Traffic flow parameter hybrid prediction method and device

Abstract

The present application relates to a kind of traffic flow parameter hybrid prediction method and device, belong to intelligent transportation technical field, solve the problem that model calibration in existing prediction method is long in time, cannot real-time reaction road sporadic situation.The method includes: selecting to be predicted section AB and obtaining the real-time traffic flow data of grid point A of to be predicted section and the historical traffic flow data of grid point B;GKT macroscopic traffic flow model is established, and the real-time traffic flow data of A is input into the model to predict the future traffic flow data of B as macroscopic traffic flow model prediction result U1;LSTM deep learning model is established and trained, and the future traffic flow data of B is predicted using model as deep learning model prediction result U2;Linear Kalman filter is used to fuse macroscopic traffic flow model prediction result U1 and deep learning model prediction result U2 as the final prediction result U3 of the future traffic flow parameter of B.The advantage of macroscopic and deep learning model is concentrated and complementary.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method and apparatus for mixed prediction of traffic flow parameters. Background Technology

[0002] Road traffic networks possess spatiotemporal, self-organizing, and stochastic characteristics. Analyzing the dynamic evolution patterns and traffic operation trends of road traffic networks helps to comprehensively understand their complex characteristics and effectively grasp the formation mechanisms of traffic bottlenecks and accidents. Constructing a model of traffic dynamic evolution patterns under hybrid traffic networks enables rapid identification of traffic behavior characteristics, providing technical support for further prediction and assessment of road traffic network security risks.

[0003] Traffic flow (simulation) models are used to characterize the features of complex traffic flow systems in order to understand, describe, and predict traffic flow. It is a fundamental tool for analyzing and experimentally studying traffic systems. Traffic flow models are no longer limited to traditional areas such as traffic system design, testing, management, and personnel training. With the surge in research on intelligent vehicles and intelligent transportation systems, they are also used to evaluate and predict the state of traffic systems.

[0004] Macroscopic traffic flow models use a mathematical model to describe traffic dynamics. This model assesses unmeasured areas based on real-time input data. This model is typically based on empirical relationships and is essentially a partial differential equation derived from vehicle conservation laws, describing density evolution based on flow gradients. Macroscopic traffic flow models have been widely used in traffic state assessment due to their advantages: First, this method explains traffic mechanisms, expands on observational data, and provides additional information. Therefore, it can predict accurate traffic conditions with less data. Second, it has higher interpretability. This means that even if the prediction is inaccurate, the cause can be found within a certain confidence interval. Third, it can be directly integrated with traffic control practices, such as using model predictive control. However, macroscopic traffic flow models also have disadvantages: inaccurate or uncalibrated models can lead to poor traffic state assessment results. Therefore, in practical applications, macroscopic traffic flow model traffic state assessment methods must be carefully selected and calibrated. In such cases, checking the effectiveness of a model or calibrating a model requires a large amount of data.

[0005] Traffic parameter prediction methods based on big data and machine learning have attracted considerable research interest in recent years. Big data analytics can help users reach their destinations via the most suitable routes and in the shortest time, while simultaneously improving their safety. Traffic management departments can predict traffic flow by rapidly collecting and analyzing massive amounts of current and historical traffic data. Big data analytics can effectively predict traffic accidents. Big data analytics primarily addresses three issues: data storage, data analysis, and data management. Machine learning is the most popular modeling and analysis theory in the big data ecosystem because it makes it easy to obtain models from large amounts of data. Machine learning models can be categorized into supervised learning, unsupervised learning, reinforcement learning, deep learning, and entity-based algorithms. Labeled training data is used for supervised learning algorithms. Among all supervised learning models, linear regression, decision trees, neural networks, and support vector machines are the most commonly used.

[0006] Data-driven traffic flow models can be categorized into those based on historical data and those based on real-time big data. Traffic state assessment methods that heavily rely on historical data utilize statistical or machine learning methods to identify relationships between historical data and then assess traffic conditions based on these relationships and real-time data. This means it doesn't require the explicit prior knowledge modeled in macro-level traffic flow models. This method typically requires a large amount of historical data. Historical data-based traffic flow models have the following advantages: less time is spent on model selection and calibration. Disadvantages include: First, being based on historical data means the model may fail when unexpected events occur or when predicting relatively long trends. Second, the computational cost for training and learning is very high. Third, this method can be considered a "black box," meaning theoretical derivation is not possible. Summary of the Invention

[0007] Based on the above analysis, the embodiments of the present invention aim to provide a method and apparatus for mixed prediction of traffic flow parameters, in order to solve the problems of long model calibration time and inability to reflect occasional road conditions in real time in traditional single traffic prediction methods.

[0008] On one hand, embodiments of the present invention provide a method for hybrid prediction of traffic flow parameters, comprising: selecting a road segment AB to be predicted and acquiring real-time traffic flow data at grid point A and historical traffic flow data at grid point B of the road segment to be predicted; establishing a GKT macro-traffic flow model and inputting the real-time traffic flow data at grid point A into the GKT macro-traffic flow model to predict the future traffic flow data at grid point B as the macro-traffic flow model prediction result U1; establishing and training an LSTM deep learning model, and using the trained LSTM deep learning model to predict the future traffic flow data at grid point B as the deep learning model prediction result U2; and using a linear Kalman filter to fuse the macro-traffic flow model prediction result U1 and the deep learning model prediction result U2 into a final prediction result U3 of the future traffic flow parameters of grid point B.

[0009] The beneficial effects of the above technical solution are as follows: By fusing the prediction results of the macro model and the deep learning model, the advantages of the macro model and the deep learning model are fully combined. At the same time, the two models can complement each other and effectively improve the shortcomings of their individual predictions. This is of great significance for improving prediction accuracy, robustness and real-time performance.

[0010] Further improvements to the above method include selecting the road segment AB to be predicted and obtaining real-time traffic flow data at grid point A and historical traffic flow data at grid point B of the road segment to be predicted, which involves: selecting the road segment AB to be predicted, provided that there are no entrance or exit ramps in the middle of the road segment AB; uniformly dividing the road segment AB to be predicted into i grids using i+1 grid points, wherein the i+1 grid points include grid point A and grid point B; and collecting real-time traffic flow data at grid point A and obtaining historical traffic flow data at grid point B, wherein the real-time traffic flow data includes density and speed.

