Model-data dual-drive highway traffic state prediction method and system

Abstract

The invention discloses a model-data dual-drive highway traffic state prediction method and system, and the method comprises the steps: firstly obtaining a road network topological structure and detector layout information, and carrying out the preprocessing of the obtained historical main line traffic flow, historical vehicle speed data and historical ramp flow data of a detector point location; for each kind of historical data, training a corresponding long and short-term memory network as a prediction model; then calibrating a second-order macroscopic traffic flow model by using the three types of preprocessed historical data to obtain global parameters and traffic fundamental diagram parameters with the optimal model-data fitting degree; and respectively inputting the existing main line traffic flow time sequence, the vehicle speed time sequence and the existing ramp flow data of the detector point location into the corresponding prediction models to obtain a traffic state prediction value. And finally, a road network traffic state predictor based on a second-order macroscopic traffic flow model and an extended Kalman filter obtains a whole road network traffic state prediction result at each moment.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic management, and in particular relates to a model-data dual-driven highway traffic state prediction method and system. Background Art

[0002] The rapid growth of car ownership is an inevitable product of social and economic development. According to existing public data, as of the end of September 2022, the national motor vehicle ownership reached 412 million, including 315 million cars. The growth of road mileage can never catch up with the growth of traffic demand, and road construction alone cannot fundamentally solve the problem of traffic congestion. This is a proven basic law of the evolution of traffic demand and supply. Road traffic congestion seriously affects travel efficiency, driving safety and traffic environment, and brings huge losses to the social economy. As an important part of the road traffic network, the traffic status of highways is becoming increasingly complex. In order to improve the operating efficiency of the highway network, alleviate traffic congestion, and optimize resource allocation, traffic status prediction has become a key task for smart highway traffic control.

[0003] Model-based traffic prediction uses a macro traffic flow model to characterize the dynamic characteristics of traffic flow, thereby achieving traffic flow state prediction. However, traffic flow models often involve certain ideal assumptions, and real data is needed for parameter calibration and model verification before application. In addition, model-based prediction methods are difficult to capture the random characteristics of traffic flow. In the case of multi-step prediction, the model prediction error will accumulate due to the lack of real-time data input. On the other hand, data-driven traffic prediction methods do not require an analytical description of traffic flow characteristics and can better capture the random characteristics of traffic flow, but they are highly dependent on the quality and quantity of data. When the training data is insufficient, the difference between the training data and the test data is large, and the training data noise interference is strong, the prediction performance is poor.

[0004] In view of this, it is urgent to develop an innovative model-data dual-driven traffic state prediction method. This method is based on the traffic state prediction results of the highway network, and deduces the congestion process of the entire road network, thereby providing basic conditions for smart highway traffic control, which is of great significance for alleviating traffic congestion, improving traffic efficiency, and optimizing the operation of the transportation system. Summary of the invention

[0005] The purpose of the present invention is to solve the problems existing in the prior art and provide a model-data dual-driven highway traffic state prediction method and system to accurately and efficiently predict the dynamic changes of the traffic volume and speed of the entire road network. The method of the present invention includes technologies such as dynamic modeling of highway network traffic flow, holographic identification of network traffic state based on traffic state estimation, model-data dual-driven traffic state prediction, and real-time dynamic deduction of large-scale road network traffic operation.

[0006] In order to achieve the above-mentioned invention object, the present invention specifically adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a model-data dual-driven highway traffic state prediction method, comprising the following steps:

[0008] S1: Obtain the road network topology and detector layout information, and pre-process the historical mainline traffic flow, historical vehicle speed data, and historical ramp flow data of the acquired detector points, and obtain the pre-processed historical mainline traffic flow, pre-processed historical vehicle speed data, and pre-processed historical ramp flow data;

[0009] S2: training a first long short-term memory network on the preprocessed historical main line traffic flow, using the trained first long short-term memory network as the flow prediction model of the detector point, training a second long short-term memory network on the preprocessed historical vehicle speed data, using the trained second long short-term memory network as the speed prediction model of the detector point, training a third long short-term memory network on the preprocessed historical ramp flow data, using the trained third long short-term memory network as the ramp flow prediction model of the detector point;

[0010] S3: Use the preprocessed historical mainline traffic flow, preprocessed historical vehicle speed data, and preprocessed historical ramp flow data in S1 to calibrate the second-order macro traffic flow model, and obtain the global parameters and traffic basic diagram parameters with the best model-data fit;

[0011] S4: for the current moment, input the existing main line traffic flow time series of the detector point into the traffic flow prediction model to obtain the main line traffic flow prediction result for the future period, input the existing vehicle speed time series of the detector point into the speed prediction model to obtain the vehicle speed prediction result for the future period, input the existing ramp flow data of the detector point into the ramp flow prediction model in step 1 to obtain the ramp flow prediction result for the future period;

[0012] S5: At the current moment, the traffic flow prediction results, vehicle speed prediction results and ramp flow prediction results of the future time period are used as traffic state prediction values, and the traffic state prediction values ​​of all detector points are used to drive the road network traffic state predictor based on the second-order macro traffic flow model and the extended Kalman filter to obtain the traffic state prediction result of the entire road network corresponding to the current moment;

[0013] S6: After obtaining the traffic status prediction result of the entire road network at the current moment, re-execute steps S4 and S5 for the next moment to obtain a new traffic status prediction result of the entire road network, and complete the highway traffic status prediction.

