An intelligent robust control method for a mixed-disturbance-oriented heterogeneous connected vehicle platoon

Abstract

The application discloses a kind of intelligent robust control methods for mixed disturbance-oriented heterogeneous networked vehicle fleet, the method first comprehensively input delay, uncertain parameter and external disturbance, establishes the third-order longitudinal dynamics model of vehicle i.Secondly based on the third-order longitudinal dynamics model established, select control target, design upper coordination control strategy based on deep reinforcement learning algorithm.Finally based on the output of upper coordination control strategy, design sliding mode following controller and multi-objective disturbance observer, complete vehicle fleet control.The application increases auxiliary control input to compensate complex disturbance, avoids the problem of difficult training while not sacrificing the overall performance of vehicle fleet, and significantly improves the adaptability and response speed to complex environment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent driving, specifically relating to an intelligent robust control method for heterogeneous connected vehicle fleets facing mixed disturbances. Background Technology

[0002] In the development of intelligent connected vehicle fleets, effective vehicle control strategies are crucial for achieving safe and efficient road traffic. Despite significant progress in autonomous driving technology in recent years, existing fleet control methods still have limitations in handling input delays, uncertain parameters, and external disturbances in complex traffic environments. Particularly in longitudinal dynamics control, existing technologies often struggle to balance stability, safety, ride comfort, and efficient following.

[0003] Furthermore, traditional control strategies, such as model-based control and PID control, often exhibit insufficient adaptability in complex and dynamically changing driving environments. To overcome these challenges, deep reinforcement learning offers an effective solution. It leverages machine learning techniques to optimize long-term, complex decision-making processes, making it particularly suitable for handling high-dimensional control tasks. However, deep reinforcement learning methods alone still have room for improvement in ensuring real-time response and system robustness. Against this backdrop, the introduction of sliding mode control becomes particularly important. As an effective nonlinear control strategy, sliding mode control is renowned for its strong robustness to uncertainties and dynamic systems. Therefore, combining deep reinforcement learning and sliding mode control can not only optimize the control strategy of intelligent connected vehicle fleets but also significantly improve adaptability and response speed to complex environments. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes an intelligent robust control method for heterogeneous intelligent connected electric vehicle (HEV) platoons under complex disturbances. This method aims to significantly improve the overall performance of HEV platoons in complex road environments, particularly in terms of stability, safety, comfort, and efficiency, through a combination of deep reinforcement learning and sliding mode control. This hierarchical control method not only optimizes the control structure but also provides an important technological foundation for the future development of intelligent transportation systems. The specific design follows these steps:

[0005] Step 1: Taking into account input delay, uncertain parameters, and external disturbances, the third-order longitudinal dynamics model of vehicle i is established as follows:

[0006]

[0007] Where p i (t), v i (t), a i (t) and a i,des(t) represents the position, velocity, acceleration, and desired acceleration of vehicle i, respectively; τ n It is the nominal value of the vehicle's engine lag parameter; u i (t) and D i (t) represents the control input and lumped disturbance (including input delay, uncertain parameters and external disturbance) used by vehicle i to compensate for uncertainties.

[0008] Step 2: Based on the third-order longitudinal dynamics model, select the control objective and design an upper-level coordinated control strategy based on deep reinforcement learning algorithm.

[0009] Step 2.1: Select the control target.

[0010] The actual distance d between vehicle i and the vehicle in front at time t i (t) is represented as:

[0011] d i (t)=p i-1 (t)-p i (t)-L i

[0012] Among them, L i Let be the length of vehicle i. During operation, the internal stability control objective is: to reduce... Where, d des This represents the desired fixed distance between adjacent vehicles.

