Online learning graph convolution method for real-time prediction of road network traffic under emergencies

The online learning graph convolution method addresses the challenge of predicting road network traffic changes during sudden accidents by using a real-time updated graph convolutional neural network to adapt model parameters, enhancing prediction accuracy and guiding traffic management.

CN118097934BActive Publication Date: 2024-12-03SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410053579.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-15
Publication Date
2024-12-03
Estimated Expiration
2044-01-15

AI Technical Summary

Technical Problem

Existing traffic flow prediction models struggle to accurately predict road network traffic changes during sudden major accidents due to the unpredictability of accident timing and location, damage to data collection equipment, and lack of sufficient historical data, making it difficult to achieve satisfactory short-term prediction results.

Method used

An online learning graph convolution method using a real-time updated graph convolutional neural network that initializes and updates model parameters based on historical data from road network detectors to quickly adapt to traffic flow changes after emergencies, employing an adaptive graph convolution model with a parameter update mechanism.

Benefits of technology

The method enables rapid and accurate prediction of road network traffic changes post-emergency, guiding traffic evacuation and route optimization by leveraging real-time data updates and historical data patterns, improving prediction accuracy significantly.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118097934B_ABST
    Figure CN118097934B_ABST
Patent Text Reader

Abstract

The present invention discloses an online learning graph convolution method for real-time prediction of road network flow in emergencies. Through different parameter training modes before and after the accident, a more accurate prediction of road network flow after the accident is achieved. The road network flow characteristics for the case where no accident occurs can be trained through the detector data before the accident. However, after the accident, the road network flow characteristics change significantly, and the model parameters trained before the accident are no longer applicable to the road network flow characteristics when the accident occurs. Therefore, after the accident occurs, a real-time parameter update method is adopted to update the parameters of the adaptive graph convolution network in real time, so that the model can recognize the characteristics of the accident and achieve a more accurate prediction of the highway network traffic flow after the accident. The method for predicting highway network traffic flow after a major traffic accident based on a real-time parameter update graph convolution network proposed in the present invention has a high prediction accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to an online learning graph convolution method for real-time prediction of road network flow under emergencies, and belongs to the technical field of traffic flow prediction. Background Art

[0002] In recent years, with the continuous expansion of traffic scale and travel demand, the frequency and number of accident types of sudden accidents and abnormal conditions in urban traffic environments have also been increasing. Accurate and timely prediction of the traffic status of urban road networks after sudden major accidents is conducive to optimizing driving routes for travelers, and assisting relevant departments to reasonably allocate traffic resources and reduce social property losses. It has great positive and practical significance for the construction of smart cities and the development of a harmonious society.

[0003] However, existing traffic flow prediction models are mainly designed for traffic scenarios under normal conditions. Other prediction models designed for sudden major accidents (for example, bridge collapse or temporary road closure causing long-term loss of traffic capacity of road sections) are difficult to fully mine data. This is because, on the one hand, the time and place of the sudden accident are difficult to determine, and the impact on the distribution and operation of traffic flow is unknown, resulting in completely different regular characteristics of traffic network flow distribution before and after the sudden accident. It is difficult to obtain the change pattern after the sudden incident through the data pattern before the sudden incident, and to predict the flow distribution trend after the incident.

[0004] On the other hand, unlike regular traffic accidents which last for a short time, sudden major traffic accidents may cause some data collection equipment to be damaged. Therefore, it is difficult to obtain sufficient historical data support and a sufficient number of training samples in sudden events. The above two factors make it difficult to achieve satisfactory results in the short-term traffic flow prediction task considering sudden accidents. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an online learning graph convolution method for real-time prediction of road network traffic under emergencies, and to use a real-time updated graph convolutional neural network to quickly and accurately respond to changes in road network traffic flow caused by major accidents, and to timely and accurately guide traffic evacuation after major accidents and optimize travelers' driving routes.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] The online learning graph convolution method for real-time prediction of road network traffic under emergencies includes the following steps:

[0008] Step 1, obtaining historical traffic flow data of road network detectors deployed near the road section where the emergency occurred, and preprocessing the historical traffic flow data using a preset time step;

[0009] Step 2: Use the preprocessed historical data to train the adaptive graph convolution model to obtain a trained model, and store the parameters of the trained model and the optimizer parameters;

[0010] Step 3: After an emergency occurs, the parameters of the model trained in step 2 and the optimizer parameters are used as the initial parameters of the adaptive graph convolution prediction model after the event occurs. The initialized adaptive graph convolution prediction model is used to predict the road network flow of the current time step to obtain the prediction result, and the real-time road network flow of the current step and the prediction result are used to update the parameters of the adaptive graph convolution prediction model.

[0011] Step 4: Scroll the time to the next step length, obtain the real-time road network flow of the next step length, return to step 3, and use the adaptive graph convolution prediction model with updated parameters to make predictions again until the emergency event is handled.

[0012] As a preferred solution of the present invention, in the step 1, traffic flow data of road network detectors deployed on highways within a radius of 16 kilometers centered on the section where the emergency occurred are obtained two weeks before the emergency occurs, including: detector number, date, vehicle passing time and flow information, a fixed time step is set, and the flow information is aggregated according to the time step to obtain pre-processed historical data.

[0013] As a preferred solution of the present invention, in step 2, the formula of the adaptive graph convolution model is as follows:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019] in, is the adjacency matrix learned from the data, E is the node embedding matrix, z t is the updated gating matrix, r t is the updated gating matrix, is the hidden state matrix, h t 、h t-1 are the hidden states at time t and t-1 respectively; σ is the activation function, i.e., the Sigmoid function; X :,t are all the node features of the input, W z , Wr , are the training weight parameters, b z 、b r , They are the bias parameters of training respectively;

[0020] The preprocessed historical data is divided into training set, validation set and test set in a ratio of 6:2:2, and the adaptive graph convolution model is trained. After the training is completed, the model parameter file, model optimizer parameter and hyperparameter file are saved.

[0021] As a preferred solution of the present invention, in step 3, the parameters of the adaptive graph convolution prediction model are updated using the gradient descent criterion, and the formula is as follows:

[0022]

[0023] in, m t =β1m t-1 +(1-β1)g t , β1 and β2 are the exponential decay rates of momentum estimation, β1∈(0,1), β2∈(0,1), m t 、m t-1 、v t-1 and v t They are all intermediate variables for calculation; η is the step size, η=0.003; ε is the minimum value, ε=10 -8 ; g t is the gradient, is the calculated gradient, L t is the calculated loss function, W θ '(t+1),W θ '(t) are the training parameters at time t+1 and time t respectively.

[0024] A computer device comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, the steps of the online learning graph convolution method for real-time prediction of road network traffic under emergency events are implemented.

[0025] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of an online learning graph convolution method for real-time prediction of road network traffic under emergencies.

[0026] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0027] 1. The present invention introduces a real-time parameter update mechanism, which enables the graph convolutional model to respond and learn the road network traffic characteristics after a major accident more quickly, and can better predict the traffic changes after a major accident, thereby guiding the traffic evacuation and route selection of travelers after the accident.

[0028] 2. Different from the previous single-section research on accidents, this paper introduces a graph convolutional neural network to reflect the changes in road network traffic after major accidents.

