the area, the more severe the upstream or lateral spread of the congestion caused. After acquiring the area data recorded in multiple preset time periods, the area values ​​are sorted, and the largest area value is recorded as the reference feature value of the traffic bottleneck node. The purpose of selecting the maximum area value is to capture the maximum impact that the bottleneck can cause throughout the entire observation period, as a stable indicator to measure its severity, thereby avoiding deviations in the assessment results due to instantaneous fluctuations in traffic flow. All traffic bottleneck nodes, their corresponding reference feature values, and occurrence times are summarized to form a structured overall traffic flow feature dataset that describes the overall traffic flow characteristics of the target highway, providing input for subsequent dynamic trend analysis.

[0023] S4. After obtaining the overall traffic flow characteristic dataset, perform a discreteness analysis to reveal the dynamic evolution of traffic congestion. Arrange the reference characteristic values ​​of all traffic bottleneck nodes in the dataset according to the time sequence of occurrence to form a time series. Calculate the difference between adjacent reference characteristic values ​​in the time series. This difference directly reflects the rate of change of congestion severity in the time dimension. According to the preset weighting rules, a weighting factor is determined for the calculated difference value. The purpose of this rule is to make the analysis process closer to the actual traffic situation. For example, the difference value that occurs during the commuting peak hours, such as 7:00-9:00 on weekdays, is assigned a weighting factor of 1.5; while the difference value that occurs near important transportation hubs, such as highway entrances and exits or service areas, is assigned a weighting factor of 1.2. The weighting factor is multiplied by the corresponding difference value to obtain the set of feature values ​​to be classified. The values ​​in this set quantify the severity of traffic state changes at a specific time and place, and together they constitute the feature index describing the trend of traffic flow changes, which is used for subsequent congestion state classification.

[0024] S5. To identify the period with the highest congestion risk from a continuous time series, it is necessary to extract the concentrated traffic congestion intervals. This step performs a grouping and classification process on the set of feature values ​​to be classified. Based on the similarity of the numerical values, each value in the set is automatically divided into a preset number of categories, and a discrete feature score level is assigned to each category. The feature values ​​are divided into three categories and assigned discrete feature scores of level 1, 2, and 3, respectively, to represent the stable, aggravated, and rapidly deteriorating congestion status. After obtaining multiple discrete feature scores, these scores are used to extract traffic congestion concentration areas. Specifically, the discrete feature scores are arranged in their corresponding time order; the specific time point in this sorted score sequence where the numerical change exceeds a preset abrupt change threshold is identified, and this point is used as the critical value of the fluctuation abrupt change point, which determines the boundary between normal fluctuations and critical transitions in traffic conditions. This threshold is a key parameter for judging normal fluctuations in traffic flow and critical transitions between states. Abrupt changes in fluctuation indicate that the traffic system transitions from a relatively stable state to another state. Based on this, the areas on both sides of the point are respectively identified as periods of severe congestion and periods of normal congestion, which together constitute the concentrated traffic congestion area that requires close attention.

[0025] S6. Within the identified traffic congestion concentration areas, further identify the congestion start point and risk time point. The purpose is to accurately locate the specific time when congestion begins to form, thereby improving the accuracy and timeliness of the early warning. The specific processing method is as follows: During periods of severe congestion and ordinary congestion, a sliding time window method is used to continuously monitor the current time point and several subsequent time points within the preset time window. For example, it includes three future time points, i.e., a 15-minute time window, and then calculates the difference between the discrete feature score at the current time point and the discrete feature score at subsequent time points; If the difference consistently exceeds the preset score change threshold and the change direction is continuously increasing, it indicates that the congestion level is experiencing a rapid and continuous increase. This is a strong signal that congestion is about to erupt fully. The preset score change threshold is used to identify the turning point from quantitative change to qualitative change. Once this condition is met, the current time point is determined as the congestion start point, and this congestion start point is recorded as a risk time point, serving as the core time basis for prediction output and control decisions.

