Vehicle longitudinal model predictive control method fusing long short-term memory network and Kalman filter

Abstract

The invention discloses a vehicle longitudinal model predictive control method fusing a long short-term memory network and a Kalman filter. The method comprises the following steps: establishing a vehicle longitudinal dynamic model considering the speed change of a front vehicle as a predictive model of an MPC control framework; using an LTSM model to predict the acceleration value of the front vehicle at each moment according to the historical acceleration of the front vehicle, substituting the acceleration value into an MPC control framework, and calculating to obtain a corresponding control quantity; if the front vehicle is an HDV, calculating a vehicle state estimation value at each moment, performing Kalman filtering processing, and substituting the vehicle state estimation value into an MPC control framework; if the front vehicle is the CAV, obtaining a vehicle state actual value in real time and substituting the vehicle state actual value into an MPC control framework; and longitudinally controlling the vehicle according to the control quantity of each control period. According to the method, the problems that the control performance of the intelligent networked automobile is reduced and the stability becomes poor under the condition that the environment and the model parameters are uncertain are solved, and the running efficiency of the automobile is improved while safety and comfort are guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of networked autonomous driving vehicles and traffic control, and in particular to a vehicle longitudinal model predictive control method integrating a long short-term memory network and a Kalman filter. Background Art

[0002] As one of the important technologies for autonomous vehicle control, vehicle platoon control inevitably has a lot of mismatches between the real object and the model due to external (environmental) disturbances, parameter changes and neglected dynamic factors, which often has a great impact on the performance of the control system. How to ensure that all vehicles in the platoon move at a consistent speed while maintaining an ideal distance between vehicles under the conditions of considering factors such as measurement noise and model uncertainty in the real driving environment has always been the research focus of autonomous driving technology.

[0003] The rapid development of CAVs has attracted widespread attention. As an important part of future transportation, CAVs can monitor the surrounding environment in real time, quickly identify potential dangers, and take corresponding measures to avoid accidents through advanced sensors, radars, cameras and other equipment. CAVs can also communicate with traffic infrastructure and other vehicles in real time to optimize driving routes and reduce traffic congestion. They have obvious advantages over traditional vehicles in improving driving safety and efficiency.

[0004] However, for the vehicle platoon following problem, if the vehicle in front of the CAV is also a CAV, the state information can be accurately obtained through the inter-vehicle V2V communication, and the error can be ignored; if the vehicle in front of the CAV is a human-driven vehicle (HDV), the CAV estimates the state of the HDV in front only by measuring with sensors, which will cause certain deviations in the control process. Therefore, it is necessary to estimate the vehicle state of the CAV at the next moment while considering this deviation. Although the current technology takes into account the process noise and measurement noise in the vehicle state update process and uses Kalman filtering to fuse and reduce the uncertainty of state estimation, it does not take into account that when the vehicle in front is a HDV, its acceleration is unmeasurable at the time of state update; at the same time, most of the current research regards the acceleration of the vehicle in front as a constant in the MPC prediction framework, which will also cause a certain degree of control accuracy reduction. Summary of the invention

[0005] Technical problem to be solved by the present invention: In view of the above-mentioned problems in the prior art, a vehicle longitudinal model predictive control method integrating a long short-term memory network and a Kalman filter is provided, which uses a long short-term memory network to predict the acceleration of the preceding vehicle, and uses the predicted acceleration of the preceding vehicle to calculate the optimal control input and state update of the model, while using a Kalman filter to reduce the uncertainty and measurement noise caused by the prediction, thereby achieving accurate prediction of the state.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A vehicle longitudinal model predictive control algorithm integrating long short-term memory network and Kalman filter comprises the following steps:

[0008] Establish a vehicle longitudinal dynamics model that takes into account the speed change of the leading vehicle and use it as a prediction model for the MPC control framework;

[0009] Use the long short-term memory network model to predict the acceleration value of the preceding vehicle at each moment based on the preceding vehicle's historical acceleration information;

[0010] Substitute the acceleration prediction value of the preceding vehicle at each moment into the prediction model of the MPC control framework to calculate the control quantity at the corresponding moment and add it to the control sequence;

[0011] If the preceding vehicle is a HDV, the estimated value of the vehicle state at each moment is calculated, and then the estimated value of the vehicle state at each moment is processed by Kalman filtering to obtain the optimal estimated value of the vehicle state at each moment. Finally, the optimal estimated value of the vehicle state at each moment is substituted into the prediction model of the MPC control framework.

[0012] If the vehicle in front is a connected autonomous vehicle (CAV), the actual value of the vehicle state at each moment is obtained in real time, and then the actual value of the vehicle state at each moment is substituted into the prediction model of the MPC control framework;

[0013] The vehicle is longitudinally controlled according to a specified control amount in a control sequence in each control cycle.

