Method for dynamically optimizing energy consumption of cooperative drop-off vehicles in airport drop-off area under mixed traffic environment

Abstract

The application discloses an airport drop-off area cooperative drop-off vehicle energy consumption dynamic optimization method in a mixed traffic environment, and relates to the technical field of intelligent transportation, comprising: constructing a drop-off time prediction model to predict the expected time of each vehicle arriving at each drop-off position and output a predicted drop-off time matrix of vehicle-parking space combinations; constructing a vehicle average energy consumption model under a mixed traffic flow state, calculating the instantaneous energy consumption of autonomous electric vehicles CAV and manually driven oil vehicles HDV respectively, and optimizing energy consumption in combination with a following mode; adopting a multi-agent deep deterministic policy gradient MADDPG algorithm, combining the vehicle average energy consumption model and the following model to construct a reinforcement learning system for optimizing vehicle fleet operation; and generating an optimal vehicle fleet cooperation strategy through the reinforcement learning system according to the real-time updated traffic state, so as to realize the comprehensive optimization of energy consumption, safety and passenger comfort. The application can effectively improve the energy efficiency of the airport drop-off area and relieve the traffic pressure during peak hours.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method for dynamic optimization of energy consumption of vehicles dropping off passengers in airport drop-off areas under mixed traffic conditions. Background Technology

[0002] With the advancement of technology, autonomous driving technology is gradually being applied to various transportation scenarios. However, the efficiency of airport drop-off areas, as transportation hubs, has long been a pain point in passenger experience. Currently, most airport drop-off areas still rely primarily on traditional manually driven vehicles, lacking efficient management and scheduling support for autonomous vehicles. This single-mode approach often leads to inefficiencies during peak hours, mainly because traditional manually driven vehicles are highly unpredictable, with drivers' behavior being uncertain and potentially resulting in prolonged temporary stops, lane obstruction, and other issues that reduce overall traffic flow.

[0003] Meanwhile, resource waste and chaos also exist. The mixed operation of manually driven and autonomous vehicles in mixed traffic conditions, without dedicated scheduling and guidance schemes, leads to empty vehicle occupation, redundant dispatching, and vehicle queues, further exacerbating congestion and resulting in a poor passenger experience. Passengers need to quickly complete boarding and alighting operations in limited spaces. In peak traffic scenarios such as airport drop-off areas, relying solely on manual management methods and traditional facility designs cannot meet the diverse transportation needs of the future. Therefore, in mixed driving and autonomous driving scenarios, the lack of intelligent guidance and collaboration mechanisms easily leads to chaos and wasted time.

[0004] Insufficient existing technology is also a limiting factor. Currently, most airport drop-off areas are designed only for traditional manually driven vehicles, lacking facilities or mechanisms to provide dedicated services for autonomous vehicles. Because the dwell time of manually driven vehicles is not fixed, and the precise scheduling capabilities of autonomous vehicles are difficult to utilize, resource utilization in mixed scenarios decreases. At the same time, parking areas and dynamic scheduling systems generally lack data connectivity and intelligent means. The random parking locations and times of manually driven vehicles, coupled with the inability of autonomous vehicles to adjust their behavior based on real-time information, cause scheduling conflicts and affect passenger travel efficiency.

[0005] Therefore, developing a dynamic optimization method for energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions is of great significance for improving the collaborative efficiency of autonomous and manual vehicles, reducing dwell time, optimizing passenger boarding and alighting experience, enhancing airport service quality, maximizing the utilization rate of drop-off area resources, and reducing operating costs. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention proposes a dynamic optimization method for energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions. This method is used to manage and control vehicles in the drop-off area in real time, guide and form an optimal fleet, and conduct cooperative drop-off to meet the goal of minimizing energy consumption.

[0007] To achieve the above objectives, this invention provides a method for dynamic optimization of energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions, including:

[0008] A drop-off time prediction model is constructed to predict the estimated time for each vehicle to arrive at each drop-off location, and the predicted drop-off time matrix of vehicle-parking space combinations is output. The drop-off time prediction model is performed by a deep learning network.

