Ecological driving method, system and equipment in roundabout scene and medium

Abstract

The invention discloses an ecological driving method, system and device in a roundabout scene and a medium, and the method comprises the steps: obtaining the initial state data of each network-connected vehicle in the roundabout scene, building an ecological driving model based on a KAN network, and setting a composite reward function, the ecological driving model comprises a control main network and a target network which are the same in structure, the composite reward function comprises a safety function, a driving efficiency function, a driving comfort function, a low-speed reward and punishment function, an energy consumption function and a destination arrival reward and punishment function; and carrying out training optimization on the ecological driving model based on the initial state data, inputting the real-time state space into the trained ecological driving model, outputting an optimal safety reference speed, and carrying out driving control on each networked vehicle based on the optimal safety reference speed. The method is suitable for an intelligent network connection environment, the problems of energy conservation and emission reduction are concerned, and the driving efficiency, safety, driving comfort, energy-saving effect and the like of the network connection vehicle passing through the roundabout are obviously improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic driving, and in particular relates to an ecological driving method, system, equipment and medium in a roundabout scenario. Background Art

[0002] In the complex urban traffic environment, traditional intersections have gradually failed to meet the needs of urban development. Although roundabouts are an alternative solution to alleviate intersection congestion, they can reduce traffic conflicts and improve traffic efficiency, but their common "butterfly effect" still limits this advantage: when vehicles in the roundabout do not give way to vehicles outside, local traffic jams will spread rapidly, causing frequent vehicle starts and stops, increased energy consumption, and ultimately causing congestion in the entire traffic system. This phenomenon is often difficult to detect and deal with in a timely manner, thus exacerbating the congestion problem.

[0003] In order to solve the congestion problem of roundabouts, researchers have proposed a centralized traffic management method. This method uses a centralized controller to guide vehicles through roundabouts safely and efficiently through real-time control. However, with the increase in the number of vehicles, the centralized management system is prone to overload due to the increased complexity of processing information, resulting in a decrease in its reliability and robustness. In recent years, with the development of connected and automated vehicle (CAV) technology, vehicles can receive surrounding traffic information through V2X (vehicle to everything) communication. With the help of V2X communication equipment on smart connected vehicles, researchers use deep reinforcement learning to perceive surrounding traffic information, train smart connected vehicles, control the throttle, brake or steering wheel of connected vehicles, and effectively solve the problem of connected vehicles passing through roundabouts safely and efficiently in multi-vehicle environments. Through deep reinforcement learning training, the system can respond more flexibly in complex traffic environments, and its reliability and robustness are improved. Although these methods have made progress in improving the efficiency, safety and comfort of vehicles passing through roundabouts, they have not fully considered the needs of ecological driving such as energy saving and emission reduction. Summary of the invention

[0004] The purpose of the present invention is to provide an eco-driving method, system, device and medium in a roundabout scenario to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides an eco-driving method in a roundabout scenario, comprising:

[0006] Acquire the initial state data of each connected vehicle in the roundabout scenario, wherein the initial state data includes the longitudinal speed, lateral speed, acceleration, speed difference and distance difference between the connected vehicle and the front and rear vehicles;

[0007] An eco-driving model is constructed based on the KAN network and a composite reward function is set, wherein the eco-driving model includes a main control network and a target network, wherein the main control network and the target network have the same structure and both include an Actor-KAN network and a Critic-KAN network connected in sequence; the composite reward function includes a safety function, a driving efficiency function, a driving comfort function, a low-speed reward and punishment function, an energy consumption function, and a destination arrival reward and punishment function;

[0008] Inputting the initial state data as the initial state space into the eco-driving model for training and optimization to obtain a trained eco-driving model;

[0009] The real-time state space is input into the trained ecological driving model, and the optimal safety reference speed is output, and driving control of each connected vehicle is performed based on the optimal safety reference speed.

