Traffic congestion prediction method based on dynamic risk field

Abstract

The application discloses a traffic congestion prediction method based on a dynamic risk field, comprising the following steps: step (1), data acquisition and preprocessing; step (2), calculation of a basic risk field; step (3), calculation of a comprehensive risk field; step (4), calculation of the influence of specific events or conditions on the risk field; and step (5), calculation of the influence of dynamic driving behavior on the risk field. The method solves the technical bottlenecks of traditional prediction methods in real-time performance, calculation efficiency, generalization ability and interpretability, realizes real-time performance improvement and calculation resource optimization, generalization ability enhancement and low data dependency, spatial risk quantification and interpretability optimization, hardware compatibility and energy saving, and safety performance collaborative optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of traffic congestion emergency handling, and particularly relates to a traffic congestion prediction method based on a dynamic risk field. BACKGROUND

[0002] With the acceleration of urbanization and the complexity of transportation systems, traffic congestion has become a global problem. Traditional traffic management methods mainly rely on passive response strategies, which are difficult to cope with dynamic changes in traffic flow characteristics.

[0003] In the prior art, traffic congestion prediction is mainly divided into two categories:

[0004] 1. Time series prediction model based on deep learning

[0005] This kind of method is represented by recurrent neural network (RNN) and long short-term memory network (LSTM), which realizes prediction by capturing the time series characteristics of traffic data.

[0006] (1) LSTM model: LSTM network performs well in short-term traffic flow prediction, but it relies on the statistical law of historical data, and the error increases significantly (prediction error rises by 40%) under sudden events (such as accidents, extreme weather);

[0007] (2) Hybrid deep learning architecture: Attention-LSTM model optimizes time series feature extraction through attention mechanism, but the model complexity is high (GPU memory occupancy reaches 4.8GB), and the black box characteristics lead to a lack of explainability in the decision-making process.

[0008] 2. Risk field model based on physical driving

[0009] This kind of model can quantify traffic risk through spatial field intensity distribution, for example:

[0010] (1) Driver risk field (DRF) is a risk field model based on Gaussian distribution, which can quantify micro-vehicle interaction behavior, but it is not extended to macro-road network level congestion prediction;

[0011] (2) Static risk field model is a car-following risk field model that simulates vehicle interaction through fixed parameters, but it cannot dynamically adapt to traffic flow mutations (such as peak period flow fluctuations, dynamic driving behavior or specific events, etc.), resulting in a decrease in prediction accuracy.

[0012] However, the prior art has many limitations, such as:

[0013] (1) Real-time and computational efficiency contradiction

[0014] Deep learning models (such as LSTM) have a response latency of up to 480ms under high load scenarios due to their recursive computation structure, making it difficult to meet the needs of real-time decision-making; while traditional risk field models, although fast in response, rely on static parameters and cannot capture dynamic traffic flow characteristics (such as a surge in local risks caused by sudden accidents).

[0015] (2) Insufficient fusion of spatiotemporal features

[0016] Existing risk field models mostly focus on a single dimension (space or time). For example, in Chinese patent CN117315944A, the LWR-CA model combines flow density and simulation rules, but does not achieve dynamic fusion of multi-source data.

[0017] For example, in Chinese patent CN119274338A, although the multimodal data method improves prediction accuracy by dividing the road into units, it does not introduce a physical driving mechanism, which limits the generalization ability under sudden events.

[0018] (3) Lack of interpretability and generalization ability

[0019] Deep learning models, due to their black-box nature, struggle to explain risk formation mechanisms, hindering the formulation of management strategies; existing models require parameter recalibration (such as road capacity and vehicle density thresholds) when applied across different scenarios, exhibiting poor generalization ability.

[0020] Therefore, the core problem with existing technologies lies in how to construct predictive models that combine physical interpretability, dynamic adaptability, and efficient computational capabilities. Specific challenges include:

[0021] (1) Dynamic field strength modeling: A dynamic parameter update mechanism (such as adaptive adjustment of Gaussian kernel function bandwidth) needs to be designed to reflect traffic flow changes in real time;

[0022] (2) Multi-source data fusion: It is necessary to integrate physical-driven models (such as risk field equations) and data-driven methods (such as deep learning) to balance model accuracy and computational efficiency;

[0023] (3) Edge computing optimization: It is necessary to compress the model size, reduce GPU memory, improve computing efficiency, and support real-time deployment on low-load devices;

[0024] The aforementioned problems have hindered the proactive transformation of traffic congestion prediction. To address the issues of poor real-time performance, low computational efficiency, and insufficient interpretability in existing technologies, this invention provides a dynamic risk field model that achieves the following through a spatiotemporally parallel field strength superposition framework and a lightweight computational path:

[0025] (1) Dynamic parameter update mechanism: Based on real-time traffic data, adaptively adjust field strength parameters to improve the response accuracy of emergencies (e.g., the prediction error of accident scenarios is reduced by 26.2%).