[0014] Furthermore, when establishing a vehicle longitudinal dynamics model that takes into account the change in the speed of the preceding vehicle, it includes:

[0015] For each vehicle in the connected autonomous driving vehicle fleet, a corresponding single-vehicle longitudinal dynamic model is established by considering the speed change of the vehicle in front of each vehicle, and then the single-vehicle longitudinal dynamic model of each vehicle is integrated to obtain the dynamic model of the connected autonomous driving vehicle fleet, and used as the vehicle longitudinal dynamic model considering the speed change of the vehicle in front;

[0016] Discretize the longitudinal dynamic model of each vehicle, set the prediction time domain of the model equal to the control time domain and consider the acceleration change of the leading vehicle to obtain the corresponding state prediction equation;

[0017] Taking comfort, efficiency and control input as optimization control objectives, the objective function and constraints of the vehicle longitudinal dynamics model considering the speed change of the leading vehicle are constructed.

[0018] Furthermore, the expression of the longitudinal dynamic model of the single vehicle is as follows:

[0019]

[0020] Among them, X is the system state variable, U is the system control input, d is the system disturbance, and u n is the system input, A, B1, B2 are the state space equation coefficient matrices, a n-1 is the acceleration of the vehicle in front of it, a n ,v n are the difference between the actual distance between the vehicle and the front vehicle and the expected distance, the relative speed with the front vehicle, the vehicle's own acceleration, speed, n is the vehicle's serial number in the connected autonomous driving vehicle fleet, τ n is the vehicle's engine time constant, t d is the expected time interval.

[0021] Furthermore, the expression of the dynamic model of the networked autonomous driving vehicle fleet is as follows:

[0022]

[0023] Among them, X P =(Δe1 s ,Δe1 v ,a1,v1,Δe2 s ,Δe2 v ,a2,v2,…,Δe N s ,Δe N v ,a N ,v N ), is the system state variable of each vehicle in the connected autonomous vehicle fleet, where Δe N s ,Δe N v ,a N ,v N are the difference between the actual distance between the Nth vehicle in the connected autonomous driving vehicle fleet and the vehicle in front and the expected distance, the relative speed with the vehicle in front, the vehicle’s own acceleration, and the speed; A P 、B1 P 、B2 P are the coefficient matrices of the augmented state space equation, U P =(u1,u2,…,u N ), is the system control input for each vehicle in the connected autonomous driving vehicle fleet, where u N is the system input for the Nth vehicle in the connected autonomous vehicle fleet, d p =(a0,a1,…,a N-1), is the system disturbance of each vehicle in the connected autonomous driving vehicle fleet, where a N-1 is the acceleration of the Nth vehicle in the connected autonomous vehicle fleet.

[0024] Furthermore, the state prediction equation is expressed as follows:

[0025]

[0026] in, are the discretized state space equation coefficient matrices, express The jth power, A j-1 express j-1th power, I is the 4×4 identity matrix, t s is the sampling time of the system, u(k|k) represents the predicted state input at time k, u(k+j-1|k) represents the predicted control input at time k for time k+j-1, x(k+j|k) represents the state prediction of the system at time k for time k+j, w(j) is the interference input; w(k+j-1|k) represents the acceleration of the preceding vehicle at the j-1th time point in the future, j=1,2…,N.

[0027] Furthermore, after obtaining the corresponding state prediction equation, the method further includes: changing the state prediction equation to relax the assumption that the acceleration of the preceding vehicle remains unchanged in the prediction time domain, and the changed state prediction equation is expressed as follows:

[0028]

[0029] y(k+j|k)=Cx(k+j|k)

[0030] Among them, u(k-1) is the optimal control quantity at time k-1, Δu(k|k) represents the state input increment at time k, Δu(k+j-1|k) represents the predicted control input increment from time k to time k+j-1, y(k+j|k) represents the measurable output of the system, and C is the unit matrix.

[0031] Furthermore, the expression of the objective function is as follows:

[0032]

[0033] The constraint condition expression is as follows:

[0034]

[0035] where y ref (k+N|k) is the reference trajectory, represents a quadratic function, Q and R are the error and input weighting matrices respectively, ε is the slack variable, p is the weight of the slack variable, Δu(k+j|k) represents the predicted control input increment from time k to time k+j, are the lower and upper limits of the artificially set parameters corresponding to the measurable output of the system, y min ,y max are the minimum and maximum values ​​of the output that can be measured by the system, are the lower and upper limits of the artificially set parameters corresponding to the predictive control input, u min 、u max are the minimum and maximum values ​​corresponding to the predicted control input, They are the lower and upper limits of the artificially set parameters corresponding to the predictive control input increment, Δu min , Δu max are the minimum and maximum values ​​corresponding to the predictive control input increment respectively.