[0009] A vehicle average energy consumption model is constructed under mixed traffic flow conditions. The instantaneous energy consumption of autonomous electric vehicles (CAV) and manually driven gasoline vehicles (HDV) is calculated respectively, and the energy consumption is optimized by combining the car-following mode.

[0010] A reinforcement learning system for optimizing fleet operation is constructed by using the multi-agent deep deterministic policy gradient (MADDPG) algorithm, combined with the vehicle average energy consumption model and the car-following model. The system aims to minimize the average energy consumption of the fleet and dynamically adjusts the vehicle acceleration strategy and parking space allocation.

[0011] Based on real-time updated traffic conditions, the reinforcement learning system generates optimal fleet cooperation strategies to achieve comprehensive optimization of energy consumption, safety, and passenger comfort.

[0012] Preferably, the predicted drop-off time matrix of the vehicle-parking space combination includes:

[0013] By deploying roadside units in the airport drop-off area, vehicle characteristics, parking space characteristics, weather information and timestamp data are acquired in real time and preprocessed to construct a training sample set, wherein each sample in the training sample set is a combination of vehicle characteristics and parking space characteristics.

[0014] Based on the drop-off time prediction model, the drop-off time of each vehicle to each parking space is predicted using the training sample set. The drop-off time prediction model is trained on the combination data of vehicle features and parking space features through a multilayer perceptron, and outputs the predicted drop-off time matrix of the vehicle-parking space combination.

[0015] Preferably, the vehicle characteristics include vehicle type, traffic flow, congestion level, occupancy status of the target parking space, time until vacancy, and location of the parking space; the weather information includes temperature, precipitation, and wind speed; and the timestamp includes date characteristics.

[0016] Preferably, the training process of the drop-off time prediction model includes:

[0017] Vehicle types are encoded using one-hot encoding to convert time features into periodic features;

[0018] Using the mean squared error (MSE) as the loss function, the model parameters are updated through backpropagation and the Adam optimizer to obtain the trained passenger drop-off time prediction model.

[0019] Preferably, the predicted drop-off time matrix for the vehicle-parking space combination is:

[0020]

[0021] In the formula, To predict the drop-off time matrix, N is the number of vehicles, M is the number of parking spaces, and T′ is... NM This is the predicted time for the Nth vehicle to drop off passengers at the Mth parking space.

[0022] Preferably, constructing the vehicle average energy consumption model under the mixed traffic flow state includes:

[0023] For the aforementioned autonomous electric vehicle (CAV), calculate the drag power loss, motor power loss, regenerative braking power, and auxiliary power loss;

[0024] For the aforementioned manually driven HDV fuel vehicle, fuel consumption is calculated based on engine torque, speed, and transient correction coefficient;

[0025] By combining the intelligent driver IDM car-following model, the energy consumption differences between connected autonomous vehicles following connected autonomous vehicles in CC mode, connected autonomous vehicles following human-driven vehicles in CH mode, human-driven vehicles following connected autonomous vehicles in HC mode, and human-driven vehicles following human-driven vehicles in HH mode are distinguished.

[0026] Preferably, the reward function R of the multi-agent deep deterministic policy gradient (MADDPG) algorithm is:

[0027] R=-(αR fuel +βR cooperation +γR comfort );

[0028] In the formula, R fuel Let R be the energy consumption function. cooperation For cooperative functions, R comfort Let α, β, and γ be the comfort reward function, where α, β, and γ are all weighting coefficients.

[0029] Preferably, the method further includes updating parking space status and vehicle dynamic information in real time through a cloud platform, and issuing optimal following mode and parking space allocation instructions to vehicles through roadside units.

[0030] Compared with the prior art, the present invention has the following advantages and technical effects:

[0031] (1) Energy consumption optimization: This invention can significantly reduce energy consumption by dynamically adjusting the dispatching strategy of drop-off vehicles, especially during peak hours, reducing unnecessary waiting and queuing, thereby improving energy utilization efficiency.

