A distributed resource game scheduling method for a smart parking lot

Abstract

The application discloses a kind of distributed resource game scheduling methods for wisdom parking lot, including, through the edge computing node being deployed in the periphery of parking lot, the vehicle dynamic and parking stall static data in its jurisdiction area are collected, the vehicle macroscopic distribution mathematical model of group driving situation is constructed using kernel density estimation method;On each edge computing node, based on the vehicle macroscopic distribution model and total driving cost function, construct mean field game problem, by solving the game problem, obtain the equilibrium solution of the optimal behavior of vehicle group;Edge computing node according to the equilibrium solution, for single or multiple vehicle terminals in jurisdiction area resolve optimal scheduling instruction and issue;The application is by the global scheduling problem is decomposed into mean field game problem solved on distributed edge node, overcome dimension disaster, realized to super large scale vehicle group nearly global optimal real-time resource scheduling, significantly improve the operation efficiency and user experience of wisdom parking lot.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology for intelligent parking, and in particular to a distributed resource game scheduling method for smart parking lots. Background Technology

[0002] With the acceleration of urbanization and the surge in car ownership, the urban "parking difficulty" problem has become a common issue restricting traffic efficiency and affecting residents' quality of life. To address this challenge, smart parking systems based on the Internet of Things (IoT) and mobile communication technologies have emerged. These systems use sensors such as geomagnetic sensors and video cameras to achieve real-time perception of parking space status, providing a data foundation for dynamic vehicle guidance and efficient allocation of parking resources. Against this backdrop, various resource allocation methods have emerged, with the mainstream methods typically employing two approaches: First, a global optimization method based on centralized cloud computing, where the status information of all vehicles and parking lots in the city is aggregated to a central server, and unified calculations and command issuance are performed using operations research algorithms. Second, a game theory model is introduced, treating drivers seeking parking spaces as rational game participants, and achieving distributed optimal allocation of resources by finding the equilibrium point of the game.

[0003] However, when facing ultra-large-scale, high-concurrency scheduling scenarios in first-tier cities, existing technologies have revealed their fundamental limitations. For traditional game theory models, when the number of vehicles participating in the game reaches hundreds of thousands or even millions at the city level, traditional game theory solutions fall into the "curse of dimensionality." The computational complexity increases exponentially with the number of participating vehicles, making it impossible to complete calculations within decision windows of seconds or even minutes, thus lacking real-time scheduling capabilities. Furthermore, due to the traditional centralized architecture, the concurrent uploading of a large amount of vehicle terminal status data places enormous pressure on the central server's communication bandwidth and computational load. The inherent communication latency also causes scheduling instructions to lag significantly behind the rapidly changing actual road conditions, and the smart parking system is susceptible to single-point-of-failure risks, making reliability difficult to guarantee.

[0004] Therefore, there is an urgent need in this field for a new scheduling paradigm that can overcome the dual constraints of computational complexity and intelligent parking system architecture, in order to meet the stringent requirements of future smart cities for large-scale, low-latency, and high-reliability parking resource scheduling. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a distributed resource game-theoretic scheduling method for smart parking lots to solve the problems mentioned in the background section.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a distributed resource game scheduling method for smart parking lots, comprising:

[0008] Based on edge computing nodes, status data of multiple vehicle terminals within their coverage area are obtained, and a mathematical model of the macroscopic distribution of vehicle groups is established as a representation of the driving status of vehicle groups.

[0009] Based on the edge computing nodes and the evolutionary dynamics of the mathematical model, the game problem of the interaction between the driving status of a single vehicle and the driving status of the vehicle group is solved to obtain an equilibrium solution.

[0010] The edge computing node generates and sends corresponding optimal scheduling instructions to at least one vehicle terminal based on the equilibrium solution.

[0011] As a preferred embodiment of the distributed resource game scheduling method for smart parking lots described in this invention, the mathematical model for establishing the macroscopic distribution of the vehicle group adopts a statistical estimation method, which uses an adaptive smoothing bandwidth, wherein the magnitude of the smoothing bandwidth is inversely proportional to the local vehicle density.

[0012] As a preferred embodiment of the distributed resource game-theoretic scheduling method for smart parking lots described in this invention, the step of solving the game problem of the interaction between the driving behavior of a single vehicle and the group of vehicles includes:

[0013] The game problem is defined by the total driving cost of the vehicle;

[0014] The total driving cost includes congestion cost, the magnitude of which is determined by the local vehicle density in the mathematical model.

