Highway traffic short-time prediction method and system based on multi-source heterogeneous data

Abstract

This invention discloses a method and system for short-term prediction of highway traffic flow based on multi-source heterogeneous data, belonging to the field of traffic information and computer technology. It aims to solve the problem of traditional traffic flow prediction models failing due to data distribution offsets and causal gaps during emergencies. This invention identifies the congestion tail upstream of the accident location and dynamically delineates the influence radius to construct a dynamic subnet. It integrates GPS harmonic mean speed and speed variation coefficient to estimate road resistance, establishes a time-varying information penetration rate model to divide drivers into informed and uninformed groups, and weights their path selection probabilities to obtain a dynamic diversion rate. Then, it combines queuing dynamics simulation to update the queue length. When the queue tail blocks upstream nodes, a blockage is applied, and the road resistance is reset and iterated until convergence. The output is the predicted traffic flow to the accident point and the predicted queue length. This invention can predict peak traffic flow and queue spread in advance, assisting in traffic control and rescue dispatch.

Description

Technical Field

[0001] This invention relates to the fields of traffic information and computer technology. More specifically, this invention relates to a method and system for short-term prediction of highway traffic flow based on multi-source heterogeneous data. Background Technology

[0002] Highway traffic flow forecasting is a fundamental task in traffic management and control. Accurate short-term traffic flow forecasting plays an important role in congestion early warning, emergency response, and improving road network operation efficiency. With the widespread adoption of ETC gantry systems, vehicle-mounted GPS devices, and roadside sensing facilities, highways now have the capability to collect multi-source heterogeneous data, providing richer data support for traffic flow forecasting.

[0003] Existing traffic flow prediction methods can be mainly divided into the following categories: The first category is traditional time series analysis methods, such as the differential ensemble moving average autoregressive model and its variants. These extrapolate based on the statistical regularities of historical traffic flow data, offering high computational efficiency but struggling to capture the complex nonlinear characteristics of traffic flow. The second category is machine learning methods, such as support vector regression and random forests. These improve prediction accuracy to some extent by constructing a nonlinear mapping relationship between features and traffic flow. The third category is deep learning methods, including recurrent neural networks, long short-term memory networks, and the recently emerging architecture combining graph neural networks and Transformers. These methods can simultaneously model the spatial topological dependence and temporal correlation of road networks, representing the current state of deep learning. Mainstream technologies include, for example, Chinese patent application CN121564994A, which describes an ETC traffic flow prediction method based on micro-trajectory reconstruction and time-frequency graph fusion. This method constructs a vehicle micro-trajectory tensor, extracts spatiotemporal features using graph attention mechanisms and Transformer encoders, and introduces frequency domain filtering and physical correction layers to enhance the robustness of prediction. Another example is Chinese patent application CN122157503A, which describes a high-speed vehicle speed calculation and trajectory prediction method based on multi-source data. This method integrates real-time ETC data, historical traffic data, and road network basic data, and predicts vehicle trajectories and exits through driving behavior normalization, spatiotemporal feature correction, and an LSTM-Attention model.

[0004] Some studies focus on traffic control in specific scenarios. For example, Chinese patent application CN115331443A, which describes a method and system for controlling traffic flow in the longitudinal section of a highway entrance ramp, formulates a graded vehicle passage guidance strategy by comparing the traffic flow at the ramp entrance with that at a certain distance ahead. This belongs to the field of ramp control.

[0005] Existing methods are all based on the fundamental assumption that historical patterns can be extrapolated, meaning that future traffic conditions are similar to those of the same historical period. This assumption holds true under normal traffic conditions, but it faces fundamental challenges in sudden traffic accident scenarios. Specifically, these challenges manifest in the following ways: accidents disrupt the assumption of a stable traffic flow data distribution. Sudden accidents can cause lane closures or a sharp drop in capacity, leading to rapid queues of upstream vehicles and a precipitous drop or drastic fluctuation in traffic flow within a short period. This results in a significant shift in data distribution. Traditional statistical models and most machine learning models assume that the training and test sets follow the same distribution; when this distribution shift occurs, their predictions become severely distorted, making it difficult to capture sudden inflection points in traffic flow. Purely data-driven models lack explicit modeling of the causal relationships of accidents, while deep learning methods, through large-scale... While models learn traffic flow patterns from historical samples, their essence remains statistical correlation rather than causal inference. When a model observes abnormal traffic flow changes, it can only treat them as numerical fluctuations and cannot determine whether the change is caused by a redistribution of traffic demand or queue overflow due to an accident ahead. This results in the model needing historical data from multiple time steps to adapt to the new situation after an accident, leading to significant lag in prediction results and failing to meet the real-time requirements of emergency response. Severe traffic accidents are low-probability events in historical data, with extremely limited sample sizes. The performance of deep learning models is highly dependent on a large number of representative training samples. For such extreme scenarios with long-tailed distributions, models struggle to effectively learn their evolutionary patterns and tend to ignore them as noise or over-smooth them, thus underestimating the amplitude and speed of traffic flow fluctuations.

[0006] Existing methods also have shortcomings in road resistance estimation and route selection behavior characterization. Road resistance functions based on cross-section detectors usually use fixed speed-flow curves, which are difficult to respond in real time to the increased speed dispersion and nonlinear decrease in traffic efficiency caused by accidents. In terms of diversion behavior prediction, most models assume that drivers have complete information and make rational decisions based on the shortest travel time, ignoring the process of accident information gradually spreading among drivers and drivers' psychological resistance to unfamiliar detour routes, resulting in biased diversion rate estimation, which in turn affects the accurate estimation of the flow rate arriving at the accident point.

[0007] Therefore, how to achieve short-term prediction of highway traffic flow in sudden accident scenarios without relying on a large number of historical accident samples, by introducing accident causal logic, dynamic road obstruction perception and information penetration process, is a technical problem that urgently needs to be solved. Summary of the Invention

[0008] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0009] To achieve these objectives and other advantages according to the present invention, a method for short-term prediction of highway traffic flow based on multi-source heterogeneous data is provided, comprising:

[0010] S1. Obtain the accident location and road network topology, traverse upstream from the accident location, identify the congestion tail location based on the real-time average speed of the road segment, determine the dynamic influence radius, take the toll station or toll gantry within the radius upstream of the congestion tail location as the boundary node set, extract the connection path from the boundary node set to the accident location, remove the far-end nodes whose free-flow travel time exceeds the limit, and form a dynamic subnet.