[0026] (2) Spatiotemporal parallel architecture: It integrates physical driving equations and data-driven feature extraction, while optimizing time series prediction (MAE reduced to 6.533) and spatial risk quantification (high-risk area identification accuracy reaches 87%).

[0027] (3) Lightweight computing framework: Through a fixed topology weight optimization algorithm, the GPU memory usage is compressed to 1.2GB and the response time is shortened to 8.5ms, supporting edge device deployment.

[0028] By integrating physics-driven risk field theory with data-driven deep learning, a dynamic traffic congestion prediction model was constructed, achieving high-precision, low-latency, and highly interpretable traffic risk quantification and early warning. Summary of the Invention

[0029] In view of this, the present invention provides a traffic congestion prediction method based on dynamic risk fields.

[0030] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0031] The traffic congestion prediction method based on dynamic risk fields mainly includes the following steps:

[0032] Step (1): Data acquisition and preprocessing;

[0033] Step (2): Calculate the basic risk field;

[0034] Step (3): Calculate the comprehensive risk field;

[0035] Step (4): Calculate the impact of a specific event or condition on the risk field;

[0036] Step (5): Calculate the impact of dynamic driving behavior on the risk field.

[0037] Preferably, in step (1), data collection includes: real-time traffic flow, vehicle speed, road topology, meteorological data, and historical congestion records;

[0038] Data preprocessing includes: using linear interpolation to handle missing values, and Z-score standardization to normalize the data, thereby improving the stability of model training.

[0039] Preferably, in step (2), the basic risk field refers to the static risk distribution, and the calculation formula is as follows:

[0040]

[0041] Where R(x,y,t) represents the traffic congestion risk value at time t and location (x,y), αi is the influence coefficient of the i-th traffic congestion event, and σ is the bandwidth parameter of the Gaussian kernel function, which controls the range of risk diffusion.

[0042] Preferably, in step (3), the formula for calculating the comprehensive risk field is as follows:

[0043] E(x,y,t)=R(x,y)+Dmotion(x,y,t)+Dmarker(x,y,t);

[0044] Where E(x,y,t) represents the comprehensive risk field, that is, the comprehensive traffic congestion risk at location (x,y) and time t; R(x,y) represents the basic risk field, that is, the basic congestion risk without dynamic driving behavior and specific events or conditions.

[0045] Preferably, in step (4), the formula for calculating the impact of a specific event or condition on the risk field is as follows:

[0046]

[0047] Preferably, in step (5), the calculation formula for the impact of dynamic driving behavior on the risk field is as follows:

[0048]

[0049] Where Q(t) is the traffic flow at time t, c is the road capacity, β is the weighting coefficient of the traffic flow influence, (xv,yv) are the vehicle position coordinates, and σv is the spatial standard deviation of the vehicle influence.

[0050] The present invention achieves the following technical effects compared to the prior art:

[0051] The method of this invention solves the technical bottlenecks of traditional prediction methods in terms of real-time performance, computational efficiency, generalization ability and interpretability, and achieves improved real-time performance and optimized computational resources, enhanced generalization ability and low data dependence, spatial risk quantification and optimized interpretability, hardware compatibility and energy saving, and coordinated optimization of security and performance. Attached Figure Description

[0052] Figure 1 This is a dynamic risk field diagram for congestion prediction in this invention;

[0053] Figure 2 This is a comparison diagram of the model architecture of the present invention;

[0054] Figure 3 This is a comparison chart of CPU usage in this invention;

[0055] Figure 4This is a comparison chart of the response times of the present invention;

[0056] Figure 5 This is a comparison chart of the accuracy across datasets for this invention;

[0057] Figure 6 This is a scatter plot showing the safety-efficiency tradeoff of the present invention.