[0013] To ensure vehicle safety, a minimum distance d is set. min To avoid collisions between the main vehicle and the vehicle in front; and to set a maximum distance d. max To prevent other vehicles from cutting into the queue. The safety control objective is:

[0014] d mmin <d i (t) <d max

[0015] To achieve driving comfort, the maximum acceleration value 'a' is set. max and the minimum acceleration a min The target for controlling ride comfort is:

[0016] a min i，des (t) max

[0017] Define the state error of vehicle i as:

[0018]

[0019] ​​

[0020]

[0021] The objective of efficient convoy following control is to reduce and

[0022] Step 2.2: In the Guide Actor-Critic algorithm, the input is the state variable. Including the actual distance d between vehicle i and the vehicle in front at time t i (t) and the expected distance d des Deviation between The speed deviation of vehicle i from the vehicle in front, the acceleration deviation of vehicle i from the vehicle in front, and the speed deviation of vehicle i from the lead vehicle. The output is the control variable Con. i (t), which is the expected acceleration a of vehicle i at time t. i，des (t).

[0023] To meet the above control objectives, the reward function Reward... i (S i (t), Con i (t) includes the reward portion R all and the punishment part P all The reward portion is designed as follows:

[0024] R all =α1*R string +α2*R flow +α3*R safety

[0025] Here, α1, α2, and α3 are the weights corresponding to the reward component. R string R flow and R safety The reward items corresponding to satisfying queue stability, driving following efficiency, and driving comfort are defined as follows:

[0026]

[0027]

[0028] R safety =|a i (t)|

[0029] Where k1, k2, k3, and k4 are weight parameters, and e is an exponential function.

[0030] Penalty item P all To penalize excessive spacing deviations during training and promote the algorithm to learn better spacing preservation strategies, it is designed as follows:

[0031]

[0032] Therefore, the reward function is expressed as:

[0033] Reward i (S i (t), Con i (t))=R all +P all

[0034] Step 2.3: Obtain weight parameters using a second-order homogeneous vehicle fleet as the training object. The second-order longitudinal dynamics model of the vehicle is represented as follows:

[0035]

[0036] Where, p i，ul (t), v i，ul (t) and a i，ul (t) represents the position, velocity, and acceleration of vehicle i at time t during the training process.

[0037] Step 3: Based on the output of the upper-level coordinated control strategy, design a sliding mode follower controller and a multi-objective disturbance observer to complete the fleet control. To optimize the observer's performance, a genetic algorithm for solving multi-objective optimization problems is selected to determine the optimal observer gain.

[0038] Step 3.1: Define the expected acceleration tracking error of the lower-level vehicle i as e. i，a (t), the sliding surface is s i (t).

[0039] To satisfy vehicle state constraints while ensuring the system state quickly approaches the sliding surface, the approach law is chosen as follows:

[0040]

[0041] Where k and β are positive parameters; B(v i (t)) and B(a i (t) are respectively about v i (t) and a i The barrier function of (t) is expressed as:

[0042] B(v i (t))=-η1 ln(v i (t)-v min )-ζ1(v max -v i (t))

[0043] B(a i(t))=-η2 ln(a i (t)-a min )-ζ2(a max -a i (t))

[0044] Where, η m and ζ m (m = 1, 2) is the growth rate used to control the function as it approaches the boundary, satisfying:

[0045]

[0046]

[0047] Where, η m0 and ζ m0 They are η m and ζ m The initial value; ∈1 and ∈2 are the decay rates; z i (t) represents the independent variable, z min and z max They represent z respectively i The maximum and minimum values ​​of (t).

[0048] Based on the sliding surface and the reaching law, the expression for the control input can be obtained as follows:

[0049]

[0050] in, For lumped disturbance D i The estimated value of (t) can be obtained by the following extended state observer:

[0051]

[0052] Where l1, l2, and l3 are the observer gains; The acceleration is an estimate, and the acceleration estimation error is...

[0053] Step 3.2: To optimize the observer's performance, a genetic algorithm for solving multi-objective optimization problems is selected to determine the optimal observer gain.

[0054] In continuous-time systems, we construct an observer parameter optimization problem based on NSGA-II.