[0011] S2. Obtain the GPS trajectory of vehicles within the dynamic subnet, and calculate the harmonic mean speed and speed variation coefficient of each road segment through map matching and speed smoothing filtering. Combine real-time observation and historical benchmark data to estimate the travel time of each road segment, which is used as the road resistance of the dynamic subnet.

[0012] S3. Based on the dynamic subnet and road resistance, candidate paths from upstream nodes to the accident point are generated. The generalized impedance of each candidate path is calculated, and a time-varying information penetration rate model is established. Drivers are divided into an unaware group and an unaware group. The selection results of the two groups are weighted according to the information penetration rate to obtain the dynamic diversion rate of each upstream node. The generalized impedance includes travel time, unfamiliarity penalty and accident risk perception. The information penetration rate increases from the time of the accident. The unaware group selects routes according to historical habits, while the unaware group selects routes discretely according to the generalized impedance.

[0013] S4. Calculate the effective demand based on the original demand of the upstream node and the dynamic diversion rate. Perform queuing dynamics simulation based on the remaining capacity of the accident point and update the queue length. If the tail of the queue blocks the upstream flow, cut off the corresponding demand and reset the relevant road blockage. Return to S1~S3 to iterate until convergence, and output the predicted flow rate and queue length of the accident point.

[0014] Preferably, the method for identifying the congestion tail position in step S1 is as follows: traverse each road segment from the accident location upstream, and determine the end position of the continuous road segment sequence where the real-time average speed of the road segment is lower than the preset speed threshold as the congestion tail position; if there is no road segment with a speed lower than the preset speed threshold at present, then the accident location is used as the congestion tail position.

[0015] Preferably, the dynamic influence radius in step S1 is determined according to the following formula:

[0016]

[0017] Among them, D max L represents the dynamic influence radius. queue Indicates the current queue length, v f T represents the free-flow velocity. lookahead This indicates the preset prediction horizon.

[0018] Preferably, the harmonic average velocity in step S2 is calculated using the following formula:

[0019]

[0020] in, N represents the harmonic average speed, and N represents the road segment e. ij The number of valid GPS samples retained within the current time window after map matching and velocity smoothing filtering, v k This represents the instantaneous speed of the k-th sample vehicle;

[0021] The velocity variation coefficient is calculated using the following formula:

[0022]

[0023] Among them, CV ij σ represents the velocity variation coefficient. v,ij This represents the standard deviation of the vehicle speeds in the valid GPS sample.

[0024] For road segments where the number of effective GPS samples is lower than the preset minimum sample threshold, spatiotemporal kriging interpolation or the harmonic mean velocity of the corresponding road segment is estimated based on the upstream road segment velocity wave propagation model.

[0025] Preferably, the estimated travel time for each road segment in step S2 is calculated using the following formula:

[0026]

[0027] Among them, t ij (t) represents the travel time, L ij Indicates road segment e ij Length, CV represents the harmonic average velocity at the current moment. ij (t) represents the velocity variation coefficient at the current moment, CV ref Q represents the baseline coefficient of variation under free-flow conditions. hist Q represents the historical traffic volume during the same period. obs Q represents the real-time observed flow rate. cap ω1 represents the traffic capacity of the road segment, ω2 represents the first weighting coefficient, and ω2 represents the second weighting coefficient.

[0028] Preferably, in step S3, the candidate paths are generated by performing the K-shortest path algorithm on each upstream node to the accident location, and each candidate path is marked as either passing through the accident location or being a detour path.

[0029] Preferably, the time-varying information penetration rate model in step S3 is as follows:

[0030]

[0031] in, This represents the information penetration rate at time t indicating that the driver was aware of the accident, where t represents the current time. acc Let λ represent the time of the accident, λ represent the information diffusion rate constant, and β represent the information diffusion rate constant. VMS β represents the coverage coefficient of the variable message sign. Nav This indicates the coverage coefficient of the navigation app.

[0032] Preferably, in step S3, the probability that the uninformed group chooses path k based on historical habits is... It is a preset fixed probability distribution;

[0033] The specific method by which the informed group makes discrete choices based on generalized impedance is as follows: using a Logit model, the probability of the informed group choosing path k is... Calculate using the following formula:

[0034]

[0035] Among them, C i,k (t) represents the generalized impedance of the k-th candidate path from upstream node i to the accident location at time t, θ represents the preset sensitivity coefficient, and K represents the total number of candidate paths;

[0036] The mixed weighting method is as follows:

[0037] The final selection probability of candidate path k Calculate using the following formula:

[0038]

[0039] The final selection probabilities of each detour path are summed to obtain the dynamic diversion rate η of upstream node i at time t. i (t):

[0040]

[0041] Among them, R bypass This represents the set of detour paths among the candidate paths.

[0042] Preferably, the specific method for queuing dynamics simulation in step S4 is as follows:

[0043] Calculate the net inflow rate Q net (t):

[0044]

[0045] in, Indicates that upstream node i has a time delay τ i The last effective demand to arrive, C inc (t) represents the remaining traffic capacity at the accident site at time t;

[0046] Update queue length L queue (t+Δt):

[0047]

[0048] Where Δt represents the time step, l veh This represents the average distance between the front ends of the vehicles.

[0049] Determine if the updated queue tail blocks the upstream node v i The specific method is as follows:

[0050] Calculate upstream node v i Position from the end of the congestion P tail Topological distance Dist(v) between i P tail If the following formula is satisfied:

[0051]

[0052] Then determine the upstream node v i When blocked, the effective demand is truncated and modified, and the modified actual demand... for:

[0053]

[0054] Where, γ block The blocking coefficient has a value range of [0, 1], and is 0 when complete blocking is achieved.