[0010] Optionally, the training optimization process of the ecological driving model specifically includes:

[0011] Step 1: input the initial state space into the control main network to obtain an initial action space, obtain an initial safety reference speed based on the initial action space, apply the initial safety reference speed to each connected vehicle and obtain the next moment state space of each connected vehicle, store the initial state space, initial action space, reward function value and next moment state space in an experience pool, and repeat step 1 until the experience pool reaches a preset saturation state;

[0012] Step 2: When the experience pool reaches a preset saturation state, the Critic-KAN network in the control main network obtains the state space from the experience pool for evaluation, obtains a first evaluation result, and updates the parameters of the Actor-KAN network in the control main network based on the first evaluation result;

[0013] Step 3: The target network obtains the next state space at the current moment from the experience pool, and calculates the action space at the next moment and the second evaluation result; and updates the parameters of the ecological driving model based on the first evaluation result and the second evaluation result;

[0014] Repeat steps 1 to 3 for iterative update until the preset number of iterations is reached, stop the iteration, and obtain the trained ecological driving model.

[0015] Optionally, the process of acquiring the initial safety reference speed specifically includes:

[0016] The state data is input as an initial state space into the Actor-KAN network in the control main network, an initial action is output, the initial action is scaled to obtain an initial action space, and an initial safety reference speed is obtained based on the initial action space.

[0017] Optionally, the step 2 specifically includes:

[0018] When the experience pool reaches a preset saturation state, the state space and the current action space at the current moment are input into the Critic-KAN network in the main control network for evaluation to obtain a first evaluation result, and the Actor-KAN network in the main control network is parameter updated based on the first evaluation result to obtain an Actor-KAN network with updated parameters.

[0019] Optionally, the step 3 specifically includes:

[0020] The Actor-KAN network in the target network obtains the next state space at the current moment from the experience pool, and calculates the action space and reward function value at the next moment based on the obtained next state space;

[0021] The next state space, the next action space and the corresponding reward function value are input into the Critic-KAN network in the target network for evaluation to obtain the second evaluation result;

[0022] With the goal of minimizing the mean square error between the second evaluation result and the first evaluation result, updating the Critic-KAN network in the control main network to obtain an updated control main network;

[0023] Based on the soft update strategy, the parameters in the updated control main network are updated to the target network to obtain the updated target network.

[0024] An eco-driving system in a roundabout scenario, comprising:

[0025] A data acquisition module is used to obtain the initial state data of each connected vehicle in the roundabout scenario, wherein the initial state data includes the longitudinal speed, lateral speed, acceleration, speed difference and distance difference with the front and rear vehicles of each connected vehicle;

[0026] A model training module is used to construct an eco-driving model according to a KAN network and set a composite reward function, wherein the eco-driving model includes a control main network and a target network, wherein the control main network and the target network have the same structure and both include an Actor-KAN network and a Critic-KAN network connected in sequence; the composite reward function includes a safety function, a driving efficiency function, a driving comfort function, a low-speed reward and punishment function, an energy consumption function, and a destination arrival reward and punishment function; the initial state data is input as an initial state space into the eco-driving model for training optimization to obtain a trained eco-driving model;

[0027] The eco-driving control module is used to input the real-time state space into the trained eco-driving model, output the optimal safety reference speed, and perform driving control on each connected vehicle based on the optimal safety reference speed.

[0028] An electronic device includes a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute an ecological driving method in a roundabout scenario.

[0029] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the ecological driving method in a roundabout scenario.

[0030] The technical effects of the present invention are:

[0031] The present invention provides an ecological driving method based on the AC-KAN network in a roundabout scenario. The method is not only applicable to the intelligent network environment, but also pays attention to the problem of energy conservation and emission reduction, so that the driving efficiency, safety, driving comfort, energy-saving effect, etc. of the networked vehicle through the roundabout are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0033] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0034] Figure 1 It is an overall framework diagram in an embodiment of the present invention;

[0035] Figure 2is a diagram of the AC-KAN network structure in an embodiment of the present invention;

[0036] Figure 3 Schematic diagram of the structure of the key parts inside the KAN network in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0038] It should be understood that the terms described in the present invention are only for describing special embodiments and are not intended to limit the present invention. In addition, for the numerical range in the present invention, it should be understood that each intermediate value between the upper and lower limits of the scope is also specifically disclosed. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the described range is also included in the present invention. The upper and lower limits of these smaller ranges can be independently included or excluded in the scope.

[0039] Unless otherwise indicated, all technical and scientific terms used herein have the same meanings as those of ordinary skill in the art generally understood by the present invention. Although the present invention describes only preferred methods, any method similar or equivalent to that described herein may also be used in the implementation or testing of the present invention. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods related to the documents. In the event of a conflict with any incorporated document, the content of this specification shall prevail.