[0014] Preferably, in step S3, the compact state equation form of the second-order macroscopic traffic flow model is as follows:

[0015] x(k+1)=f[x(k),d(k),p,ξ(k)]

[0016] where x(k+1) and x(k) represent the state vectors of the k+1th and kth discrete time periods, respectively; f represents the second-order macroscopic traffic flow model; d(k) represents the disturbance vector of the kth discrete time period; ξ(k) represents the system noise vector of the kth discrete time period; p is the model parameter vector, which consists of the global parameters τ, v, k, δ, φ and the traffic basic graph parameters v f ,ρ cr ,q cap The specific forms are as follows:

[0017] p=[τ,v,k,δ,φ,v f ,ρ cr ,q cap ]

[0018] Where τ represents the relaxation time parameter; v represents the steady-state expectation parameter; κ represents the stability parameter; δ represents the confluence parameter; φ represents the lane number reduction parameter; v f represents the free flow velocity, ρ cr represents the critical density; q cap Indicates traffic capacity.

[0019] Preferably, in step S3, the global parameters are relaxation time parameters, steady-state expectation parameters, stability parameters, confluence parameters and lane number reduction parameters.

[0020] Preferably, in step S3, the basic traffic map parameters are free flow speed, critical density and traffic capacity.

[0021] Based on the above scheme, each step can be implemented in the following preferred specific manner.

[0022] Preferably, in step S5, the analytical expression of the road network traffic state predictor is as follows:

[0023]

[0024] In the formula, represents the one-step forward estimation of the system state x(k) based on the traffic detection vector y(k) of the kth discrete period; f1, g are both nonlinear vector functions; K(k) represents the Kalman gain matrix of the kth discrete period.

[0025] In a second aspect, the present invention provides a model-data dual-driven highway traffic state prediction system, which includes:

[0026] The data acquisition and processing module is used to obtain the road network topology and detector layout information, and pre-process the historical mainline traffic flow, historical vehicle speed data and historical ramp flow data of the acquired detector points, and obtain the pre-processed historical mainline traffic flow, pre-processed historical vehicle speed data and pre-processed historical ramp flow data;

[0027] A model acquisition module is used to train a first long short-term memory network on the preprocessed historical main line traffic flow, use the trained first long short-term memory network as the flow prediction model of the detector point, train a second long short-term memory network on the preprocessed historical vehicle speed data, use the trained second long short-term memory network as the speed prediction model of the detector point, train a third long short-term memory network on the preprocessed historical ramp flow data, and use the trained third long short-term memory network as the ramp flow prediction model of the detector point;

[0028] A model calibration module is used to calibrate the second-order macro traffic flow model using the preprocessed historical main line traffic flow, the preprocessed historical vehicle speed data, and the preprocessed historical ramp flow data to obtain the global parameters and traffic basic diagram parameters with the best model-data fit;

[0029] A prediction module is used to input the existing main line traffic flow time series of the detector point into the traffic flow prediction model for the current moment to obtain the main line traffic flow prediction result for the future period, input the existing vehicle speed time series of the detector point into the speed prediction model to obtain the vehicle speed prediction result for the future period, and input the existing ramp flow data of the detector point into the ramp flow prediction model for the step to obtain the ramp flow prediction result for the future period;

[0030] The result acquisition module uses the traffic flow prediction results, vehicle speed prediction results and ramp flow prediction results of the future time period as the traffic state prediction value at the current moment, and uses the traffic state prediction values ​​of all detector points to drive the road network traffic state predictor based on the second-order macro traffic flow model and the extended Kalman filter to obtain the traffic state prediction result of the entire road network corresponding to the current moment;

[0031] The result updating module is used to re-execute the prediction module and the result obtaining module for the next moment after obtaining the traffic status prediction result of the entire road network at the current moment, obtain the new traffic status prediction result of the entire road network, and complete the highway traffic status prediction.

[0032] In a third aspect, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the model-data dual-driven highway traffic status prediction method as described in any one of the schemes in the first aspect above.

[0033] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the model-data dual-driven highway traffic status prediction method as described in any of the schemes in the first aspect above is implemented.

[0034] In a fifth aspect, the present invention provides a computer electronic device comprising a memory and a processor;

[0035] The memory is used to store computer programs;

[0036] The processor is used to implement the model-data dual-driven highway traffic status prediction method as described in any of the solutions of the first aspect when executing the computer program.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] Existing machine learning traffic state prediction methods are almost all purely data-driven and do not rely on physical models. These methods are suitable for realizing traffic state prediction for detection points with rich historical data; however, they are mostly beyond the reach of non-detection points with missing historical data. In view of this, the method of the present invention combines machine learning methods and second-order macro traffic flow models, not only giving play to the advantages of machine learning methods in data-driven state prediction, but also making use of the large spatiotemporal dynamic representation capabilities of second-order macro traffic flow models to strive to break through the traffic state prediction problem of non-detection points. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the prediction steps of the method of the present invention;

[0040] Figure 2 A schematic diagram of the traffic status of the entire highway network of the present invention;

[0041] Figure 3 This is a topological diagram of an experimental road network in a certain megacity in China adopted by the present invention;

[0042] Figure 4 It is a spatiotemporal flow diagram of the traffic status prediction of the entire road network of the present invention;

[0043] Figure 5 It is a time-space diagram of the predicted speed of the traffic status of the entire road network of the present invention;

[0044] Figure 6It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0045] In order to make the above-mentioned purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation mode of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0046] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for the purpose of distinguishing descriptions, and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features.

[0047] In this embodiment, first, for the points where detectors are deployed in the highway network, LSTM is used to extract the time characteristics of the detection data (historical main line traffic flow, historical vehicle speed data, and historical ramp flow data) to accurately capture the evolution of traffic flow, speed, and ramp flow. By learning the time characteristics of the three historical data, a prediction model corresponding to each type of historical data is constructed. Secondly, the second-order macro traffic flow model METANET is used to model the traffic flow of the entire highway network under investigation, the model parameters are calibrated using all existing data, and the traffic status of the road network is reproduced based on the model. Finally, a road network traffic status predictor is designed based on the second-order macro traffic flow model METANET and the extended Kalman filter EKF. The road network traffic status predictor can use the detection data from limited detection points at any time to reconstruct the entire road network traffic status at the current moment in real time, and realize the traffic status prediction of the entire road network in the future period.