[0055] This invention provides a system for the aforementioned method of short-time traffic flow prediction for highways based on multi-source heterogeneous data, comprising:

[0056] The data acquisition module is used to acquire accident locations and road network topology, as well as vehicle GPS trajectory data, historical benchmark data, variable message signs and information release status of navigation apps;

[0057] The dynamic subnet construction module is used to traverse upstream from the accident location, identify the congestion tail position based on the real-time average speed of the road segment, determine the dynamic influence radius, take the toll station or toll gantry within the radius upstream of the congestion tail position as the boundary node set, extract the connecting path from the boundary node set to the accident location, and remove the far-end nodes whose free-flow travel time exceeds the limit to form a dynamic subnet.

[0058] The road resistance calculation module is used to obtain the GPS trajectory of vehicles in the dynamic subnet, and calculate the harmonic mean speed and speed variation coefficient of each road segment through map matching and speed smoothing filtering. It integrates real-time observation and historical benchmark data to estimate the travel time of each road segment as the road resistance of the dynamic subnet.

[0059] The route selection module generates candidate routes from upstream nodes to the accident point based on dynamic subnets and road resistance, calculates the generalized impedance of each candidate route, establishes a time-varying information penetration rate model, divides drivers into an unaware group and an informed group, and weights the selection results of the two groups according to the information penetration rate to obtain the dynamic diversion rate of each upstream node. The generalized impedance includes travel time, unfamiliarity penalty and accident risk perception. The information penetration rate increases from the time of the accident. The unaware group selects routes according to historical habits, while the informed group selects routes discretely according to the generalized impedance.

[0060] The traffic prediction and feedback module is used to calculate the effective demand based on the original demand of the upstream nodes and the dynamic diversion rate, perform queuing dynamics simulation based on the remaining capacity of the accident point and update the queue length. If the tail of the queue blocks the upstream flow, the corresponding demand is cut off and the relevant road resistance is reset, triggering the dynamic subnet construction module, road resistance calculation module and path selection module to iterate in sequence until convergence, and outputting the traffic prediction result and queue length prediction result of the accident point.

[0061] The present invention has at least the following beneficial effects:

[0062] First, this invention identifies the position of the tail end of the queue in real time and determines the dynamic influence radius by combining the current queue length with the preset prediction horizon. This allows the delineation of the accident influence range to be automatically adjusted according to the congestion status. This avoids the problem of missing remote detour paths due to an excessively small fixed range, and also avoids the problem of introducing irrelevant nodes and causing computational redundancy due to an excessively large range. At the same time, it eliminates remote nodes with excessively long free-flow travel times to further reduce the subnet size, effectively controlling the consumption of computing resources while ensuring prediction accuracy.

[0063] Secondly, this invention uses harmonic mean speed to characterize road segment traffic efficiency, which can reasonably highlight the impact of low-speed congestion on travel time and effectively suppress the interference of instantaneous speed fluctuations on mean estimation. It combines speed variation coefficient to quantify traffic flow dispersion and integrates historical benchmark data to correct real-time observations, so that road resistance remains stable and reliable even when GPS samples are sparse or traffic flow is close to saturation, providing continuous and reliable input for subsequent route selection.

[0064] Third, this invention uses a time-varying information penetration rate model to depict the gradual process of accident information from occurrence to widespread dissemination. Drivers are divided into an unaware group and an informed group. The unaware group makes decisions based on historical habits, while the informed group makes logit discrete choices based on generalized impedance including travel time, unfamiliarity penalty, and accident risk perception. The dynamic diversion rate is obtained through mixed weighting, which makes the prediction of detour behavior closer to the influence of information lag and psychological impedance in actual traffic, and avoids the diversion rate deviation caused by the assumption of perfect rationality.

[0065] Fourth, this invention couples queuing dynamics simulation with diversion results. By determining whether the tail of the queue blocks the upstream node, a blocking coefficient is applied to the effective demand to cut off the flow rate. The path resistance between the cut-off node and the accident point is reset to the maximum value, triggering an iterative process of dynamic subnet re-delineation, path resistance re-estimation, and diversion rate recalculation. This allows the queuing overflow effect to be passed on in a chain in the system, and the final output arrival flow and queue length prediction results remain self-consistent throughout the entire process.

[0066] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0067] Figure 1 A flowchart of a method for short-time prediction of highway traffic flow based on multi-source heterogeneous data, one of the technical solutions of the present invention;

[0068] Figure 2 This is a flowchart illustrating the construction process of a dynamically reachable network, one of the technical solutions of the present invention.

[0069] Figure 3 This is a flowchart illustrating the robust road resistance calculation process of one of the technical solutions of the present invention;

[0070] Figure 4 This is a flowchart illustrating the path selection process for one of the technical solutions of the present invention.

[0071] Figure 5 This is a flowchart illustrating the traffic prediction and feedback process of one of the technical solutions of the present invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to examples, so that those skilled in the art can implement it based on the description.

[0073] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0074] This invention provides a method for short-term prediction of highway traffic flow based on multi-source heterogeneous data, comprising:

[0075] S1. Obtain the accident location and road network topology, traverse upstream from the accident location, identify the congestion tail location based on the real-time average speed of the road segment, determine the dynamic influence radius, take the toll station or toll gantry within the radius upstream of the congestion tail location as the boundary node set, extract the connection path from the boundary node set to the accident location, remove the far-end nodes whose free-flow travel time exceeds the limit, and form a dynamic subnet.

[0076] S2. Obtain the GPS trajectory of vehicles within the dynamic subnet, and calculate the harmonic mean speed and speed variation coefficient of each road segment through map matching and speed smoothing filtering. Combine real-time observation and historical benchmark data to estimate the travel time of each road segment, which is used as the road resistance of the dynamic subnet.

[0077] S3. Based on the dynamic subnet and road resistance, candidate paths from upstream nodes to the accident point are generated. The generalized impedance of each candidate path is calculated, and a time-varying information penetration rate model is established. Drivers are divided into an unaware group and an unaware group. The selection results of the two groups are weighted according to the information penetration rate to obtain the dynamic diversion rate of each upstream node. The generalized impedance includes travel time, unfamiliarity penalty and accident risk perception. The information penetration rate increases from the time of the accident. The unaware group selects routes according to historical habits, while the unaware group selects routes discretely according to the generalized impedance.