[0015] As a preferred embodiment of the distributed resource game scheduling method for smart parking lots described in this invention, the total driving cost further includes the parking space competition cost:

[0016] The cost of competing for parking spaces is directly proportional to the real-time price of the target parking lot and inversely proportional to the expected number of available parking spaces in the parking lot calculated based on the mathematical model.

[0017] As a preferred embodiment of the distributed resource game scheduling method for smart parking lots described in this invention, the evolution dynamics of the mathematical model includes a random influence term simulating the uncertainty of driving behavior, and the random influence term is set to be consistent with the direction of the road where the vehicle is currently located.

[0018] As a preferred embodiment of the distributed resource game scheduling method for smart parking lots described in this invention, the process of obtaining the equilibrium solution is accelerated in the following ways:

[0019] Before each solution iteration begins, a pre-trained deep neural network is used to predict the macroscopic distribution of the vehicle group in the new round based on the macroscopic distribution of the vehicle group in the previous round.

[0020] The predicted macroscopic distribution of the new vehicle group is used as the initial condition for this round of iterative calculation.

[0021] As a preferred embodiment of the distributed resource game scheduling method for smart parking lots described in this invention, the step of generating and issuing corresponding optimal scheduling instructions to at least one vehicle terminal includes:

[0022] Based on the equilibrium solution, a suggested driving route and a speed range matching the driving route are planned for the vehicle terminal within a future preset time period. The route and speed range are then encapsulated into a structured data format and sent out.

[0023] As a preferred embodiment of the distributed resource game scheduling method for smart parking lots described in this invention, the method further includes a cross-regional collaboration step:

[0024] Between edge computing nodes of two adjacent geographic regions, vehicle flow rate information on their shared boundary is exchanged;

[0025] When each edge computing node solves a game problem, it uses the received vehicle flow rate information as the boundary condition for calculating the game problem.

[0026] As a preferred embodiment of the distributed resource game scheduling method for smart parking lots described in this invention, the method further includes an adaptive optimization step for the parameters in the mathematical model:

[0027] Through the central cloud platform, based on scheduling performance metrics collected from one or more edge computing nodes, the weight parameters of the total travel cost in the game problem are automatically adjusted using a sequential optimization algorithm, and the adjusted weight parameters are then sent to the edge computing nodes.

[0028] As a preferred embodiment of the distributed resource game scheduling method for smart parking lots described in this invention, the status data of the vehicle terminal also includes the vehicle's remaining battery power or fuel level, the user's expected parking duration, and user level information. The total driving cost defined in the game problem is adjusted based on the consumption cost of the remaining battery power or fuel level in the vehicle terminal's status data and the priority of the user level information.

[0029] Compared with existing technologies, the beneficial effects of the invention are:

[0030] 1. This invention transforms the scheduling problem of a large number of vehicles from a game problem whose computational complexity increases exponentially with the number of vehicles into a game problem that describes the interaction between individuals and the group's macro-level situation, independent of the number of participants (vehicles). This fundamentally solves the "curse of dimensionality" problem of traditional game models in ultra-large-scale scenarios, enabling this method to handle the concurrent scheduling of millions of vehicles at the city level and has excellent scalability.

[0031] 2. By defining congestion costs coupled with local vehicle density and parking space competition costs related to expected parking space occupancy rates in the game model, the vehicle resource scheduling results can reflect the macro traffic situation, guide vehicles to actively avoid potential congestion and ineffective parking space search, improve the overall utilization efficiency of parking resources and user experience, and ensure efficient solution of the model at edge nodes by introducing an acceleration mechanism for equilibrium solutions, taking into account both the advanced nature of the scheduling strategy and the feasibility of engineering applications.

[0032] 3. Furthermore, by adopting a distributed edge computing architecture, the game-solving task is pushed down to edge computing nodes closer to the user, realizing local closed-loop data processing and decision-making. This reduces the dependence on communication bandwidth and computing load of the central cloud platform, shortens the end-to-end latency from state awareness to instruction issuance, and improves the real-time performance and effectiveness of scheduling instructions. At the same time, the distributed edge computing architecture avoids the risk of single point of failure and enhances the robustness and reliability of the entire smart parking system. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0034] Figure 1 This is a flowchart illustrating the overall process of a distributed resource game scheduling method for smart parking lots according to an embodiment of the present invention. Detailed Implementation

[0035] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0036] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0037] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0038] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0039] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0040] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0041] Example 1

[0042] Reference Figure 1This is the first embodiment of the present invention, which provides a distributed resource game scheduling method for smart parking lots, including:

[0043] S1. Based on edge computing nodes, obtain the status data of multiple vehicle terminals within their coverage area, establish a mathematical model of the macroscopic distribution of vehicle groups, and use it as a representation of the driving status of vehicle groups.

