Method and system for intelligently grading ecological driving strategies of intelligent network-connected vehicles

Abstract

The invention relates to a networked vehicle ecological driving intelligent grade division method and system, which are used for providing individual element support for scale deployment of ecological driving strategies and improving road traffic efficiency. According to the method, preliminary division of ecological driving levels is realized by controlling the number of information extracted by an intelligent agent on state features and limiting the degree of freedom of decision making of a control strategy on road driving; for the ecological driving grade intelligent agents subjected to preliminary division, setting the number of information for extracting different states for the intelligent agent grades of different grades, and setting two reward functions, namely a self basic reward and a reward sum of the self basic reward and the interference of the self vehicle to surrounding vehicles; a vehicle internal energy management strategy is set, energy consumption solving and power distribution are carried out on an upper-layer vehicle longitudinal control value, and an energy consumption solving value participates in reward function calculation; and selecting corresponding reward functions according to the divided grades, and training in groups to obtain final different ecological driving grade intelligent agents after completion.

Description

Technical Field

[0001] This invention relates to the field of eco-driving for new energy vehicles, and more particularly to a method for classifying the intelligent levels of eco-driving strategies for intelligent connected vehicles. Background Technology

[0002] Most existing research on eco-driving focuses on evaluating the driving behavior of the vehicle itself or surrounding vehicles, lacking an assessment of the impact of intelligent vehicle driving behavior on traffic flow, making it difficult to simulate road traffic efficiency. Performance optimization applied in scenarios essentially pushes eco-driving strategies in a self-serving direction, that is, optimizing only for the vehicle itself. While this may optimize the vehicle's driving efficiency, energy consumption, and stability, it often focuses too much on the vehicle itself, easily deteriorating the driving efficiency, energy consumption, and stability of surrounding vehicles. For example, in a congested scenario, a sudden deceleration might be taken at a certain moment to achieve better energy consumption, potentially causing following vehicles to decelerate sharply, thus worsening the driving stability of surrounding vehicles.

[0003] While multi-agent optimization for specific scenarios introduces the expression of individual behavioral differences in the traffic system, most still optimize macro-level operational indicators (such as average system delay / average speed, road segment throughput, total system energy consumption / total emissions, and safety constraints) in a centralized or global reward manner. It fails to start from individual behavioral patterns to hierarchically characterize ecological driving goals, and also lacks a mechanism to explicitly incorporate the interactive effects of vehicle behavior on other traffic participants, such as speed disturbances and obstruction deceleration, into evaluation and learning objectives in an interpretable way. Therefore, although multi-agent analysis can distinguish different intelligent vehicles in traffic flow, it still does not discuss the impact of surrounding vehicles. Summary of the Invention

[0004] To address the lack of an individual behavior pattern-based approach in existing technologies, this invention aims to propose a method for classifying intelligent levels in an ecological driving strategy for intelligent connected vehicles. By classifying ecological driving levels, the method obtains the impact of different levels of intelligent vehicles on local traffic flow. This provides support for understanding the impact of intelligent vehicles of different levels mixed in different proportions and existing with different penetration rates on local and overall traffic flow, as well as discussing the impact of intelligent output actions on macro-traffic flow. This facilitates more effective traffic management by relevant departments based on the impact effects.

[0005] To achieve the aforementioned objectives, an intelligent connected vehicle ecological driving strategy intelligent level classification method is proposed. This method achieves preliminary ecological driving level classification by controlling the amount of information extracted from state features by the intelligent agent and limiting the decision-making freedom of the control strategy during road driving. For the intelligent agents at the preliminary ecological driving level, different information extraction quantities are set for different levels of the intelligent agent, and two reward functions are defined: a basic reward for the agent itself and a reward sum of the basic reward and the reward for the vehicle's interference with surrounding vehicles. Energy consumption is calculated and power is allocated for the longitudinal control values ​​of the upper-level vehicle, with the energy consumption solution participating in the reward function calculation. For each level classification, the corresponding reward function is selected, and the agents are trained in groups to obtain the final intelligent agents at different ecological driving levels.