[0032] (2) Improve travel efficiency: This method effectively improves the vehicle traffic efficiency in the drop-off area by optimizing the vehicle stopping order and cooperative drop-off strategy, reduces the queuing time of vehicles during peak hours, thereby speeding up the boarding and alighting of passengers and improving travel efficiency.

[0033] (3) Adaptive and intelligent: This invention uses deep learning and reinforcement learning algorithms, which can adjust itself in real time according to the changing traffic conditions, making the drop-off process more flexible, intelligent and efficient;

[0034] (4) Improve passenger experience: The optimized drop-off process not only improves energy efficiency, but also ensures passenger comfort and safety, providing passengers with a more convenient travel experience;

[0035] (5) Adapting to mixed traffic environment: Considering the mixed traffic situation of electric vehicles and fuel vehicles, the system can perform differentiated optimization for different types of vehicles to ensure that both electric vehicles and traditional fuel vehicles can achieve the best cooperative drop-off strategy in mixed traffic environment. Attached Figure Description

[0036] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0037] Figure 1 This is a flowchart of the dynamic optimization method for energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions, according to an embodiment of the present invention.

[0038] Figure 2 This is a schematic diagram of a reinforcement learning agent according to an embodiment of the present invention. Detailed Implementation

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0041] This invention proposes a method for dynamic optimization of energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions, such as... Figure 1 ,include:

[0042] A drop-off time prediction model is constructed to predict the estimated time for each vehicle to arrive at each drop-off location, and the predicted drop-off time matrix of vehicle-parking space combinations is output. The drop-off time prediction model is performed by a deep learning network.

[0043] A vehicle average energy consumption model is constructed under mixed traffic flow conditions. The instantaneous energy consumption of autonomous electric vehicles (CAV) and manually driven gasoline vehicles (HDV) is calculated respectively, and the energy consumption is optimized by combining the car-following mode.

[0044] A reinforcement learning system for optimizing fleet operation is constructed by using the multi-agent deep deterministic policy gradient (MADDPG) algorithm, combined with the vehicle average energy consumption model and the car-following model. The system aims to minimize the average energy consumption of the fleet and dynamically adjusts the vehicle acceleration strategy and parking space allocation.

[0045] Based on real-time updated traffic conditions, the reinforcement learning system generates optimal fleet cooperation strategies to achieve comprehensive optimization of energy consumption, safety, and passenger comfort.

[0046] This embodiment utilizes big data and intelligent technologies to effectively improve the energy efficiency of airport drop-off areas through systematic learning and optimization, alleviate traffic pressure during peak hours, and enhance the travel experience for passengers, demonstrating broad application prospects.

[0047] Furthermore, the predicted drop-off time matrix for the vehicle-parking space combination includes:

[0048] By deploying roadside units in the airport drop-off area, vehicle characteristics, parking space characteristics, weather information and timestamp data are acquired in real time and preprocessed to construct a training sample set, wherein each sample in the training sample set is a combination of vehicle characteristics and parking space characteristics.

[0049] Based on the drop-off time prediction model, the drop-off time of each vehicle to each parking space is predicted using the trained sample set. The drop-off time prediction model is trained on the combined data of vehicle features and parking space features through a multilayer perceptron and outputs the predicted drop-off time matrix of the vehicle-parking space combination.

[0050] Specifically, vehicle characteristics, parking space characteristics, weather information, and timestamp data are obtained through roadside units to predict travel time.

[0051] In this embodiment, vehicle feature X vehicle Includes: Vehicle type v j ∈0,1 (CAV or HDV); Current traffic flow Congestion level Parking space characteristics include: the occupancy status of the target parking space. i ∈0,1: Idle (0) or occupied (1); time remaining before vacancy. The time it takes for a parking space to become available after it has been occupied; the location of the parking space. The specific location of the parking space on the road. Weather information w includes weather-related features such as temperature, precipitation, and wind speed. Timestamp h includes date features such as weekday, weekend, and holiday.