[0036] Furthermore, the estimated value of the vehicle state at each moment is expressed as follows:

[0037]

[0038] in is the estimated value of the state variable at time k, u(k-1) is the optimal control variable at time k-1, ω(k-1) is the acceleration of the preceding vehicle at time k-1, x kal (k-1) is the optimal estimate of the state variable at time k-1 after Kalman filtering.

[0039] Furthermore, when the vehicle state estimation value at each moment is processed by Kalman filtering, the following steps are included:

[0040] Calculate the covariance of the optimal estimate of the state variable at the previous moment and calculate the Kalman gain at the current moment. The expression is as follows:

[0041]

[0042] Among them, K K is the Kalman gain at time k, is the covariance P(k-1) of the optimal estimate of the state variable at time k-1 and the covariance Q of the Gaussian noise θ The sum, R ξ is the measured value covariance, C is the proportionality coefficient of the state variable, Update the matrix for the state;

[0043] The measured value of the vehicle state at the current moment is obtained, and the estimated value of the vehicle state at the current moment is corrected according to the measured value of the vehicle state at the current moment and the Kalman gain to obtain the optimal estimated value of the vehicle state at the current moment. The expression is as follows:

[0044]

[0045] Among them, x kal (k) is the optimal estimate of the state variable at time k, is the estimated value of the state variable at time k, is the measured value of the state variable at time k.

[0046] Furthermore, the designated control amount in the control sequence of each control cycle specifically refers to the first control amount in the control sequence.

[0047] Compared with the prior art, the advantages of the present invention are:

[0048] (1) Compared with the traditional vehicle platoon longitudinal control method, the present invention uses the LSTM method to predict acceleration and uses the predicted acceleration to update the control quantity during the MPC calculation process, thereby improving the control accuracy.

[0049] (2) Compared with the traditional vehicle platoon longitudinal control method, the model of the present invention is established on the scenario of a CAV platoon following an HDV, and it is assumed that the acceleration information of the preceding vehicle can be obtained in real time through V2V communication between CAV vehicles, while the acceleration information of the preceding vehicle cannot be obtained in real time between CAV and HDV, which is more in line with the actual situation.

[0050] (3) Compared with the traditional vehicle platoon longitudinal control method, the present invention uses the Kalman filter algorithm to filter the calculation errors caused by measurement noise and uncertainty of the preceding vehicle parameters, thereby reducing the uncertainty of the model.

[0051] (4) Compared with the traditional vehicle platoon longitudinal control method, the present invention significantly improves the operation efficiency of the fleet while taking into account driving safety, comfort and energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of an embodiment of the present invention.

[0053] Figure 2 It is a control structure diagram of the present invention.

[0054] Figure 3 It is a schematic diagram of an example scenario of the present invention.

[0055] Figure 4 It is the acceleration curve of the leading vehicle within the simulation time period of the present invention.

[0056] Figure 5 This is an acceleration curve diagram of four following vehicles using the K-MPC method of the present invention.

[0057] Figure 6 This is the acceleration curve diagram of four following vehicles using the L-MPC method of the present invention.

[0058] Figure 7 This is the acceleration curve diagram of four following vehicles using the KL-MPC method of the present invention.

[0059] Figure 8 This is a curve diagram of the gap error between four following vehicles using the K-MPC method of the present invention.

[0060] Fig. 9 This is a curve diagram of the gap error between four following vehicles using the L-MPC method of the present invention.

[0061] Fig.10 This is a curve diagram of the gap error between four following vehicles using the KL-MPC method of the present invention. DETAILED DESCRIPTION

[0062] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0063] This embodiment proposes a vehicle longitudinal model predictive control method (KL-MPC) that integrates a long short-term memory network (LSTM) and a Kalman filter. This method solves the problem of decreased control performance and poor stability of intelligent connected vehicles in the presence of environmental and model parameter uncertainties. By rewriting the vehicle longitudinal dynamics equation, the assumption that the acceleration of the leading vehicle remains unchanged in the prediction time domain is relaxed, and the LSTM model is used to predict the acceleration value of the leading vehicle in the future time window and substitute it into the MPC framework to obtain the optimal control quantity. In the state update process, a Kalman filter is designed to estimate the state variables to eliminate the interference of measurement noise and model parameter uncertainty, thereby improving the control accuracy of the model. The method flow of this embodiment is as follows: Figure 1 As shown, the specific steps include:

[0064] S1) establishing a vehicle longitudinal dynamics model taking into account the speed change of the leading vehicle and using it as a prediction model for the MPC control framework;

[0065] S2) Using the LSTM model to predict the acceleration value of the preceding vehicle at each moment in the future based on the preceding vehicle's historical acceleration information;