[0011] Based on further improvements to the above method, a GKT macro-traffic flow model is established, and the real-time traffic flow data at grid point A is input into the GKT macro-traffic flow model to predict the future traffic flow data at grid point B as the macro-traffic flow model prediction result U1. This includes: the grid length x is L / i, and the grid length is used as the spatial step, where L is the length of the road segment AB to be predicted; the time step is determined according to the CFL condition, where the time step is not greater than x / V0, and V0 is the maximum expected speed; the density and speed at grid point A are input into the GKT macro-traffic flow model; the grid points are reconstructed spatially using the 5th-order WENO format, and advanced temporally using the Runge-Kutta format to predict the future traffic flow data at grid point B as the macro-traffic flow model prediction result U1.

[0012] Based on the further improvement of the above method, before reconstructing the grid points in space using the 5th-order WENO scheme, the conservation expression of the GKT macroscopic traffic flow model is as follows:

[0013]

[0014] Based on the conservation expression, we obtain: variable U = [ρ, ρu] T Flux f(U) = [ρu, ρu 2 +θρ] T , source item The conservation expression is then transformed into:

[0015]

[0016] According to the chain rule, we can obtain: The Jacobian matrix A is diagonalized as follows: Where Λ is a diagonal matrix composed of eigenvalues, and the following variables are set: V = R -1 U; then By decoupling the original model equations through a series of changes, a single-wave equation is obtained. In the single-wave equation, each variable is independent of the others, so that the error is controlled within a relatively reasonable range.

[0017] Based on further improvements to the above method, a 5th-order WENO format is used to spatially reconstruct the grid points, and the Runge-Kutta format is used to advance the prediction in time to obtain the future traffic flow data at grid point B. The prediction result U1 of the macro-traffic flow model includes: finding the estimated value. calculate Decoupling into single-wave equations results in smaller oscillations in the GKT macroscopic traffic flow model, i.e. Using V i+b The estimated values ​​obtained by performing WENO reconstruction are as follows: The optimal weights are selected as follows: Smoothing factor IS corresponding to positive flux + for:

[0018]

[0019] Smoothing factor IS corresponding to negative flux - for:

[0020]

[0021] The weighting coefficients are calculated as follows:

[0022] The positive flux at position i+1 / 2 is:

[0023]

[0024] The negative flux at position i+1 / 2 is:

[0025]

[0026] That is, the result

[0027]

[0028] Obtain the accurate estimate U i+1 / 2 =R i+1 / 2 V i+1 / 2

[0029] Similarly, we can obtain V i-1 / 2 Then, the third-order Runge-Kutta scheme is used for time-advanced solution, as follows:

[0030] according to get:

[0031]

[0032] This equation is solved using the third-order Runge-Kutta method:

[0033] V (1) =V n +ΔtL(V n )

[0034]

[0035]

[0036] The value of variable V is obtained by advancing the time, and the prediction result U1 of the GKT macroscopic traffic flow model is obtained based on the value of variable V.

[0037] The grid points are reconstructed spatially using a 5th-order WENO format, and then advanced temporally using a Runge-Kutta format to predict the future traffic flow data at grid point B. The macro-traffic flow model prediction result U1 includes: finding the estimated value. calculate Decoupling into single-wave equations results in smaller oscillations in the GKT macroscopic traffic flow model, i.e. Using V i+b Reconstruct WENO to obtain V i+1 / 2 The estimated value, where WENO reconstruction is to make V i+1 / 2 The estimated value was determined for all templates (V) i-2 V i-1 V i V i+1 Vi+2 Reconstructed value (q) on ) j A linear (weight ω) j ) combination (∑ω j q j ), but when V i+1 / 2 When the surface is not smooth, using linear weights alone is insufficient; a smoothing factor (IS) must be introduced. j Improved weighting coefficient ω j :

[0038] Optimal weights are selected respectively: To determine the weight coefficients, i.e., the weights ω j Introducing a smoothing factor, corresponding to the three positive fluxes, namely, the smoothing factor IS corresponding to the positive flux. + for:

[0039]

[0040] Smoothing factor IS corresponding to negative flux - for:

[0041]

[0042] The smoothing factor IS for the above positive and negative fluxes + IS - The calculation is to obtain α j Thus, the weighting coefficient ω is calculated. j Here, ε represents a very small number, and the weighting coefficients are calculated as follows:

[0043] The positive flux (where "positive" refers to upstream traffic flow information, specifically the left-hand value in this application) at spatial grid point i+1 / 2 (i.e., the midpoint between grid point i and grid point i+1) is:

[0044]

[0045] The negative flux (where "negative" refers to the downstream information volume of the traffic flow) at spatial grid point i+1 / 2 (i.e., the midpoint between grid point i and grid point i+1) is:

[0046]

[0047] In the WENO reconstruction, the accurate estimate at point i+1 / 2 is obtained through the above positive and negative fluxes, i.e., the result at the spatial grid point i+1 / 2 is... Where i refers to grid point i of road segment, and j refers to the optimal weight C. j Smoothing factor IS j Flux q j The subscript (j = 0, 1, 2).

[0048] Obtain the accurate estimate U i+1 / 2 =R i+1 / 2 V i+1 / 2 ,

[0049] Similarly, we can obtain V i-1 / 2 Then, a time-advanced solution is obtained using the third-order Runge-Kutta scheme, as follows: According to get:

[0050]

[0051] This equation is solved using the third-order Runge-Kutta method:

[0052] V (1) =u n +ΔtL(u n )

[0053]

[0054]

[0055] The value of variable V is obtained by advancing the time, and the prediction result U1 of the GKT macro traffic flow model (i.e., variable U in the above formula) is obtained based on the value of variable V.

[0056] Further improvements to the above method involve establishing and training an LSTM deep learning model, and using the trained LSTM deep learning model to predict future traffic flow data at grid point B as the deep learning model prediction result U2. This includes: decomposing the input data into trend components, periodic components, and residual components based on local weighted regression, wherein the input data includes date, time, location, weather, accident, and traffic control data; using a composite deep learning network model composed of a one-dimensional convolutional neural network and a bidirectional long short-term memory network to predict the trend components, periodic components, and residual components to obtain trend component prediction results, periodic component prediction results, and residual component prediction results; and using a multilayer perceptron neural network to fuse the trend component prediction results, periodic component prediction results, and residual component prediction results into future traffic flow data at grid point B for prediction as the deep learning model prediction result U2, wherein the deep learning model prediction result U2 includes traffic flow speed, vehicle flow density, and vehicle volume.

[0057] Further improvements to the above method include fusing the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 into a final prediction result U3 of the future traffic flow parameters for grid point B using a linear Kalman filter. This involves: using the macroscopic traffic flow model prediction result U1 as the test input of the linear Kalman filter; using the deep learning model prediction result U2 as the prediction input of the linear Kalman filter; and automatically adjusting the weights of the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 so that the final prediction result U3 is close to the actual traffic flow parameters.