[0052] Numerical features are standardized or normalized to ensure that features of different dimensions have a similar impact on the model. For each vehicle j and each corresponding parking position i, a training sample set is constructed by combining each vehicle with each parking position.

[0053] For example, N cars and M parking spaces will generate N*M samples. Each sample includes a set of vehicle features, parking space features, and the target drop-off time. The feature combination X ji Represented as:

[0054] X ji =[f j ,v j ,c j ,o i ,t vec,i ,l i [,w,h] (1);

[0055] In the formula, f j For the current traffic flow, t vec，i Let be the distance to the i-th parking space and the vacancy time.

[0056] At the same time, for vehicle type v j Transformed into v through one-hot encoding j ∈0,1 2 ; Convert hourly features into periodic features to better capture the periodic effects of time; Use Cartesian product to combine the features of vehicles and parking spaces to form a feature combination matrix. Where d is the feature dimension.

[0057] Multilayer perceptron (MLP) is used to predict drop-off time for each vehicle-parking space combination. The input samples are combined features. Where d is the feature dimension. Each hidden layer of the drop-off time prediction model is calculated using the following formula:

[0058] H (l) =φ(W (l) H H(l-1) +b (l) (2);

[0059] In the formula, H (l) This is the output of the l-th layer; Let n be the weight matrix of the l-th layer; nl is the number of neurons in the l-th layer of the neural network; H is the output of the hidden layer; Here, φ is the bias vector of the l-th layer; φ is the activation function; and the initial input is H. (0) =X ji H(l-1) represents the output of the (l-1)th layer. The final output is a scalar representing the predicted drop-off time T′. ji .

[0060] Furthermore, the training process of the drop-off time prediction model includes:

[0061] Vehicle types are encoded using one-hot encoding to convert time features into periodic features;

[0062] Using the mean squared error (MSE) as the loss function, the model parameters are updated through backpropagation and the Adam optimizer to obtain the trained passenger drop-off time prediction model.

[0063] Specifically, the mean squared error (MSE) is used as the loss function L:

[0064]

[0065] In the formula, T ji For actual drop-off time, T′ ji To predict passenger drop-off times.

[0066] The gradient is calculated using the backpropagation algorithm, and the model parameters are updated using the Adam optimizer:

[0067]

[0068] In the formula, η is the learning rate, L is the loss function, and b (l) is the bias vector of the l-th layer.

[0069] Furthermore, the predicted drop-off time matrix for the vehicle-parking space combination is as follows:

[0070]

[0071] In the formula, To predict the drop-off time matrix, N is the number of vehicles, M is the number of parking spaces, and T′ is... NM This is the predicted time for the Nth vehicle to drop off passengers at the Mth parking space.

[0072] Furthermore, constructing a vehicle average energy consumption model under mixed traffic flow conditions includes:

[0073] For autonomous electric vehicles (CAVs), calculate the drag power loss, motor power loss, regenerative braking power, and auxiliary power loss.

[0074] For manually driven gasoline vehicles (HDV), fuel consumption is calculated based on engine torque, speed, and transient correction coefficient.

[0075] By combining the intelligent driver IDM car-following model, the energy consumption differences between connected autonomous vehicles following connected autonomous vehicles in CC mode, connected autonomous vehicles following human-driven vehicles in CH mode, human-driven vehicles following connected autonomous vehicles in HC mode, and human-driven vehicles following human-driven vehicles in HH mode are distinguished.