[0078] S4. Calculate the effective demand based on the original demand of the upstream node and the dynamic diversion rate. Perform queuing dynamics simulation based on the remaining capacity of the accident point and update the queue length. If the tail of the queue blocks the upstream flow, cut off the corresponding demand and reset the relevant road blockage. Return to S1~S3 to iterate until convergence, and output the predicted flow rate and queue length of the accident point.

[0079] In the above technical solution, step S1, through real-time congestion tail identification and dynamic impact radius calculation, can adaptively delineate the accident impact subnet, avoiding redundancy caused by full road network calculation and preventing the omission of long-distance detour routes, thus balancing prediction accuracy and computational efficiency; step S2 adopts a road resistance estimation method combining harmonic mean speed and speed variation coefficient, effectively suppressing noise interference in GPS data, and by introducing historical benchmark correction, ensuring the impedance reliability of sparse road segments, providing robust input for route selection; step S3 distinguishes between informed and uninformed drivers through a time-varying information penetration rate model. Furthermore, psychological resistance such as unfamiliarity penalty and accident risk perception is introduced to make the calculation of diversion rate more consistent with the phenomenon of real decision lag and information asymmetry, thereby improving the rationality of detour behavior prediction. Step S4 couples queue dynamics with diversion results, and triggers iterative feedback through queue tail blockage judgment and road resistance reset, so that the dynamic subnet and diversion rate are automatically updated as congestion spreads, forming a closed-loop correction mechanism. The final output arrival flow and queue length prediction have higher timeliness and accuracy. Each step is interconnected and progressively advanced, and together they achieve a stable and interpretable solution for short-term prediction of highway flow under accident scenarios.

[0080] In another technical solution, the method for identifying the congestion tail position in step S1 is as follows: traverse each road segment from the accident location upstream, and determine the end position of the continuous road segment sequence where the real-time average speed of the road segment is lower than the preset speed threshold as the congestion tail position; if there is no road segment with a speed lower than the preset speed threshold, then the accident location is used as the congestion tail position.

[0081] In the above technical solution, by traversing each road segment upstream from the accident location and judging continuous low-speed road segments based on speed thresholds, the actual boundary of the congestion wave propagating upstream can be accurately captured, avoiding misjudgments caused by instantaneous speed fluctuations in individual road segments. At the same time, when no obvious congestion has formed, the accident location is directly used as the tail, ensuring that the algorithm can still start the subsequent process normally in non-congested scenarios, thus balancing the accuracy of congestion tail identification with the robustness of the method.

[0082] In another technical solution, the dynamic influence radius in step S1 is determined by the following formula:

[0083]

[0084] Among them, D max L represents the dynamic influence radius. queue Indicates the current queue length, v f T represents the free-flow velocity. lookahead This indicates the preset prediction horizon.

[0085] In the above technical solution, the determination of the dynamic influence radius takes into account both the current queue length and the distance that the free flow can reach within the predicted horizon. It reflects both the physical occupancy caused by the congestion and the range that the congestion may spread upstream in the future. This allows the boundary of the dynamic subnet to be adaptively adjusted with the duration of the accident and the evolution of the congestion, avoiding the omission of far-end diversion paths due to the range being too small or the calculation redundancy caused by the range being too large.

[0086] In another technical solution, the harmonic average velocity in step S2 is calculated according to the following formula:

[0087]

[0088] in, N represents the harmonic average speed, and N represents the road segment e. ij The number of valid GPS samples retained within the current time window after map matching and velocity smoothing filtering, v k This represents the instantaneous speed of the k-th sample vehicle;

[0089] The velocity variation coefficient is calculated using the following formula:

[0090]

[0091] Among them, CV ij σ represents the velocity variation coefficient. v,ij This represents the standard deviation of the vehicle speeds in the valid GPS sample.

[0092] For road segments where the number of effective GPS samples is lower than the preset minimum sample threshold, spatiotemporal kriging interpolation or the harmonic mean velocity of the corresponding road segment is estimated based on the upstream road segment velocity wave propagation model.

[0093] In the above technical solutions, the harmonic mean speed gives higher weight to low-speed vehicles, which can more realistically reflect the decline in road segment traffic efficiency under congestion. The speed variation coefficient characterizes the dispersion of traffic flow speed, providing a quantitative indicator for identifying unstable flow states. For road segments with sparse GPS samples, estimation is performed through spatiotemporal Kriging interpolation or speed wave propagation models, ensuring the integrity and continuity of road resistance data in the road network space, so that subsequent route selection is based on reliable data.

[0094] In another technical solution, the estimated travel time for each road segment in step S2 is calculated using the following formula:

[0095]

[0096] Among them, t ij (t) represents the travel time, L ij Indicates road segment e ij Length, CV represents the harmonic average velocity at the current moment. ij (t) represents the velocity variation coefficient at the current moment, CV ref Q represents the baseline coefficient of variation under free-flow conditions. hist Q represents the historical traffic volume during the same period. obs Q represents the real-time observed flow rate. cap ω1 represents the traffic capacity of the road segment, ω2 represents the first weighting coefficient, and ω2 represents the second weighting coefficient.

[0097] In the above technical solution, the travel time estimation formula introduces the ratio of the speed variation coefficient to the baseline variation coefficient, taking into account the additional time loss caused by speed fluctuations. At the same time, it uses the difference between historical traffic flow and real-time traffic flow to correct road resistance, which can capture hidden delays when congestion has not yet fully formed but traffic flow is close to saturation. This makes the road resistance function more sensitive to different traffic conditions and improves the accuracy of impedance estimation.

[0098] In another technical solution, in step S3, the candidate paths are generated by performing the K-shortest path algorithm on each upstream node to the accident location, and each candidate path is marked as either passing through the accident location or being a detour path.

[0099] In the above technical solution, multiple candidate paths are generated by the K-shortest path algorithm and clearly marked whether they pass through the accident location and whether they are detour paths. This not only ensures the diversity of the driver's path selection space, but also provides clear path attribute labels for distinguishing between straight and detour behaviors in the subsequent diversion rate calculation, which facilitates the quantification of the detour ratio and enhances the interpretability of diversion behavior.