[0044] It should be noted that edge computing nodes are deployed at communication base stations or roadside units;

[0045] Specifically, edge computing nodes establish communication links with all vehicle terminals connected to the smart parking system within their coverage area through wireless communication networks (such as 5G-V2X or cellular networks). Each vehicle terminal will periodically or actively report its status data to its respective edge computing node when its status changes significantly.

[0046] Specifically, each vehicle terminal has a built-in on-board unit that integrates a GPS / BeiDou positioning module, an inertial measurement unit (IMU), and a communication module. The on-board unit collects its own status data at a preset high frequency (e.g., 1-10Hz).

[0047] Furthermore, when the edge computing node receives the reported state data, it organizes it into a multi-dimensional state vector to comprehensively describe the individual state of the i-th vehicle at time t.

[0048] Preferably, the state vector includes at least the vehicle's real-time geographical coordinates (latitude and longitude), driving speed, and direction. Furthermore, to support more refined scheduling strategies, the state vector may also include the vehicle's remaining fuel or battery power (for electric vehicles), the driver's preset expected parking duration, and pre-defined user level information. Therefore, the state vector V... i The specific form of (t) can be defined as:

[0049] V i (t)={ID i ,(lon i ,lat i ),v i ,θ i ,a i E i ,T i ,L user,i}

[0050] Among them, ID i Let be a unique anonymous identifier for the i-th vehicle, (lon i ,lat i Let v represent the latitude and longitude coordinates of the i-th vehicle. iLet θ represent the instantaneous velocity of the i-th vehicle. i Represented as the heading angle, a i Expressed as acceleration (optional), E i T represents the remaining fuel or battery percentage of the i-th vehicle. i L represents the expected parking time set for the i-th vehicle. user,i This represents the user level information for the i-th vehicle.

[0051] Furthermore, at the beginning of each scheduling cycle (e.g., 1–5 seconds), the edge computing nodes aggregate the state vectors of all vehicles received within that cycle, forming a discrete set of sample points (i.e., individual vehicle data) S(t) = {V1(t), V2(t), ..., V...}. M (t)}, where M is the total number of vehicles in the current area. The network time protocol (NTP) is used to ensure that the timestamps of all sample point data are synchronized, thereby constructing an accurate state snapshot;

[0052] It needs to be explained that using traditional gridded statistics limits model accuracy to the grid size and results in numerous zero values ​​in sparsely populated areas, failing to accurately reflect potential driving probabilities. While fixed-bandwidth kernel density estimation (KDE) cannot simultaneously adapt to both highly densely populated areas like CBDs and sparsely populated areas like suburbs, the fixed large bandwidth in dense areas leads to excessive smoothing, obscuring crucial traffic congestion details (such as density differences between adjacent lanes). Conversely, in sparse areas, the fixed small bandwidth generates numerous spurious, disconnected density spikes, distorting the model. Therefore, to address these issues, this invention employs a kernel density estimation-based method... Density Estimation (KDE) is a statistical estimation method. The core idea of ​​this scheme is to treat the state vector of each vehicle terminal as a sample point in the state space, and to superimpose a smooth kernel function (such as a Gaussian kernel function) on each sample point. Finally, the summation of all kernel functions is obtained to obtain a continuous and smooth probability density function. This probability density function is the mathematical model of the macroscopic distribution of the vehicle group. Through this model, the probability density of vehicle appearance at any location and state can be intuitively reflected, that is, the macroscopic degree of vehicle aggregation.

[0053] Furthermore, edge computing nodes first use a global fixed bandwidth h pilot (For example, based on the average width of the road network), a standard kernel density estimate is performed on the individual vehicle dataset. The purpose is not to obtain a final mathematical model, but to quickly generate a rough "pilot" density field that reflects the approximate density distribution; subsequently, for each vehicle sample point V in the individual vehicle dataset...i (t)(its state is x) i ), calculate its local smoothing bandwidth h i , that h i The calculation follows the principle of being inversely proportional to the local vehicle density; while edge computing nodes use bandwidth h tailored for each sample point. i A second kernel density estimation is performed to finally generate a continuous probability density field m(x,t) that can describe the macroscopic behavior of the group, where x is a point (i.e., a state) in the state space. This probability density field is the mathematical model of the macroscopic distribution of the vehicle group. In addition, physically, this model can be discretized into a fine grid covering the entire jurisdiction of the edge node, where each grid cell stores the vehicle density value at the location of the jurisdiction.