[0006] In one embodiment of the above technical solution, the feature acquisition step for the intelligent agent used for decision-making includes: obtaining a gridded relative velocity matrix based on the vehicle from the surrounding traffic flow. and a gridded road location matrix based on the vehicle A spatial matrix is ​​constructed, from which a spatial feature matrix is ​​extracted. This matrix is ​​then combined with the feature tensor of the state space obtained based on road behavior safety and traffic light efficiency. The features are then stitched together; the stitched features are then reduced and dimensionality lowered.

[0007] In one embodiment of the above technical solution, the state space , Let be the relative speed difference between the vehicle and the vehicle in front at the current moment. Let be the difference in relative acceleration between the vehicle and the vehicle in front at the current moment. This represents the relative distance between the vehicle and the vehicle in front at the current moment. The time from the traffic light in front of the vehicle until it turns green again. This is the remaining time of the green light at this moment.

[0008] In one embodiment of the above technical solution, the levels are initially divided into four levels: Level 1 agents only consider vehicles within a distance threshold in their current lane during state feature extraction, extracting the state information of all surrounding vehicles as the state input for their own vehicle; Level 2 agents extract the state feature information of all vehicles within a distance threshold in front of and behind their current lane, and extract the traffic flow operation state as input through feature processing; Level 3 and Level 4 agents perceive vehicles in surrounding lanes and extract the state feature information of all vehicles in both surrounding lanes and their own lane within a distance threshold as input; the reward functions for Level 1, Level 2, and Level 3 agents are as follows: , Let the objective function of the agent be . Rewards related to the energy consumption of intelligent agents. Its coefficient, The reward for movement during state transitions is greater the distance the agent moves forward. As a penalty for collision, This represents the difference between the vehicle's current speed and the average speed of the current traffic flow. With its coefficient, this reward enables the agent to maintain a relatively reasonable speed control range during the initial stage after a traffic light and during the driving process, thereby reducing the traffic interference of the vehicle on surrounding vehicles; the reward function for a Level 4 agent is as follows: , Penalties for following vehicles slowing down due to longitudinal control and for following vehicles slowing down when merging into a lane during lane-changing decisions. Its coefficient.

[0009] In one embodiment of the above technical solution, an adaptive fuzzy neural network is used to formulate the shape of the fuzzy control membership function for the power system at different intelligence levels, adjust the power distribution to manage the energy flow, and extend the life of the battery and fuel cell.

[0010] The intelligent connected vehicle ecological driving intelligent level classification system implemented according to the above-mentioned method and technical solution includes the following modules: A preliminary classification module is configured to control the amount of information extracted by the intelligent agent from state features and limit the decision-making freedom of the control strategy on the road to achieve preliminary classification of ecological driving levels; a setting module is configured to set different information extraction quantities for different levels of intelligent agents in the preliminary ecological driving level classification, and to set two reward functions: a basic reward and a sum of the basic reward and the reward for interference from surrounding vehicles; an energy consumption solution module is configured to solve for the energy consumption of the longitudinal control values ​​of the upper-level vehicles and perform power allocation, wherein the energy consumption solution value participates in the reward function calculation; and a training module selects the corresponding reward function for each classification level, groups the agents for training, and obtains the final intelligent agents of different ecological driving levels. The beneficial technical effects of this disclosure are: the vehicle intelligent ecological driving strategy proposed in this disclosure can provide individual element support for the large-scale deployment of ecological driving strategies, and is an extension of reinforcement learning ecological driving strategies oriented towards interaction, which plays a crucial role in improving road traffic efficiency. Attached Figure Description

[0011] Figure 1 , one A schematic diagram of the method flow in one implementation method.

[0012] Figure 2 , one A schematic diagram of the network structure for processing traffic flow characteristics of eco-driving in one implementation method.