[0066] S3) Substituting the acceleration prediction value of the preceding vehicle at each moment into the prediction model of the MPC control framework, calculating the optimal control amount at the corresponding moment and adding it to the control sequence;

[0067] S4) If the preceding vehicle is a HDV, the vehicle state estimation value at each moment is calculated, and then the vehicle state estimation value at each moment is processed by Kalman filtering to obtain the optimal vehicle state estimation value at each moment, and finally the optimal vehicle state estimation value at each moment is substituted into the prediction model of the MPC control framework;

[0068] If the vehicle in front is a connected autonomous vehicle (CAV), the actual value of the vehicle state at each moment is obtained in real time, and then the actual value of the vehicle state at each moment is substituted into the prediction model of the MPC control framework;

[0069] S5) Rolling time domain optimization, performing longitudinal control on the vehicle according to the specified control amount in the control sequence of each control cycle, thereby achieving longitudinal control of the vehicle.

[0070] Each step is described in detail below.

[0071] In step S1 of this embodiment, when establishing a vehicle longitudinal dynamics model taking into account the change in the speed of the preceding vehicle, the following steps are included:

[0072] S1.1) For each vehicle in the networked autonomous driving vehicle fleet, a corresponding single-vehicle longitudinal dynamic model is established by considering the speed change of the vehicle in front of each vehicle. The expression of the single-vehicle longitudinal dynamic model is as follows:

[0073]

[0074] Among them, X is the system state variable, U is the system control input, d is the system disturbance, and u n is the system input, A, B1, B2 are the state space equation coefficient matrices, a n-1 is the acceleration of the vehicle in front of it, a n ,v n are the difference between the actual distance between the vehicle and the front vehicle and the expected distance, the relative speed with the front vehicle, the vehicle's own acceleration, speed, n is the vehicle's serial number in the connected autonomous driving vehicle fleet, τ n is the vehicle's engine time constant, t d The expected time interval

[0075] S1.2) integrating the longitudinal dynamic model of each vehicle to obtain the dynamic model of the networked autonomous driving vehicle fleet, and using it as the longitudinal dynamic model of the vehicle considering the speed change of the leading vehicle;

[0076] Consider Figure 2 For the longitudinal following scenario shown in the figure, for a CAV fleet consisting of N vehicles, the state variables of the system can be defined as X P =(Δe1 s ,Δe1v ,a1,v1,Δe2 s ,Δe2 v ,a2,v2,…,Δe N s ,Δe N v ,a N ,v N ), where Δe N s ,Δe N v ,a N ,v N are the difference between the actual distance and the expected distance between the Nth vehicle in the connected autonomous driving vehicle fleet and the vehicle in front, the relative speed with the vehicle in front, the vehicle’s own acceleration, and the speed; the control input is defined as U P =(u1,u2,…,u N ), where u N is the system input of the Nth vehicle in the connected autonomous vehicle fleet, and the system disturbance is d p =(a0,a1,…,a N-1 ), a0 represents the acceleration of the leading vehicle followed by the convoy, a N-1 is the acceleration of the front vehicle of the Nth vehicle in the connected autonomous vehicle fleet. Therefore, the matrix form of the dynamic model of the CAV fleet with N vehicles can be expressed as:

[0077]

[0078]

[0079] Among them, A P 、B1 P 、B2 P are the coefficient matrices of the augmented state space equations

[0080] S1.3) Based on the longitudinal dynamics model of a single vehicle, the longitudinal dynamics model of each vehicle is discretized using the forward Euler method. The prediction time domain of the model is set equal to the control time domain and the acceleration change of the leading vehicle is considered to obtain the corresponding state prediction equation, which is expressed as follows:

[0081]

[0082] in, are the coefficient matrices of the discretized state space equation, express The jth power, A j-1 express j-1th power, I is the 4×4 identity matrix, t sis the sampling time of the system, u(k|k) represents the predicted state input at time k, u(k+j-1|k) represents the predicted control input at time k for time k+j-1, x(k+j|k) represents the state prediction of the system at time k for time k+j, w(j) is the interference input; w(k+j-1|k) represents the acceleration of the preceding vehicle at the j-1th time point in the future, j=1,2…,N;

[0083] After establishing the vehicle longitudinal dynamics model considering the change in the speed of the leading vehicle, this embodiment relaxes the assumption that the acceleration of the leading vehicle remains unchanged in the prediction time domain. Assuming that the control process is implemented by incremental control Δu, the predicted control input at time k to time k+j is the sum of the predicted control input at time k to time k+j-1 and the incremental predicted control input at time k to time k+j, and the expression is as follows:

[0084] u(k|k)=Δu(k|k)+u(k-1)

[0085] u(k+j|k)=u(k+j-1|k)+Δu(k+j|k)

[0086] Among them, u(k|k) represents the predicted state input at time k, Δu(k|k) represents the state input increment at time k, u(k-1) is a known quantity at time k, and so on.