[0054] Specifically, the principle that the local smoothing bandwidth is inversely proportional to the local vehicle density can be expressed mathematically as follows:

[0055] h i =h0×(G / m pilot (x i )) α

[0056] Where h0 represents the reference bandwidth, m pilot (x i ) indicates that the sample point is at x i The "pilot" density value found at the query point is G, which is the geometric mean of the "pilot" density of all sample points and a normalization factor; α is a sensitivity parameter (usually taken as 0.5), used to control the degree of drastic change in local smoothing bandwidth with local vehicle density.

[0057] It should be noted that, based on this principle, areas located in high-density regions (i.e., m...) pilot For vehicles with high (value), their h i It will be very small, thus preserving the details of the region; while in sparse regions (i.e., m) pilot For vehicles with a small value, their h i This will be larger, thus forming a smooth connection and compensating for the lack of data;

[0058] S2. Based on the edge computing nodes and the evolutionary dynamics of the mathematical model, solve the game problem of the interaction between the driving status of a single vehicle and the driving status of a group of vehicles, and obtain the equilibrium solution.

[0059] Specifically, based on the mathematical model of the macroscopic distribution of the vehicle group constructed in step S1, a complex game problem involving N participants that was originally computationally infeasible is reduced to an efficient optimization control problem that describes the interaction between the driving status of a single vehicle and the vehicle group. This process is called Mean Field Game (MFG). When the number of game participants N in the smart parking system approaches infinity (in the context of smart cities), the impact of each individual on the entire smart parking system can be ignored. Therefore, the environment of the individual can be approximated by the average effect of all other individuals (i.e., the probability density field m(x,t)). The game problem can be decoupled and reduced in dimension from a complex problem with N people coupled to two core equations: the inverse-time Hamilton-Jacobi-Bellman (HJB) equation and the forward-time Fokker-Planck-Kolmogorov (FPK) equation.

[0060] Specifically, the inverse Hamilton-Jacobi-Bellman equation describes the individual (single vehicle) optimality. The equation solves for a value function u(x,t), which means that the minimum cumulative cost that a vehicle in state x at time t can obtain if it always adopts the optimal strategy from this moment to the future end time (cycle C).

[0061] It should be noted that this equation is solved from the future (t=C) to the present (t=0) because the final cumulative cost (e.g., whether parking is successful) is determined in the future;

[0062] Specifically, the forward-moving Fokker-Planck-Kolmogorov equation describes the evolution of the swarm (vehicle swarm). This equation describes how the density distribution (probability density field m(x,t)) of the entire vehicle swarm evolves over time, driven by all vehicles adopting the optimal strategy given by the HJB equation, starting from the current time t=0 and proceeding forward.

[0063] It should be noted that since the HJB equation and the FPK equation are coupled to each other through m and u, the solution of the HJB equation depends on m (because the congestion cost is determined by m), and the evolution of the FPK equation depends on the optimal strategy obtained from the HJB equation. Therefore, it is necessary to find a set of (u,m) that can satisfy both equations at the same time, which is the equilibrium solution.

[0064] Furthermore, an instantaneous operating cost function L(x,β,m) is constructed for the vehicle through edge computing nodes, where β represents the vehicle's control strategy (such as the selected vehicle speed and direction), and the total driving cost is the integral of this instantaneous operating cost function over the entire driving time domain, specifically including:

[0065] Congestion cost aims to directly reflect the impact of a vehicle group on an individual vehicle. Edge computing nodes, based on the mathematical model of the macroscopic distribution of the vehicle group obtained in step S1, map the local vehicle density at the current vehicle location to a cost value. This mapping is a monotonically increasing function, C. cong =f(m(x,t)), where f is a preset congestion penalty function; when the route chosen by a vehicle passes through a high-density area indicated by m(x,t), the congestion cost it bears will increase sharply, thereby incentivizing vehicles to actively avoid potential or existing congestion when planning routes;

[0066] Parking space competition cost mainly targets the ultimate goal of vehicles—parking. Edge computing nodes dynamically evaluate the attractiveness of each parking lot. That is, for a target parking lot, the parking space competition cost is directly proportional to the real-time price of the parking lot and inversely proportional to the expected number of available parking spaces. The expected number of available parking spaces is a key predictor. This predictor is not simply equal to the current number of available parking spaces, but is dynamically predicted through a mathematical model of the macroscopic distribution of vehicle groups.