[0013] Figure 3 , one A schematic diagram of the surrounding vehicle interference attention mechanism in one implementation method. Detailed Implementation

[0014] This invention addresses the application of intelligent ecological driving strategies in the context of connected vehicle development, and the future road traffic system will be composed of intelligent vehicles embedded with different strategies. It proposes an ecological driving level classification method to provide individual behavior models for predicting traffic flow development based on different elements. This method relies on a reinforcement learning framework, integrating vehicle driving state with surrounding vehicle and traffic signal information. It pre-classifies the degrees of freedom of action output for agents of different intelligence levels, sets targeted reward function components to train the agents to achieve the level classification goal, and incorporates an attention mechanism to visualize agent decision-making. This has practical significance for analyzing the individual element behavior of traffic systems under large-scale deployment.

[0015] The implementation steps of the technical solution of this invention are briefly described below: S1. Based on V2V and V2I communication, the status of surrounding vehicles and traffic light information are obtained. For the status of surrounding vehicles in traffic flow, CNN is used to reduce the dimensionality of features, clarify the status information of the ecological driving strategy, and add an attention mechanism to the state extraction layer to encode the features of surrounding vehicles, so as to realize the visualization of the basis for subsequent decision-making behavior.

[0016] S2. By controlling the amount of information extracted by the intelligent agent from state features and limiting the decision-making freedom of the control strategy on the road, the initial classification of ecological driving level is achieved.

[0017] S3. For the ecological driving level intelligent agents initially classified in S2, different information extraction quantities are set for different levels of intelligent agents, and two reward functions are set: one is the agent's own basic reward, and the other is the agent's own basic reward plus a reward for interference from surrounding vehicles. These are then trained in groups. The interference reward is calculated when a surrounding vehicle enters the feature extraction distance range of the agent in traffic flow.

[0018] The formula for calculating the basic reward is as follows: ,in, Let the objective function of the agent be . Rewards related to the energy consumption of intelligent agents. Its coefficient. In order to fairly reflect energy consumption information during the extraction of ecological driving control strategies, the energy consumption reward is adopted as the ratio of the equivalent energy consumption of the vehicle in the previous transition to the distance the agent moves during the state transition. This setting can effectively improve the forward tendency problem of the agent during training. This will serve as the movement reward during the state transition, and it is also the main reward during agent training. The greater the distance the agent moves forward, the greater the reward. As a collision penalty, the negative maximum value is taken during the training process to make the agent learn illegal operations. This represents the difference between the vehicle's current speed and the average speed of the current traffic flow. With its coefficient, this reward enables the agent to maintain a relatively reasonable speed control range during the initial stage after a traffic light and during the driving process, thereby reducing the traffic interference of the vehicle on surrounding vehicles.

[0019] S4. Set the vehicle's internal energy management strategy, solve for the energy consumption of the upper-level vehicle longitudinal control value and allocate power. The energy consumption solution value is used as the specific value of the upper-level driving control layer sub-reward function in S3 for training.

[0020] S5. Based on the initial classification of levels in S2, select the corresponding reward function, group and train them, and obtain the final intelligent agents of different ecological driving levels.

[0021] S6. By comprehensively considering the energy consumption optimization performance, traffic efficiency, and overall impact of road behavior on traffic flow operation of different levels of intelligent agent strategies, the overall impact assessment results are used as the verification of the proposed decision.