[0048] like Figure 1 As shown, in a preferred implementation of the present invention, the above-mentioned model-data dual-driven highway traffic state prediction method includes the following steps S1 to S6. The specific implementation process is described in detail below.

[0049] 1. Historical data acquisition and processing

[0050] S1: Obtain the road network topology and detector layout information, and preprocess the historical mainline traffic flow, historical vehicle speed data, and historical ramp flow data of the acquired detector points to obtain the preprocessed historical mainline traffic flow, preprocessed historical vehicle speed data, and preprocessed historical ramp flow data.

[0051] It should be noted that in step S1 of the present invention, the gantry detectors on the highway are used to obtain data. These detectors record the time when the vehicle passes through the gantry, providing a valuable data source. For the gantry detectors on the main line, the single vehicle trajectory information is first extracted using the time when the vehicle passes through the gantry:

[0052] {e i ,g i,1 ,g i,2 ,…,g i,n ,t i,1 ,t i,2 ,…,t i,n}

[0053] In the formula, e i is the license plate number of the i-th vehicle; g i,1 ,g i,2 ,…,g i,n are the numbers of the 1st, 2nd, ..., nth gantry that the i-th vehicle passes through; t i,1 ,t i,2 ,…,t i,n are the times when the i-th vehicle passes through the 1st, 2nd, …, nth gantry respectively.

[0054] For each bicycle trajectory information, the license plate of the i-th vehicle can be counted into the n-th gantry g when passing through the n-th gantry i,n The detected traffic flow. The location d of the n-1th gantry can be obtained by using the road topology information i,n-1 The position d between the nth gantry i,n Then the average speed of the section between the two gantries when the i-th vehicle passes can be calculated. Here, for simplification, the average speed of this section is included in the g of the n-th gantry when the i-th vehicle passes through the n-th gantry. i,n Detected speed:

[0055]

[0056] In the formula, v i,n represents the speed of the i-th vehicle passing through the n-th gantry; d i,n d i,n-1 They represent the spatial positions of the nth and n-1th gantry respectively through which the i-th vehicle passes; t i,n ,t i,n-1 They represent the time when the i-th vehicle passes through the n-th and n-1-th gantry respectively.

[0057] Through time-based aggregation processing, the historical mainline traffic flow q and historical vehicle speed data v of each gantry detector on the mainline can be obtained. Then, the ramp flow corresponding to the toll station can be obtained using the gantry data of the toll station. For ramps that do not include toll stations, the mainline gantry data can be used to reverse the missing ramp flow. The main idea is that vehicles cannot suddenly disappear or appear: if the i-th vehicle first appears in g i,1 gantry, then it must enter the main line from the nearest upstream entrance ramp; conversely, if the i-th vehicle last appears at g i,n If the vehicle has a gantry, it must leave the main line from the nearest downstream exit ramp. The historical ramp flow data r can be obtained by using the reverse ramp flow and the ramp flow recorded by the toll station.

[0058] Since the gantry detector cannot directly provide available mainline traffic flow, vehicle speed data, and ramp flow data, it is necessary to preprocess the historical mainline traffic flow, historical vehicle speed data, and historical ramp flow data at the detector point. In this embodiment, due to temporary failure of the detector and other reasons, a small amount of data is missing in the historical mainline traffic flow, historical vehicle speed data, and historical ramp flow data. In order to complete the missing data, this embodiment uses linear interpolation and historical averaging methods to preprocess the above three types of historical data. Linear interpolation refers to the use of adjacent time period data in the current day's data to interpolate and complete the missing data points; the historical averaging method refers to the completion of missing data points based on the average value of non-missing data in the same time period of other days.

[0059] 2. Prediction of limited detection point status

[0060] In the present invention, for the target road network, firstly, the long short-term memory network LSTM is used to extract the time characteristics of the traffic flow of all detector points, and predict the traffic status of these detector points (including all boundary conditions of the road network). Then the second-order macro traffic flow model METANET is introduced to model the traffic status of the entire road network, explicitly considering the impact of entrance and exit ramps on the main line flow, and using the extended Kalman filter EKF to design a road network traffic status predictor based on online estimation of model parameters. Finally, with the support of the long short-term memory network, the integrated deduction and prediction of the spatiotemporal dynamics of the traffic flow of the entire road network is realized. This process is as follows:

[0061] S2: Train the first LSTM network on the preprocessed historical main line traffic flow, and use the trained first LSTM network as the traffic prediction model for the detector point; train the second LSTM network on the preprocessed historical vehicle speed data, and use the trained second LSTM network as the speed prediction model for the detector point; train the third LSTM network on the preprocessed historical ramp flow data, and use the trained third LSTM network as the ramp flow prediction model for the detector point.

[0062] It should be noted that in step S2 of the present invention, by combining the traffic history status data (including main line traffic flow, vehicle speed data and ramp flow), a long short-term memory network (LSTM) including an input layer, an intermediate layer and an output layer is constructed, and the trained long short-term memory network is used as a prediction model for actual prediction tasks, realizing full connection and loop processing of data while realizing accurate prediction of the flow and vehicle speed at the limited detection points on the main line of the road network and the entrance and exit ramps. When training the long short-term memory network, in order to optimize the performance of the long short-term memory network, in this embodiment, the number of neural units in the layer, the number of network iterations, the discard function and other parameters can also be adjusted.

[0063] In step S2 of this embodiment, the implementation method of the long short-term memory network is specifically introduced to facilitate those skilled in the art to understand the present invention.

[0064] The long short-term memory network (LSTM) is the most representative type of recurrent neural network. Its core concept is a grid-based basic unit containing a gate structure. The gate structure mainly has three types: forget, input, and output. It automatically learns and decides which information to save or forget during training, realizing selective processing of information.