[0087] Thus, the vehicle prediction equation is rewritten as follows:

[0088]

[0089] Dw(j),j=1,2…,N

[0090] Among them, u(k-1) is the optimal control quantity at time k-1, Δu(k|k) represents the state input increment at time k, Δu(k+j-1|k) represents the predicted control input increment from time k to time k+j-1, y(k+j|k) represents the measurable output of the system, and C is the unit matrix;

[0091] S1.4) Taking comfort, efficiency and control input as optimization control objectives at the same time, the objective function and constraint conditions of the vehicle longitudinal dynamics model considering the speed change of the leading vehicle are constructed.

[0092] Specifically, this embodiment minimizes the loss of vehicle efficiency while considering improving vehicle comfort. Comfort, efficiency and control input are simultaneously used as optimization control objectives, and the objective function expression of the model is as follows:

[0093]

[0094] To ensure the safe driving of vehicles, the head distance between adjacent vehicles must meet the minimum safe distance; to ensure the operating efficiency of vehicles, the relative speed between vehicles must also be controlled within a range; due to the constraints of controller performance, the change in the control amount of each vehicle must also be less than a fixed value; of course, boundary constraints must also be imposed on the operating speed and acceleration. Finally, in order to prevent the situation where too strict constraints make it impossible to find a feasible solution, a relaxation vector is introduced to soften the original strict constraints. The constraint expression of the model is as follows:

[0095]

[0096] Among them, y ref (k+N|k) is the reference trajectory, represents a quadratic function, Q and R are the error and input weighting matrices respectively, ε is the slack variable, p is the weight of the slack variable, Δu(k+j|k) represents the predicted control input increment from time k to time k+j, are the lower and upper limits of the artificially set parameters corresponding to the measurable output of the system, y min ,y max are the minimum and maximum values ​​of the output that can be measured by the system, are the lower and upper limits of the artificially set parameters corresponding to the predictive control input, u min 、u max are the minimum and maximum values ​​corresponding to the predicted control input, They are the lower and upper limits of the artificially set parameters corresponding to the predictive control input increment, Δu min , Δu max are the minimum and maximum values ​​corresponding to the predictive control input increment respectively.

[0097] Step S2 of this embodiment uses the LSTM model to predict the acceleration value of the preceding vehicle in the future time window based on the preceding vehicle's historical acceleration information. Specifically, it includes:

[0098] Firstly, an LSTM model is established. LSTM can learn the implicit information and internal data characteristics in the long-term historical acceleration, realize accurate prediction of the disturbance, and reduce the MPC control error.

[0099] The LSTM model is then trained using the acceleration time series of the vehicle ahead in car-following driving. Specifically, a training data set is generated from historical acceleration data and divided into a training set and a test set. The acceleration features are learned through the training set data, and the prediction effect is tested on the test set. The prediction length of the LSTM prediction model is the same as the prediction domain length in MPC.

[0100] Finally, in each control time period, the historical acceleration data of the preceding vehicle within a fixed time window is input into the trained LSTM model to predict the acceleration value of the preceding vehicle at each moment in the future prediction time domain.

[0101] In step S3 of this embodiment, the predicted acceleration of the preceding vehicle at each moment is substituted into the MPC control framework to obtain the optimal control quantity at the moment. Specifically, the obtained acceleration prediction value is used as the disturbance term d in the above formula. p =(a0,a1,…,a N-1 ), under the premise of satisfying the constraints and minimizing the objective function, the objective function is a convex function and the constraints are linear. This problem can be regarded as a quadratic programming convex optimization problem, which is solvable and has an optimal solution. Finally, by solving this quadratic programming convex optimization problem, the optimal control input at this moment can be obtained as the optimal control quantity.

[0102] In step S4 of this embodiment, when the front vehicle is an HDV vehicle, the front vehicle acceleration a n-1 At time k-1, it cannot be measured but can only be estimated. Therefore, the estimated state calculated during the state update process is There will inevitably be a calculation error between the actual state x(k), let the error In addition, since there is no communication between the vehicles, there will be a measurement noise in the process of measuring the vehicle status. Assuming that the measurement noise Therefore, the Kalman filter algorithm is used to correct these errors and reduce the interference they cause. The relevant calculation and derivation process is as follows:

[0103] In this embodiment, the vehicle state estimation value at each moment is expressed as follows:

[0104]

[0105] in is the estimated value of the state variable at time k, u(k-1) is the optimal control variable at time k-1, ω(k-1) is the acceleration of the preceding vehicle at time k-1, and x kal (k-1) is the optimal estimate of the state variable at time k-1 after Kalman filtering.