[0058] The beneficial effects of the above technical solution are as follows: The hybrid model that integrates the macro model and the deep learning model uses the macro model prediction result as the system state quantity and the deep learning model prediction result as the system detection quantity through the Kalman filter, thereby obtaining the final traffic information prediction result.

[0059] Based on a further improvement of the above method, the final prediction result U3 of the future traffic flow parameters of grid point B, which is obtained by fusing the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 using a linear Kalman filter, includes: the test input U1 of the linear Kalman filter is expressed by the following formula:

[0060] U1(n|n-1)=AU1(n-1|n-1)+Bu n +w n ,

[0061] The error covariance is: P(n|n-1)=AP(n-1|n-1)A T +Q,

[0062] The predictor input U2 of the linear Kalman filter is expressed by the following formula:

[0063] U2(n) = HU1(n) + v(n),

[0064] The Kalman gain is calculated using the following formula:

[0065]

[0066] The optimal estimate U3 is calculated using the following formula:

[0067] U3 = U1 + K g (n)(U2-HU1),

[0068] Update the error covariance using the following formula:

[0069] P(n|n)=[1-K g (n)H]P(n|n-1),

[0070] Among them, u n Let w be the state control variable at time n, A be the state transition matrix, B be the control matrix, and w be the control variable. n Let be the system process noise at time n, Q be the system process covariance, H be the measurement system parameters, v(n) be the measurement noise at time n, and R be the measurement noise covariance.

[0071] Based on further improvements to the above method, both the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 are vectors that change over time.

[0072] On the other hand, an embodiment of the present invention provides a traffic flow parameter hybrid prediction device, comprising: a grid point acquisition module, used to select a road segment AB to be predicted and acquire real-time traffic flow data at grid point A and historical traffic flow data at grid point B of the road segment to be predicted; a GKT macro traffic flow model, used to establish a GKT macro traffic flow model and input the real-time traffic flow data at grid point A into the GKT macro traffic flow model to predict the future traffic flow data at grid point B as the macro traffic flow model prediction result U1; an LSTM deep learning model, used to establish and train an LSTM deep learning model, and use the trained LSTM deep learning model to predict the future traffic flow data at grid point B as the deep learning model prediction result U2; and a fusion module, used to fuse the macro traffic flow model prediction result U1 and the deep learning model prediction result U2 using a linear Kalman filter to obtain the final prediction result U3 of the future traffic flow parameters of grid point B.

[0073] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0074] 1. By fusing the prediction results of macroscopic models and deep learning models, the advantages of both models are fully combined. At the same time, the two models can complement each other and effectively improve the shortcomings of their individual predictions. This is of great significance for improving prediction accuracy, robustness and real-time performance.

[0075] 2. This hybrid model can obtain more accurate traffic prediction results with relatively less data, based on an understanding of real-time traffic flow characteristics. The proposed technical solution aligns with the concept of intelligent transportation, is conducive to promoting the construction of an improved intelligent transportation system, and provides new ideas and solutions for urban brains.

[0076] 3. The hybrid model, integrating macroscopic and deep learning models, uses a Kalman filter to combine the macroscopic model's predictions as system state variables and the deep learning model's predictions as system detection variables, thus obtaining the final traffic information prediction result. Specifically, the Kalman filter has two inputs: a prediction variable and a test variable. U2 is the prediction result from the deep learning model, serving as the Kalman filter's prediction input. U1 is the test variable input from the GKT model. The Kalman filter's role is to balance the prediction results of these two models, automatically adjusting their weights to make the output result closer to the actual traffic parameters.

[0077] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0078] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0079] Figure 1 A flowchart of a traffic flow parameter hybrid prediction method according to an embodiment of the present invention;

[0080] Figure 2 This is an overall flowchart of the traffic flow parameter hybrid prediction method according to an embodiment of the present invention;

[0081] Figure 3 This is a schematic diagram of the road segment AB to be predicted and the grid division according to an embodiment of the present invention;

[0082] Figure 4 This is a structural diagram of the GKT macroscopic traffic flow model according to an embodiment of the present invention;

[0083] Figure 5 This is an illustration of how, according to an embodiment of the present invention, the grid points are reconstructed spatially using a 5th-order WENO format and advanced temporally using a Runge-Kutta format.

[0084] Figure 6 This is a nonlocal approximation diagram of piecewise linear interpolation according to an embodiment of the present invention;

[0085] Figure 7 A block diagram of a deep learning traffic flow model based on traffic data according to an embodiment of the present invention; and

[0086] Figure 8 This is a block diagram of a traffic flow parameter mixing prediction device according to an embodiment of the present invention. Detailed Implementation

[0087] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0088] This application first studies the identification of dynamic evolution patterns and situational analysis of road traffic networks, aiming to combine the advantages of macro-traffic flow models and data-based traffic flow models in order to achieve the goal of accurately and in real-time predicting changes in traffic parameters.

[0089] The main principle of prediction algorithms is to combine numerical simulation, traffic models, real-time data, and historical data to predict the evolution of future traffic conditions. Designing fast, scalable, and accurate road traffic prediction tools is key to overcoming the shortcomings of existing traffic management information systems in predictive capabilities. Hybrid traffic flow models, which combine the advantages of both macro-level traffic flow models and data-driven traffic flow models, are of great significance for improving the accuracy, robustness, and real-time performance of predictions.

[0090] refer to Figure 1 A specific embodiment of the present invention discloses a method for hybrid prediction of traffic flow parameters, comprising: in step S102, selecting the road segment AB to be predicted and obtaining real-time traffic flow data at grid point A and historical traffic flow data at grid point B of the road segment to be predicted; in step S104, establishing a GKT macro-traffic flow model and inputting the real-time traffic flow data at grid point A into the GKT macro-traffic flow model to predict the future traffic flow data at grid point B as the macro-traffic flow model prediction result U1; in step S106, establishing and training an LSTM deep learning model, and using the trained LSTM deep learning model to predict the future traffic flow data at grid point B as the deep learning model prediction result U2; and in step S108, using a linear Kalman filter to fuse the macro-traffic flow model prediction result U1 and the deep learning model prediction result U2 into the final prediction result U3 of the future traffic flow parameters of grid point B.

[0091] Compared with existing technologies, the traffic flow parameter hybrid prediction method provided in this embodiment integrates the prediction results of macroscopic models and deep learning models, which fully combines the advantages of macroscopic models and deep learning models. At the same time, the two models can complement each other and effectively improve the shortcomings of their single prediction. This is of great significance for improving prediction accuracy, robustness and real-time performance.