[0065] The LSTM functional unit involves the following calculation formula:

[0066] F t =σ(W f ·[H t-1 ,X t ]+b f )

[0067] I t =σ(W i ·[H t-1 ,X t ]+b i )

[0068]

[0069] O t =σ(W o ·[H t-1 ,X t ]+b o )

[0070] H t =O t *tanh(C t )

[0071] Where: F t represents the forget gate; σ represents the sigmoid activation function; W frepresents the weight matrix of the forget gate; H t-1 Indicates the hidden state of the previous moment; X t Represents the external input at the current moment; b f Represents the bias matrix of the forget gate; I t represents the input gate; W i represents the weight matrix of the input gate; b i represents the bias matrix of the input gate; represents the candidate memory state at the current moment; tanh represents the hyperbolic tangent activation function; W c The weight matrix representing the candidate memory state; b c The bias matrix representing the candidate memory state; C t Indicates the memory status at the current moment; C t-1 Indicates the memory status of the previous moment; O t represents the output gate; W o represents the weight matrix of the output gate; b o represents the bias matrix of the output gate; H t It represents the hidden state at the current moment and also the output at the current moment.

[0072] Among them, the forget gate is used to selectively forget the memory of the previous moment; the input gate is used to control how much new memory needs to be input at the current moment; the selection of candidate memory depends on the output of the previous moment and the input of the current moment; the output gate is used to control how much internal memory is output at the current moment.

[0073] LSTM has both short-term and long-term information memory. The so-called short-term memory means that the information extracted at each moment can be stored in certain network parameters, but will be overwritten and updated at the next moment. Long-term memory means that LSTM has the ability to store the key information captured at each moment in other network parameters for a longer time interval. The specific storage method is controlled by the forget gate and input gate. In contrast, traditional recurrent neural networks only have short-term memory.

[0074] S3: Use the preprocessed historical mainline traffic flow, preprocessed historical vehicle speed data, and preprocessed historical ramp flow data in S1 to calibrate the second-order macro traffic flow model METANET to obtain the global parameters with the best model-data fit and the best traffic basic diagram parameters.

[0075] It should be noted that, in the calibration process of the second-order macroscopic traffic flow model in step S3 of the present invention, for the existing road network, relying on the second-order macroscopic traffic flow model METANET, a model parameter calibration problem is constructed, and the downhill simplex method, genetic algorithm and other optimization algorithms are used to solve and obtain the optimal global parameters and traffic basic graph parameters.

[0076] In step S3 of the present embodiment, the global parameters are relaxation time parameters, steady-state expectation parameters, stability parameters, confluence parameters and lane number reduction parameters, and the basic traffic map parameters are free flow speed, critical density and capacity. After the second-order macroscopic traffic flow model METANET is calibrated in S3, the global parameters and the basic traffic map parameters are fixed. When this second-order macroscopic traffic flow model is used later, the optimal global parameters are directly used. Unlike the global parameters, the basic traffic map parameters need to be continuously updated according to the main line traffic flow, vehicle speed data and ramp flow data at the latest moment. In the prediction process, the basic traffic map parameters at each moment are different, so the basic traffic map parameters can be regarded as unknown parameters.

[0077] Specifically, the present invention uses the second-order macro traffic flow model METANET to describe the dynamics of highway traffic flow. In space, a long highway is divided into N sections with a length of Δ i In terms of time, the total time domain under investigation is divided into a series of discrete time periods k=0,1,… of length T.

[0078] For the i-th section, the equation of the second-order macroscopic traffic flow model is as follows:

[0079] Conservation equations:

[0080]

[0081] The flow diversion equation for the exit ramp is:

[0082] s i (k) = β i (k) q i-1 (k)

[0083] Transport equation:

[0084]

[0085] Dynamic velocity equation:

[0086]

[0087] Steady-state velocity-density relationship equation:

[0088]

[0089] Basic traffic graph equation:

[0090] Q(ρ)=ρ·V(ρ)

[0091] Where: i (k+1),ρ i(k) represents the density of the i-th segment in the k+1th and kth discrete time periods respectively; T represents the length of the discrete time period; Δ i Indicates the segment length; λ i represents the number of lanes in the i-th section; q i-1 (k), q i (k) represents the flow of the i-1th and i-th sections in the kth discrete time period respectively; r i (k) represents the incoming flow of the i-th section in the k-th discrete time period; s i (k) represents the outgoing flow of the i-th section in the k-th discrete time period; β i (k) represents the turn-out rate of the i-th section in the k-th discrete time period; v i (k), v i (k+1) represents the speed of the i-th segment in the k-th and k+1-th discrete time periods respectively; represents the traffic noise of the i-th segment in the k-th discrete period; V[ρ i (k)] represents the velocity calculated based on the density of the i-th segment at the k-th discrete time period; v i-1 (k) represents the speed of the i-1th segment in the kth discrete time period; ρ i+1 (k) represents the density of the i+1th segment in the kth discrete time period; Δλ represents the change in the number of lanes (the difference between the number of lanes in the current segment and the number of lanes in the next segment); represents the velocity noise of the i-th segment in the k-th discrete time period; V(ρ) represents the velocity calculated based on density; Q(ρ) represents the flow calculated based on density; v f represents the free flow velocity, ρ cr represents the critical density; a represents the amplitude of change of density with respect to speed; ρ represents the density of any discrete period in any segment; τ represents the relaxation time parameter, υ represents the steady-state expectation parameter, k represents the stability parameter, δ represents the confluence parameter, and φ represents the lane number reduction parameter.