[0029] 3. Due to the randomness and severity of major sudden accidents, historical data of such accidents are often missing. The different training parameter modes before and after the accident of the present invention can make the model reflect the characteristics of road network traffic changes when the accident occurs, based on the original learned road network traffic characteristics, without relying on historical major sudden accident data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of an online learning graph convolution method for real-time prediction of road network traffic under emergencies according to the present invention;

[0031] Figure 2 It is the prediction model block diagram;

[0032] Figure 3 This is the collapse location of the I-35W bridge in the embodiment of the present invention;

[0033] Figure 4 is the position of the detector selected for deployment in the embodiment of the present invention;

[0034] Figure 5 It is a flow prediction comparison diagram of the S574 detector after an accident in an embodiment of the present invention;

[0035] Figure 6 is a flow prediction comparison diagram of the S61 detector after an accident in an embodiment of the present invention;

[0036] Figure 7 It is a flow prediction comparison chart of the S569 detector after an accident in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0038] like Figure 1 , Figure 2 As shown in FIG. 1 , it is a flow chart of the online learning graph convolution method for real-time prediction of road network traffic under emergency events proposed by the present invention, and the specific steps are as follows:

[0039] (1) Data collation and extraction: The data of the road network detectors are collated and the field data extracted include: detector ID, vehicle passing time, and traffic data.

[0040] (2) Data division: Aggregate data according to time granularity. According to step (1), the flow information of the deployed road network detector data is aggregated according to time granularity.

[0041] (3) Pre-training the model using historical data: Input the historical road network detector data sorted in step (2) into the adaptive graph convolution (AGCRN) model. The historical data time selection range can be selected according to different prediction accuracies. Generally, 2 weeks of historical data is sufficient. The specific operation is as follows:

[0042] (31) Divide the historical data into training set, validation set, and test set;

[0043] (32) The divided data set is input into the adaptive graph convolution model to train the model, and the trained model parameters and optimizer parameters are stored.

[0044] The AGCRN model is one of the best performing deep learning models for traffic prediction problems in traditional scenarios, the traditional graph convolutional neural network (GCN). The core formula of GCN is as follows:

[0045]

[0046] Among them, I N is the identity matrix; A is the adjacency matrix; D is the degree matrix; X is the input feature matrix; Θ and b are both learnable parameter matrices; Z is the learned graph. Unlike previous GCN networks that require matrices A and D as input, AGCRN learns the intrinsic hidden associations between nodes through data adaptive learning.

[0047]

[0048] Among them, E is the node embedding learnable matrix, which is used to identify the adjacency relationship between nodes.

[0049] Another difference is that AGCRN decomposes the learnable parameter matrices Θ and b into the product of node embedding matrices E and W to reduce the number of parameters that need to be calibrated. Formula (1) can be rewritten as:

[0050]

[0051] Where E is the node embedding matrix and W is the weight matrix. They are all learnable parameter matrices. Combining formulas (2) and (3) gives the final formula of AGCRN:

[0052]

[0053] Among them, σ is the activation function, that is, the Sigmoid function. t is the updated gating matrix, the larger it is, the more information from the previous time step is retained. t is the update gating matrix, which is used to compute the predicted hidden state. is the hidden state matrix. :,t are all the node features of the input. W and b represent the corresponding weight parameter matrix and the vector of bias values, respectively.

[0054] Assume that there is traffic flow data from N detectors in a traffic road network, which can be represented as T, where is the number of detectors in the road network at time X :,t traffic flow data, C is the characteristic dimension of the acquired traffic data. If C = 1, the analysis is only performed from the perspective of traffic flow.

[0055]

[0056] Among them, when X :,t =X (a):,t When an unexpected event occurs, the training batch is 1. :,t =X (n):,t Under normal traffic conditions, the training batch is H, H>>1.

[0057] The training batches were altered before and after major events, and equation (5) was substituted into equation (4) to obtain the encoding process.

[0058] The accident prediction problem can be solved by using formula (5) for classification prediction. For the prediction problem after a major accident, represents the predicted value of the road network node characteristics at time t+1 after the major accident, that is, the output of the single-step traffic flow prediction model, and then It can be decoded and obtained by the following equation:

[0059]

[0060] Among them, h :,t is the encoding result, which can be obtained from formula (4), W θ is the set of parameters that the model learns and continuously updates, namely W and b. (a):,t+1 is the true value of each node feature of the road network after a major accident, that is, the predicted label value. The objective function or gradient descent criterion of the method proposed in the present invention is:

[0061]

[0062] Among them, g t is the calculated gradient. The process of calculating the learning rate involves using the Adam optimizer to calculate the learning rate:

[0063]

[0064] Among them, β1, β2∈(0,1) are the exponential decay rates of momentum estimation, β1=0.9, β2=0.999; g t ⊙g t ;m t and v t To calculate intermediate variables; Represents the power of β1 and β2. Parameter update process:

[0065]

[0066] Among them, η is the default step value, η = 0.003; ε is the minimum value, ε = 10 -8 .