[0067] Specifically, edge computing nodes calculate the density of vehicle flow (obtained by m(x,t)) that will flow into the parking lot in the future, thereby predicting the probability of actual available parking spaces when vehicles arrive. If a large number of vehicles are heading towards the same parking lot, the expected number of available parking spaces will drop sharply, causing the cost of parking space competition to soar. This "deterrence" of some vehicles in the game and guides them to other parking lots with less competition, thus achieving balanced utilization of parking space resources.

[0068] In addition, in order to provide differentiated services, the total driving cost will be adjusted according to the specific condition of the vehicle to obtain a personalized cost;

[0069] Specifically, for cars with low remaining battery or fuel percentage, a consumption cost parameter will be added to their total driving cost for personalized costs. This consumption cost parameter will be given higher weight to congested road sections or road sections with uphill sections that require frequent acceleration and deceleration, guiding them to choose smoother and more energy-efficient routes. For users with higher user levels, the weight of time cost in their total driving cost will be higher, or a discount coefficient will be obtained in the above congestion cost calculation to reflect the service priority of their user level.

[0070] Specifically, the total operating cost can be expressed in its complete form as:

[0071] L=ω c ×C cong +ω p ×Cparking +C personalized

[0072] Where, ω c and ω p C is a weight parameter that can be adjusted by the central cloud platform. personalized The cost of personalization consists of the following components:

[0073] Energy cost C energy This cost is related to the vehicle's remaining energy E. remain (Fuel or battery level) is related to the control strategy (acceleration), which is specifically a penalty function that is activated when the remaining energy is below a preset threshold (e.g., 20%). C energy =ω e ×(1 / E remain )×(α accel ) 2 , where α accel It is the acceleration component of the vehicle, ω e It is an energy consumption cost weight, which means that for vehicles with low battery / fuel levels, any violent acceleration will lead to a sharp increase in costs, thereby guiding them to choose a smoother, more energy-efficient driving strategy.

[0074] User priority adjustment C priority This is achieved through a discount factor related to user level, with the following mapping: high-level users δ = [0.1, 0.6), level users δ = [0.6, 0.8), and low-level users δ = [0.8, 1.0]. This discount factor is directly applied to the congestion cost and time-related cost components of the total cost; that is, the adjusted total travel cost becomes L′ = δ × (ω c ×C cong )+ω p ×C parking +C personalized This makes it easier for high-level users to perceive congestion and time in game theory calculations, making it easier for them to obtain faster paths, thus reflecting service priority.

[0075] Furthermore, the edge computing nodes employ a fixed-point iterative numerical method to solve the aforementioned HJB-FPK coupled equations to obtain an equilibrium solution. The process is as follows:

[0076] S201. Given an initial macroscopic distribution of vehicles m0(x,t);

[0077] S202, m k Substituting (x,t) (the distribution of the k-th iteration) into the total travel cost, we solve the HJB equation in reverse from t=C to obtain the optimal value function u. k (x,t) and the corresponding optimal control strategy (i.e., optimal velocity field) γk (x,t);

[0078] S203, Apply the optimal control strategy γ k Substituting (x,t) into the FPK equation and solving it forward from t=0, we obtain a new macroscopic vehicle distribution m. k+1 (x,t);

[0079] S204, Compare m k+1 With m k If the difference between the two (e.g., L2 norm) is less than a preset threshold, the iteration is considered to have converged and the equilibrium solution has been found; otherwise, let k = k + 1 and return to S202.

[0080] Furthermore, to meet the real-time requirements of parking in smart city scenarios, this invention introduces an acceleration mechanism:

[0081] To reduce the number of iterations for the equilibrium solution, this invention introduces an acceleration step before the iteration begins. The edge computing nodes utilize a pre-trained deep neural network, whose input is defined as the macroscopic vehicle distribution m that converged in the previous scheduling cycle (e.g., 5 seconds ago). final (t-Δt), and the output of the neural network is defined as a prediction result of the initial distribution m0(x,t) of the current scheduling cycle, and the prediction result is used as the initial condition for iterative calculation;

[0082] Specifically, the training data for this deep neural network comes from historical scheduling data or high-fidelity simulation data;

[0083] Specifically, the central cloud platform periodically collects and stores complete scheduling logs from each edge computing node over a period of time (e.g., the past month). Each log entry contains the initial macroscopic distribution m of vehicles for a scheduling cycle. initial and equilibrium distribution m final , will (m initial ,m final Yes, as the training dataset for the neural network, the macroscopic distribution of vehicles is discretized into a two-dimensional matrix of H×W corresponding to the jurisdiction of the edge computing nodes, where H and W are the height and width of the grid resolution.