[0058] Figure 7 This is a box plot showing the minimum clearance of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] This invention discloses a traffic congestion prediction method based on a dynamic risk field, which mainly includes the following steps:

[0061] Step (1): Data acquisition and preprocessing;

[0062] Data collection includes: real-time traffic flow, vehicle speed, road topology, weather data, and historical congestion records;

[0063] Data preprocessing includes: using linear interpolation to handle missing values, and Z-score standardization and normalization of data to improve model training stability;

[0064] Step (2): Calculate the basic risk field;

[0065] The basic risk field R refers to the static risk distribution, and its calculation formula is as follows:

[0066]

[0067] Where R(x,y,t) represents the traffic congestion risk value at time t and location (x,y), αi is the influence coefficient of the i-th traffic congestion event, and σ is the bandwidth parameter of the Gaussian kernel function, which controls the range of risk diffusion;

[0068] Step (3): Calculate the comprehensive risk field;

[0069] The formula for calculating the overall risk field E is as follows:

[0070] E(x,y,t)=R(x,y)+Dmotion(x,y,t)+Dmarker(x,y,t);

[0071] Where E(x,y,t) represents the comprehensive risk field, that is, the comprehensive traffic congestion risk at location (x,y) and time t; R(x,y) represents the basic risk field, that is, the basic congestion risk without dynamic driving behavior and specific events or conditions.

[0072] Step (4): Calculate the impact of a specific event or condition on the risk field;

[0073] For a specific event (such as an accident), its impact can be represented as a Gaussian function, which has a high value near the location and time of the event, and decays rapidly with increasing distance and time. The formula for calculating the impact of a specific event or condition on the risk field is as follows:

[0074]

[0075] Step (5): Calculate the impact of dynamic driving behavior on the risk field;

[0076] The impact of dynamic driving behavior includes changes in vehicle speed and fluctuations in traffic flow. The dynamic driving behavior field is combined with real-time vehicle speed and traffic flow fluctuations to adjust the risk field. The calculation formula for the impact of dynamic driving behavior on the risk field is as follows:

[0077]

[0078] Where Q(t) is the traffic flow at time t, c is the road capacity, β is the weighting coefficient of the traffic flow influence, (xv,yv) are the vehicle position coordinates, and σv is the spatial standard deviation of the vehicle influence.

[0079] Among them, the comprehensive risk field E is obtained by adding R, Dmotion and Dmarker. It comprehensively considers various factors and their interactions, providing a complete view of driving risks.

[0080] A visualized 3D surface plot showing the distribution of traffic congestion risks, such as Figure 1 As shown, the X-axis represents the traffic flow risk factor, the Y-axis represents the road capacity risk factor, and the Z-axis represents the congestion risk value.

[0081] The changes in color and height in the diagram visually reflect the degree of congestion risk under different combinations of risk factors.

[0082] Example 1: Real-time performance improvement and computing resource optimization

[0083] Employing a spatially parallel processing architecture, such as Figure 2 As shown, multidimensional spatial data is processed simultaneously through gridded parallel computing units, replacing the chain-like temporal dependency structure of traditional LSTM.

[0084] A lightweight physics-driven algorithm is designed, based on a linear classification threshold (0.5X+0.3Y+0.2Z=70) of risk field strength (X), parallel computing throughput (Y), and resource allocation density (Z), to achieve second-level dynamic updates of risk values ​​(data points are refreshed every 5 seconds).

[0085] Experimental data, as shown in Table 1, show that the CPU utilization rate of the risk field model (15-30%) is significantly lower than that of LSTM (60-90%), the single iteration time is only 8.5ms (LSTM is 120ms), and the batch processing speed reaches 12,000 samples / second (LSTM is 800 samples / second).

[0086] Table 1:

[0087]

[0088] CPU usage comparison chart, such as Figure 3 As shown.

[0089] Under high load scenarios, the latency of the risk field model remained stable within 9ms, while the latency of LSTM soared to 480ms, verifying its millisecond-level response capability. The response time is comparable to... Figure 4 As shown.

[0090] It is evident that this breakthrough overcomes the real-time limitations of traditional deep learning models caused by complex recursive computations, thus meeting the timeliness requirements of intelligent transportation systems for dynamic decision-making.

[0091] Example 2: Enhanced generalization ability and low data dependency

[0092] This risk field quantification framework, driven by physical equations, characterizes the interactions of traffic elements (such as vehicle density and road geometry) through field strength, rather than relying on statistical data. It employs Monte Carlo simulation to generate standardized risk values ​​(0–100) and combines this with a dynamic threshold adjustment mechanism to adapt to different traffic network parameters.

[0093] Cross-dataset testing, such as Figure 5 As shown, the accuracy of the risk field model remained stable (0.82→0.85) under different data distributions, while that of LSTM decreased significantly (0.88→0.55) due to data bias.

[0094] In the training data coverage experiment, the risk field model still maintained high accuracy (0.85) when the coverage was less than 60%, verifying its low dependence on data representativeness.