[0100] In another technical solution, the time-varying information penetration rate model in step S3 is as follows:

[0101]

[0102] in, This represents the information penetration rate at time t indicating that the driver was aware of the accident, where t represents the current time. acc Let λ represent the time of the accident, λ represent the information diffusion rate constant, and β represent the information diffusion rate constant. VMS β represents the coverage coefficient of the variable message sign. Nav This indicates the coverage coefficient of the navigation app.

[0103] In the above technical solution, the time-varying information penetration rate model uses an exponential growth form to describe the process of accident information gradually spreading from the moment of occurrence. Combined with the coverage coefficients of variable message signs and navigation apps, it can reflect the differentiated impact of different information dissemination methods on the proportion of drivers who are aware of the information. This makes the division between aware and unaware drivers in the diversion rate calculation closer to the actual information dissemination pattern and avoids the bias of simply assuming that all or all drivers are completely unaware.

[0104] In another technical solution, in step S3, the probability of the uninformed group choosing path k based on historical habits is... It is a preset fixed probability distribution;

[0105] The specific method by which the informed group makes discrete choices based on generalized impedance is as follows: using a Logit model, the probability of the informed group choosing path k is... Calculate using the following formula:

[0106]

[0107] Among them, C i,k (t) represents the generalized impedance of the k-th candidate path from upstream node i to the accident location at time t, θ represents the preset sensitivity coefficient, and K represents the total number of candidate paths;

[0108] The mixed weighting method is as follows:

[0109] The final selection probability of candidate path k Calculate using the following formula:

[0110]

[0111] The final selection probabilities of each detour path are summed to obtain the dynamic diversion rate η of upstream node i at time t. i (t):

[0112]

[0113] Among them, R bypass This represents the set of detour paths among the candidate paths.

[0114] In the above technical solution, the uninformed group chooses a route based on fixed historical habits, reflecting the inertial behavior pattern under information lag; the informed group makes discrete choices based on generalized impedance using the Logit model, reflecting rational decision-making guided by real-time information. By mixing and weighting the selection probabilities of the two groups through information penetration rate, it can continuously characterize the transition stage from the initial occurrence of the accident to the widespread dissemination of information. Finally, the probability of detour routes is accumulated to obtain the dynamic diversion rate, so that the diversion prediction changes smoothly with the information diffusion time, which is more in line with the evolution characteristics of actual driving behavior.

[0115] In another technical solution, the specific method for queuing dynamics simulation in step S4 is as follows:

[0116] Calculate the net inflow rate Q net (t):

[0117]

[0118] in, Indicates that upstream node i has a time delay τ i The last effective demand to arrive, C inc (t) represents the remaining traffic capacity at the accident site at time t;

[0119] Update queue length L queue (t+Δt):

[0120]

[0121] Where Δt represents the time step, l veh This represents the average distance between the front ends of the vehicles.

[0122] Determine if the updated queue tail blocks the upstream node v i The specific method is as follows:

[0123] Calculate upstream node v i Position from the end of the congestion P tail Topological distance Dist(v) between i P tail If the following formula is satisfied:

[0124]

[0125] Then it is determined that the upstream node vi is blocked, its effective demand is truncated and corrected, and the corrected actual demand is... for:

[0126]

[0127] Where, γ block The blocking coefficient has a value range of [0, 1], and is 0 when complete blocking is achieved.

[0128] In the above technical solution, the queuing dynamics simulation updates the queue length through the net inflow rate and uses topological distance to determine whether the tail of the queue is blocking the upstream node. Once blocked, a blocking coefficient is applied to the effective demand of the node to cut off the blockage, directly quantifying the inhibitory effect of physical queuing on upstream flow. At the same time, the road resistance between the blocked node and the accident point is reset to the maximum value and triggers the iterative feedback of subnet redrawing and road resistance recalculation, so that the queuing overflow effect can be transmitted in a chain in the dynamic subnet and diversion rate, ensuring that the prediction results remain self-consistent and convergent during the congestion spread process.

[0129] This invention provides a system for short-time traffic flow prediction of highways based on multi-source heterogeneous data, comprising:

[0130] The data acquisition module is used to acquire accident locations and road network topology, as well as vehicle GPS trajectory data, historical benchmark data, variable message signs and information release status of navigation apps;

[0131] The dynamic subnet construction module is used to traverse upstream from the accident location as the downstream reference, identify the congestion tail position based on the real-time average speed of the road segment, determine the dynamic influence radius based on the current queue length and the preset prediction horizon, take all toll stations or toll gantries within the dynamic influence radius extended upstream from the congestion tail position as the boundary node set, extract all topological connectivity paths from the boundary node set to the accident location, and remove far-end nodes whose free-flow travel time exceeds the preset time limit to form a dynamic subnet.

[0132] The road resistance calculation module is used to obtain the GPS trajectory of vehicles in the dynamic subnet, and calculate the harmonic mean speed and speed variation coefficient of each road segment through map matching and speed smoothing filtering. It also estimates the travel time of each road segment by integrating real-time observation data and historical benchmark data, which serves as the road resistance of the dynamic subnet.

[0133] The route selection module is used to generate candidate routes from each upstream node to the accident point based on the dynamic subnet and road resistance, calculate the generalized impedance of each candidate route, establish a time-varying information penetration rate model, divide drivers into an unaware group and an informed group, and weight the selection results of the unaware group and the informed group according to the information penetration rate to obtain the dynamic diversion rate of each upstream node. The generalized impedance includes travel time, unfamiliarity penalty and accident risk perception. The information penetration rate increases from the time of the accident. The unaware group selects routes based on historical habits, and the informed group makes discrete selections based on the generalized impedance.

[0134] The traffic prediction and feedback module is used to obtain the original demand of each upstream node, calculate the effective demand by combining the dynamic diversion rate, perform queuing dynamics simulation based on the remaining capacity of the accident point and the current queue length, update the queue length, and if the tail of the queue blocks the upstream node from merging, the effective demand of the corresponding upstream node is truncated by a blocking coefficient, and the road resistance of each segment between the truncated node and the accident point is reset to the maximum value. This triggers the dynamic subnet construction module, road resistance calculation module and path selection module to re-define the dynamic subnet, estimate the road resistance and calculate the diversion rate in sequence, iterate until convergence, and output the traffic prediction result and queue length prediction result of the accident point.