[0106] When the vehicle state estimation value at each moment is processed by Kalman filtering, the following steps are included:

[0107] Calculate the covariance of the optimal estimate of the state variable at the previous moment and calculate the Kalman gain at the current moment. The expression is as follows:

[0108]

[0109] Among them, KK is the Kalman gain at time k, Represents the covariance of the predicted value, specifically the covariance P(k-1) of the optimal estimate of the state variable at time k-1 and the Gaussian noise covariance Q θ The sum, R ξ is the measured value covariance, C is the proportional coefficient of the state variable, Update the matrix for the state;

[0110] The measured value of the vehicle state at the current moment is obtained, and the estimated value of the vehicle state at the current moment is corrected according to the measured value of the vehicle state at the current moment and the Kalman gain to obtain the optimal estimated value of the vehicle state at the current moment. The expression is as follows:

[0111]

[0112] Among them, x kal (k) is the optimal estimate of the state variable at time k, is the estimated value of the state variable at time k, is the measured value of the state variable at time k.

[0113] like Figure 3 As shown, first, the optimal control input u(k-1) at time k-1 is obtained through the MPC controller. In the state update process, there is an inaccurate state calculation and an inaccurate state observable Then, the Kalman filter is used to filter the noisy state variables, and the state variables obtained by Kalman filtering are used as parameters to update the Kalman filter. Finally, the state variables x obtained by Kalman filtering are used kal (k) Substitute the prediction model into the MPC control framework to enter a new round of rolling optimization to obtain the optimal system input at the next moment.

[0114] In step S5 of this embodiment, since MPC solves an optimization problem in each control cycle and obtains a control sequence, only the first control quantity of this sequence is executed, and then this process is repeated in the next cycle. Through the rolling optimization method, MPC can adaptively respond to changes in the system, thereby achieving precise control of the vehicle queue.

[0115] The effect of the method of this embodiment is verified by specific experiments below.

[0116] Assume that the vehicle engine time constant is t = 0.4s, the vehicle body length is l = 3m, and the system sampling time is t s = 0.1s. The constraints of the system are specified according to actual needs, as shown below:

[0117] Table 1. Constraint parameter settings

[0118]

[0119] Set the weight of the slack variable to p=3, and set the output and input slack variable coefficients according to Table 3-2.

[0120] Table 2. Slack variable coefficient settings

[0121]

[0122] Considering the calculation error and measurement error of the vehicle during driving, and assuming that they all obey the normal distribution, the specific parameters of each error term are set as follows.

[0123] Table 3. Error variable settings

[0124]

[0125] In addition, the system output matrix C = diag (1, 1, 1, 1) is set, that is, the system state is directly output. The fixed time interval safety distance algorithm is used, and the minimum safety distance d0 = 5m is set. Then the controller optimization parameters of the model are set as follows.

[0126] Table 4. Control parameter settings

[0127]

[0128] In this embodiment, several indicators are selected to evaluate the performance of the CAV queue under different control strategies:

[0129] (1) The maximum value of the square of the relative speed of the following vehicle (e v ) 2 max and the maximum value of the square of the gap error (e s ) 2 max , and its calculation formula is:

[0130]

[0131] Where T is the total simulation time, which is 150s here.

[0132] (2) The sum of the squares of the relative speeds of the following vehicles is sum(e v ) 2 The sum of squares of the gap error is sum(e s ) 2 , and its calculation formula is:

[0133]

[0134] In order to demonstrate the significance of the KL-MPC method, a vehicle platoon consisting of five vehicles including a leading vehicle is considered and a simulation experiment is performed using Python. The leading vehicle uses the LSTM test set data in the above-mentioned dataset as its acceleration value, and the initial states of all vehicles are set to zero. The simulation time is 150 seconds. Figure 4 It is the acceleration curve of the leading vehicle during the simulation period.

[0135] The acceleration curves of the four following vehicles are shown in the figure. Figure 5 is the acceleration curve of the K-MPC method, Figure 6 is the acceleration curve of the L-MPC method, Figure 7 is the acceleration curve of the KL-MPC method.