[0055] Compared with the prior art, the advantages of the present invention are as follows:

[0056] (1) Taking into account input delay, uncertain parameters and external disturbances, we use second-order homogeneous vehicle teams as training objects and third-order heterogeneous vehicle teams as control objects. We increase auxiliary control input to compensate for complex disturbances, which avoids the problem of training difficulties without sacrificing the overall performance of the vehicle teams.

[0057] (2) Combining deep reinforcement learning and sliding mode control significantly improves adaptability and response speed to complex environments. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is the control framework of the method described in this invention;

[0060] Figure 2 This is a schematic diagram of the Guide Actor-Critic algorithm described in this invention;

[0061] Figure 3 This is a schematic diagram of the barrier function used in this invention within a certain range;

[0062] Figure 4 It is the trajectory curve of all following vehicles during the driving process on the simulated road;

[0063] Figure 5 It is the tracking result of the actual acceleration and expected acceleration of the following vehicle during driving on a simulated road. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] A smart and robust control method for heterogeneous connected vehicle fleets with mixed disturbances, the control framework is as follows: Figure 1 As shown, it includes the following steps:

[0066] Step 1: Taking into account input delay, uncertain parameters, and external disturbances, the third-order longitudinal dynamics model of vehicle i is established as follows:

[0067]

[0068] Where p i (t), v i (t), a i (t) and a i,des (t) represents the position, velocity, acceleration, and desired acceleration of vehicle i, respectively; τ n It is the nominal value of the vehicle's engine lag parameter; u i (t) is the control input used by vehicle i to compensate for uncertainties. D i (t) is the lumped disturbance, expressed as:

[0069]

[0070] Where, τ i ω is the actual engine lag parameter of vehicle i; ie (t) and ω id (t) represents the external disturbance and the hysteresis term caused by the input delay, respectively.

[0071] Step 2: Based on the third-order longitudinal dynamics model established in Step 1, select the control objective and design the upper-level coordinated control strategy.

[0072] Step 2.1: Select the control target

[0073] The actual distance d between vehicle i and the vehicle in front at time t i (t) is represented as:

[0074] d i (t)=p i-1 (t)-p i (t)-L i

[0075] Among them, L i Let be the length of vehicle i. During operation, the internal stability control objective is: to reduce... Where, d des This represents the desired fixed distance between adjacent vehicles.

[0076] To ensure vehicle safety, a minimum distance d is set. min To avoid collisions between the main vehicle and the vehicle in front; and to set a maximum distance d. max To prevent other vehicles from cutting into the queue. The safety control objective is:

[0077] d min <d i (t) <d max

[0078] To achieve driving comfort, the maximum acceleration value 'a' is set. max =2m / s2 and the minimum acceleration a min = -2m / s 2 The target for controlling ride comfort is:

[0079] a min i，des (t) max

[0080] Define the state error of vehicle i as:

[0081]

[0082]

[0083]

[0084] The objective of efficient convoy following control is to reduce and

[0085] Step 2.2, as follows Figure 2 As shown, in the Guide Actor-Critic algorithm, the input is the state variable. Including the actual distance d between vehicle i and the vehicle in front at time t i (t) and the expected distance d des Deviation between The speed deviation of vehicle i from the vehicle in front, the acceleration deviation of vehicle i from the vehicle in front, and the speed deviation of vehicle i from the lead vehicle. The output is the control variable Con. i (t), which is the expected acceleration a of vehicle i at time t. i，des (t).

[0086] To meet the above control objectives, the reward function Reward... i (S i (t), Con i (t) includes the reward portion R all and the punishment part P all The reward portion is designed as follows:

[0087] R all =α1*R string +α2*R flow +α3*R safety

[0088] Here, α1, α2, and α3 are the weights of the reward component. R string R flow and R safety The reward items corresponding to satisfying queue stability, driving following efficiency, and driving comfort are defined as follows: ​​

[0089]

[0090]

[0091] R safety =|a i (t)|

[0092] Among them, k1, k2, k3, and k4 are weight parameters.