[0022] The following will be combined with the appendix Figure 1-3 This application provides a clear and complete description of how the technical solution in this case is implemented. Obviously, the described implementation methods are only a part of the implementation methods in this case, not all of them. Based on the implementation methods in this case, all other implementation methods obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0023] This invention proposes an eco-driving intelligence level classification method, which is an extension of reinforcement learning-based eco-driving strategies oriented towards interaction. The indicators related to the interaction between the vehicle and surrounding vehicles are represented based on an attention mechanism. In this reinforcement learning process, convolutional neural networks are used for feature dimensionality reduction of multi-source heterogeneous traffic information. Eco-driving levels are pre-classified by restricting the state extraction dimension and constraining the action degrees of freedom. Then, different types of agents are trained using different group reward functions to obtain eco-driving strategies with different model parameter distributions. The lower-level energy management strategy is implemented based on the upper-level speed solution value. In this process, the attention mechanism can provide the difference in attention weights between the vehicle and surrounding vehicles, providing a visualization tool for distinguishing eco-driving strategies. The obtained different strategies are then examined within traffic flow to evaluate their road behavior's interference with traffic flow and their own energy consumption optimization performance. This method can provide individual element support for the large-scale deployment of eco-driving strategies.

[0024] The overall flowchart is as follows: Figure 1 As shown, this includes traffic information acquisition and feature extraction, as well as the definition and classification of intelligent levels.

[0025] (I) Traffic Information Acquisition and Feature Extraction In traffic information acquisition and feature extraction, features of surrounding vehicle traffic flow information and key information tensors in the state space are acquired based on the intelligent traffic environment. The two are then concatenated as vectors, and the interaction weights between surrounding vehicles and the vehicle are obtained using a dual-head attention mechanism. These interaction weights are used to visualize the interactions between these vehicles at that moment.

[0026] The state space consists of quantities representing road safety and traffic light efficiency. The key information tensor in the state space is denoted as I. ,in Let be the relative speed difference between the vehicle and the vehicle in front at the current moment. Let be the difference in relative acceleration between the vehicle and the vehicle in front at the current moment. This represents the relative distance between the vehicle and the vehicle in front at the current moment. This represents the time from when the traffic light in front of the vehicle turns green until the next green light. If the traffic light in front is currently green, fill with 0. This is the remaining time of the green light at this moment. If the traffic light phase is not green at this moment, it is filled with the maximum green light phase duration.

[0027] Obtain a gridded relative speed matrix based on the vehicle from the surrounding traffic flow information. and a gridded road location matrix based on the vehicle Grid the relative velocity matrix and gridded road location matrix After feature dimensionality reduction, the vector is concatenated with the key information tensor I in the state space. The concatenated vector is then used as the state input for defining and classifying the intelligence level.

[0028] To effectively utilize traffic flow information, the state space for feature extraction via convolutional neural networks is defined as follows:

[0029]

[0030]

[0031] In the formula, Indicates the vehicle's current speed. These are the lane index position relative to the vehicle and the gridded road position index relative to the vehicle, respectively. To satisfy the boundary communication conditions and the information margin range of all surrounding vehicles. Let i represent the surrounding vehicle at lane index position i relative to its own vehicle and gridded road position j relative to its own vehicle, with a speed of . . Indicates the vehicle's speed.

[0032] The attention mechanism described above specifically employs an embedded dual-head attention mechanism to visualize the basis for the agent's action output. This dual-head attention mechanism consists of two single-head attention mechanisms with different feature weights. The single-head attention mechanism extracts traffic state features as follows. The concatenated vector is denoted as... The multilayer perceptron (MLP) is used for the following preprocessing to unify the dimensions, as shown in the equation: .

[0033] The vector obtained after encoding Attention calculations are performed as follows: , , .

[0034] in, These represent Query, key, and value, respectively. These are their parameter matrices. In the attention module, the self-carrier will output the query. Used to access other vehicles This represents the attention weight of the ego vehicle towards surrounding vehicles. Set denoted as Each entity value is assigned an attention weight. The overall process of attention mechanism is as follows: .

[0035] Attention Matrix Calculate by dot product, then by coefficient. Scaling is used to normalize the output. Softmax is chosen as the activation function to ensure the representation of the neural network output and to help visualize the attention weights in subsequent processes.

[0036] The dual-head attention mechanism splices two attention heads together. ,in Join the two attention points together. To output the matrix, after processing multi-source traffic information, the attention module outputs the attention weights of surrounding vehicles (summing up to 1), which can be used to visualize the interactions of these vehicles at that moment. The process of dual-head attention for extracting and visualizing road vehicle information is described in the reference section. Figure 2 .