[0092] In the following text, refer to Figures 1 to 7The steps of the traffic flow parameter mixing prediction method according to embodiments of the present invention will be described in detail.

[0093] In step S102, the road segment AB to be predicted is selected, and real-time traffic flow data at grid point A and historical traffic flow data at grid point B of the road segment to be predicted are obtained. Specifically, refer to... Figure 2 The process of selecting the road segment AB to be predicted and obtaining real-time traffic flow data at grid point A and historical traffic flow data at grid point B of the road segment to be predicted includes: selecting the road segment AB to be predicted, and ensuring that there are no entrance or exit ramps in the middle of the road segment AB; uniformly dividing the road segment AB to be predicted into i grids using i+1 grid points, where i+1 grid points include grid point A and grid point B; and collecting real-time traffic flow data at grid point A and obtaining historical traffic flow data at grid point B, where the real-time traffic flow data includes density and speed.

[0094] In step S104, a GKT macroscopic traffic flow model is established, and the real-time traffic flow data at grid point A is input into the GKT macroscopic traffic flow model to predict the future traffic flow data at grid point B, which is then used as the macroscopic traffic flow model prediction result U1. Specifically, refer to... Figure 4 and Figure 5 A GKT macro-traffic flow model was established, and real-time traffic flow data at grid point A was input into the GKT macro-traffic flow model to predict future traffic flow data at grid point B. The prediction result U1 of the macro-traffic flow model includes: a grid length x of L / i, with the grid length serving as the spatial step size, where L is the length of the road segment AB to be predicted; a time step determined according to the CFL condition, where the time step is no greater than x / V0, where V0 is the maximum expected speed; inputting the density and speed at grid point A into the GKT macro-traffic flow model; spatially reconstructing the grid points using a 5th-order WENO format; and advancing the data temporally using a Runge-Kutta format to predict future traffic flow data at grid point B, which is then used as the macro-traffic flow model prediction result U1. The macro-traffic flow model prediction result U1 is a vector that varies over time.

[0095] The U value, which is the prediction result U1 of the GKT macroscopic traffic flow model, can be obtained from the V value through the following steps:

[0096] (1) The original conservation expression is diagonalized by using the Jacobian matrix, and the original equation is transformed into a single-wave equation by decoupling, so that the variable U can be solved by solving the variable V.

[0097] (2) In order to obtain an accurate estimate of the variable V, we use Weno to reconstruct it spatially, thus obtaining V. i+1 / 2 and V i-1 / 2The accurate estimate is then obtained by advancing the time step using the third-order Runge-Kutta method to obtain the V value at each time step.

[0098] (3) Since V = R -1 If U is a matrix, then U = RV, and the matrix R is known during the diagonalization process.

[0099] Before reconstructing the grid points spatially using the 5th-order WENO scheme, the following steps are included:

[0100] The conservation expression for the GKT macroscopic traffic flow model is:

[0101]

[0102] According to the conservation expression:

[0103] Variable U=[ρ,ρu] T Flux f(U) = [ρu, ρu 2 +θρ] T , source item

[0104] The conservation expression is then transformed into:

[0105]

[0106] According to the chain rule, we can obtain:

[0107]

[0108] The Jacobian matrix A is diagonalized as follows:

[0109]

[0110] Where Λ is a diagonal matrix composed of eigenvalues, with the following variables set:

[0111] V=R -1 U;

[0112] So

[0113] By decoupling the original model equations through a series of changes, a single-wave equation is obtained. The variables in the single-wave equation are independent of each other, so that the error is controlled within a relatively reasonable range.

[0114] The grid points are reconstructed spatially using a 5th-order WENO format, and then advanced temporally using a Runge-Kutta format to predict the future traffic flow data at grid point B. The macro-traffic flow model prediction result U1 includes:

[0115] The grid points are reconstructed spatially using a 5th-order WENO format, and then advanced temporally using a Runge-Kutta format to predict the future traffic flow data at grid point B. The macro-traffic flow model prediction result U1 includes:

[0116] Finding the estimated value

[0117] calculate

[0118] Decoupling into single-wave equations results in smaller oscillations in the GKT macroscopic traffic flow model, i.e.

[0119] Using V i+b Reconstruct WENO to obtain V i+1 / 2 The estimated value, where WENO reconstruction is to make V i+1 / 2 The estimated value was determined for all templates (V) i-2 V i-1 V i V i+1 V i+2 Reconstructed value (q) on ) j A linear (weight ω) j ) combination (∑ω j q j ), but when V i+1 / 2 When the surface is not smooth, using linear weights alone is insufficient; a smoothing factor (IS) must be introduced. j Improved weighting coefficient ω j :

[0120] Optimal weights are selected respectively: To determine the weight coefficients, i.e., the weights ω j Introducing a smoothing factor, corresponding to the three positive fluxes, namely, the smoothing factor IS corresponding to the positive flux. + for:

[0121]

[0122] Smoothing factor IS corresponding to negative flux - for:

[0123]

[0124] The smoothing factor IS for the above positive and negative fluxes + IS - The calculation is to obtain α j Thus, the weighting coefficient ω is calculated. j Here, ε represents a very small number, and the weighting coefficients are calculated as follows:

[0125] The positive flux (where "positive" refers to upstream traffic flow information, specifically the left-hand value in this application) at spatial grid point i+1 / 2 (i.e., the midpoint between grid point i and grid point i+1) is:

[0126]

[0127] The negative flux (where "negative" refers to the downstream information volume of the traffic flow) at spatial grid point i+1 / 2 (i.e., the midpoint between grid point i and grid point i+1) is:

[0128]

[0129] In the WENO reconstruction, the accurate estimate at point i+1 / 2 is obtained through the above positive and negative fluxes, i.e., the result at the spatial grid point i+1 / 2 is... Where i refers to grid point i of road segment, and j refers to the optimal weight C. j Smoothing factor IS j Flux q j The subscript (j = 0, 1, 2).

[0130] Obtain the accurate estimate U i+1 / 2 =R i+1 / 2 V i+1 / 2 ,

[0131] Similarly, we can obtain V i-1 / 2 Then, a time-advanced solution is obtained using the third-order Runge-Kutta scheme, as follows: According to get:

[0132]

[0133] This equation is solved using the third-order Runge-Kutta method:

[0134] V (1) =u n +ΔtL(u n )

[0135]

[0136]

[0137] The value of variable V is obtained by advancing the time, and the prediction result U1 of the GKT macro traffic flow model (i.e., variable U in the above formula) is obtained based on the value of variable V.