[0084] Preferably, the U-Net architecture is adopted because its unique encoder-decoder structure and skip connections are particularly suitable for handling image-to-image conversion tasks, and are highly compatible with the task of predicting one distribution from another in this invention;

[0085] Specifically, the neural network takes the converged macroscopic vehicle distribution of the previous scheduling cycle (i.e., a single-channel H×W image) as input and outputs a prediction of the balanced distribution for the current scheduling cycle (also a single-channel H×W image). The encoder extracts deep, abstract features from the distribution image through successive convolutional and max-pooling layers. The decoder uses upsampling (e.g., transposed convolution) and convolutional layers to gradually restore these abstract features to their original resolution and uses skip connections to supplement the shallow detail features (road contours) of the corresponding encoder layers to ensure high fidelity in the predicted distribution. This neural network is trained offline using supervised learning. For each sample (m) in the training set... initial ,m final ), to m initial As input, we obtain the predicted value, and the goal of offline training is to minimize the difference between the predicted value and the true convergence value m. final The difference between l can be addressed by using the loss function, which can be the mean absolute error (MAE) or the mean squared error (MSE) between the two. The weight parameters of the network are continuously adjusted through the backpropagation algorithm and the gradient descent optimizer (such as Adam) until the loss function converges. After offline training is completed, the model is distributed to each edge computing node for inference.

[0086] It should be noted that, due to the spatiotemporal continuity of traffic flow evolution, the prediction result will be very close to the actual equilibrium solution, thereby reducing the number of iterations from tens or even hundreds to just a few, so as to ensure that the entire solution process can be completed within seconds with the limited computing power of the edge nodes.

[0087] Furthermore, to make the mathematical model of the macroscopic distribution of vehicle groups more closely resemble the uncertainty of driver behavior in reality, this invention introduces a random influence term (i.e., diffusion term) into the evolution dynamics of the FPK equation. This random influence term is not isotropic (i.e., the randomness is the same in all vehicle directions), but anisotropic. Its diffusion direction is set to be consistent with the direction of the road where the vehicle is currently located. This means that the mathematical model considers that the random deviation behavior of vehicles is more likely to be moving back and forth along the road direction (e.g., random fluctuations in following distance) rather than perpendicular to the road direction (e.g., suddenly veering onto the sidewalk). By setting this random influence term, the evolution of the mathematical model becomes more consistent with physical reality and common sense about driving, thereby improving the accuracy of the equilibrium solution.

[0088] It should be noted that when multiple edge computing nodes jointly cover a city, the problem of inter-regional coordination needs to be addressed; when solving the FPK equation for this region, the boundary conditions of its computational domain also need to be defined; this invention mainly achieves this by exchanging vehicle flow rate information on their shared boundary between adjacent edge computing nodes.

[0089] Specifically, when an edge computing node informs the neighboring region of the vehicle throughput it has calculated, the vehicle throughput flowing to the neighboring region, it also receives the vehicle throughput that will flow into its own region from the neighboring region. The vehicle flow rate information exchanged is used as the Neumann boundary condition of the FPK equation, which ensures that even though the computation is distributed, the macroscopic vehicle distribution in each region is continuous and conserved at the boundary, thereby obtaining a globally consistent scheduling view.

[0090] S3. The edge computing node generates and sends the corresponding optimal scheduling instruction to at least one vehicle terminal based on the equilibrium solution.

[0091] It should be noted that the task of this step is to transform the equilibrium solution into a micro-scheduling instruction that can be executed and personalized for each specific vehicle terminal, so as to ensure that the issued instructions are both optimal and practical.

[0092] Furthermore, before issuing instructions, the edge computing node first determines the optimal parking target for each vehicle terminal that needs to be scheduled. This process does not recommend the nearest or cheapest parking lot, but is achieved by parsing the value function in step S2.

[0093] Specifically, since this value function represents the minimum expected total cost of a vehicle located at (x,t) to the end of the scheduling cycle under equilibrium conditions, all candidate parking lots P1, P2, ..., P within the jurisdiction of the edge computing node are obtained. q The geographic coordinates of the parking lots are given, where q represents the number of candidate parking lots, and the value function u(P) is used to query these parking lots at the end of the scheduling cycle. q According to the definition of the value function, the minimum parking lot P is obtained. optimal This means that after comprehensively considering the expected travel time to reach the parking lot, congestion costs, parking space competition costs, and vehicle personalization costs, the global optimal choice for the vehicle terminal is made. This selection process implicitly avoids parking space competition, because a parking lot that is expected to be crowded has already increased its corresponding u value in the parking space competition cost.