[0092] The equations of the above second-order macroscopic traffic flow model can also be written in a compact state equation form:

[0093] x(k+1)=f[x(k),d(k),p,ξ(k)]

[0094] Where x(k+1) and x(k) represent the state vectors of the k+1th and kth discrete time periods, respectively, which contain the average speed and density of each section; f represents the second-order macroscopic traffic flow model; d(k) represents the disturbance vector of the kth discrete time period, which is used to characterize the boundary conditions such as the main line entrance flow and speed, the incoming flow, and the outgoing rate; ξ(k) represents the system noise vector of the kth discrete time period; p is the model parameter vector, which is composed of the global parameters and the traffic basic map parameters. The specific form is as follows:

[0095] p=[τ,v,κ,δ,φ,v f ,ρ cr ,q cap ]

[0096] In the formula, q cap Indicates traffic capacity.

[0097] It should also be noted that the above global parameters take the same value for all segments.

[0098] S4: For the current moment, the existing main line traffic flow time series of the detector point is input into the traffic prediction model of step S2, and the dynamic evolution of the main line traffic flow at the detector point in the future time period is predicted to obtain the main line traffic flow prediction result for the future time period. The existing vehicle speed time series of the detector point is input into the speed prediction model of step S2, and the dynamic evolution of the vehicle speed at the detector point in the future time period is predicted to obtain the vehicle speed prediction result for the future time period. The existing ramp flow data at the detector point is input into the ramp flow prediction model of step S2, and the dynamic evolution of the ramp flow at the detector point in the future time period is predicted to obtain the ramp flow prediction result for the future time period.

[0099] It should be noted that in step S4 of the present invention, the input of the long short-term memory network is the traffic status data at the current moment, which includes the main line traffic flow and main line vehicle speed data at different moments at the locations of each gantry detector on the main line of the expressway, and also includes ramp flow data and weather information at different moments, and the output is the main line traffic flow prediction value at different locations at different future moments within the future prediction step, and the main line vehicle speed prediction value and ramp flow prediction value at different locations at different future moments within the future prediction step can also be output. Among them, the input traffic history status data can be calculated according to the method in the above S1 through the latest data recorded by the gantry detectors on the main line of the expressway and at the ramp of the toll station.

[0100] S5: At the current moment, the traffic flow prediction results, vehicle speed prediction results and ramp flow prediction results of the future time period are used as traffic state prediction values. The traffic state prediction values ​​of all detector points are used to drive the road network traffic state predictor based on the second-order macro traffic flow model METANET and the extended Kalman filter EKF to obtain the traffic state prediction result of the entire road network corresponding to the current moment.

[0101] It should be noted that in step S5 of the present invention, the road network traffic state predictor used has an adaptive calibration function for the basic map parameters of the traffic flow, which can ensure the accuracy and robustness of the estimation and prediction performance. In constructing the road network traffic state predictor, the present invention uses an extended Kalman filter. The extended Kalman filter is an effective means of establishing a logical association between traffic states between detection points and non-detection points through a model. Considering that in the compact state equation of the second-order macroscopic traffic flow model, the basic traffic map parameters are unknown, and some elements in the disturbance vector d of each discrete time period are undetectable (such as some entrance and exit ramps are not equipped with detectors), these unknown parts are also regarded as system states, and the compact state equation of the second-order macroscopic traffic flow model can be rewritten in the following form:

[0102] x1(k+1)=f1[x1(k),ξ1(k)]

[0103] Where x1(k+1) represents the state vector of the k+1th discrete period; ξ1(k) represents the system noise vector of the kth discrete period; f1 is a nonlinear vector function; represents the state vector; the subscript tr represents the matrix transpose; x represents the state vector in the compact state equation of the second-order macroscopic traffic flow model; d1 represents the undetected boundary condition; p1 represents the unknown traffic basic diagram parameter.

[0104] In addition, the traffic state prediction values ​​of all detector points are also written in the following state equation form:

[0105]

[0106] Where y(k) represents the traffic detection vector of the kth discrete period; represents the estimated value of the state vector in the kth discrete period; η(k) represents the detection noise in the kth discrete period; g is a nonlinear vector function.

[0107] Based on the above two equations, the road network traffic state predictor can be designed based on the extended Kalman filter, and its analytical expression is as follows:

[0108]

[0109] In the formula, represents the one-step forward estimation of the system state x(k) based on the traffic detection vector y(k) of the kth discrete period; K(k) represents the Kalman gain matrix of the kth discrete period.

[0110] Specifically, the analytical expression of the road network traffic state predictor contains two items, state estimation based on the traffic flow model and correction based on real-time detection data. Since the calculation process of the above analytical expression is recursive, the Kalman gain matrix K(k) of the kth discrete period is related to the observation data of the k-1th, k-2th, ... discrete periods in history. Given the traffic detection vector y(k) of the kth discrete period, the above analytical expression realizes state estimation by minimizing the covariance of the estimation error:

[0111]

[0112] In the formula, E represents the expectation.

[0113] 3. Model-data dual-driven traffic status prediction for the entire road network

[0114] S6: After obtaining the traffic status prediction result of the entire road network at the current moment, re-execute steps S4 and S5 for the next moment to obtain a new traffic status prediction result of the entire road network, and complete the highway traffic status prediction.

[0115] It should be noted that in step S6 of the present invention, an innovative model-data dual-driven method is adopted to give full play to the advantages of machine learning methods in fixed-point data-driven state prediction, and at the same time, with the help of the large-scale spatiotemporal dynamic characterization capabilities of the traffic flow model, an effort is made to achieve accurate prediction of traffic conditions at non-detection points.

[0116] The present invention will now use a specific example to demonstrate the application effect of the model-data dual-driven highway traffic state prediction method described in S1 to S6 in the above embodiments on a specific data set, so as to facilitate understanding of the essence of the present invention.

[0117] Example

[0118] The specific implementation process of the model-data dual-driven highway traffic status prediction method used in this embodiment is as described above and will not be repeated here.