[0138] In step S106, an LSTM deep learning model is established and trained, and the trained LSTM deep learning model is used to predict the future traffic flow data at grid point B, which is then used as the deep learning model prediction result U2. Specifically, refer to... Figure 7 The process involves establishing and training an LSTM deep learning model, and then using this trained LSTM model to predict future traffic flow data at grid point B, resulting in the deep learning model prediction result U2. This includes: decomposing the input data into trend, periodic, and residual components using local weighted regression, where the input data includes date, time, location, weather, accident, and traffic control data; using a composite deep learning network model composed of a one-dimensional convolutional neural network and a bidirectional long short-term memory network to predict the trend, periodic, and residual components; and using a multilayer perceptron neural network to fuse these predictions into a single prediction result U2, representing the future traffic flow data at grid point B. The deep learning model prediction result U2 includes traffic speed, vehicle density, and vehicle volume. The deep learning model prediction result U2 is a time-varying vector.

[0139] In step S108, the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 are fused using a linear Kalman filter to obtain the final prediction result U3 of the future traffic flow parameters for grid point B. Specifically, fusing the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 into the final prediction result U3 of the future traffic flow parameters for grid point B using a linear Kalman filter includes: using the macroscopic traffic flow model prediction result U1 as the test input of the linear Kalman filter; using the deep learning model prediction result U2 as the prediction input of the linear Kalman filter; and automatically adjusting the weights of the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 so that the final prediction result U3 is close to the actual traffic flow parameters.

[0140] The final prediction result U3 of the future traffic flow parameters for grid point B is obtained by fusing the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 using a linear Kalman filter. This prediction result includes the test input U1 of the linear Kalman filter, expressed by the following formula:

[0141] U1(n|n-1)=AU1(n-1|n-1)+Bu n +w n ,

[0142] The error covariance is: P(n|n-1)=AP(n-1|n-1)A T +Q,

[0143] The predictor input U2 of the linear Kalman filter is expressed by the following formula:

[0144] U2(n) = HU1(n) + v(n),

[0145] The Kalman gain is calculated using the following formula:

[0146]

[0147] The optimal estimate U3 is calculated using the following formula:

[0148] U3 = U1 + K g (n)(U2-HU1),

[0149] Update the error covariance using the following formula:

[0150] P(n|n)=[1-K g (n)H]P(n|n-1),

[0151] Among them, u n Let w be the state control variable at time n, A be the state transition matrix, B be the control matrix, and w be the control variable. n Let be the system process noise at time n, Q be the system process covariance, H be the measurement system parameters, v(n) be the measurement noise at time n, and R be the measurement noise covariance.

[0152] refer to Figure 8 A specific embodiment of the present invention discloses a traffic flow parameter hybrid prediction device, comprising: a grid point acquisition module 802, used to select the road segment AB to be predicted and acquire real-time traffic flow data at grid point A and historical traffic flow data at grid point B of the road segment to be predicted; a GKT macro traffic flow model 804, used to establish a GKT macro traffic flow model and input the real-time traffic flow data at grid point A into the GKT macro traffic flow model to predict the future traffic flow data at grid point B as the macro traffic flow model prediction result U1; an LSTM deep learning model 806, used to establish and train an LSTM deep learning model, and use the trained LSTM deep learning model to predict the future traffic flow data at grid point B as the deep learning model prediction result U2; and a fusion module 808, used to fuse the macro traffic flow model prediction result U1 and the deep learning model prediction result U2 using a linear Kalman filter to obtain the final prediction result U3 of the future traffic flow parameters of grid point B.

[0153] In the following text, refer to Figures 2 to 7 The method for mixed prediction of traffic flow parameters according to embodiments of the present invention will be described in detail with specific examples.

[0154] refer to Figure 2 and Figure 3 Step 1: Select road segment AB, collect traffic flow data at cross-sections A and B respectively, take cross-section A as the initial boundary according to the actual traffic flow direction, and divide road segment AB into grids evenly according to the road segment length, as follows: If there are i grids in total, then the grid length Δx is L / i, and there are i+1 grid points.

[0155] Step 2: Use the GKT (Gas-Kinetic based Traffic model) macroscopic traffic flow model to predict the future traffic flow parameters of grid point B on the road segment, and input the traffic flow information collected at grid point A on the road segment into the model.

[0156] refer to Figure 4 and Figure 5 When predicting traffic flow data at grid point B, the GKT model first divides the road segment into grid points to determine the spatiotemporal step size (Δt, Δx). After determining the spatiotemporal step size, the traffic flow data collected at grid point A, typically density and speed data, are input into the model. Then, spatially, a 5th-order WENO (Weighted Essentially Non-Oscillatory) format is used for reconstruction, and temporally, a 3rd-order Runge-Kutta method is used to advance the model to obtain the speed and density of grid point B.

[0157] The conservation form of the model is:

[0158] Since there are no entrance or exit ramps in the AB road segment studied in this paper, vehicles can only enter from A and exit from B. Therefore, the source term on the right side of the first line of the formula is 0. This line of the formula utilizes the conservation law that the number of vehicles entering and exiting the AB road segment remains constant (similar to the conservation of mass). The second line of the formula utilizes the traffic acceleration equation, which represents the driver's expected acceleration or deceleration response to traffic flow density. It can be approximately obtained by multiplying the first line of the formula by the velocity and then adding the traffic acceleration equation by the density (similar to the conservation of momentum). By inputting the velocity and density data of grid point A into this expression and performing spatiotemporal evolution, the velocity and density data of grid point B are obtained, thereby predicting the traffic flow.

[0159] The equation for traffic acceleration is as follows:

[0160]

[0161] In this equation, the first and second terms on the right-hand side describe the relaxation and desired effects, respectively, with τ and P representing the vehicle relaxation time and traffic pressure. The relaxation term indicates that the vehicle's speed u is less than the equilibrium speed V. eWhen the density (ρ) is high, the vehicle should accelerate; conversely, it should decelerate. Pressure P = P(ρ) is assumed to be an increasing function of density, meaning P′(ρ) > 0 indicates lower downstream density, allowing the vehicle to accelerate; conversely, it should decelerate. Specifically, P is ρθ, where θ is the velocity variance. The actual acceleration or deceleration of the vehicle depends on the interaction between the slack term and the traffic pressure term.

[0162] In the model conservation expression: x is the spatial step, i.e., the grid length obtained in step one; t is the time step, which must satisfy the CFL condition (Courant-Friedrichs-Lewy condition), i.e., it should not be greater than x / V0, where V0 is the maximum expected speed, and its value can be referred to Table 1 below, or it can be calibrated according to the actual situation. u is the average vehicle speed, which represents the average vehicle speed at position x at time t; ρ is the vehicle density, which represents the vehicle density at position x at time t; q is the traffic flow, which is the product of speed and density; τ represents the vehicle acceleration time, also known as the relaxation time, and its specific value can be referred to Table 1 below.