[0076] Specifically, in this embodiment, the Intelligent Driver (IDM) model in the basic traffic flow graph model divides vehicle operation into four car-following modes: CC mode (connected autonomous vehicles following connected autonomous vehicles), CH mode (connected autonomous vehicles following manually driven vehicles), HC mode (manually driven vehicles following connected autonomous vehicles), and HH mode (manually driven vehicles following manually driven vehicles). Simultaneously, based on relevant electric vehicle energy consumption models and fuel consumption models, an average vehicle energy consumption model under mixed traffic conditions is constructed. The car-following model under mixed traffic conditions is combined with and optimized into the energy consumption models of both types of vehicles. Furthermore, parameters considering passenger comfort and other relevant factors are proposed, and the constructed model covers aspects such as optimal energy consumption, safety assurance, and comfort improvement. In the optimal energy consumption model, to simplify the model, CAVs are considered to be autonomous electric vehicles, while HDVs are considered to be manually driven gasoline vehicles, resulting in the following comprehensive energy consumption model:

[0077]

[0078] S j =x j-1 -x j -l(7);

[0079]

[0080] If there are no vehicles in front of the vehicle or the distance between vehicles is too far, the vehicle is considered to have entered free-driving mode, and its acceleration model is as follows:

[0081]

[0082] In the formula, V j Let A be the speed of vehicle j. j Let V be the maximum acceleration of vehicle j. f S is the free-flow velocity, δ is the acceleration exponent, which is taken as 4 in this embodiment. j This is the actual headway. To determine the desired headway, xj-1 Let x be the position of car j-1. j Let J be the position of vehicle j, l be the vehicle length, S1 be the minimum safety clearance, and T be the distance between vehicles. s For safe headway, B j For comfortable deceleration, t is time.

[0083] Based on the following model under different conditions:

[0084] T s =τ m + Fixed safety time (10);

[0085] In the formula, τ m The reaction time depends on the car-following mode:

[0086]

[0087] In the formula, τ c τ is the reaction time of CC mode. g τ represents the reaction time in the CH mode. h The reaction time is for HC or HH mode.

[0088] Formula for instantaneous energy consumption of electric vehicles (CAVs):

[0089] P EV (t)=P r (t)+P m (t)+P g (t)+P a (12);

[0090]

[0091] P a =2(kW) (16);

[0092] In the formula, P EV (t) is the instantaneous energy consumption power function of the electric vehicle, P r (t) represents the resistance power loss, P m (t) represents the motor power loss, P g (t) represents the regenerative braking power, P a To assist in power loss, C D A is the air drag coefficient. f For the windward area, f r Where m is the rolling resistance coefficient, g is the vehicle mass, b is the acceleration due to gravity, and R is the bearing damping coefficient. t Let r be the wheel radius, r be the motor internal resistance, K be the motor constant, and η be the regenerative braking efficiency (taken as 0.3 in this embodiment).

[0093] Formula for instantaneous power consumption of gasoline-powered vehicles (HDVs):

[0094]

[0095]

[0096] E total (t)=∑ i∈CAV P EV,i (t)+∑ j∈HDV (P r1HDV (t)+P m1HDV (t)+P cor,HDV (t))(22);

[0097]

[0098] N = N CAV +N HDV (25);

[0099] In the formula, E fuel (t) is the instantaneous energy consumption function of the fuel vehicle. For the power loss due to the resistance of the oil tanker, For transmission power loss, T ICE For engine torque, ω ICE Where N is the engine speed, ρ is the total number of vehicles in the fleet, LHV is the lower heating value of the fuel, and η is the fuel density. engine For engine efficiency, β i,j C is the transient correction factor. r Here, ρ is the gravity correction factor, ρ is the fuel density, A is the frontal area, θ is the road inclination angle, δ is the weighting parameter, and P is the frontal area. cor,HDV (t) is the fuel consumption correction function for gasoline vehicles, E total (t) is the total instantaneous energy consumption power function, P EV,i (t) represents the instantaneous energy consumption power of the electric vehicle, E total,accum For total energy consumption, E avg For average energy consumption, T total N represents the total time. CAV N represents the number of trams. HDV η represents the number of oil tankers. T For thermal efficiency.

[0100] Furthermore, the reward function R of the multi-agent deep deterministic policy gradient (MADDPG) algorithm is:

[0101] R=-(αR fuel +βR cooperation +γR comfort (26);

[0102] In the formula, Rfuel Let R be the energy consumption function. cooperation For cooperative functions, R comfort Let α be the comfort reward function, and β be the weight coefficients corresponding to each function.