[0119] This embodiment further Figure 3 The above method is explained and verified by taking the road network shown in the figure as an example. The road network is 9.06 kilometers long, including 6 entrance ramps and 7 exit ramps. There are 17 detectors deployed along the whole road network, of which the blue detectors are used as input detectors to provide driving data for traffic state prediction, and the gray detectors are used to evaluate the state prediction results.

[0120] The data set used in this embodiment includes traffic flow and speed data of the mainline detector, as well as traffic data of the entrance and exit ramps. The time granularity of these data is 5 minutes, and the date span is from February 18, 2022 to March 4, 2022, a total of 15 days. Since there are a lot of missing data in the early morning time period of the three original data of traffic flow, speed data, and flow data, in order to facilitate testing and research, this embodiment only selects 6:00 to 23:55 every day as the available time period for analysis.

[0121] After data preprocessing, this embodiment obtains 6480 valid data, which are used to construct the training set, validation set and test set required for the prediction task. The training set is used for model training and parameter tuning, the validation set is used for hyperparameter selection, and the test set is used for model performance evaluation. This embodiment selects the last day of data from the above data set (8:00-23:55 on March 4) as the test set, and the other data are used as training sets and validation sets.

[0122] In order to verify the effectiveness of the method of the present invention, this example compares the following three prediction methods:

[0123] G1: The traditional ARIMA method is used to predict the road boundary conditions (in-turn flow, out-turn rate) and the flow and speed at the mainline detector, and the linear interpolation method is used to obtain the traffic state estimation of the points without detectors in the road network;

[0124] G2: Use LSTM to predict road boundary conditions (in-turn flow, out-turn rate) and the flow and speed at the mainline detector, and use linear interpolation to obtain traffic state estimates at points without detectors in the road network;

[0125] G3: Use LSTM to predict road boundary conditions (in-turn flow, out-turn rate) and the flow and speed at the mainline detector, and use this to drive the road network traffic status predictor based on METANET and EKF to predict the traffic status of any detector-free point in the road network, which is the method of the present invention.

[0126] Specifically, G1's ARIMA is a classic statistical learning method based on time series, which can analyze the periodic trend of time series data; G2's LSTM is a special recurrent neural network that can effectively transmit and express the long-term correlation of time series and learn the patterns and features in time series data more accurately; G3 uses machine learning methods and macro traffic flow physical models together. In contrast, G1-G2 rely on simple linear interpolation to achieve traffic status prediction at non-detection points. So far, no related work has been reported on G3. The driving data of G1 and G2 comes from Figure 3 The input detectors shown in Figure 2, G3 additionally uses the flow data of the entrance and exit ramps. The evaluation detectors of G1-G3 are also shown in Figure 2. Figure 3 .

[0127] Using the above training set and validation set, the LSTM hyperparameters are selected through grid search as follows: the number of network layers is 2, the feature dimension of the hidden layer is 128, the random inactivation rate (dropout) is 0.3, the learning rate (learning rate) is 0.001, the optimizer is AdamW, the weight decay coefficient (weight decay) is 0.001, the number of training times is 600, the input sequence step length (sequence length) is 12 (that is, the backtracking time of historical data is 60 minutes), and the prediction step lengths are 1, 2, 3, 4, and 6 respectively (that is, the prediction time is 5 minutes, 10 minutes, 15 minutes, 20 minutes, and 30 minutes respectively).

[0128] The traffic flow data on March 3 were used to calibrate the parameters of the second-order macro traffic flow model METANET, and the results are shown in Table 1. Further research shows that these METANET global parameters are not very sensitive to data from other dates.

[0129] Table 1. METANET model global parameters

[0130]

[0131] The experimental result evaluation indicators of this embodiment are mean absolute percentage error (MAPE) and root mean square error (RMSE), and their expressions are as follows:

[0132]

[0133]

[0134] in, represents the predicted value of the model at time t, y t represents the actual observation value at time t, and m represents the number of time series points to be predicted. For a given prediction task (i.e., the same m and y t ,t=1,…,m). RMSE and MAPE represent the absolute and relative accuracy of prediction respectively. The evaluation calculation based on the above experimental results only involves Figure 3 8 evaluation detectors in .

[0135] The global prediction performance results of the three methods G1–G3 are shown in Table 2, where the prediction time 0min refers to the estimation of the current traffic status of all non-detection points based on the detection data at the current time. The traffic flow prediction results of the whole road network of method G3 are shown in Table 2. Figure 4 The speed prediction results of the entire road network are shown in Figure 5 shown.

[0136] Table 2. Performance evaluation of traffic state prediction methods

[0137]

[0138]

[0139] As can be seen from Table 2, in terms of traffic flow prediction, the prediction errors of G1 and G2 are significantly higher than those of the method of the present invention, indicating that it is not appropriate to rely solely on spatial linear interpolation to obtain the traffic flow at non-detection points without considering the traffic flow at the entrance and exit ramps. Since there are always entrance and exit ramps in the highway network, the traffic flow in the main line section is not only affected by the traffic flow in the upstream and downstream sections, but also by the outflow and inflow of the adjacent ramps. Due to the introduction of the flow conservation equation, the prediction performance of the method G3 of the present invention is significantly improved. In terms of speed prediction, both the ARIMA-based G1 and the LSTM-based G2 rely on simple linear interpolation to achieve traffic state prediction at non-detection points. The method G3 of the present invention uses a machine learning method and a second-order macro traffic flow model in combination, and uses the large spatiotemporal dynamic characterization capability of the second-order macro traffic flow model METANET and the sequential tracking and deduction capability of the extended Kalman filter EKF to accurately describe the evolution process of traffic flow and achieve accurate traffic state prediction at non-detection points in the highway network. Therefore, the speed prediction accuracy of the method G3 of the present invention is significantly ahead of G1–G2.