[0095] It is evident that this overcomes the insufficient generalization ability of traditional models due to data bias, and is applicable to diverse scenarios such as urban roads and highways.

[0096] Example 3: Spatial Risk Quantification and Interpretability Optimization

[0097] By visually identifying congestion hotspots (such as commercial areas and intersections) through field strength mapping, it supports dynamic risk level classification (low / high risk decision threshold surface); it integrates the physical coupling mechanism of multi-source data (traffic flow, vehicle speed, weather), such as the interaction coefficient (1.2) between rainy days and morning rush hour to quantify the superposition effect of environment and time.

[0098] Safety-efficiency trade-off experiments, such as Figure 6 As shown, the risk field model achieves controllable optimization by adjusting the risk threshold (slope a = -0.32), while LSTM cannot actively balance the two due to its black-box characteristics.

[0099] It is evident that this solves the problem of opaque decision-making logic caused by the black-box structure of traditional models, providing a transparent and intervention-friendly basis for traffic management.

[0100] Example 4: Hardware Compatibility and Energy Efficiency

[0101] Fixed topology weight calculation is used instead of the recursive weight update of LSTM, reducing GPU memory usage (1.2GB vs. 4.8GB). Distributed computing resource scheduling is supported, and hardware overload is avoided through dynamic resource allocation density (Z = throughput × log(risk intensity)).

[0102] Hardware resource comparisons show that the risk field model reduces GPU memory requirements by 75% and CPU utilization by 50%, making it suitable for edge computing deployment. This reduces the model's reliance on high-performance hardware and lowers the deployment costs of intelligent transportation systems.

[0103] Example 5: Collaborative Optimization of Safety and Efficiency

[0104] The minimum vehicle gap distribution index is introduced to dynamically adjust vehicle spacing through risk fields, thereby suppressing the risk of sudden braking and rear-end collisions. The optimal path planning strategy is generated in real time based on the risk value to avoid high-risk areas (such as construction sections).

[0105] Box plot analysis, such as Figure 7 As shown, the median minimum gap of the risk field model is improved by 23% compared to LSTM, and the accident risk is reduced by 18%. Experiments verify that the prediction accuracy during peak hours reaches 0.82±0.04 (LSTM is 0.65±0.04), supporting dynamic lane control and variable speed limit decisions. This addresses the shortcomings of traditional models in balancing safety, achieving the dual goals of congestion mitigation and accident prevention.

[0106] In summary, this invention achieves technological breakthroughs in core dimensions such as real-time performance, generalization, interpretability, and energy efficiency through spatial parallel architecture, physics-driven algorithms, and multi-dimensional risk quantification mechanisms, providing an efficient, reliable, and low-cost congestion prediction solution for intelligent transportation systems.

[0107] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A traffic congestion prediction method based on dynamic risk fields, characterized in that, The main steps include: Step (1): Data acquisition and preprocessing; Step (2): Calculate the basic risk field; Step (3): Calculate the impact of a specific event or condition on the risk field; Step (4): Calculate the impact of dynamic driving behavior on the risk field; Step (5): Calculate the overall risk field; In step (3), the formula for calculating the impact of a specific event or condition on the risk field is as follows: In step (4), the calculation formula for the impact of dynamic driving behavior on the risk field is as follows: Where Q(t) is the traffic flow at time t, c is the road capacity, β is the weighting coefficient of the traffic flow's influence, (x v ,y v ) represents the vehicle's position coordinates, σ v It is the spatial standard deviation affected by vehicles; In step (5), the formula for calculating the comprehensive risk field is as follows: E(x,y,t)=R(x,y,t)+D motion (x,y,t)+D marker (x,y,t); Where E(x,y,t) represents the comprehensive risk field, that is, the comprehensive traffic congestion risk at location (x,y) and time t; R(x,y,t) represents the basic risk field, that is, the basic congestion risk without dynamic driving behavior and specific events or conditions.

2. The traffic congestion prediction method based on dynamic risk field according to claim 1, characterized in that, In step (1), data collection includes: real-time traffic flow, vehicle speed, road topology, meteorological data, and historical congestion records; Data preprocessing includes: using linear interpolation to handle missing values, and Z-score standardization to normalize the data, thereby improving the stability of model training.

3. The traffic congestion prediction method based on dynamic risk field according to claim 1, characterized in that, In step (2), the basic risk field refers to the static risk distribution, and the calculation formula is as follows: Where R(x,y,t) represents the traffic congestion risk value at time t and location (x,y), α i σ is the impact coefficient of the i-th traffic congestion event, and σ is the bandwidth parameter of the Gaussian kernel function, which controls the range of risk diffusion.

Images