[0040] It will be apparent to those skilled in the art that various modifications and variations may be made to the specific embodiments of the present invention description without departing from the scope or spirit of the present invention. Other embodiments derived from the present invention description will be apparent to those skilled in the art. The present application description and examples are exemplary only.

[0041] The words “include,” “including,” “have,” “contain,” etc. used in this article are open-ended terms, meaning including but not limited to.

[0042] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0043] Embodiment 1

[0044] like Figure 1 - Figure 3As shown, in this embodiment, an eco-driving method in a roundabout scenario is provided, including: obtaining initial state data of each connected vehicle in the roundabout scenario, the initial state data including the longitudinal speed, lateral speed, acceleration, speed difference and distance difference between the connected vehicles and the front and rear vehicles; constructing an eco-driving model based on the KAN network and setting a composite reward function, the eco-driving model including a control main network and a target network, the control main network and the target network have the same structure, and both include an Actor-KAN network and a Critic-KAN network connected in sequence; the composite reward function includes a safety function, a driving efficiency function, a driving comfort function, a low-speed reward and punishment function, an energy consumption function and a destination arrival reward and punishment function; inputting the initial state data as the initial state space into the eco-driving model for training optimization to obtain a trained eco-driving model; inputting the real-time state space into the trained eco-driving model, outputting an optimal safety reference speed, and performing driving control on each connected vehicle based on the optimal safety reference speed.

[0045] This embodiment provides an eco-driving method based on the AC-KAN network in a roundabout scenario, which can not only improve the driving efficiency, safety and driving comfort of the vehicle, but also effectively reduce energy consumption.

[0046] This embodiment mainly includes the following steps:

[0047] Step 1: Design the state space. In the roundabout scenario, the longitudinal speed, lateral speed, acceleration, and speed difference and distance between the connected vehicle and the front and rear vehicles are set as the state space.

[0048] Step 2: Design an eco-driving model. Design an eco-driving model based on the AC-KAN network and use the state space of step 1 as the input of the model.

[0049] Step 3: Design a composite reward function. A reward function is developed for the eco-driving model in step 2, which includes a safety function, a driving efficiency function, a driving comfort function, a low-speed reward and penalty function, an energy consumption function, and a destination-reaching reward and penalty function.

[0050] Step 4: Design safe actions. Set the acceleration of the connected vehicle as the action space and use it as the output of the eco-driving model in step 2. Combine the safe speed and anti-idling strategy to scale the action and obtain the safe reference speed to achieve safe control of the connected vehicle.

[0051] First, in the intelligent networked roundabout scenario, this embodiment uses V2X communication technology to obtain the traffic information of the networked vehicle and parses the state space. Before the experience pool is saturated, the ecological driving model based on the AC-KAN network is in the experience accumulation period. Only the initial network is used to calculate the input state space to obtain the action output, and the action is scaled in combination with the safe speed and anti-idling strategy to obtain the safe reference speed, so as to realize the intelligent control of the networked vehicle. At the same time, the state space, action space, reward function, next state space, round end flag and other information parsed from the traffic information are stored in the experience pool to accumulate experience. When the experience pool is saturated, the ecological driving model based on the AC-KAN network enters the sampling learning period. Through the deep calculation of the input features such as the state space, action space, reward function, next state space, etc., the network is continuously updated to generate the action space with the highest evaluation value, and the optimal safe reference speed of the networked vehicle in the roundabout scenario is obtained by combining the safe speed and anti-idling strategy. Repeat the above process to finally obtain the optimal ecological driving model.

[0052] The specific implementation process of this embodiment includes:

[0053] Step 1: Design the state space. In the roundabout scenario, the longitudinal speed, lateral speed, acceleration, and speed difference and distance between the connected vehicle and the front and rear vehicles are set as the state space. The state space is expressed as follows:

[0054]

[0055] in, is the longitudinal speed of the connected vehicle, is the lateral speed of the connected vehicle, is the acceleration of the connected car, is the speed difference between the connected vehicle and the vehicle in front, is the speed difference between the connected vehicle and the vehicle behind, is the distance difference between the connected vehicle and the vehicle in front, is the distance difference between the connected vehicle and the vehicle behind, n is the number of model iterations, and i is the current time step. This information can be obtained through V2X technology and GPS (Global Positioning System), and the real-time data is guaranteed.