[0140] Similar to other macro traffic flow models, the second-order macro traffic flow model METANET describes the dynamic evolution of highway traffic flow through the flow conservation equation. Combined with machine learning methods, the road network traffic state predictor proposed in this invention can realize the global traffic state prediction of the highway network, providing a solid and powerful data foundation for congestion evolution, traffic control, etc.

[0141] In summary, if Figure 2As shown in the figure, the current methods for predicting the traffic state of highway networks are almost all purely data-driven and do not rely on physical models. Most of these methods focus on a limited number of detector points and attempt to simply extend the traffic state prediction of the detection points to the entire road network through spatial linear interpolation. Although the prediction performance of the detection points is acceptable, it will inevitably lead to a large global traffic state prediction error. Compared with the prior art, the present invention combines machine learning methods and second-order macro traffic flow models for the first time, not only giving play to the advantages of machine learning methods in data-driven state prediction, but also relying on the large spatiotemporal dynamic characterization ability of the second-order macro traffic flow model to strive to break through the traffic state prediction problem of non-detection points. First, the second-order macro traffic flow model is used to establish a logical association of traffic states between limited detection points and a large number of non-detection points; then, a road network traffic state predictor is designed based on the second-order macro traffic flow model and the extended Kalman filter, and the traffic state prediction of the entire road network is realized under the drive of the machine learning traffic state prediction of limited points. And further verified on a real road network data set of a megacity in China, it is proved that the method of the present invention can effectively improve the prediction accuracy of road network flow and speed. In addition, most current traffic prediction methods do not pay attention to the topological changes of the highway network. If this is taken into account and the traffic conservation equation is introduced, the accuracy of traffic prediction can be greatly improved.

[0142] It should also be noted that the model-data dual-driven highway traffic state prediction method in the above embodiment can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a model-data dual-driven highway traffic state prediction system corresponding to the model-data dual-driven highway traffic state prediction method provided in the above embodiment, such as Figure 6 As shown, it includes:

[0143] The data acquisition and processing module is used to obtain the road network topology and detector layout information, and pre-process the historical mainline traffic flow, historical vehicle speed data and historical ramp flow data of the acquired detector points, and obtain the pre-processed historical mainline traffic flow, pre-processed historical vehicle speed data and pre-processed historical ramp flow data;

[0144] A model acquisition module is used to train a first long short-term memory network on the preprocessed historical main line traffic flow, use the trained first long short-term memory network as the flow prediction model of the detector point, train a second long short-term memory network on the preprocessed historical vehicle speed data, use the trained second long short-term memory network as the speed prediction model of the detector point, train a third long short-term memory network on the preprocessed historical ramp flow data, and use the trained third long short-term memory network as the ramp flow prediction model of the detector point;

[0145] A model calibration module is used to calibrate the second-order macro traffic flow model using the preprocessed historical main line traffic flow, the preprocessed historical vehicle speed data, and the preprocessed historical ramp flow data to obtain the global parameters and traffic basic diagram parameters with the best model-data fit;

[0146] A prediction module is used to input the existing main line traffic flow time series of the detector point into the traffic flow prediction model for the current moment to obtain the main line traffic flow prediction result for the future period, input the existing vehicle speed time series of the detector point into the speed prediction model to obtain the vehicle speed prediction result for the future period, and input the existing ramp flow data of the detector point into the ramp flow prediction model for the step to obtain the ramp flow prediction result for the future period;

[0147] The result acquisition module uses the traffic flow prediction results, vehicle speed prediction results and ramp flow prediction results of the future time period as the traffic state prediction value at the current moment, and uses the traffic state prediction values ​​of all detector points to drive the road network traffic state predictor based on the second-order macro traffic flow model and the extended Kalman filter to obtain the traffic state prediction result of the entire road network corresponding to the current moment;

[0148] The result updating module is used to re-execute the prediction module and the result obtaining module for the next moment after obtaining the traffic status prediction result of the entire road network at the current moment, obtain the new traffic status prediction result of the entire road network, and complete the highway traffic status prediction.

[0149] It is understandable that the model-data dual-driven highway traffic state prediction method described in S1 to S6 above can be substantially implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the model-data dual-driven highway traffic state prediction method provided in the above embodiment, which includes a computer program / instruction. When the computer program / instruction is executed by a processor, the model-data dual-driven highway traffic state prediction method described in the above embodiment can be implemented.

[0150] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the model-data dual-driven highway traffic state prediction method provided in the above embodiment, which includes a memory and a processor;

[0151] The memory is used to store computer programs;

[0152] The processor is used to implement the model-data dual-driven highway traffic status prediction method in the above embodiment when executing the computer program.

[0153] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention.

[0154] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the model-data dual-driven highway traffic status prediction method provided in the above embodiment, and a computer program is stored on the storage medium. When the computer program is executed by the processor, the model-data dual-driven highway traffic status prediction method in the above embodiment can be implemented.

[0155] It is understandable that the above storage medium may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage. The storage medium may also be a U disk, a mobile hard disk, a magnetic disk or an optical disk, etc., which can store program codes.