[0163] θ is the velocity variance, and to ensure that the velocity variance disappears when the average velocity approaches 0, an empirical formula is established: θ = A(ρ)u 2 A(ρ) is the density-dependent variance factor, which can be approximated by the Fermi function, as follows:

[0164]

[0165] Where: A0 and A0+2δA are the approximate variance factors for free traffic and congested traffic, respectively; ρ cr It is the critical density for the transition from free flow to congested flow; while δρ shows the transition width, which, along with A0 and δA, are constant values, as shown in Table 1 below.

[0166] V e This is the dynamic equilibrium velocity, obtained by subtracting the necessary deceleration term from the maximum velocity term. It depends not only on the local density and average velocity, but also on the nonlocal density and average velocity. The specific expression is as follows:

[0167]

[0168] in, V0 represents the vehicle's maximum desired speed; ρ max It is the maximum density, also known as the blockage density; This expresses a Boltzmann-like collision term function. To ensure nonlocality, we define the expected position x in front of the actual position x. a The required value is calculated at this point, i.e.:

[0169] ρ a(x,t)=ρ(x a ,t),u a (x,t)=u(x a ,t),θ a =A(ρ a )u a 2 ,

[0170] x a =x+γ(1 / ρ max +Tu),

[0171] Where, ρ a u a and θ a In nonlocal x a The expected term measured at the point; γ∈[1,2] is the interaction influence factor, which indicates that the driver has a certain expected response to braking; T represents the braking time, that is, the safe forward travel time. The specific values ​​of γ and T can be found in Table 1.

[0172] Table 1 Typical values ​​of GKT model parameters under general conditions

[0173]

[0174] Note: Some parameter values ​​may vary depending on actual road conditions and need to be specifically calibrated.

[0175] In specific calculations, non-local terms need to be approximated using interpolation methods, as follows:

[0176] At x = iΔx, there is a expected distance s. a =x a -x, s a It is also the minimum safe distance during vehicle operation, which is approximated locally through piecewise linear interpolation, such as... Figure 6 As shown:

[0177]

[0178] Where: U = [ρ, q] T These are the variables in the above model equations; in addition, the layer function... This represents the largest integer not greater than the value in the square brackets.

[0179] Step 3: Based on the decoupling of the macroscopic variables of the GKT model, the model is numerically solved. First, it is reconstructed spatially using the 5th-order WENO scheme, and then advanced temporally using the 3rd-order Runge-Kutta scheme, as detailed below:

[0180] Based on the conservation expression of the GKT model in step two, we can obtain: variable U = [ρ, ρu] TFlux f(U) = [ρu, ρu 2 +θρ] T , source item

[0181] The original equation can then be transformed into:

[0182]

[0183] According to the chain rule, we can obtain:

[0184]

[0185] The Jacobian matrix A can be diagonalized, i.e.

[0186]

[0187] Where Λ is a diagonal matrix composed of eigenvalues, and a new variable is set as follows:

[0188] V=R -1 U

[0189] So,

[0190]

[0191] By decoupling the original model equations through this series of changes, a single-wave equation is obtained, in which each variable is independent of the others, thereby controlling the error within a relatively reasonable range.

[0192] Then, the grid points are reconstructed using the WENO scheme, and the solution is obtained through time-progression using the Runge-Kutta scheme. The specific steps are as follows:

[0193] Step 1: Find a less precise one The estimated value can generally be 100%.

[0194] Step 2: Calculation at this time

[0195] Step 3: Decouple into single-wave equations to reduce model oscillations, i.e.

[0196] Step 4: Using V i+b We reconstruct WENO to obtain V i+1 / 2 The estimated values ​​are as follows:

[0197] Optimal weights are selected respectively:

[0198] Introducing a smoothing factor, the three positive fluxes (right-skewed values) are:

[0199]

[0200] The three corresponding negative fluxes (left-skewed values) are:

[0201]

[0202] The weighting coefficients are calculated as follows:

[0203] The positive flux (right-biased value) at position i+1 / 2 is:

[0204]

[0205] The negative flux (left-biased value) at position i+1 / 2 is:

[0206]

[0207] That is, the result

[0208]

[0209] Step 5: Obtain the accurate estimate U i+1 / 2 =R i+1 / 2 V i+1 / 2

[0210] Step 6: Similarly, we can obtain V i-1 / 2 Then, the third-order Runge-Kutta scheme is used for time-advanced solution, as follows:

[0211] according to We can obtain:

[0212]

[0213] Solving this equation using the third-order TVD (Total Variation Diminishing) type Runge-Kutta method yields:

[0214] V (1) =u n +ΔtL(u n )

[0215]

[0216]

[0217] By advancing the time step, the value of variable V can be obtained, which in turn allows us to obtain the macroscopic traffic flow parameter prediction result U1. U1 is the prediction result of the macroscopic traffic flow model, or in other words, U1 is the prediction output of the macroscopic traffic flow model. This output is one input to the Kalman filter. The other input to the Kalman filter is the output of the deep learning model.

[0218] Step 4: Use an LSTM deep learning model to predict future traffic flow parameters at grid point B on road segment. The traffic flow information collected at grid point B is used as input to the model. Specifically:

[0219] This study uses deep learning algorithms to establish a data-driven traffic flow model. In complex traffic environments, time series data such as vehicle speed, density, and flow rate exhibit certain characteristics. The STL (Seasonal and Trend Decomposition using Loess) method is employed to decompose the variables in the input vector into features, obtaining more predictable trend and periodic components. STL is a common algorithm in time series decomposition, decomposing the input data into trend components, periodic components, and residuals based on LOESS. It is further divided into inner and outer loops. The inner loop mainly performs trend fitting and periodic component calculation, while the outer loop is mainly used to adjust robust weights. The trend component refers to the general trend of the time series, while the periodic component refers to the cyclical characteristics with a fixed time period. The remaining components are the residuals. An inner loop is nested within an outer loop. The trend component and the cycle component are updated once each iteration of the inner loop. Each iteration of the outer loop includes the calculation of the inner loop and the robustness weights. These weights will be used in the next inner loop to reduce the impact of temporary, outlier points on the trend component and the cycle component.

[0220] Step 5: Solve the LSTM model to obtain the traffic flow parameter deep learning model prediction result U2.