[0094] Furthermore, after determining the optimal target parking lot (i.e., P) optimal Edge computing nodes need to plan a specific driving path for the vehicle terminal from the vehicle's current location.

[0095] Furthermore, edge computing nodes view the value function as a dynamic, time-evolving cost map, where the value of each point represents the expected cost of traveling to that point, while viewing the optimal speed field as a guiding wind field, representing the optimal driving direction and speed of the vehicle at each location;

[0096] Preferably, the edge computing nodes use an improved A* algorithm based on MFG (mean field game) for path planning;

[0097] Furthermore, the A* algorithm starts with the vehicle's current position and ends with the optimal parking lot.

[0098] It should be noted that, in the traditional A* algorithm, the cost from node A to node B is usually its physical distance. However, in this invention, the cost is defined as the total cost of traversing the road segment, which is calculated based on the integral or difference of the value function on the road segment. That is, the selected path tends to follow the direction of the fastest gradient descent of the value function, so that the planned path can actively avoid areas with high value function (i.e., high expected cost), such as congested road segments.

[0099] It should be noted that, in the traditional A algorithm, the heuristic function is the estimated distance from the current node to the destination (such as Euclidean distance or Manhattan distance). However, in this invention, the heuristic function is directly set as the value function of the node h(n) = u(n, t′), where t′ is the estimated time to reach node n. This directly gives the optimal expected cost from the point to the destination, improving the search efficiency and path optimality of the A algorithm.

[0100] Furthermore, through the improved A* algorithm described above, the edge computing node generates a suggested driving route for the vehicle terminal, consisting of geographic coordinate points. At the same time, for each segment on the route, the edge computing node queries the optimal speed field of that segment at the corresponding travel time, uses the speed of the optimal speed field as the suggested speed, and sets a reasonable speed range around it (e.g., ±10% of the suggested speed) to provide a certain degree of driving flexibility.

[0101] Furthermore, in order to ensure that the vehicle terminal can accurately parse and execute instructions, the edge computing node will encapsulate the planned path and speed information into a lightweight, standardized structured data format.

[0102] Preferably, the data format is JSON. The edge computing node encrypts the encapsulated JSON data packet and sends it to the target vehicle terminal through the low-latency, high-reliability wireless communication link (such as the PC5 interface of 5G-V2X or the Uu interface of cellular network) that has been established between it and the vehicle terminal.

[0103] Specifically, after receiving the instruction, the vehicle terminal's onboard unit decrypts and parses it, highlights the suggested driving route on the in-vehicle navigation screen, and informs the driver of the suggested speed range for the current road segment through voice or visual prompts. For vehicles with assisted driving or autonomous driving functions, this instruction can be used as a high-priority input for its path planning and speed control modules. At the same time, the edge computing node continuously receives status data reported by the vehicle terminal to monitor whether it has deviated from the suggested driving route. If the vehicle deviates significantly from the route (for example, the driver chooses another road for personal reasons), or in the next scheduling cycle (for example, after 5 seconds), a new and better equilibrium solution will be generated. The edge computing node will repeat the path planning process, generate a new optimal scheduling instruction, and issue it to achieve continuous, dynamic, and adaptive guidance for the vehicle until it successfully reaches the target parking lot.

[0104] In addition, to quantify the quality of scheduling strategies, a set of key performance indicators (KPIs) were defined on the central cloud platform. These KPIs include at least: average vehicle search time in the area (the average time from when a vehicle enters the area to when it successfully parks), parking space turnover rate (the average number of times a parking space is used per unit time), load balancing between edge nodes (measured by the variance of the number of vehicles processed by each node), and user satisfaction (indirectly statistically measured by user app feedback or command acceptance rate).

[0105] Specifically, after each edge computing node completes each scheduling cycle, it calculates the aforementioned KPIs within its jurisdiction and reports these statistics, along with the corresponding cost weight parameters, to the central cloud platform periodically during non-peak business periods (such as early morning) in the form of lightweight messages.

[0106] Preferably, the central cloud platform uses Bayesian optimization as the sequential optimization algorithm. The reason for choosing this algorithm is that evaluating the performance of weight parameters (i.e., running and observing KPIs in a real or simulation environment) is very costly, while Bayesian optimization is particularly good at finding optimal solutions on "expensive black-box functions".

[0107] S301. The central cloud platform first uses the collected historical data (weight parameters and KPIs) to build a probabilistic proxy model (such as Gaussian process regression). This model can predict the performance of KPIs that may be generated by any new set of weight parameters and give the uncertainty of the prediction.