[0135] In the above technical solution, the system is implemented in a modular way through the organic composition of a data acquisition module, a dynamic subnet construction module, a road resistance calculation module, a path selection module, and a traffic prediction and feedback module. A data flow closed loop is formed between the modules. When queuing is blocked, the traffic prediction and feedback module actively triggers the upstream module to recalculate iteratively, ensuring the system's adaptive adjustment capability in dynamic traffic scenarios, which facilitates engineering deployment and real-time operation.

[0136] Example 1

[0137] like Figures 1 to 5 As shown, a short-term traffic flow prediction method for highways based on multi-source heterogeneous data includes the following four modules:

[0138] The first step is to construct a dynamic subnet to determine which road sections are affected by the accident. The affected area automatically expands as congestion spreads, ensuring that long-distance detours are not missed and that computing power is not wasted on calculating irrelevant areas.

[0139] The second step is robust road resistance calculation, which calculates how long it will take to travel this road, eliminates data noise interference, discards instantaneous acceleration, and calculates the harmonic mean speed and speed variation coefficient of the road segment.

[0140] The third step is to select routes based on information perception, predict how many vehicles will detour, simulate the lag in real human decision-making, simulate information penetration, and obtain the diversion rate by mixing probabilities.

[0141] The fourth step is to predict and feedback the supply and demand coupling of traffic flow, correct the actual number of vehicles that can reach the destination, handle physical queuing and blockage, feed back the new traffic flow status after the blockage to the first and second steps, recalculate the road resistance of the affected road sections, and repeat the above process until the system state converges.

[0142] Module 1: Dynamic Reachability Network Construction. This module dynamically defines the impact range of an accident based on the real-time location of the tail of congestion.

[0143] Input: Accident location P inc Real-time average speed of road segments Free flow velocity v f ;

[0144] step:

[0145] Congestion tail recognition, from P inc Traverse upstream along the road segments, identify consecutive sequences of road segments where speeds decrease significantly, and define the congestion tail position P. tail To meet The farthest upstream point, such as v threshold =40km / h, if there is no traffic congestion, then P tail =P inc ; where v threshold This indicates the preset speed threshold and the real-time average speed of the road segment. and They represent the same physical quantity.

[0146] Dynamic influence radius calculation, calculating the potential maximum influence distance D max :

[0147]

[0148] Among them, L queue For the current estimated queue length, T lookahead To predict the horizon, for example, set to 45 minutes, the dynamic boundary node set V boundary : All distances P tail Upstream D max scope

[0149] Toll booths or toll gantries within the area;

[0150] Subnet construction, extraction of V boundary To P inc All topological connection paths, excluding those whose travel time exceeds T even in a completely free-flow state. max limit The farthest node, such as T max limit Set to 90 minutes, relax the restrictions to prevent missed judgments; among them, T max limit This indicates the preset maximum travel time threshold.

[0151] Output:

[0152] Dynamic subnets Its range automatically expands as congestion spreads;

[0153] Module 2, Robust Road Resistance Calculation: This module utilizes real-time GPS trajectory data, considering speed and acceleration characteristics, to calculate the real-time impedance (travel time) of each road segment in the reachable network. It focuses on solving the GPS acceleration noise problem and constructs a road resistance function that is insensitive to outliers.

[0154] Input: Dynamic subnet Cleaned GPS trajectory data and historical baseline data from the same period;

[0155] step:

[0156] Data cleaning and smoothing: High-precision map matching is performed using a hidden Markov model; Kalman filtering is applied to smooth the trajectory and speed of a single vehicle to remove abrupt noise; and the speed variation coefficient is calculated.

[0157] Road segment feature aggregation, for road segment e ij Calculate within the time window Δt:

[0158] Harmonized average speed better reflects the impact of low-speed vehicles;

[0159] Speed ​​stability index, also known as stability index; where represents free flow, and represents congestion or frequent starts and stops;

[0160] The path resistance function is integrated using a weighted fusion strategy, combining real-time observations with historical benchmarks.

[0161]

[0162] Where ω1 represents the first weighting coefficient and ω2 represents the second weighting coefficient, which are dynamically adjusted with the sample size. When the sample size is small, ω2 increases, depending on historical data; when the sample size is large, ω1 increases, depending on real-time data. CV ref The first term is the baseline coefficient of variation under free-flow conditions, and the second term is used to capture the implicit delays caused by excessive flow but incomplete blockage.

[0163] Sparse data interpolation, if road segment e ij Sample size The estimation is performed using spatiotemporal kriging interpolation or a propagation model based on the velocity wave of the upstream road segment. ;

[0164] Output: Robust segment travel time t ij (t);

[0165] Module 3, Path Selection Model under Information Perception Delay: This module simulates the driver's path selection behavior based on real-time road obstruction, calculates the path travel time to the accident point and the vehicle diversion rate (detour ratio), and introduces information penetration rate and psychological resistance.

[0166] Input: Subnet with road resistance, Variable Message Sign (VMS) / Navigation guidance information release status I(t), 0 indicates unreleased, 1 indicates released, route type attribute, direct highway and detour via national highway;

[0167] step:

[0168] K-shortest path generation: For each OD pair, generate K candidate paths, and mark whether each path passes through an accident point. (or a detour route) );

[0169] Generalized impedance calculation considers not only time, but also psychological impedance and information factors:

[0170]

[0171] Among them, I unfam The Risk penalty is higher when taking a non-highway detour due to path unfamiliarity. i,k For path risk perception, if an accident is known, the path risk value passing through the accident point is extremely high, where α T The impedance weighting factor α represents the travel time. fam α represents the weighting coefficient of the unfamiliarity penalty. risk The weighting coefficient represents the perceived risk of an accident; where T i,k (t) represents the travel time from upstream node i to the accident location via the k-th candidate path, i.e., the travel time t of each segment contained in that path. ij The sum of (t).