[0221] refer to Figure 7The Kalman filter employs a composite deep learning network model consisting of a one-dimensional convolutional neural network and a bidirectional long short-term memory network to learn and predict each feature component. Finally, a multilayer perceptron neural network (MLP) is used to intelligently fuse the prediction results of each component, forming a data-driven traffic flow model. The final output is the traffic flow parameter deep learning model prediction result U2. The Kalman filter has two inputs: a prediction and a test. U2 is the prediction result given by the deep learning model, serving as the prediction input to the Kalman filter. U1 is the test input to the Kalman filter, derived from the prediction result of the GKT model. The Kalman filter's role is to balance the prediction results of these two models, automatically adjusting their weights to make the output result closer to the actual traffic parameters.

[0222] Step Six: Use U1 as the state variable of the Kalman filter system and U2 as the detection variable of the Kalman filter system (the Kalman filter is a mature theory with two inputs and one output. The two inputs are the system state prediction and the detection variable, and the output is the fusion result. This patent uses Kalman filtering to fuse the prediction results of two traffic flow parameter prediction models. The results of these two traffic flow models are used as the two inputs of the Kalman filter. Among them, the deep learning model LSTM is used as the system state prediction input of the Kalman filter; the GKT model is used as the measurement input of the Kalman filter). Using a linear Kalman filter, the macroscopic model prediction result U1 (both the macroscopic model result U1 and the deep learning model prediction result U2 are one-dimensional quantities that change with time. If time is considered as a dimension, then U1 and U2 are two-dimensional quantities. For example, the traffic flow velocity v of a certain road segment fluctuates with time t, and a traffic flow velocity fluctuation image can be drawn with t as the horizontal axis and v as the vertical axis) and the deep learning model prediction result U2 are fused into a new prediction result U3, which is the final prediction result of the future traffic flow parameters of the road segment grid point B. Specifically:

[0223] 1. System prediction

[0224] (1) Prior estimation (system state description): U1(n|n-1)=AU1(n-1|n-1)+Bu n +w n

[0225] Among them, u n Let w be the state control variable at time n, A be the state transition matrix, B be the control matrix, and w be the control variable. n Let n be the system process noise at time n.

[0226] (2) Error covariance: P(n|n-1)=AP(n-1|n-1)A T +Q

[0227] Where Q is the system process covariance.

[0228] 2. Establish measurement equations

[0229] U2(n) = HU1(n) + v(n)

[0230] Where H represents the measurement system parameters, and v(n) represents the measurement noise at time n.

[0231] 3. Generate the optimal estimate

[0232] (1) Calculate the Kalman gain: Where R is the measurement noise covariance. The optimal estimate U3 can be calculated by combining the observed values.

[0233] (2) Corrected estimate: U3 = U1 + K g (n)(U2-HU1).

[0234] 4. Update error covariance

[0235] P(n|n)=[1-K g (n)H]P(n|n-1).

[0236] The current value is predicted using the previous best result (prior estimate), while the current value is corrected using the observed values ​​to obtain the optimal result. The weights of the estimated value and the observed value are determined separately using a Kalman filter, and the two are integrated into a corrected value.

[0237] This technique employs a second-order macroscopic model of nonlocal gas dynamics, which includes a second-order partial differential equation describing the dynamic changes in velocity. The second-order gas motion traffic flow model (GKT) is solved using the relaxation algorithm. The relaxation algorithm is an accurate and robust high-order finite volume method for solving macroscopic traffic flow models using numerical integration. Its spatial discretization uses a fifth-order weighted non-oscillatory interpolation method, while the time integration uses a third-order explicit-implicit Runge-Kutta method.

[0238] Traffic parameter prediction methods based on big data and machine learning have attracted considerable research interest in recent years. Big data analytics can help users reach their destinations via the most suitable routes and in the shortest time, while simultaneously improving their safety. Traffic management departments can predict traffic flow by rapidly collecting and analyzing massive amounts of current and historical traffic data. Big data analytics can effectively predict traffic accidents. Big data analytics primarily addresses three issues: data storage, data analysis, and data management. Machine learning is the most popular modeling and analysis theory in the big data ecosystem because it makes it easy to obtain models from large amounts of data. Machine learning models can be categorized into supervised learning, unsupervised learning, reinforcement learning, deep learning, and entity-based algorithms. Labeled training data is used for supervised learning algorithms. Among all supervised learning models, linear regression, decision trees, neural networks, and support vector machines are the most commonly used.

[0239] The hybrid traffic flow parameter prediction method integrating macroscopic and deep learning models has the following advantages: This invention overcomes a series of problems in traditional single traffic prediction methods, such as long model calibration time and inability to reflect occasional road conditions in real time. It provides a hybrid traffic flow parameter prediction method integrating macroscopic and deep learning models. This hybrid model uses a Kalman filter to use the macroscopic model prediction results as system state variables and the deep learning model prediction results as system detection variables, thus obtaining the final traffic information prediction result. It integrates the prediction results of macroscopic and deep learning models, fully concentrating the advantages of both models. Simultaneously, the two models complement each other, effectively improving the shortcomings of individual predictions. This is of great significance for improving prediction accuracy, robustness, and real-time performance. Therefore, this hybrid model can obtain more accurate traffic prediction results with relatively less data based on an understanding of real-time road traffic flow characteristics. The proposed technical solution aligns with the concept of intelligent transportation, is conducive to promoting the construction of intelligent transportation systems, and provides new ideas and solutions for urban brains.

[0240] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0241] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting traffic flow parameters using a hybrid approach, characterized in that, include: Select the road segment AB to be predicted and obtain the real-time traffic flow data at grid point A and the historical traffic flow data at grid point B of the road segment to be predicted; A GKT macro-traffic flow model is established, and the real-time traffic flow data at grid point A is input into the GKT macro-traffic flow model to predict the future traffic flow data at grid point B as the macro-traffic flow model prediction result U1. The grid points are reconstructed spatially using the 5th-order WENO format, and advanced temporally using the Runge-Kutta format to predict the future traffic flow data at grid point B as the macro-traffic flow model prediction result U1. An LSTM deep learning model is established and trained, and the trained LSTM deep learning model is used to predict the future traffic flow data at grid point B as the deep learning model prediction result U2. Specifically, the input data is decomposed into trend components, periodic components, and residual components based on local weighted regression; a composite deep learning network model composed of a one-dimensional convolutional neural network and a bidirectional long short-term memory network is used to predict the trend components, periodic components, and residual components to obtain trend component prediction results, periodic component prediction results, and residual component prediction results; and a multilayer perceptron neural network is used to fuse the trend component prediction results, periodic component prediction results, and residual component prediction results into the future traffic flow data at grid point B for prediction as the deep learning model prediction result U2. The macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 are fused into a final prediction result U3 of the future traffic flow parameters of grid point B using a linear Kalman filter. This fusion process includes: using the macroscopic traffic flow model prediction result U1 as the test input of the linear Kalman filter; using the deep learning model prediction result U2 as the prediction input of the linear Kalman filter; and automatically adjusting the weights of the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 to make the final prediction result U3 approximate the actual traffic flow parameters.