[0056] Step 2: Design an eco-driving model. Design an eco-driving model based on the AC-KAN network and use the state space of step 1 as the input of the model. Figure 2 The following is a schematic diagram of the AC-KAN network structure of this embodiment. The specific steps are as follows:

[0057] (1) Experience accumulation period. The Actor-KAN network obtains the state space of the connected vehicle at the current time step 1 from the roundabout scene. And the original action value is calculated by the KAN network. In order to ensure that the action value is within the appropriate range, the action value is scaled to obtain the action space in step 4 The safe reference speed is obtained by combining the safe speed and anti-idling strategy, and is applied to the connected vehicles in the roundabout scene. Store the state space obtained from the roundabout scene Action Space Reward function value r i n , the next state space And the number of iterations is marked as done. Before the experience pool is saturated, the current state space is equal to the next state space, that is, Realize the experience accumulation cycle process.

[0058] in, Indicates that the model performs an action After that, the state space obtained from the roundabout scene is recorded as done indicates the number of model iterations. When done = True, the model will proceed to the next iteration. When done = Flash, the model will continue to accumulate experience.

[0059] (2) Sampling learning period. When the experience pool is saturated, the Critic-KAN network obtains the state space of the current time step from the experience pool. Obtaining action space from Actor-KAN network Calculate the state-action pair The goal of the Critic-KAN network is to evaluate Next execution action space Based on the value feedback information of the Critic-KAN network, the Actor-KAN network actor The function updates its own parameters to generate the action space corresponding to the next time step. The value is larger.

[0060] in, The value is the Critic-KAN network's estimate of the value of the current state and action, reflecting the expected cumulative reward that can be obtained by executing the action under the current strategy. The value is calculated as follows, The schematic formula for the output of the Critic-KAN network is:

[0061]

[0062] Among them, L actor The formula is as follows, N is the training batch size, and its purpose is to maximize value:

[0063]

[0064] (3) Target network learning period. When the experience pool is saturated, the target Actor-KAN network obtains the next state space of the current time step from the experience pool. The scaled target action space is calculated through the KAN network The target Critic-KAN network obtains the next state space of the current time step from the experience pool and the reward function value Get the target action space from the target Actor-KAN network Calculate the goal for this state-action-reward pair value.

[0065] Among them, the target The value is calculated based on the actual reward and the estimated value of the target network, representing a more accurate expected reward, which is used to guide the Critic-KAN network to learn so that its estimate is closer to the true value. The value is calculated as follows: The schematic formula for the target Critic-KAN network output is:

[0066]

[0067] (4) Network update period. The ecological driving model is updated through L critic Function Minimization Values ​​and Goals The mean square error of the values ​​is used to update the parameters of the Critic-KAN network, thereby improving its evaluation accuracy of the action space quality. The target Actor-KAN network and the target Critic-KAN network use a soft update strategy to gradually update the parameters of the Actor-KAN network and the Critic-KAN network to the target network with a smaller step size τ to ensure the stability of the training. Repeat the above steps to obtain the optimal ecological driving model.

[0068] Among them, L critic The formula is as follows, N is the training batch size, y is the intermediate variable, and γ is the discount factor, which is generally 0.99:

[0069]

[0070]

[0071] Among them, the formula for parameter soft update is as follows, θ target is the target network parameter, θcurrent is the current network parameter, step size τ=0.005:

[0072] θ target ←τθ current +(1-τ)θ target (7)

[0073] Among them, the internal key structure of the KAN network included in all the above networks is shown as follows Figure 3 The environment adaptive KAN network construction method in this embodiment has the following specific implementation steps:

[0074] (1) Designing adaptive nodes: Obtaining the state space of step 1 before building the KAN network and the action space of step 4 If the Actor-KAN network only uses the state space The number of dimensions is the number of nodes in the input layer, and the state space As input value, if the Critic-KAN network uses the state space Plus action space The dimension and number of nodes as input layer, state space and action space The number of nodes in the middle layer is the number of nodes in the input layer plus 1, and the number of layers in the middle layer uses a 4-layer structure. These settings are more in line with the construction form of a multi-layer KAN network. The number of nodes in the output layer uses the action space The number of dimensions.