[0037] The advantages of applying attention mechanisms can be summarized as follows: (1) It can effectively enhance the correlation between training objectives and rewards.

[0038] (2) This feature processing method can extract information of key vehicles in the traffic flow matrix, assign attention weights to vehicles in front and behind and to the vehicle itself, and enhance the interpretability of the decision-making basis of the control network.

[0039] (3) This processing method can effectively adaptively adjust the weights of the feature extraction network to effectively analyze the rationality of the road behavior of intelligent vehicles and the feasibility of intelligent level classification.

[0040] (II) Definition and Classification of Intelligence Levels To clarify the control process of the algorithm utilizing traffic information, the upper-level control decision-making process is as follows, with the specific network structure as follows: Figure 3 As shown.

[0041] The feature extraction process begins by extracting features from the spatial matrix.

[0042]

[0043] In the formula, It integrates spatial features of the relative velocity matrix and relative position matrix based on the vehicle, where V and S are velocity and position features. After the spatial features are defined and stitched together in three dimensions, the output is performed through the following control framework.

[0044]

[0045] in splicing feature output network, This is a convolutional neural network targeting spatial features. After extracting high-dimensional features through two-dimensional convolution and applying regularization methods, the output is given. It is worth noting that... Add the feature tensor I of the state space in (1) to the middle, and... Input after splicing Complete the reduction and dimensionality reduction of features. The mapping from current features to actions enables speed planning and lane-changing decisions for connected vehicle ecosystem driving strategies.

[0046] Eco-driving levels are pre-classified by restricting the acceleration and lane-changing degrees of freedom of the intelligent agent in driving decisions. The classification is based on two criteria: first, restrictions on the range of features; and second, restrictions on the degrees of freedom of actions. As follows: Level 1: The agent can consider the driving style of the vehicle ahead, and in state feature extraction, it only considers vehicles within a certain distance threshold in its current lane, extracting all surrounding vehicle state information as its own state input. Furthermore, considering its own energy consumption and safety constraints, the agent is restricted to longitudinal movement only, achieving continuous control of longitudinal acceleration. It does not consider the impact of its own actions on following vehicles.

[0047] Level 2: The intelligent agent can extract the state feature information of all vehicles within the distance threshold in front of and behind the vehicle's current lane, and extract the traffic flow operation status as input through feature processing. Compared with Level 1, it enhances the consideration of traffic information, but the degree of freedom of action is still limited to the longitudinal range.

[0048] Level 3: The agent can perceive vehicles in surrounding lanes and extract state feature information of all vehicles in surrounding lanes (including its own lane) within a distance threshold. It uses a hybrid state feature extraction method to achieve more comprehensive motion freedom for longitudinal control and lane-changing decisions, but still does not consider the impact on surrounding vehicles.

[0049] Level 4: Building upon Level 3, this level calculates the differences in speed models regarding the impact of the vehicle's road behavior (acceleration, deceleration, lane changing) on ​​surrounding vehicles. These parameters are then incorporated into the vehicle's behavior penalty, and the penalty coefficient is further refined based on Level 4 to account for the interference from surrounding vehicle behavior.

[0050] The reward function differs for ecological driving intelligent agents classified at different levels.

[0051] The reward function for Level 1, 2, and 3 agents is applicable to the absence of interference and includes their own basic reward.

[0052] For a Level 4 intelligent driving agent, the reward function consists of its own reward plus a sub-function for interference terms. This applies to situations where there is interference from surrounding vehicles. The definition is as follows:

[0053] Among them, adding As a corresponding penalty for the intelligent vehicle's road behavior and interference from surrounding vehicles, the main considerations are the following: the following vehicle's obstruction and deceleration behavior caused by longitudinal control, and the following vehicle's deceleration penalty when merging into the lane during lane-changing decisions. The coefficients are used to define the different behaviors of vehicles in the engineering process that are disturbed by surrounding vehicles. The coefficients are then divided into intervals to obtain the reinforcement learning ecological driving strategy that distinguishes the driving level four. A reward function consisting of itself and interference.