2. The traffic flow parameter hybrid prediction method according to claim 1, characterized in that, Selecting the road segment AB to be predicted and obtaining the real-time traffic flow data at grid point A and the historical traffic flow data at grid point B of the road segment to be predicted includes: Select the road segment AB to be predicted, and there are no entrance or exit ramps in the middle of the road segment AB to be predicted; The road segment AB to be predicted is uniformly divided into i grids using i+1 grid points, wherein the i+1 grid points include grid point A and grid point B; and Real-time traffic flow data is collected at grid point A, and historical traffic flow data at grid point B is obtained. The real-time traffic flow data includes density and speed.

3. The traffic flow parameter hybrid prediction method according to claim 1, characterized in that, Establishing a GKT macro-traffic flow model and inputting the real-time traffic flow data at grid point A into the GKT macro-traffic flow model includes: The grid length x is L / i, and the grid length is used as the spatial step size, where L is the length of the road segment AB to be predicted; The time step is determined based on the CFL condition, wherein the time step is no greater than x / V0, and V0 is the maximum desired velocity; Input the density and velocity at grid point A into the GKT macroscopic traffic flow model; 4. The traffic flow parameter hybrid prediction method according to claim 1, characterized in that, Before reconstructing the grid points spatially using the 5th-order WENO scheme, the following steps are included: The conservation expression of the GKT macroscopic traffic flow model is: ； Based on the conservation expression, we obtain: variable flux , source item ; The conservation expression is then transformed into: ； According to the chain rule, we can obtain: ； The Jacobian matrix A is diagonalized as follows: ； in, It is a diagonal matrix composed of eigenvalues, with the following variables set: ； So By decoupling the original model equation through a series of changes, a single-wave equation is obtained. In the single-wave equation, each variable is independent of the other, so that the error is controlled within a relatively reasonable range. u is the average speed of the vehicle. It is vehicle density; Indicates the vehicle's acceleration time. It is the velocity variance.

5. The traffic flow parameter hybrid prediction method according to claim 4, characterized in that, The grid points are reconstructed spatially using a 5th-order WENO format, and then advanced temporally using a Runge-Kutta format to predict the future traffic flow data at grid point B. The macro-traffic flow model prediction result U1 includes: Finding the estimated value ; calculate ; Decoupling into single-wave equations results in smaller oscillations in the GKT macroscopic traffic flow model, i.e. ; use The estimated value obtained by reconstructing WENO is: Optimal weights are selected respectively: Smoothing factor corresponding to positive flux IS + for: Smoothing factor corresponding to negative flux IS - for: The weighting coefficients are calculated as follows: The positive flux at position i+1 / 2 is: The negative flux at position i+1 / 2 is: That is, the result Obtain accurate estimates Similarly, we can obtain Then, the third-order Runge-Kutta scheme is used for time-advanced solution, as follows: according to get: This equation is solved using the third-order Runge-Kutta method: The value of variable V is obtained by advancing the time, and the prediction result U1 of the GKT macroscopic traffic flow model is obtained based on the value of variable V.

6. The traffic flow parameter hybrid prediction method according to claim 1, characterized in that, The input data includes date, time, location, weather, accident, and traffic control data; The prediction result U2 of the deep learning model includes traffic flow speed, vehicle flow density, and vehicle volume.

7. The traffic flow parameter hybrid prediction method according to claim 1, characterized in that, The final prediction result U3 of the future traffic flow parameters of grid point B, obtained by fusing the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 using a linear Kalman filter, includes: The test input U1 of the linear Kalman filter is expressed by the following formula: ， The error covariance is: , The predictor input U2 of the linear Kalman filter is expressed by the following formula: ， The Kalman gain is calculated using the following formula: ， The optimal estimate U3 is calculated using the following formula: ， Update the error covariance using the following formula: ， in, Let A be the state control variable at time n, A be the state transition matrix, and B be the control matrix. Let Q be the system process noise at time n, Q be the system process covariance, and H be the measurement system parameters. Let R be the measurement noise at time n, and let R be the measurement noise covariance.

8. The traffic flow parameter hybrid prediction method according to any one of claims 1 to 6, characterized in that, Both the macro-traffic flow model prediction result U1 and the deep learning model prediction result U2 are time-varying vectors.

9. A traffic flow parameter hybrid prediction device, characterized in that, include: The grid point acquisition module is used to select the road segment AB to be predicted and acquire the real-time traffic flow data at grid point A and the historical traffic flow data at grid point B of the road segment to be predicted. The GKT macro traffic flow model is used to establish the GKT macro traffic flow model and input the real-time traffic flow data at grid point A into the GKT macro traffic flow model to predict the future traffic flow data at grid point B as the macro traffic flow model prediction result U1. The grid points are reconstructed spatially using the 5th-order WENO format and advanced temporally using the Runge-Kutta format to predict the future traffic flow data at grid point B as the macro traffic flow model prediction result U1. An LSTM deep learning model is used to build and train an LSTM deep learning model, and to predict future traffic flow data at grid point B using the trained LSTM deep learning model as the deep learning model prediction result U2. Specifically, the input data is decomposed into trend components, periodic components, and residual components based on local weighted regression; a composite deep learning network model composed of a one-dimensional convolutional neural network and a bidirectional long short-term memory network is used to predict the trend components, periodic components, and residual components to obtain trend component prediction results, periodic component prediction results, and residual component prediction results; and a multilayer perceptron neural network is used to fuse the trend component prediction results, periodic component prediction results, and residual component prediction results into the future traffic flow data at grid point B for prediction as the deep learning model prediction result U2. The fusion module is used to fuse the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 into a final prediction result U3 of the future traffic flow parameters of grid point B using a linear Kalman filter. The fusion of the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 into the final prediction result U3 of the future traffic flow parameters of grid point B using a linear Kalman filter includes: using the macroscopic traffic flow model prediction result U1 as the test input of the linear Kalman filter; using the deep learning model prediction result U2 as the prediction input of the linear Kalman filter; and automatically adjusting the weights of the macroscopic traffic flow model prediction result U1 and the deep learning model prediction result U2 so that the final prediction result U3 is close to the actual traffic flow parameters.

Images