[0136] As can be seen from the figure, the acceleration of the following vehicle using the K-MPC method fluctuates frequently. The reason is that the four following vehicles cannot obtain the acceleration information of the vehicle in front of them in advance, so they use the acceleration of the vehicle in front of them at the previous moment instead of the acceleration in the future prediction time domain, which causes the control amount calculated each time to fluctuate greatly, thereby causing the acceleration to oscillate. Using the L-MPC method, the first following vehicle cannot accurately obtain the information of the leading vehicle, which increases the uncertainty of the model during the control process, causing the system control amount to oscillate violently, increasing the fluctuation of the acceleration. However, the acceleration change amplitude of all vehicles in the KL-MPC method is the smallest and the acceleration value is maintained in a relatively low range. In addition, the acceleration fluctuation amplitude of the four vehicles decreases with the upstream direction of the queue. From the results in the figure, it can be basically concluded that the method has chord stability, that is, the acceleration change of the leading vehicle will not be amplified with the upstream direction of the queue. Therefore, compared with the K-MPC method and the L-MPC method, the KL-MPC method can effectively reduce the fluctuation of the actual acceleration of the vehicle, indicating that the Kalman filter combined with the LSTM prediction model can effectively filter out external interference signals, which can not only improve the anti-interference ability of the control system, but also improve the ride comfort performance.

[0137] The following vehicle gap error values ​​of the four following vehicles during the simulation period are shown in the figure. Figure 8 is the gap error curve of the K-MPC method, Fig. 9 is the gap error curve of the L-MPC method, Fig.10 is the gap error curve of the KL-MPC method.

[0138] As can be seen from the figure, the vehicle gap error curve of the K-MPC method is similar to its acceleration curve, with small fluctuations and more frequent fluctuations. The error curve of the first following vehicle of the L-MPC method is relatively chaotic, while the vehicle gap error curve of the KL-MPC method combines the advantages of the two methods, with small fluctuations and regular shapes. Furthermore, the sum(e s ) 2 is 49.92, (e s ) 2 max is 0.30. The sum of the first following vehicle of L-MPC (e s ) 2 is 55.89, (e s ) 2 max is 0.32. The sum(e s ) 2 is 48.06, (e s ) 2 max The result shows that the introduction of LSTM to predict the acceleration of the preceding vehicle is more in line with the actual situation than assuming that the preceding vehicle's speed remains unchanged in the prediction time domain. On this basis, the addition of the Kalman filter to filter the calculation error and measurement error can derive a more reasonable control quantity, thereby achieving the effect of reducing the relative spacing error between vehicles.

[0139] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter, characterized in that: The following steps are involved: Establish a vehicle longitudinal dynamics model that takes into account the speed change of the leading vehicle and use it as a prediction model for the MPC control framework; Use the long short-term memory network model to predict the acceleration value of the preceding vehicle at each moment based on the preceding vehicle's historical acceleration information; Substitute the acceleration prediction value of the preceding vehicle at each moment into the prediction model of the MPC control framework to calculate the control quantity at the corresponding moment and add it to the control sequence; If the preceding vehicle is a HDV, the estimated value of the vehicle state at each moment is calculated, and then the estimated value of the vehicle state at each moment is processed by Kalman filtering to obtain the optimal estimated value of the vehicle state at each moment. Finally, the optimal estimated value of the vehicle state at each moment is substituted into the prediction model of the MPC control framework. If the vehicle in front is a connected autonomous vehicle (CAV), the actual value of the vehicle state at each moment is obtained in real time, and then the actual value of the vehicle state at each moment is substituted into the prediction model of the MPC control framework; The vehicle is longitudinally controlled according to a specified control amount in a control sequence in each control cycle.

2. The vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter according to claim 1 is characterized in that: When establishing a vehicle longitudinal dynamics model that takes into account the speed change of the preceding vehicle, it includes: For each vehicle in the connected autonomous driving vehicle fleet, a corresponding single-vehicle longitudinal dynamic model is established by considering the speed change of the vehicle in front of each vehicle, and then the single-vehicle longitudinal dynamic model of each vehicle is integrated to obtain the dynamic model of the connected autonomous driving vehicle fleet, and used as the vehicle longitudinal dynamic model considering the speed change of the vehicle in front; Discretize the longitudinal dynamic model of each vehicle, set the prediction time domain of the model equal to the control time domain and consider the acceleration change of the leading vehicle to obtain the corresponding state prediction equation; Comfort, efficiency and control input are taken as optimization control objectives at the same time, and the objective function and constraint conditions of the vehicle longitudinal dynamics model considering the speed change of the leading vehicle are constructed.

3. The vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter according to claim 2 is characterized in that: The expression of the longitudinal dynamic model of the single vehicle is as follows: Among them, X is the system state variable, U is the system control input, d is the system disturbance, and u n is the system input, A, B1, B2 are the state space equation coefficient matrices, a n-1 is the acceleration of the vehicle in front of it, a n ,v n are the difference between the actual distance between the vehicle and the front vehicle and the expected distance, the relative speed with the front vehicle, the vehicle's own acceleration, speed, n is the vehicle's serial number in the connected autonomous driving vehicle fleet, τ n is the vehicle's engine time constant, t d is the expected time interval.