[0172] Time-varying information penetration rate model, defining time. The probability P that the driver is aware of the accident info (t):

[0173]

[0174] Among them, t acc Let λ be the time of the accident, λ be the information diffusion rate constant, and β be the time of the accident. VMS， β Nav These are the coverage coefficients for variable message signs and navigation apps, respectively, with values ​​ranging from 0 to 1;

[0175] Mixed choice probability, drivers are divided into informed and uninformed groups, with the uninformed group accounting for 1-P. infoBased solely on historical habits or free-flow time selection, there is a tendency to follow the original path; the informed group (percentage P) info Based on real-time generalized impedance C i,k (t) The probability of choosing path k after performing Logit selection:

[0176]

[0177]

[0178] Flow split rate calculation: ;

[0179] Output: Dynamic splitting rate η considering information lag i (t);

[0180] Module 4, Supply and Demand Coupling and Queue Feedback-based Traffic Forecasting: This module combines traffic demand and diversion results to predict traffic flow to the accident site in the future, solves the queue overflow problem, and achieves self-correction of traffic forecasting.

[0181] Input: Upstream node requirements Diversion rate η i (t), Remaining traffic capacity at the accident site C inc (t), Current queue length L queue (t);

[0182] step:

[0183] Effective demand calculation:

[0184]

[0185] Queuing dynamics simulation simplifies bottleneck model:

[0186] Calculate the net inflow rate:

[0187] Update queue length:

[0188] Where Δt is the time step, l veh This represents the average distance between the front ends of the vehicles.

[0189] An overflow feedback mechanism is used to determine whether the tail of the queue is blocking the upstream node v. i The inflow, if If so, the valid demand for that node is truncated. The corrected demand is as follows:

[0190]

[0191] Where, γ block This is the blocking coefficient, which ranges from [0, 1], and is 0 when completely blocking.

[0192] If a truncation occurs, it is necessary to return to Module 1 to recalculate the road resistance of the affected road segment. At this time, the road resistance approaches infinity, and Module 3 is rerun to force diversion or prevent entry until the system state converges or the maximum number of iterations is reached. The maximum number of iterations is usually 2-3 times.

[0193] Final traffic forecast:

[0194]

[0195] Actual arrival traffic is limited by traffic capacity and the ability to dissipate queues;

[0196] in, Figure 5 In this embodiment, "calculating effective traffic demand" refers to "effective demand calculation," and "calculating net increment" refers to "calculating net inflow rate." Figure 5 The meaning of "system convergence? flow change < threshold" is as follows: the system is considered to have converged when the change in the predicted flow from the accident point obtained in two adjacent iterations is less than the preset threshold ε.

[0197] Output:

[0198] The corrected flow rate curve from the accident point, dynamic queue length prediction, and actual cutoff status of each upstream node.

[0199] Table 1 shows the parameter calibration guide;

[0200] Table 1 Parameter Calibration Guide

[0201]

[0202] Among them, α in Table 1 fam This represents the weighting coefficient for the unfamiliarity penalty.

[0203] This embodiment, through dynamic subnet construction, robust road resistance calculation, path selection under information perception, and closed-loop iteration of supply-demand coupled queuing feedback, can output the predicted peak traffic flow and queue length at the accident point 15-60 minutes in advance. The prediction error rate is significantly reduced compared to traditional methods. It can quantitatively evaluate the impact of different information dissemination strategies, such as the timing of variable message sign releases and the impact of navigation APP penetration on diversion rates, providing data support for traffic management departments to formulate optimal control plans. Based on the dynamic prediction results of queue spread, it can guide rescue forces to be deployed to key blockage points in advance, effectively preventing secondary accidents. Each step has clear physical meaning and behavioral logic support. Unlike data-driven black-box models, its prediction results have good interpretability, making them easy for decision-makers to understand, trust, and adopt.

[0204] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. Other modifications can be easily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for short-term prediction of highway traffic flow based on multi-source heterogeneous data, characterized in that, include: S1. Obtain the accident location and road network topology, traverse upstream from the accident location, identify the congestion tail location based on the real-time average speed of the road segment, determine the dynamic influence radius, take the toll station or toll gantry within the radius upstream of the congestion tail location as the boundary node set, extract the connection path from the boundary node set to the accident location, remove the far-end nodes whose free-flow travel time exceeds the limit, and form a dynamic subnet. S2. Obtain the GPS trajectory of vehicles within the dynamic subnet, and calculate the harmonic mean speed and speed variation coefficient of each road segment through map matching and speed smoothing filtering. Combine real-time observation and historical benchmark data to estimate the travel time of each road segment, which is used as the road resistance of the dynamic subnet. S3. Based on the dynamic subnet and road resistance, candidate paths from upstream nodes to the accident point are generated. The generalized impedance of each candidate path is calculated, and a time-varying information penetration rate model is established. Drivers are divided into an unaware group and an unaware group. The selection results of the two groups are weighted according to the information penetration rate to obtain the dynamic diversion rate of each upstream node. The generalized impedance includes travel time, unfamiliarity penalty and accident risk perception. The information penetration rate increases from the time of the accident. The unaware group selects routes according to historical habits, while the unaware group selects routes discretely according to the generalized impedance. S4. Calculate the effective demand based on the original demand of the upstream node and the dynamic diversion rate. Perform queuing dynamics simulation based on the remaining capacity of the accident point and update the queue length. If the tail of the queue blocks the upstream flow, cut off the corresponding demand and reset the relevant road blockage. Return to S1~S3 to iterate until convergence, and output the predicted flow rate and queue length of the accident point.

2. The method for short-term prediction of highway traffic flow based on multi-source heterogeneous data according to claim 1, characterized in that, The method for identifying the congestion tail position in step S1 is as follows: traverse each road segment from the accident location upstream, and determine the end position of the continuous road segment sequence where the real-time average speed of the road segment is lower than the preset speed threshold as the congestion tail position; if there is no road segment with a speed lower than the preset speed threshold, then the accident location is used as the congestion tail position.