[0108] S302. The central cloud platform can define a collection function to determine the next set of weight parameters to be tested. The collection function is required to balance "exploration" (trying parameter regions with high uncertainty, which may discover new optimal solutions) and "utilization" (fine-tuning near the currently known optimal solutions).

[0109] S303, the central cloud platform will send the new weight parameters selected by the acquisition function to one or a group of edge nodes for "A / B testing", and collect their performance indicators after running for a period of time;

[0110] S304. Add the new (weight parameters and KPIs) data points to the historical dataset, update the surrogate model, and then repeat S302 to S304 sequentially and iteratively to approximate the weight parameter combination that optimizes the overall KPIs.

[0111] Specifically, once the optimization algorithm converges or reaches the preset number of iterations, the central cloud platform will distribute the found optimal weight parameters as a global configuration to all edge computing nodes to update the total driving cost function in their S2 step.

[0112] It should be noted that, through the collaborative architecture of "edge computing + central optimization", this invention achieves continuous learning and adaptive evolution of the scheduling model.

[0113] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0118] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A distributed resource game-theoretic scheduling method for smart parking lots, characterized in that, include: Based on edge computing nodes, status data of multiple vehicle terminals within their coverage area are obtained, and a mathematical model of the macroscopic distribution of vehicle groups is established as a representation of the driving status of vehicle groups. The mathematical model for establishing the macroscopic distribution of the vehicle group adopts a statistical estimation method, which uses an adaptive smoothing bandwidth, wherein the magnitude of the smoothing bandwidth is inversely proportional to the local vehicle density. Based on the edge computing nodes and the evolutionary dynamics of the mathematical model, the game problem of the interaction between the driving status of a single vehicle and the driving status of the vehicle group is solved to obtain an equilibrium solution. The game problem of solving the interaction between the driving behavior of a single vehicle and the group of vehicles includes: The game problem is defined by the total driving cost of the vehicle; The total driving cost includes congestion cost, the magnitude of which is determined by the local vehicle density in the mathematical model; The evolutionary dynamics of the mathematical model include a stochastic influence term that simulates the uncertainty of driving behavior, and the stochastic influence term is set to be consistent with the direction of the road where the vehicle is currently located. The edge computing node generates and sends corresponding optimal scheduling instructions to at least one vehicle terminal based on the equilibrium solution.

2. The distributed resource game-theoretic scheduling method for smart parking lots as described in claim 1, characterized in that, The total operating cost also includes the cost of competing for parking spaces: The cost of competing for parking spaces is directly proportional to the real-time price of the target parking lot and inversely proportional to the expected number of available parking spaces in the parking lot calculated based on the mathematical model.

3. The distributed resource game-theoretic scheduling method for smart parking lots as described in claim 1, characterized in that, The process of obtaining the equilibrium solution is accelerated in the following ways: Before each solution iteration begins, a pre-trained deep neural network is used to predict the macroscopic distribution of the vehicle group in the new round based on the macroscopic distribution of the vehicle group in the previous round. The predicted macroscopic distribution of the new vehicle group is used as the initial condition for this round of iterative calculation.

4. The distributed resource game scheduling method for smart parking lots as described in claim 1, characterized in that, The generation and issuance of corresponding optimal scheduling instructions to at least one vehicle terminal includes: Based on the equilibrium solution, a suggested driving route and a speed range matching the driving route are planned for the vehicle terminal within a future preset time period. The route and speed range are then encapsulated into a structured data format and sent out.

5. The distributed resource game scheduling method for smart parking lots as described in claim 1, characterized in that, The method also includes a cross-regional collaboration step: Between edge computing nodes of two adjacent geographic regions, vehicle flow rate information on their shared boundary is exchanged; When each edge computing node solves a game problem, it uses the received vehicle flow rate information as the boundary condition for calculating the game problem.

6. The distributed resource game scheduling method for smart parking lots as described in claim 1, characterized in that, The method also includes an adaptive optimization step for the parameters in the mathematical model: Through the central cloud platform, based on scheduling performance metrics collected from one or more edge computing nodes, the weight parameters of the total travel cost in the game problem are automatically adjusted using a sequential optimization algorithm, and the adjusted weight parameters are then sent to the edge computing nodes.

7. The distributed resource game scheduling method for smart parking lots as described in claim 1, characterized in that, The vehicle terminal's status data also includes the vehicle's remaining battery or fuel level, the user's expected parking time, and user level information. The total driving cost defined in the game problem is adjusted based on the consumption cost of the remaining battery or fuel level in the vehicle terminal's status data and the priority of the user level information.

Images