[0103] Specifically, this embodiment employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm, combined with a vehicle energy consumption model and a car-following model, to construct a reinforcement learning system for optimizing fleet operation. Its main objective is to minimize the average energy consumption of the fleet while ensuring the safety and comfort of traffic flow.

[0104] By dynamically adjusting the acceleration strategy for each vehicle, reinforcement learning can optimize the following three key metrics in complex mixed traffic environments: energy consumption: reducing the energy consumption of electric vehicles and the fuel consumption of gasoline vehicles; safety: ensuring reasonable headway between vehicles to avoid collision risks; and comfort: reducing abrupt acceleration changes and improving the riding experience.

[0105] During the deployment phase, the trained reinforcement learning system can perceive changes in the traffic environment in real time (such as changes in vehicle speed and fluctuations in traffic density), and generate the optimal acceleration adjustment strategy for each vehicle, dynamically achieving the functions of reducing overall energy consumption, adapting to dynamic traffic environments, and ensuring safety and comfort.

[0106] The reinforcement learning system in this embodiment effectively integrates vehicle dynamics, energy consumption models, and the traffic environment to construct a highly efficient and intelligent fleet operation optimization system. Its functions cover energy optimization, safety assurance, and comfort enhancement, providing strong technical support for the optimized operation of hybrid fleets in complex traffic environments.

[0107] The following is a detailed process (e.g.) Figure 2 ):

[0108] For a mixed fleet, the state space is:

[0109]

[0110] In the formula, V j Let j be the speed of vehicle j. Let j be the acceleration of vehicle j. Let τ be the rate of change of acceleration of vehicle j. m For the vehicle's follow-car mode, S j T represents the actual headway between the vehicle and the vehicle in front. j Let x be the estimated time vector from car j to each parking space in the estimated time matrix T. j Let Type be the position of the j-th car. j For vehicle type, p compliance,j Assuming compliance rate, CAV is considered to be 1 and HDV to be a random value between 0 and 1.

[0111] Action Space A: The actions of vehicle j are divided into parking space allocation and adjustment of the following mode:

[0112] Assigned to a unique parking space * :

[0113]

[0114] In the formula, E actual,j,i T′ is the energy consumption of vehicle j allocated to parking space i. ji To predict passenger drop-off times.

[0115] At the same time, due to τ c <τ g <τ h The vehicle will prioritize adjusting to the following mode with a smaller head-on distance.

[0116] Based on this rule, establish the reward function R:

[0117] R=-(αR fuel +βR cooperation +γR comfort (29);

[0118]

[0119] In the reward function, R fuel R is the energy consumption function; cooperation R is a cooperative function designed to encourage vehicles adjacent to a target to cooperate in dropping off passengers; comfort For the comfort reward function, use the rate of change of acceleration. For reference, a small change in acceleration can improve passenger comfort, thus providing a reward; Penalty non-adiacent For the vehicle following penalty function, For vehicles The horizontal coordinate, For vehicles The horizontal coordinate, For vehicles The vertical coordinate, For vehicles The vertical coordinate, d adj β is the set adjacent distance threshold. comfort This is a comfort weighting coefficient. Let be the rate of change of acceleration of vehicle j in the x-direction. Let a be the rate of change of acceleration of vehicle j in the y direction. x,j Let a be the acceleration of vehicle j in the x direction. y,j Let be the acceleration of vehicle j in the y direction.

[0120] Reinforcement learning employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm.

[0121] Network update process:

[0122]

[0123] y j =R+γ l Q′ j (S′,A′;θ′ Q (36);

[0124]

[0125] In the formula, L is the loss function; y j The target Q value; Q j To evaluate the network function, let θ represent the estimated Q-value obtained by taking action A in state S; S is the current state, A is the current action, and θ is the current action. Q To evaluate the parameters of the network (parameters to be optimized), γ l Let Q′ be the discount factor. j Let Q be the Q-value of the target comment network, S′ be the next state, A′ be the next action, and θ′ be the next action. Q The parameters of the target comment network, The gradient of the policy network (Actor Network) Let Q be the gradient of the Q-value function with respect to the action. The policy function π j The gradient with respect to the parameter θ.