4. The vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter according to claim 3 is characterized in that: The expression of the dynamic model of the networked autonomous driving vehicle fleet is as follows: A i 4×4 =A;i=1,2,…,N Among them, X P =(Δe1 s ,Δe1 v ,a1,v1,Δe2 s ,Δe2 v ,a2,v2,…,Δe N s ,Δe N v ,a N ,v N ), is the system state variable of each vehicle in the connected autonomous vehicle fleet, where Δe N s ,Δe N v ,a N ,v N are the difference between the actual distance between the Nth vehicle in the connected autonomous driving vehicle fleet and the vehicle in front and the expected distance, the relative speed with the vehicle in front, the vehicle’s own acceleration, and the speed; A P 、B1 P 、B2 P are the coefficient matrices of the augmented state space equation, U P =(u1,u2,…,u N ) is the system control input for each vehicle in the connected autonomous vehicle fleet, where u N is the system input for the Nth vehicle in the connected autonomous vehicle fleet, d p =(a0,a1,…,a N-1 ), is the system disturbance of each vehicle in the connected autonomous driving vehicle fleet, where a N-1 is the acceleration of the Nth vehicle in the connected autonomous vehicle fleet.

5. The vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter according to claim 4 is characterized in that: The state prediction equation is expressed as follows: D=[A j-1 B2 A j-2 B2…B2]; in, are the coefficient matrices of the state space equation after discretizing the continuous time. express The jth power, A j-1 express j-1th power, I is the 4×4 identity matrix, t s is the sampling time of the system, u(k|k) represents the predicted state input at time k, u(k+j-1|k) represents the predicted control input at time k for time k+j-1, x(k+j|k) represents the state prediction of the system at time k for time k+j, w(j) is the interference input; w(k+j-1|k) represents the acceleration of the preceding vehicle at the j-1th time point in the future, j=1,2…,N.

6. The vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter according to claim 5 is characterized in that: After obtaining the corresponding state prediction equation, the method further includes: changing the state prediction equation to relax the assumption that the acceleration of the preceding vehicle remains unchanged in the prediction time domain. The expression of the changed state prediction equation is as follows: Among them, u(k-1) is the optimal control quantity at time k-1, Δu(k|k) represents the state input increment at time k, Δu(k+j-1|k) represents the predicted control input increment from time k to time k+j-1, y(k+j|k) represents the measurable output of the system, and C is the unit matrix.

7. The vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter according to claim 6 is characterized in that: The expression of the objective function is as follows: The constraint condition expression is as follows: where y ref (k+N|k) is the reference trajectory, represents a quadratic function, Q and R are the error and input weighting matrices respectively, ε is the slack variable, p is the weight of the slack variable, Δu(k+j|k) represents the predicted control input increment from time k to time k+j, are the lower and upper limits of the artificially set parameters corresponding to the measurable output of the system, y min ,y max are the minimum and maximum values ​​of the output that can be measured by the system, are the lower and upper limits of the artificially set parameters corresponding to the predictive control input, u min 、u max are the minimum and maximum values ​​corresponding to the predicted control input, They are the lower and upper limits of the artificially set parameters corresponding to the predictive control input increment, Δu min , Δu max are the minimum and maximum values ​​corresponding to the predictive control input increment respectively.

8. The vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter according to claim 7 is characterized in that: The estimated value of the vehicle state at each moment is expressed as follows: in is the estimated value of the state variable at time k, u(k-1) is the optimal control variable at time k-1, ω(k-1) is the acceleration of the preceding vehicle at time k-1, x kal (k-1) is the optimal estimate of the state variable at time k-1 after Kalman filtering.

9. The vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter according to claim 1, characterized in that: When the vehicle state estimation value at each moment is processed by Kalman filtering, the following steps are included: Calculate the covariance of the optimal estimate of the state variable at the previous moment and calculate the Kalman gain at the current moment. The expression is as follows: Among them, K K is the Kalman gain at time k, is the covariance P(k-1) of the optimal estimate of the state variable at time k-1 and the covariance Q of the Gaussian noise θ The sum, R ξ is the measured value covariance, C is the proportional coefficient of the state variable, Update the matrix for the state; The measured value of the vehicle state at the current moment is obtained, and the estimated value of the vehicle state at the current moment is corrected according to the measured value of the vehicle state at the current moment and the Kalman gain to obtain the optimal estimated value of the vehicle state at the current moment. The expression is as follows: Among them, x kal (k) is the optimal estimate of the state variable at time k, is the estimated value of the state variable at time k, is the measured value of the state variable at time k.

10. The vehicle longitudinal model predictive control method integrating long short-term memory network and Kalman filter according to claim 1, characterized in that: The designated control quantity in the control sequence of each control cycle specifically refers to the first control quantity in the control sequence.

Images