[0054] The self-reward function and the self + interference reward function are applied to the corresponding ecological driving agents for group training, resulting in final ecological driving level agents. Specifically, the upper-level velocity and acceleration values ​​are obtained from the output of the trained ecological driving agents, and the vehicle's driving power requirement is calculated based on the vehicle's longitudinal dynamics formula.

[0055]

[0056]

[0057] in, Indicates the driving force of a car. m Represents the quality of fuel cell hybrid buses. This represents the vehicle's driving gradient angle. The value represents the vehicle's drag coefficient, A represents the vehicle's frontal area, and V represents the vehicle's speed. This represents the overall vehicle rotational drag coefficient.

[0058] An adaptive fuzzy neural network is used to formulate the shape of the fuzzy control membership function for the power system at different intelligence levels, adjusting power distribution to manage energy flow and extend the lifespan of the battery and fuel cell. Energy distribution is performed by considering the operation of energy-consuming components within the vehicle; the energy distribution process is as follows:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] in, For fuel cell hydrogen consumption, Hydrogen has a low calorific value. For fuel cell output power, This is a lookup function for fuel cell efficiency; Characterizing the degree of fuel cell lifespan degradation, All are coefficients. For the number of start-stop cycles of the fuel cell, For the high power output time of fuel cells, For the low power output time of fuel cells, For the degree of change in fuel cell output power, The permissible lifespan degradation level of the fuel cell; The battery current is given by SOC, which characterizes the battery's state of charge. The lookup table formula for the open-circuit voltage corresponding to the battery's state of charge (SOC) is as follows. This refers to the resistance corresponding to the battery's state of charge (SOC). Relevant data can be found in a table. This refers to the battery's output power. This is the initial SOC value of the battery. Battery capacity; For the instantaneous hydrogen consumption of the battery, This is a lookup table function for the efficiency corresponding to the battery's output power. A lookup table function for the efficiency corresponding to the battery charging power. This refers to the average output efficiency of the fuel cell. This is the equivalent hydrogen consumption of the battery. Due to battery life loss, For coefficients, For activation energy, The molar gas constant, For temperature, For the battery to accumulate charge, The exponential coefficient; Due to battery capacity loss, The rated charge / discharge rate of the battery. This refers to the actual charge / discharge rate of the battery. For total battery loss, All are coefficients.

[0068] In summary, this disclosure proposes a method for classifying the intelligent driving strategy of intelligent connected vehicles into different levels. This method achieves initial classification of the ecological driving level by controlling the amount of information extracted from state features by the intelligent agent and limiting the decision-making freedom of the control strategy during road driving. For the intelligent agents at the initially classified ecological driving levels, different amounts of information extracted from different states are set for different agent levels, and two reward functions are defined: a basic reward for the agent itself and a sum of the basic reward and the reward for the vehicle's interference with surrounding vehicles. Energy consumption is calculated for the longitudinal control values ​​of the upper-level vehicle, and power allocation is performed, with the energy consumption solution participating in the reward function calculation. For each classified level, a corresponding reward function is selected, and the agents are trained in groups to obtain the final intelligent agents at different ecological driving levels.

[0069] Based on the above-described method and technical solution, those skilled in the art can implement a corresponding intelligent connected vehicle ecological driving intelligent level classification system. For example, the intelligent connected vehicle ecological driving intelligent level classification system includes: a preliminary classification module, configured to control the amount of information extracted by the intelligent agent from state features and limit the decision-making freedom of the control strategy on the road to achieve a preliminary classification of ecological driving levels; a setting module, configured to set different amounts of information extracted from different states for the ecological driving level intelligent agents after preliminary classification, and to set two reward functions: a basic reward for the agent itself and a sum of the basic reward and the reward for interference from surrounding vehicles; an energy consumption solution module, configured to solve for the energy consumption of the longitudinal control values ​​of the upper-level vehicles and perform power allocation, wherein the energy consumption solution value participates in the reward function calculation; and a training module, which selects the corresponding reward function for each classification level, groups the agents for training, and obtains the final intelligent agents of different ecological driving levels after completion.