[0026] S7. Based on the identified risk points and the analyzed traffic congestion concentration areas, execute prediction output and control decisions, output the predicted congestion areas and their corresponding time periods as prediction results, which are used for external information dissemination and traffic participant guidance. On this basis, in order to achieve intelligent traffic control, further execute decision steps to determine whether there are time points in the corresponding time period where the traffic flow change value exceeds the preset traffic flow change threshold. The traffic flow change value here reflects the self-regulation capability of the road network. If the traffic efficiency of downstream road sections improves, resulting in a significant net outflow of traffic from congested areas, it indicates that the congestion is likely to dissipate on its own. If such a point in time exists, it means that the current process can end without external intervention; Conversely, if such a time point does not exist, it indicates that the congestion has stabilized or continued to worsen and cannot recover on its own. In this case, the risk time point and the predicted congestion area are output as the area to be regulated and the regulation period, and this information is sent to the traffic management platform to initiate corresponding diversion or control measures. This embodiment achieves intelligent processing across the entire chain, from data collection to congestion prediction and control decisions, through precise location, quantitative analysis, and dynamic trend prediction of traffic bottleneck nodes. This not only improves the accuracy of predictions but also incorporates a mechanism for judging the self-regulating capacity of the road network, avoiding unnecessary intervention and thus improving operational efficiency.

[0027] Example 2 Please refer to Figure 2 This embodiment discloses a highway traffic congestion prediction system, including the following modules: Traffic flow feature acquisition module: The traffic flow characteristic information of the target highway is obtained, and an overall traffic flow characteristic dataset is constructed based on this. In the execution process, the route of the target highway is logically divided into grids, for example, each 100-meter segment is divided into a grid unit. The vehicle target information associated in each grid unit is continuously collected, such as the number of vehicles, average speed, headway, etc. According to the preset congestion rules, it is determined whether the vehicle target information represents a continuous arrangement of vehicles forming a congestion sequence. For example, if the average vehicle speed in the grid cell is less than 20 km / h and the vehicle density exceeds 80 vehicles per kilometer, if it is determined to be yes, the corresponding grid cell is marked as congested. Then, the grid cells that are congested in space or adjacent to each other are determined as congested areas. These congested areas constitute part of the traffic flow feature information. To further refine key features, the geometric center point of the identified congested area is extracted as the node to be judged, and it is determined whether the node to be judged has a reverse blocking effect. If the existence of the node can cause vehicles to stop in the upstream direction, and the duration of the stop exceeds the preset stop duration threshold, such as 3 minutes, then the node is determined to have a reverse blocking effect and is identified as a traffic bottleneck node. The reference feature values ​​corresponding to all traffic bottleneck nodes are aggregated to form an overall traffic flow feature dataset. The reference feature values ​​are determined by obtaining the situation of the congested area associated with each traffic bottleneck node in the past preset time period, such as the area of ​​each 5-minute period in the past 24 hours. Then, these area values ​​are sorted, and the largest area value recorded is used as the reference feature value of the traffic bottleneck node at that time point. By periodically executing the above process, a total traffic flow feature dataset arranged in time series is constructed and continuously updated.