[0075] (2) Define spline basis functions: The KAN network uses B-spline basis functions as spline basis functions. Some B-spline basis functions are as follows: Figure 3 As shown, it is a function composed of a group of piecewise polynomials and has good local control and continuity. Due to the high complexity of the state space of step 1 of this embodiment, a third-order B-spline basis function is used, and its formula is as follows:

[0076]

[0077] Where, j is the number of B-spline basis functions, k is the order of the B-spline basis function, x is the function input value, t is the node of the B-spline function, B j,k (x) represents the jth B-spline basis function.

[0078] (3) KAN network learning update: In the KAN network, the output y of each layer KAN It is calculated by nonlinear combination of input value x. Each layer f in the network m The transformation is performed through a spline basis function, and the formula is as follows:

[0079]

[0080] Among them, m represents the number of layers of KAN, n KAN Represents the dimension of weight, ω j Represents the weights of the spline basis functions.

[0081] In the KAN network, the loss function L is defined as KAN , which depends on the mean square error between the output of the KAN network and the true value in the experience pool, and its formula is as follows:

[0082]

[0083] Among them, N KAN represents the batch size, h represents the sample index value, is the final output value of formula (9), is the true value in the experience pool, here is the action space

[0084] Use gradient descent to update weights ω in the KAN network j and the knot t of the spline basis function j , the formula is as follows:

[0085]

[0086]

[0087] Among them, ρ is the network learning rate, generally 0.001.

[0088] When the Actor-KAN and Critic-KAN networks and the corresponding target networks are updated, the parameters in the KAN network are trained and updated synchronously to ensure that the network learning process complements each other. Each training usually includes 30 steps of updates, and the number of steps can be adjusted according to the complexity of the task. During the training process, the KAN network optimizes the parameters by introducing regularization and sparsification techniques. Regularization helps constrain the size of the network weights to prevent overfitting, while sparsification makes the network more concise and efficient by reducing unnecessary connections or nodes. This optimization method not only improves the generalization ability of the network, but also improves the computational efficiency when deploying the ecological driving model.

[0089] Step 3: Design a composite reward function: Formulate a reward function for the eco-driving model in step 2. The function formula is as follows:

[0090]

[0091] in, is the security function, is the driving efficiency function, is the driving comfort function, is the energy consumption function, is a low-speed reward and penalty function, is the reward and penalty function for reaching the destination.

[0092] The specific description of each function is as follows:

[0093] (1) Safety function: This embodiment uses the sum of the lane deviation of the connected vehicle and TTC (vehicle collision time) to represent the safety of the current action. The formula of the safety function is as follows:

[0094]

[0095]

[0096]

[0097] Among them, α safe is a scaling factor, which is set by calculation and manual adjustment safe =10, is the lane deviation reward and penalty function, is the TTC reward and punishment function. is the lateral distance of the vehicle relative to the center of the lane, L width is the lane width. and is the distance difference and speed difference between the connected vehicle and the vehicle in front in the state space of step 1, and D is the distance threshold between the connected vehicle and the front and rear vehicles. Since the diameter of the roundabout generally does not exceed 100m, it is more appropriate to set D=50 here. When the distance to the front vehicle exceeds 50m, is close to 0, so No effect. On the contrary, when the distance to the front vehicle is less than or equal to 50m, the formula (14) is used to calculate

[0098] (2) Driving efficiency function: This embodiment encourages the connected vehicle to drive as efficiently as possible within the speed limit according to its speed. The formula of the driving efficiency function is as follows:

[0099]

[0100] Among them, α efficient is a scaling factor, which is set by calculation and manual adjustment efficient =40, v limit is the road speed limit, is the longitudinal velocity of the connected vehicle in the state space of step 1.

[0101] (3) Driving comfort function: This embodiment uses the square of the acceleration change of the connected vehicle to represent the degree of turbulence caused to passengers by the vehicle's speed change behavior. The formula of the driving comfort function is as follows:

[0102]

[0103] Among them, α comfort is a scaling factor, which is set by calculation and manual adjustment comfort =1, is the acceleration in the state space of step 1, is the acceleration of the previous time step in the state space, and t(i) represents the time step size.