[0070] Through the above description of the embodiments, those skilled in the art can clearly understand that this disclosure can be implemented using software plus necessary general-purpose hardware, or it can be implemented using dedicated hardware including dedicated integrated circuits, dedicated CPUs, dedicated memory, dedicated components, etc. Generally, any function performed by a computer program can be easily implemented using corresponding hardware, and the specific hardware structure used to implement the same function can be diverse, such as analog circuits, digital circuits, or dedicated circuits. However, for this disclosure, software implementation is more often a preferred implementation method.

[0071] It should be noted that the terms "one embodiment," "another embodiment," and "embodiment" used in this specification refer to specific features, structures, or characteristics described in connection with that embodiment, which are included in at least one embodiment described in the general description of this application. The appearance of the same expression in multiple places in the specification does not necessarily refer to the same embodiment. Furthermore, when a specific feature, structure, or characteristic is described in connection with any embodiment, the intention is to suggest that implementing such a feature, structure, or characteristic in conjunction with other embodiments also falls within the scope of this invention.

[0072] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A method for classifying the intelligent levels of an intelligent connected vehicle's ecological driving strategy, characterized in that: By controlling the amount of information extracted by the intelligent agent from state features and limiting the decision-making freedom of the control strategy on the road, a preliminary classification of ecological driving levels can be achieved. For the ecological driving level intelligent agents that have been initially classified, the amount of information extracted from different states is set for different levels of intelligent agents, and two reward functions are set, namely the self-base reward and the sum of the self-base reward and the reward for the vehicle's interference with surrounding vehicles. The energy consumption of the longitudinal control values ​​of the upper-level vehicles is solved and power is allocated. The energy consumption solution is used in the calculation of the reward function. Based on the classification level, the corresponding reward function is selected, and the groups are trained to obtain the final intelligent agents of different ecological driving levels.

2. The method according to claim 1, characterized in that, The features used by an agent for decision-making are obtained through the following steps: The gridded relative velocity matrix based on the vehicle, obtained from the surrounding traffic flow. and a gridded road location matrix based on the vehicle A spatial matrix is ​​constructed, from which a spatial feature matrix is ​​extracted. This matrix is ​​then combined with the feature tensor of the state space obtained based on road behavior safety and traffic light efficiency. To splice; The spliced ​​features are then reduced and dimensionality reduced.

3. The method according to claim 2, characterized in that, The state space , Let be the relative speed difference between the vehicle and the vehicle in front at the current moment. Let be the difference in relative acceleration between the vehicle and the vehicle in front at the current moment. This represents the relative distance between the vehicle and the vehicle in front at the current moment. The time from the traffic light in front of the vehicle until it turns green again. This is the remaining time of the green light at this moment.

4. The method according to claim 1, characterized in that, The initial classification is into four levels: In Level 1, the agent only considers vehicles within the current lane ahead of the vehicle in the state feature extraction process, and extracts all surrounding vehicle state information as the state input of the vehicle. In Level 2, the agent extracts the state feature information of all vehicles within the distance threshold before and after the current vehicle in the lane, and extracts the traffic flow operation status as input through feature processing. The agents in Level 3 and Level 4 perceive vehicles in the surrounding lanes and extract the state feature information of all vehicles in the surrounding lanes and their own lane as input within the distance threshold. The reward functions for Level 1, Level 2, and Level 3 agents are as follows: , Let the objective function of the agent be . Rewards related to the energy consumption of intelligent agents. Its coefficient, The reward for movement during state transitions is greater the distance the agent moves forward. As a penalty for collision, This represents the difference between the vehicle's current speed and the average speed of the current traffic flow. With its coefficient, this reward can enable the agent to maintain a relatively reasonable speed control range during the starting phase after a traffic light and during the driving process, thereby reducing the traffic interference of the vehicle to surrounding vehicles. The reward function for a Level 4 agent is as follows: , Penalties for following vehicles slowing down due to longitudinal control and for following vehicles slowing down when merging into a lane during lane-changing decisions. Its coefficient.