[0126] Reinforcement learning initialization: First, predict the passenger drop-off time matrix T and vehicle state S. j Actor network and Critic network. Based on the current vehicle state S j Using the policy network π j (S j Select action A j Then, parking spaces are allocated and the car-following mode is adjusted. Then, actions are executed to update the vehicle's position, speed, and acceleration:

[0127] V x,j (t+Δt)=V x,j (t)+a x,j (t)·Δt(38);

[0128] V y,j (t+Δt)=V y,j (t)+a y,j (t)·Δt(39);

[0129] xj (t+Δt)=x j (t)+V x,j (t)·Δt+0.5·a x,j (t)·(Δt) 2 (40);

[0130] y j (t+Δt)=y j (t)+V y,j (t)·Δt+0.5·a y,j (t)·(Δt) 2 (41);

[0131] S j =x j-1 -x j (42);

[0132] In the formula, V x,j (t+Δt) represents the velocity of vehicle j in the x-direction at time t+Δt, V x,j (t) represents the velocity of vehicle j in the x-direction at time t, a x,j (t) represents the acceleration of vehicle j in the x-direction at time t, Δt is the time step, and V y,j (t+Δt) represents the velocity of vehicle j in the y-direction at time t+Δt, V y,j (t) represents the velocity of vehicle j in the y-direction at time t, a y,j (t) represents the acceleration of vehicle j in the y-direction at time t, x j (t+Δt) represents the position of vehicle j in the x-direction at time t+Δt, where x j (t) represents the position of vehicle j in the x-direction at time t, and y j (t+Δt) represents the position of vehicle j in the y-direction at time t+Δt, where y j (t) represents the position of vehicle j in the y-direction at time t, and S j Let x be the relative distance between vehicle j and vehicle j-1. j-1 Let x be the position of vehicle j-1 in the x direction. j Let j be the position of vehicle j in the x direction.

[0133] The above steps are dynamically repeated, and the policy is converged to the optimal goal through reinforcement learning.

[0134] Furthermore, the method also includes updating parking space status and vehicle dynamic information in real time through a cloud platform, and issuing optimal following mode and parking space allocation instructions to vehicles through roadside units.

[0135] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic optimization of energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions, characterized in that, include: A drop-off time prediction model is constructed to predict the estimated time for each vehicle to arrive at each drop-off location, and the predicted drop-off time matrix of vehicle-parking space combinations is output. The drop-off time prediction model is performed by a deep learning network. A vehicle average energy consumption model is constructed under mixed traffic flow conditions. The instantaneous energy consumption of autonomous electric vehicles (CAV) and manually driven gasoline vehicles (HDV) is calculated respectively, and the energy consumption is optimized by combining the car-following mode. A reinforcement learning system for optimizing fleet operation is constructed by using the multi-agent deep deterministic policy gradient (MADDPG) algorithm, combined with the vehicle average energy consumption model and the car-following model. The system aims to minimize the average energy consumption of the fleet and dynamically adjusts the vehicle acceleration strategy and parking space allocation. Based on real-time updated traffic conditions, the reinforcement learning system generates the optimal fleet cooperation strategy to achieve comprehensive optimization of energy consumption, safety, and passenger comfort. For a mixed fleet, the state space is: ； In the formula, Let j be the speed of vehicle j. Let j be the acceleration of vehicle j. Let j be the rate of change of acceleration of vehicle j. This is the vehicle's follow-the-car mode. The actual headway between the vehicle and the vehicle in front. Let T be the estimated time vector from car j to each parking space in the estimated time matrix. Let j be the position of the j-th car. Vehicle type Assuming compliance rate, CAV is considered to be 1 and HDV to be a random value between 0 and 1. This represents the total number of vehicles in the fleet. Action Space A: The actions of vehicle j are divided into parking space allocation and adjustment of the following mode: Allocated to a single parking space : ； In the formula, This refers to the energy consumption of vehicle j allocated to parking space i. To predict passenger drop-off times; At the same time, due to The vehicle will prioritize adjusting to the following mode with a smaller head-on distance; for The reaction time, For the reaction time in CH mode, for The reaction time; Based on this rule, establish the reward function R: ； ； ； ； ； ； In the reward function, It is an energy consumption function; This is a cooperative function designed to encourage vehicles with adjacent destinations to cooperate in dropping off passengers. For the comfort reward function, use the rate of change of acceleration. For reference, a small change in acceleration can improve passenger comfort, and this is used as a reward. For the vehicle following penalty function, For vehicles The horizontal coordinate, For vehicles The horizontal coordinate, For vehicles The vertical coordinate, For vehicles The vertical coordinate, For the set adjacent distance threshold, This is a comfort weighting coefficient. Let be the rate of change of acceleration of vehicle j in the x-direction. Let be the rate of change of acceleration of vehicle j in the y direction. Let be the acceleration of vehicle j in the x-direction. Let be the acceleration of vehicle j in the y-direction. , , All are weighting coefficients. This represents the average energy consumption. Total energy consumption.