[0067] (4) When a major accident occurs, the real-time adaptive graph convolution prediction model is initialized: the model parameters and optimizer parameters pre-trained in step (3) are introduced into the adaptive graph convolution prediction model after the accident occurs and used as the initial parameters of the model.

[0068] (5) Real-time parameter update during an accident: Use the deployed sensors to obtain the road network traffic flow data during the accident online, and input the data into the adaptive graph convolution model initialized in step (4) in real time for prediction. After obtaining the prediction results, the parameter update process is performed. The specific operations are as follows:

[0069] (51) Calculate the loss function: The real value of the accident real-time data at the current time step (i.e., the label value) is input into the loss function, and the error result between the two is calculated in combination with the prediction result in step (5);

[0070] (52) Real-time parameter update: According to the loss function calculation result in step (51), the partial derivatives of the model parameters and optimizer parameters are solved for the loss function, and back-propagated to update the parameter values ​​in the adaptive graph convolution model.

[0071] (6) Real-time prediction of road network flow during an accident: After the model parameters are updated, scroll to the next time step. After collecting the traffic flow data for the next time step, return to step (5) and input it into the model again for prediction.

[0072] Example

[0073] This implementation case uses data from 59 detectors around the I-35W Bridge in Minnesota from July 15, 2007 to August 3, 2007. The I-35W Bridge collapsed at 6 pm on August 1, causing the loss of traffic capacity on the road section and causing serious traffic disruption. Through data collation and analysis, we explore the traffic flow of major accidents based on graph convolutional networks, realize travel path selection and node flow prediction. It mainly includes the following four stages:

[0074] Data collation and extraction: The detector data deployed in Minnesota is imported into the SPSS database. Each piece of data contains seven fields: detector number, date, passing time, flow information, average speed, road section capacity, and traffic density. Since the present invention focuses on flow extraction, only four fields of data, detector number, date, passing time, and flow information, are retained. The resulting data table is shown in Table 1. The passing time is accurate to half a minute, for example, S1009, 2007 / 7 / 30, 12:01:30AM, represents July 30, 2007, 12:01:30 am to 12:02 am, and the number of vehicles passing through detector S1009 is 2 vehicles. The flow information is aggregated according to 5 minutes, and 288 flow data are obtained for each detector every day.

[0075] Table 1 Detector data sheet

[0076]

[0077]

[0078] Bridge sections where major accidents occur, such as Figure 3 As shown in FIG. 1 , a total of 59 node detectors near the highway network are selected near the road section, and the road network detectors are selected from the main road detectors, such as highways, expressways and other road section detectors with higher road grades, such as Figure 4 shown.

[0079] Pre-training of graph convolution model: The graph convolution (AGCRN) model was pre-trained with the traffic data from 2 weeks before the accident (2007.07.15-2007.07.31), and the data was divided into training set, validation set, and test set in a ratio of 6:2:2. The training batch was 64 and the training round was 400. After the training was completed, the model parameter file, model optimizer parameter and hyperparameter file were saved.