5. The method according to claim 1, characterized in that, By using an adaptive fuzzy neural network to formulate the shape of the fuzzy control membership function for power systems at different intelligence levels, the power distribution is adjusted to manage energy flow and extend the life of batteries and fuel cells.

6. A smart connected vehicle ecological driving intelligence level classification system, characterized in that: The initial classification module is configured to achieve the initial classification of ecological driving levels by controlling the amount of information extracted by the intelligent agent from state features and limiting the decision-making freedom of the control strategy on the road. The setting module is configured to extract different amounts of information from different states for the initially divided ecological driving level intelligent agents, and to set two reward functions: the basic reward of the agent itself and the sum of the basic reward of the agent itself and the reward for interference from surrounding vehicles. The energy consumption solution module is configured to solve the energy consumption of the longitudinal control values ​​of the upper-level vehicle and perform power allocation, wherein the energy consumption solution values ​​are used in the calculation of the reward function. The training module selects the corresponding reward function for each level, groups them for training, and obtains the final intelligent agents of different ecological driving levels after completion.

7. The system according to claim 6, characterized in that, The features of an entity used for decision-making, and the steps to obtain them include: The gridded relative velocity matrix based on the vehicle, obtained from the surrounding traffic flow. and a gridded road location matrix based on the vehicle A spatial matrix is ​​constructed, from which a spatial feature matrix is ​​extracted. This matrix is ​​then combined with the feature tensor of the state space obtained based on road behavior safety and traffic light efficiency. To splice; The spliced ​​features are then reduced and dimensionality reduced.

8. The system according to claim 7, characterized in that: The state space , Let be the relative speed difference between the vehicle and the vehicle in front at the current moment. Let be the difference in relative acceleration between the vehicle and the vehicle in front at the current moment. This represents the relative distance between the vehicle and the vehicle in front at the current moment. The time from the traffic light in front of the vehicle until it turns green again. This is the remaining time of the green light at this moment.

9. The system according to claim 6, characterized in that, The initial classification is into four levels: In Level 1, the agent only considers vehicles within the current lane ahead of the vehicle in the state feature extraction process, and extracts all surrounding vehicle state information as the state input of the vehicle. In Level 2, the agent extracts the state feature information of all vehicles within the distance threshold before and after the current vehicle in the lane, and extracts the traffic flow operation status as input through feature processing. The agents in Level 3 and Level 4 perceive vehicles in the surrounding lanes and extract the state feature information of all vehicles in the surrounding lanes and their own lane as input within the distance threshold. , Let the objective function of the agent be . Rewards related to the energy consumption of intelligent agents. Its coefficient, The reward for movement during state transitions is greater the distance the agent moves forward. As a penalty for collision, This represents the difference between the vehicle's current speed and the average speed of the current traffic flow. With its coefficient, this reward can enable the agent to maintain a relatively reasonable speed control range during the starting phase after a traffic light and during the driving process, thereby reducing the traffic interference of the vehicle to surrounding vehicles. The reward function for a Level 4 agent is as follows: , Penalties for following vehicles slowing down due to longitudinal control and for following vehicles slowing down when merging into a lane during lane-changing decisions. Its coefficient.

10. The system according to claim 6, characterized in that, By using an adaptive fuzzy neural network to formulate the shape of the fuzzy control membership function for power systems at different intelligence levels, the power distribution is adjusted to manage energy flow and extend the life of batteries and fuel cells.

Images