[0104] (4) Energy consumption function: Since the vehicle used in this embodiment is a new energy electric vehicle with an energy recovery function, this embodiment calculates the reward value based on the power consumption and driving distance of the networked vehicle, using the quotient of the power consumed per unit time step and the driving distance. If the driving distance of the vehicle in this time step is less than 0.01m, that is, the vehicle is moving slowly, a fixed negative reward value of 50 is set, which can effectively solve the problem of vehicle idling. The formula of the energy consumption function is as follows:

[0105]

[0106] Among them, α energy is a coefficient of variation, which is set by calculation and manual adjustment. energy =15, is the power consumed in the current time step, is the distance traveled in the current time step.

[0107] (5) Low-speed reward and penalty function: This embodiment gives penalties based on the speed value of the connected vehicle. When the vehicle is obviously driving at a low speed, the lower the speed, the greater the penalty. The formula of the low-speed reward and penalty function is as follows:

[0108]

[0109] Among them, α idle is a coefficient of variation, which is set by calculation and manual adjustment. idle =120, is the longitudinal velocity of the connected vehicle in the state space of step 1.

[0110] (6) Reward and Penalty Function for Arriving at Destination: This embodiment first obtains the driving route of the connected vehicle, and then locks the last lane in the route, that is, the target lane. When the distance traveled by the connected vehicle in the target lane is greater than or equal to the length of the target lane, it is determined that the connected vehicle has arrived at the destination. At this moment, arrive_flash=True is set, otherwise arrive_flash=False. When the connected vehicle arrives at the destination, a fixed positive reward value is given, which is set to 50, otherwise it is 0. The formula for the reward and penalty function for arriving at the destination is as follows:

[0111]

[0112] Among them, arrive_flash represents the mark value when the connected vehicle arrives at the destination.

[0113] Step 4: Design safe actions: In this embodiment, the acceleration of the connected vehicle is set as the action space of the eco-driving model in step 2. Considering the physical meaning of safety and speed, the acceleration output by the algorithm cannot be directly applied to the movement of the vehicle in some cases and needs to be constrained. Therefore, it is necessary to scale the action value output by the eco-driving model first, as shown below:

[0114]

[0115] The scaled action space is set as follows, which is a one-dimensional vector:

[0116]

[0117] The safety action formula after constraint is as follows:

[0118]

[0119] in, is the original output value of the eco-driving model, a min , a max are the maximum deceleration and maximum acceleration of the vehicle, v limit is the road speed limit, v carfollow is the speed calculated by the default car-following model (Krauss model), is the reference speed value that the intelligent driving model finally applies to the intelligent network scene. The acceleration multiplied by the unit time is t(i). and When the connected vehicle is considered to be in idling state, the default following speed v is used. carfollow , and in other cases, safe speed limits are imposed. Repeat the above steps to train the optimal intelligent driving model.

[0120] This embodiment provides an ecological driving method based on the AC-KAN network in a roundabout scenario. The method is not only applicable to the intelligent network environment, but also pays attention to the issues of energy conservation and emission reduction, so that the driving efficiency, safety, driving comfort, energy-saving effect, etc. of the networked vehicle through the roundabout are significantly improved.

[0121] An eco-driving system in a roundabout scenario, comprising:

[0122] A data acquisition module is used to obtain the initial state data of each connected vehicle in the roundabout scenario, wherein the initial state data includes the longitudinal speed, lateral speed, acceleration, speed difference and distance difference with the front and rear vehicles of each connected vehicle;

[0123] A model training module is used to construct an eco-driving model according to a KAN network and set a composite reward function, wherein the eco-driving model includes a control main network and a target network, wherein the control main network and the target network have the same structure and both include an Actor-KAN network and a Critic-KAN network connected in sequence; the composite reward function includes a safety function, a driving efficiency function, a driving comfort function, a low-speed reward and punishment function, an energy consumption function, and a destination arrival reward and punishment function; the initial state data is input as an initial state space into the eco-driving model for training optimization to obtain a trained eco-driving model;

[0124] The eco-driving control module is used to input the real-time state space into the trained eco-driving model, output the optimal safety reference speed, and perform driving control on each connected vehicle based on the optimal safety reference speed.

[0125] An electronic device includes a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute an ecological driving method in a roundabout scenario.

[0126] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the ecological driving method in a roundabout scenario.