[0080] Comparison of prediction results:

[0081] The road network flow predicted by real-time parameter update is compared with the model before improvement. The 12 hours after the sudden accident is selected. Figure 5 , 6and 7. Use MAE, RMSE, and MAPE indicators as metrics:

[0082]

[0083]

[0084]

[0085] Table 2

[0086]

[0087] As can be seen from Table 2, all three indicators have decreased significantly, and the prediction effect has been significantly improved. From the MAPE indicator, it can be seen that the original graph convolution model has a very low accuracy in predicting road network flow after sudden accidents and cannot reflect the characteristics of the accident. The improved real-time parameter update model can achieve a prediction accuracy of more than 63%, which shows that the method for predicting road network flow after sudden traffic accidents based on real-time parameter update graph convolution network proposed in the present invention is feasible and has significantly improved accuracy compared with the original method.

[0088] Based on the same inventive concept, the implementation case of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the aforementioned online learning graph convolution method for real-time prediction of road network traffic under emergencies are implemented.

[0089] Based on the same inventive concept, the implementation case of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the aforementioned online learning graph convolution method for real-time prediction of road network traffic under emergencies.

[0090] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0092] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0093] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0094] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. An online learning graph convolution method for real-time prediction of road network traffic under emergencies, characterized by: The steps include: Step 1, obtaining historical traffic flow data of road network detectors deployed near the road section where the emergency occurred, and preprocessing the historical traffic flow data using a preset time step; In the step 1, traffic flow data of road network detectors arranged on the highway within a radius of 16 kilometers with the section where the emergency occurred as the center is obtained two weeks before the emergency occurs, including: detector number, date, vehicle passing time and flow information, a fixed time step is set, and the flow information is aggregated according to the time step to obtain pre-processed historical data; Step 2: Use the preprocessed historical data to train the adaptive graph convolution model to obtain a trained model, and store the parameters of the trained model and the optimizer parameters; Step 3: After an emergency occurs, the parameters of the model trained in step 2 and the optimizer parameters are used as the initial parameters of the adaptive graph convolution prediction model after the event occurs. The initialized adaptive graph convolution prediction model is used to predict the road network flow of the current time step to obtain the prediction result, and the real-time road network flow of the current step and the prediction result are used to update the parameters of the adaptive graph convolution prediction model. Step 4: Scroll the time to the next step length, obtain the real-time road network flow of the next step length, return to step 3, and use the adaptive graph convolution prediction model with updated parameters to make predictions again until the emergency event is handled.

2. The online learning graph convolution method for real-time prediction of road network traffic under emergencies according to claim 1 is characterized in that: In step 2, the formula of the adaptive graph convolution model is as follows: in, is the adjacency matrix learned from the data, E is the node embedding matrix, z t is the updated gating matrix, r t is the updated gating matrix, is the hidden state matrix, h t 、h t-1 are the hidden states at time t and t-1 respectively; σ is the activation function, i.e., the Sigmoid function; X :,t are all the node features of the input, W z , W r , are the training weight parameters, b z , b r , They are the bias parameters of training respectively; The preprocessed historical data is divided into training set, validation set and test set in a ratio of 6:2:2, and the adaptive graph convolution model is trained. After the training is completed, the model parameter file, model optimizer parameter and hyperparameter file are saved.

3. The online learning graph convolution method for real-time prediction of road network traffic under emergencies according to claim 1 is characterized in that: In step 3, the parameters of the adaptive graph convolution prediction model are updated using the gradient descent criterion, and the formula is as follows: in, m t =β1m t-1 +(1-β1)g t , β1 and β2 are the exponential decay rates of momentum estimation, β1∈(0,1), β2∈(0,1), m t 、m t-1 、v t-1 and v t They are all intermediate variables for calculation; η is the step size, η=0.003; ε is the minimum value, ε=10 -8 ; g t is the gradient, is the calculated gradient, L t is the calculated loss function, W θ '(t+1),W θ '(t) are the training parameters at time t+1 and time t respectively.

4. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: When the processor executes the computer program, the steps of the online learning graph convolution method for real-time prediction of road network traffic under emergency events as described in any one of claims 1 to 3 are implemented.

5. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the online learning graph convolution method for real-time prediction of road network traffic under emergency events as described in any one of claims 1 to 3 are implemented.