[0028] Congestion Risk Analysis Module: Receive and process the overall traffic flow feature dataset generated by the traffic flow feature acquisition module, perform discreteness analysis on it to obtain feature indicators, identify congestion risks when conditions are met, arrange the reference feature values ​​in the dataset in time period order, and calculate the difference between adjacent reference feature values ​​in turn. To highlight key changes, a weighting factor is determined for each difference value according to a preset weighting rule. For example, higher weights are assigned to differences occurring during peak traffic hours or differences that are inherently large. The weighting factor is then weighted and calculated with the corresponding difference values ​​to obtain a set of feature values ​​to be classified as feature indicators. The feature indicator is continuously monitored. When the overall volatility or rate of change of the feature indicator meets the preset congestion prediction triggering conditions, such as when the variance of the set of feature values ​​to be classified increases by more than a certain proportion in a short period of time, it indicates that the stability of traffic flow is declining and the risk of congestion is accumulating. At this time, the deep risk identification process will be initiated. In the deep risk identification process, clustering methods are used to process the set of feature values ​​to be classified, such as DBSCAN or K-Means algorithms. This is only a preferred implementation and does not constitute a limitation. Multiple discrete feature scores are output. By analyzing the time series of discrete feature scores, fluctuation and abrupt change points where the numerical changes exceed a preset abrupt change threshold are identified. These abrupt change points mark the critical area where the traffic state evolves from stable to congested. The areas on both sides of these fluctuation and abrupt change points are determined as traffic congestion concentration zones. Within the traffic congestion zone, identify the time points where the increase in discrete feature scores exceeds a preset score change threshold, and determine these key time points as risk points because they represent the moments when the congestion worsens most significantly.

[0029] Congestion prediction output module: The response congestion risk analysis module identifies risk points and generates final predictions and control suggestions. Upon receiving the traffic congestion concentration area and the risk points within it, the traffic congestion concentration area is directly used as the predicted highway congestion period, and the congestion area corresponding to the traffic bottleneck node with the strongest correlation to the risk point is used as the predicted congestion area. Then, the prediction results, congestion period and corresponding congestion area are output for reference by traffic management personnel or for use by downstream systems. After outputting the above prediction, an additional judgment is performed, which analyzes whether there are traffic flow changes during the predicted highway congestion period. For example, if the difference between the number of vehicles entering the area and the number of vehicles leaving the area exceeds the preset traffic flow change threshold, it indicates that the congestion is likely to dissipate on its own. If the conclusion is that there is no such time point, meaning that the congestion is expected to remain stable, the previously identified risk points and corresponding congestion areas will be further clearly output as areas and time periods to be regulated, providing a direct basis for implementing precise traffic interventions such as ramp control, variable speed limits, and information dissemination.

[0030] The aforementioned system can form a complete closed loop from data collection and risk analysis to predictive output. It can accurately identify potential traffic bottlenecks from massive traffic flow data and, through in-depth analysis of the dispersion of bottleneck characteristics, achieve accurate early prediction of highway congestion. This system can be used for intelligent traffic management and proactive congestion mitigation in complex road networks.

[0031] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand and implement the present invention. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for predicting highway traffic congestion, characterized in that, When the feature indicators obtained from the discreteness analysis of the overall traffic flow feature dataset of the target highway meet the preset congestion prediction trigger conditions, execute: Identify risk points within traffic congestion concentration zones extracted based on characteristic indicators; Based on the risk time points and traffic congestion concentration areas, the predicted highway congestion time periods and their corresponding congestion areas are output. The dispersion analysis of the overall traffic flow feature dataset of the target highway includes: arranging the reference feature values ​​in the overall traffic flow feature dataset in time period order; calculating the difference between adjacent reference feature values ​​in turn; determining a weight factor for each difference value according to a preset weighting rule; and weighting the weight factor and the difference value to obtain a set of feature values ​​to be classified as feature indicators. Identifying risk points within traffic congestion clusters extracted based on feature indicators includes: processing the set of feature values ​​to be classified using clustering methods to output multiple discrete feature scores; identifying fluctuation and abrupt change points in the discrete feature scores where the numerical changes exceed a preset abrupt change threshold, and defining the areas on both sides of the fluctuation and abrupt change points as traffic congestion clusters; and identifying, within the traffic congestion clusters, the time points where the increase in discrete feature scores exceeds a preset score change threshold, and defining these time points as risk points.

2. The method for predicting highway traffic congestion based on deep learning according to claim 1, characterized in that, The overall traffic flow feature dataset includes: Based on traffic flow characteristic information obtained from the target highway, multiple traffic bottleneck nodes are identified; The reference feature values ​​corresponding to all traffic bottleneck nodes are summarized to form an overall traffic flow feature dataset.