[0127] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An eco-driving method in a roundabout scenario, characterized in that: include: Acquire the initial state data of each connected vehicle in the roundabout scenario, wherein the initial state data includes the longitudinal speed, lateral speed, acceleration, speed difference and distance difference between the connected vehicle and the front and rear vehicles; An eco-driving model is constructed based on the KAN network and a composite reward function is set, wherein the eco-driving model includes a main control network and a target network, wherein the main control network and the target network have the same structure and both include an Actor-KAN network and a Critic-KAN network connected in sequence; the composite reward function includes a safety function, a driving efficiency function, a driving comfort function, a low-speed reward and punishment function, an energy consumption function, and a destination arrival reward and punishment function; Inputting the initial state data as the initial state space into the eco-driving model for training and optimization to obtain a trained eco-driving model; The real-time state space is input into the trained ecological driving model, and the optimal safety reference speed is output, and driving control of each connected vehicle is performed based on the optimal safety reference speed.

2. The eco-driving method in a roundabout scenario according to claim 1, characterized in that: The training optimization process of the ecological driving model specifically includes: Step 1: input the initial state space into the control main network to obtain an initial action space, obtain an initial safety reference speed based on the initial action space, apply the initial safety reference speed to each connected vehicle and obtain the next moment state space of each connected vehicle, store the initial state space, initial action space, reward function value and next moment state space in an experience pool, and repeat step 1 until the experience pool reaches a preset saturation state; Step 2: When the experience pool reaches a preset saturation state, the Critic-KAN network in the control main network obtains the state space from the experience pool for evaluation, obtains a first evaluation result, and updates the parameters of the Actor-KAN network in the control main network based on the first evaluation result; Step 3: The target network obtains the next state space at the current moment from the experience pool, and calculates the action space at the next moment and the second evaluation result; and updates the parameters of the ecological driving model based on the first evaluation result and the second evaluation result; Repeat steps 1 to 3 for iterative update until the preset number of iterations is reached, stop the iteration, and obtain the trained ecological driving model.

3. The eco-driving method in a roundabout scenario according to claim 2, characterized in that: The process of obtaining the initial safety reference speed specifically includes: The state data is input as an initial state space into the Actor-KAN network in the control main network, an initial action is output, the initial action is scaled to obtain an initial action space, and an initial safety reference speed is obtained based on the initial action space.

4. The eco-driving method in a roundabout scenario according to claim 2, characterized in that: The step 2 specifically includes: When the experience pool reaches a preset saturation state, the state space and the current action space at the current moment are input into the Critic-KAN network in the main control network for evaluation to obtain a first evaluation result, and the Actor-KAN network in the main control network is parameter updated based on the first evaluation result to obtain an Actor-KAN network with updated parameters.

5. The eco-driving method in a roundabout scenario according to claim 2, characterized in that: The step 3 specifically includes: The Actor-KAN network in the target network obtains the next state space at the current moment from the experience pool, and calculates the action space and reward function value at the next moment based on the obtained next state space; The next state space, the next action space and the corresponding reward function value are input into the Critic-KAN network in the target network for evaluation to obtain the second evaluation result; With the goal of minimizing the mean square error between the second evaluation result and the first evaluation result, updating the Critic-KAN network in the control main network to obtain an updated control main network; Based on the soft update strategy, the parameters in the updated control main network are updated to the target network to obtain the updated target network.

6. An eco-driving system for a roundabout scenario, characterized in that: include: A data acquisition module is used to obtain the initial state data of each connected vehicle in the roundabout scenario, wherein the initial state data includes the longitudinal speed, lateral speed, acceleration, speed difference and distance difference with the front and rear vehicles of each connected vehicle; A model training module is used to construct an eco-driving model according to a KAN network and set a composite reward function, wherein the eco-driving model includes a control main network and a target network, wherein the control main network and the target network have the same structure and both include an Actor-KAN network and a Critic-KAN network connected in sequence; the composite reward function includes a safety function, a driving efficiency function, a driving comfort function, a low-speed reward and punishment function, an energy consumption function, and a destination arrival reward and punishment function; the initial state data is input as an initial state space into the eco-driving model for training optimization to obtain a trained eco-driving model; The eco-driving control module is used to input the real-time state space into the trained eco-driving model, output the optimal safety reference speed, and perform driving control on each connected vehicle based on the optimal safety reference speed.

7. An electronic device, characterized in that: It includes a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute an ecological driving method in a roundabout scenario according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements an ecological driving method in a roundabout scenario as described in any one of claims 1 to 5.

Images