3. The method for short-term prediction of highway traffic flow based on multi-source heterogeneous data according to claim 1, characterized in that, The dynamic influence radius in step S1 is determined by the following formula: Among them, D max L represents the dynamic influence radius. queue Indicates the current queue length, v f T represents the free-flow velocity. lookahead This indicates the preset prediction horizon.

4. The method for short-term prediction of highway traffic flow based on multi-source heterogeneous data according to claim 1, characterized in that, In step S2, the harmonic average velocity is calculated using the following formula: in, N represents the harmonic average speed, and N represents the road segment e. ij The number of valid GPS samples retained within the current time window after map matching and velocity smoothing filtering, v k This represents the instantaneous speed of the k-th sample vehicle; The velocity variation coefficient is calculated using the following formula: Among them, CV ij σ represents the velocity variation coefficient. v,ij This represents the standard deviation of the vehicle speeds in the valid GPS sample. For road segments where the number of effective GPS samples is lower than the preset minimum sample threshold, spatiotemporal kriging interpolation or the harmonic mean velocity of the corresponding road segment is estimated based on the upstream road segment velocity wave propagation model.

5. The method for short-term prediction of highway traffic flow based on multi-source heterogeneous data according to claim 4, characterized in that, In step S2, the estimated travel time for each road segment is calculated using the following formula: Among them, t ij (t) represents the travel time, L ij Indicates road segment e ij Length, CV represents the harmonic average velocity at the current moment. ij (t) represents the velocity variation coefficient at the current moment, CV ref Q represents the baseline coefficient of variation under free-flow conditions. hist Q represents the historical traffic volume during the same period. obs Q represents the real-time observed flow rate. cap ω1 represents the traffic capacity of the road segment, ω2 represents the first weighting coefficient, and ω2 represents the second weighting coefficient.

6. The method for short-time prediction of highway traffic flow based on multi-source heterogeneous data according to claim 1, characterized in that, In step S3, candidate paths are generated by performing the K-shortest path algorithm on each upstream node to the accident location, and each candidate path is marked as either passing through the accident location or being a detour path.

7. The method for short-term prediction of highway traffic flow based on multi-source heterogeneous data according to claim 1, characterized in that, The time-varying information penetration rate model in step S3 is as follows: in, This represents the information penetration rate at time t indicating that the driver was aware of the accident, where t represents the current time. acc Let λ represent the time of the accident, λ represent the information diffusion rate constant, and β represent the information diffusion rate constant. VMS β represents the coverage coefficient of the variable message sign. Nav This indicates the coverage coefficient of the navigation app.

8. The method for short-term prediction of highway traffic flow based on multi-source heterogeneous data according to claim 1, characterized in that, In step S3, the probability that the uninformed group chooses path k based on historical habits. It is a preset fixed probability distribution; The specific method by which the informed group makes discrete choices based on generalized impedance is as follows: using a Logit model, the probability of the informed group choosing path k is... Calculate using the following formula: Among them, C i,k (t) represents the generalized impedance of the k-th candidate path from upstream node i to the accident location at time t, θ represents the preset sensitivity coefficient, and K represents the total number of candidate paths; The mixed weighting method is as follows: The final selection probability of candidate path k Calculate using the following formula: The final selection probabilities of each detour path are summed to obtain the dynamic diversion rate η of upstream node i at time t. i (t): Among them, R bypass This represents the set of detour paths among the candidate paths.

9. The method for short-term prediction of highway traffic flow based on multi-source heterogeneous data according to claim 1, characterized in that, The specific method for queuing dynamics simulation in step S4 is as follows: Calculate the net inflow rate Q net (t): in, Indicates that upstream node i has a time delay τ i The last effective demand to arrive, C inc (t) represents the remaining traffic capacity at the accident site at time t; Update queue length L queue (t+Δt): Where Δt represents the time step, l veh This represents the average distance between the front ends of the vehicles. Determine if the updated queue tail blocks the upstream node v i The specific method is as follows: Calculate upstream node v i Position from the end of the congestion P tail Topological distance Dist(v) between i P tail If the following formula is satisfied: Then it is determined that the upstream node vi is blocked, its effective demand is truncated and corrected, and the corrected actual demand is... for: Where, γ block The blocking coefficient has a value range of [0, 1], and is 0 when complete blocking is achieved.

10. A system for the short-term traffic flow prediction method for highways based on multi-source heterogeneous data as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire accident locations and road network topology, as well as vehicle GPS trajectory data, historical benchmark data, variable message signs and information release status of navigation apps; The dynamic subnet construction module is used to traverse upstream from the accident location, identify the congestion tail position based on the real-time average speed of the road segment, determine the dynamic influence radius, take the toll station or toll gantry within the radius upstream of the congestion tail position as the boundary node set, extract the connecting path from the boundary node set to the accident location, and remove the far-end nodes whose free-flow travel time exceeds the limit to form a dynamic subnet. The road resistance calculation module is used to obtain the GPS trajectory of vehicles in the dynamic subnet, and calculate the harmonic mean speed and speed variation coefficient of each road segment through map matching and speed smoothing filtering. It integrates real-time observation and historical benchmark data to estimate the travel time of each road segment as the road resistance of the dynamic subnet. The route selection module generates candidate routes from upstream nodes to the accident point based on dynamic subnets and road resistance, calculates the generalized impedance of each candidate route, establishes a time-varying information penetration rate model, divides drivers into an unaware group and an informed group, and weights the selection results of the two groups according to the information penetration rate to obtain the dynamic diversion rate of each upstream node. The generalized impedance includes travel time, unfamiliarity penalty and accident risk perception. The information penetration rate increases from the time of the accident. The unaware group selects routes according to historical habits, while the informed group selects routes discretely according to the generalized impedance. The traffic prediction and feedback module is used to calculate the effective demand based on the original demand of the upstream nodes and the dynamic diversion rate, perform queuing dynamics simulation based on the remaining capacity of the accident point and update the queue length. If the tail of the queue blocks the upstream flow, the corresponding demand is cut off and the relevant road resistance is reset, triggering the dynamic subnet construction module, road resistance calculation module and path selection module to iterate in sequence until convergence, and outputting the traffic prediction result and queue length prediction result of the accident point.

Images