[0093] Penalty item P all To penalize excessive spacing deviations during training and promote the algorithm to learn better spacing preservation strategies, it is designed as follows:

[0094]

[0095] Therefore, the reward function is expressed as:

[0096] Reward i (S i (t), Con i (t))=R all +P all

[0097] Step 2.3: Obtain weight parameters using a second-order homogeneous vehicle fleet as the training object. The second-order longitudinal dynamics model of the vehicle is represented as follows:

[0098]

[0099] Where, p i，ul (t), v i，ul (t) and a i，ul (t) represents the position, velocity, and acceleration of vehicle i at time t during the training process.

[0100] Step 3: Based on the output of the upper-level coordinated control strategy designed in Step 2, design a sliding mode follower controller and a multi-target disturbance observer to complete the fleet control.

[0101] Step 3.1, Define the expected acceleration tracking error of the lower-level vehicle i as:

[0102] e i，a (t)=a i (t)-a i，des (t)

[0103] To achieve convergence of tracking error, the sliding surface s i (t) is selected as:

[0104] s i (t)=α 1，i e i，a (t)+α2，i ∫0 t e i，a (τ)dτ

[0105] Where, α 1,i and α 2，i The parameters of the sliding surface are positive.

[0106] To satisfy vehicle state constraints while ensuring the system state quickly approaches the sliding surface, the approach law is chosen as follows:

[0107]

[0108] Where k and β are positive parameters; B(v i (t)) and B(a i (t) are respectively about v i (t) and a i The barrier function of (t) is expressed as:

[0109] B(v i (t))=-η1 ln(v i (t)-v min )-ζ1(v max -v i (t))

[0110] B(a i (t))=-η2 ln(a i (t)-a min )-ζ2(a max -a i (t))

[0111] Where, η m and ζ m (m = 1, 2) is the growth rate used to control the function as it approaches the boundary, satisfying:

[0112]

[0113]

[0114] Where, η m0 and ζ m0 They are η m and ζ m The initial value; ∈1 and ∈2 are the decay rates; z i (t) represents the independent variable, z min and z max They represent z respectively i The maximum and minimum values ​​of (t). Let v max =20, v min =0, A schematic diagram of the function within a certain range is shown below. Figure 3 As shown.

[0115] Based on the sliding surface and the reaching law, the expression for the control input can be obtained as follows:

[0116]

[0117] in, For lumped disturbance D i The estimated value of (t) can be obtained by the following extended state observer:

[0118]

[0119] Where l1, l2, and l3 are the observer gains; This is an estimate of the acceleration.

[0120] Step 3.2: To optimize the observer's performance, a genetic algorithm for solving multi-objective optimization problems is selected to determine the optimal observer gain.

[0121] In a continuous-time system, we construct an observer parameter optimization problem based on NSGA-II, with the objective function expressed as:

[0122]

[0123] J2 = T c

[0124] Where J1 represents the acceleration estimation error. The square integral of J² represents the convergence time for the estimation error to converge to a preset range. By solving this optimization problem, a set of optimal observer parameters can be obtained to reduce the estimation error and convergence time.

[0125] Example:

[0126] In this embodiment, a smart heterogeneous connected vehicle fleet consisting of one lead vehicle and four follower vehicles is constructed and controlled using the methods described in steps one through three. First, a set of convergent weight parameters is obtained using a second-order homogeneous vehicle fleet as the training object to achieve stability, driving safety, and ride comfort. Then, the acceleration of the lead vehicle is controlled between -2 m / s² and 2 m / s², and the model is replaced with the convergent weight parameters. Based on the input state, the desired acceleration of the third-order heterogeneous vehicle fleet is obtained. The trajectories of all follower vehicles are as follows... Figure 4 As shown, the distance between each following vehicle and the adjacent vehicle in front is always maintained at a safe level, with positional deviations within a bounded and reasonable range, ensuring the safety of the convoy. Figure 5In this process, the actual acceleration of each vehicle is consistent with the expected acceleration curve, and the tracking error is always controlled within 0.06m / s^2, ensuring the tracking performance of the fleet and the driving comfort.