[0156] It is understandable that the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0157] It should also be noted that those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the various embodiments provided in this application, the division of steps or modules in the system and method is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0158] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A model-data dual-driven highway traffic state prediction method, characterized in that: The following steps are involved: S1: Obtain the road network topology and detector layout information, and pre-process the historical mainline traffic flow, historical vehicle speed data, and historical ramp flow data of the acquired detector points, and obtain the pre-processed historical mainline traffic flow, pre-processed historical vehicle speed data, and pre-processed historical ramp flow data; S2: training a first long short-term memory network on the preprocessed historical main line traffic flow, using the trained first long short-term memory network as the flow prediction model of the detector point, training a second long short-term memory network on the preprocessed historical vehicle speed data, using the trained second long short-term memory network as the speed prediction model of the detector point, training a third long short-term memory network on the preprocessed historical ramp flow data, using the trained third long short-term memory network as the ramp flow prediction model of the detector point; S3: Use the preprocessed historical mainline traffic flow, preprocessed historical vehicle speed data, and preprocessed historical ramp flow data in S1 to calibrate the second-order macro traffic flow model, and obtain the global parameters and traffic basic diagram parameters with the best model-data fit; S4: for the current moment, input the existing main line traffic flow time series of the detector point into the traffic flow prediction model to obtain the main line traffic flow prediction result for the future period, input the existing vehicle speed time series of the detector point into the speed prediction model to obtain the vehicle speed prediction result for the future period, input the existing ramp flow data of the detector point into the ramp flow prediction model in step 1 to obtain the ramp flow prediction result for the future period; S5: At the current moment, the traffic flow prediction results, vehicle speed prediction results and ramp flow prediction results of the future time period are used as traffic state prediction values, and the traffic state prediction values ​​of all detector points are used to drive the road network traffic state predictor based on the second-order macro traffic flow model and the extended Kalman filter to obtain the traffic state prediction result of the entire road network corresponding to the current moment; S6: After obtaining the traffic status prediction result of the entire road network at the current moment, re-execute steps S4 and S5 for the next moment to obtain a new traffic status prediction result of the entire road network, and complete the highway traffic status prediction.

2. A model-data dual-driven highway traffic state prediction method as claimed in claim 1, characterized in that: In step S3, the compact state equation form of the second-order macroscopic traffic flow model is as follows: x(k+1)=f[x(k),d(k),p,ξ(k)] where x(k+1) and x(k) represent the state vectors of the k+1th and kth discrete time periods, respectively; f represents the second-order macroscopic traffic flow model; d(k) represents the disturbance vector of the kth discrete time period; ξ(k) represents the system noise vector of the kth discrete time period; p is the model parameter vector, which consists of the global parameters τ, v, κ, δ, φ and the traffic basic graph parameters v f ,ρ cr ,q cap The specific forms are as follows: p=[τ,v,κ,δ,φ,v f ,r cr ,q cap ] Where τ represents the relaxation time parameter; v represents the steady-state expectation parameter; κ represents the stability parameter; δ represents the confluence parameter; φ represents the lane number reduction parameter; v f represents the free flow velocity, ρ cr represents the critical density; q cap Indicates traffic capacity.

3. A model-data dual-driven highway traffic state prediction method as claimed in claim 2, characterized in that: In step S3, the global parameters are relaxation time parameter, steady-state expectation parameter, stability parameter, confluence parameter and lane number reduction parameter.

4. A model-data dual-driven highway traffic state prediction method as claimed in claim 2, characterized in that: In step S3, the basic traffic map parameters are free flow speed, critical density and traffic capacity.

5. The model-data dual-driven highway traffic state prediction method according to claim 1, characterized in that: In step S5, the analytical expression of the road network traffic state predictor is as follows: In the formula, represents the one-step forward estimation of the system state x(k) based on the traffic detection vector y(k) of the kth discrete period; f1, g are both nonlinear vector functions; K(k) represents the Kalman gain matrix of the kth discrete period.

6. The model-data dual-driven highway traffic state prediction method according to claim 1, characterized in that: In step S2, when the first long short-term memory network, the second long short-term memory network, and the third long short-term memory network are trained, the number of network layers is 2, the feature dimension of the hidden layer is 128, the random inactivation rate is 0.3, the learning rate is 0.001, the optimizer is AdamW, the weight decay coefficient is 0.001, and the number of training times is 600.

7. A model-data dual-driven highway traffic status prediction system, characterized in that: include: The data acquisition and processing module is used to obtain the road network topology and detector layout information, and pre-process the historical mainline traffic flow, historical vehicle speed data and historical ramp flow data of the acquired detector points, and obtain the pre-processed historical mainline traffic flow, pre-processed historical vehicle speed data and pre-processed historical ramp flow data; A model acquisition module is used to train a first long short-term memory network on the preprocessed historical main line traffic flow, use the trained first long short-term memory network as the flow prediction model of the detector point, train a second long short-term memory network on the preprocessed historical vehicle speed data, use the trained second long short-term memory network as the speed prediction model of the detector point, train a third long short-term memory network on the preprocessed historical ramp flow data, and use the trained third long short-term memory network as the ramp flow prediction model of the detector point; A model calibration module is used to calibrate the second-order macro traffic flow model using the preprocessed historical main line traffic flow, the preprocessed historical vehicle speed data, and the preprocessed historical ramp flow data to obtain the global parameters and traffic basic diagram parameters with the best model-data fit; A prediction module is used to input the existing main line traffic flow time series of the detector point into the traffic flow prediction model for the current moment to obtain the main line traffic flow prediction result for the future period, input the existing vehicle speed time series of the detector point into the speed prediction model to obtain the vehicle speed prediction result for the future period, and input the existing ramp flow data of the detector point into the ramp flow prediction model for the step to obtain the ramp flow prediction result for the future period; The result acquisition module uses the traffic flow prediction results, vehicle speed prediction results and ramp flow prediction results of the future time period as the traffic state prediction value at the current moment, and uses the traffic state prediction values ​​of all detector points to drive the road network traffic state predictor based on the second-order macro traffic flow model and the extended Kalman filter to obtain the traffic state prediction result of the entire road network corresponding to the current moment; The result updating module is used to re-execute the prediction module and the result obtaining module for the next moment after obtaining the traffic status prediction result of the entire road network at the current moment, obtain the new traffic status prediction result of the entire road network, and complete the highway traffic status prediction.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, it can implement the model-data dual-driven highway traffic status prediction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the model-data dual-driven highway traffic state prediction method as described in any one of claims 1 to 6 is implemented.

10. A computer electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to implement the model-data dual-driven highway traffic status prediction method as described in any one of claims 1 to 6 when executing the computer program.

Images