3. The method for predicting highway traffic congestion based on deep learning according to claim 2, characterized in that, Traffic flow characteristic information includes congested areas, and the methods for obtaining congested areas include: The route of the target highway is divided into grids, and the vehicle target information associated with each grid is collected; Based on preset congestion rules, determine whether the vehicle target information indicates that vehicles are arranged in a continuous congestion sequence; If the determination is yes, the corresponding grid is marked as congested, and one or more grids in the congested state are identified as congested areas.

4. The method for predicting highway traffic congestion based on deep learning according to claim 2, characterized in that, Based on traffic flow characteristic information obtained from the target highway, multiple traffic bottleneck nodes are identified, including: For grids within congested areas, extract their center points as nodes to be determined; Determine whether the node to be judged has a reverse blocking effect. If the node to be judged can stop vehicles in at least one direction and the duration of the stop exceeds a preset stop duration threshold, then it is determined to have a reverse blocking effect and is identified as a traffic bottleneck node.

5. The method for predicting highway traffic congestion based on deep learning according to claim 2, characterized in that, Determining the reference feature values ​​includes: For each traffic bottleneck node, obtain the area of ​​its associated congested area within multiple preset time periods; The areas within multiple preset time periods are sorted, and the maximum area value is recorded as the reference feature value of the traffic bottleneck node.

6. The method for predicting highway traffic congestion based on deep learning according to claim 1, characterized in that, After outputting the predicted highway congestion periods and their corresponding congestion areas, the method also includes: Determine whether there are any points in the predicted highway congestion period where the change in traffic flow exceeds a preset traffic flow change threshold. If no specific time point exists, the risk time point and the corresponding congested area will be output as the area to be regulated and the regulation period.

7. A highway traffic congestion prediction system based on deep learning, characterized in that, Includes the following modules: The traffic flow feature acquisition module is used to acquire traffic flow feature information of the target highway in order to construct an overall traffic flow feature dataset; The congestion risk analysis module is used to perform dispersion analysis on the overall traffic flow feature dataset to obtain feature indicators, and when the feature indicators meet the preset congestion prediction trigger conditions, it identifies the risk time points in the traffic congestion concentration area extracted based on the feature indicators.

8. A highway traffic congestion prediction system based on deep learning according to claim 7, characterized in that, Also includes: The congestion prediction output module is used to output the predicted highway congestion period and its corresponding congestion area in response to the congestion risk analysis module's identification of risk time points. The congestion prediction output module is also used to determine whether there are time points in the predicted highway congestion period where the traffic flow change value exceeds the preset traffic flow change threshold, and if there are no time points, output the risk time point and the corresponding congestion area as the area to be controlled and the control period.

9. A highway traffic congestion prediction system based on deep learning according to claim 7, characterized in that, Traffic flow characteristic information includes congested areas, and the methods for obtaining congested areas include: The route of the target highway is divided into grids, and the vehicle target information associated with each grid is collected; Based on preset congestion rules, determine whether the vehicle target information indicates that vehicles are arranged in a continuous congestion sequence; If the determination is yes, the corresponding grid is marked as congested, and one or more grids in the congested state are identified as congested areas.

10. A highway traffic congestion prediction system based on deep learning according to claim 7, characterized in that, Identification of risk timing includes: The clustering method is used to process the set of feature values ​​to be classified, which are used as feature indicators, and output multiple discrete feature scores. The fluctuation and mutation points in the discrete feature scores that exceed the preset mutation threshold are identified, and the areas on both sides of the fluctuation and mutation points are determined as traffic congestion concentration areas. In addition, within areas of concentrated traffic congestion, the time point at which the increase in discrete feature scores exceeds a preset score change threshold is identified, and this time point is determined as a risk point.

Images