[0127] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A smart robust control method for heterogeneous connected vehicle fleets facing mixed disturbances, characterized in that, The method comprises the following steps: Step 1, the third order longitudinal dynamics model of vehicle is established by integrating input delay, uncertain parameters and external disturbance. ​ The third-order longitudinal dynamic model is: ； wherein and are the position, velocity, acceleration and desired acceleration of the vehicle respectively; is the nominal value of the engine lag parameter of the vehicle ; and are the control input and lumped disturbance used to compensate for the uncertainties of the vehicle respectively; The lumped disturbance includes input time delay, uncertain parameters and external disturbance; Step 2, based on the third-order longitudinal dynamic model, a control target is selected, and a high-level coordinated control strategy based on a deep reinforcement learning algorithm is designed, and the specific process is as follows: Step 2.1, Time vehicle Actual distance to the vehicle in front is expressed as: ； in, For vehicles The vehicle length; during operation, the internal stability control objective is: to reduce... ,in, This represents the desired fixed distance between adjacent vehicles; The queue stability control objective is to set a minimum distance and a maximum distance The safety control objective is to ； Setting an acceleration maximum value and an acceleration minimum value ; the ride comfort control target is to ； Defining the state error of the vehicle is: ； ； ； The control target of the high efficiency of the vehicle fleet following is to reduce , and ; Step 2.2: In the Guide Actor-Critic algorithm, the input is the state variable. ,include Time vehicle Actual distance from the vehicle in front Spacing from the expected distance Deviation between ,vehicle The vehicle's speed deviation from the vehicle in front The vehicle's acceleration deviation from the vehicle in front Speed ​​deviation from the lead vehicle; output is a control variable. vehicles exist Expected acceleration at any moment ; Reward function Including a reward portion And a penalty portion ; wherein the reward portion is designed to: ； wherein, , and are reward part corresponding weights; and are reward items corresponding to meeting queue stability, driving following efficiency and driving comfort respectively, defined as: ； ； ； wherein , , , is a weight parameter, is an exponential function; penalty term A penalty term is designed to penalize the training process for cases where the distance deviation is too large. ； Therefore, the reward function is represented as: ； Step 2.3, taking a second-order homogeneous vehicle platoon as a training object to obtain neural network parameters as a high-level decision controller; Step 3, based on the output of the high-level coordinated control strategy, a sliding mode following controller and a multi-objective disturbance observer are designed to complete vehicle platoon control.

2. The intelligent robust control method for the hybrid-disturbance-oriented heterogeneous net-connected vehicle fleet according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 3.1, Define the lower vehicle The desired acceleration tracking error is: ； Slip surface is selected to be: ； wherein and is a positive sliding surface parameter; The approaching law is selected as: ； wherein is a positive parameter; and are barrier functions with respect to and respectively, denoted as: ； ； where and are used to control the rate of growth of the function near the boundary, satisfying: ； ； wherein and are initial values of and respectively; and are decay rates; denotes an argument, and denote maximum and minimum values of respectively; Based on the sliding surface and the approaching law, the control input expression is obtained as: ； wherein is an estimate of the collective disturbance is obtained from the following extended state observer: ； wherein, is an observer gain; is an estimated value of acceleration; Step 3.2, a genetic algorithm for solving a multi-objective optimization problem is selected to determine the optimal observer gain; In a continuous-time system, an observer parameter optimization problem based on NSGA-II is constructed, and the objective function is represented as: ； ； in, Indicates acceleration estimation error The square integral, This represents the convergence time for the estimation error to converge to a preset range. By solving this optimization problem, a set of optimal observer parameters can be obtained to reduce the estimation error and convergence time.

Images