2. The method for dynamic optimization of energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions as described in claim 1, characterized in that, The output of the predicted drop-off time matrix for the vehicle-parking space combination includes: By deploying roadside units in the airport drop-off area, vehicle characteristics, parking space characteristics, weather information and timestamp data are acquired in real time and preprocessed to construct a training sample set, wherein each sample in the training sample set is a combination of vehicle characteristics and parking space characteristics. Based on the drop-off time prediction model, the drop-off time of each vehicle to each parking space is predicted using the training sample set. The drop-off time prediction model is trained on the combination data of vehicle features and parking space features through a multilayer perceptron, and outputs the predicted drop-off time matrix of the vehicle-parking space combination.

3. The method for dynamic optimization of energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions as described in claim 2, characterized in that, The vehicle characteristics include vehicle type, traffic flow, congestion level, occupancy status of the target parking space, time until vacancy, and location of the parking space. The weather information includes temperature, precipitation, and wind speed. The timestamp includes date characteristics.

4. The method for dynamic optimization of energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions as described in claim 3, characterized in that, The training process of the drop-off time prediction model includes: Vehicle types are encoded using one-hot encoding to convert time features into periodic features; Using the mean squared error (MSE) as the loss function, the model parameters are updated through backpropagation and the Adam optimizer to obtain the trained passenger drop-off time prediction model.

5. The method for dynamic optimization of energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions as described in claim 2, characterized in that, The predicted drop-off time matrix for the vehicle-parking space combination is as follows: ； In the formula, T∈ To predict the drop-off time matrix, N is the number of vehicles and M is the number of parking spaces. This is the predicted time for the Nth vehicle to drop off passengers at the Mth parking space.

6. The method for dynamic optimization of energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions as described in claim 1, characterized in that, Constructing the vehicle average energy consumption model under the mixed traffic flow condition includes: For the aforementioned autonomous electric vehicle (CAV), calculate the drag power loss, motor power loss, regenerative braking power, and auxiliary power loss; For the aforementioned manually driven HDV fuel vehicle, fuel consumption is calculated based on engine torque, speed, and transient correction coefficient; By combining the intelligent driver IDM car-following model, the energy consumption differences between connected autonomous vehicles following connected autonomous vehicles in CC mode, connected autonomous vehicles following human-driven vehicles in CH mode, human-driven vehicles following connected autonomous vehicles in HC mode, and human-driven vehicles following human-driven vehicles in HH mode are distinguished.

7. The method for dynamic optimization of energy consumption of cooperative drop-off vehicles in airport drop-off areas under mixed traffic conditions as described in claim 1, characterized in that, The method also includes updating parking space status and vehicle dynamic information in real time through a cloud platform, and issuing optimal following mode and parking space allocation instructions to vehicles through roadside units.

Images