Parking space recommendation and reservation method

Abstract

The present invention discloses a method for recommending and reserving parking spaces, comprising: regularizing acquired multi-source heterogeneous data to generate static parking lot attribute data, real-time parking space status data, and dynamic user request data; constructing an objective basic potential field using fuzzy reasoning based on a fuzzy rule set that can adaptively evolve based on performance feedback; modeling users' long-term and short-term preferences using methods such as hierarchical Bayesian and attention mechanisms to generate a user subjective preference modulator; and constructing a system total potential energy function by integrating the objective potential field, subjective modulator, and a multi-agent repulsive potential field with dynamically adjustable parameters. The method then determines and outputs a conflict-avoiding collaborative guidance strategy for all users by solving the minimum value of the function. The present invention can provide efficient, personalized parking guidance with macro-cooperation capabilities, significantly improving user experience and parking lot operation efficiency.

Description

Technical Field

[0001] The present invention relates to a parking recommendation method, in particular to a parking space recommendation and reservation method. Background Art

[0002] With the acceleration of global urbanization and the continued rise in car ownership, parking difficulties have become a widespread problem that hinders urban traffic efficiency and impacts residents' quality of life. Traditional parking lot management and guidance methods often cause drivers to waste a significant amount of time on ineffective detours and searches within the lot. This not only severely degrades the user's parking experience, but also leads to additional fuel consumption, exhaust emissions, and traffic congestion within the lot, exacerbating the overall burden on urban transportation.

[0003] To address these issues, existing technologies have explored a range of approaches. Initial technologies primarily focused on one-way information dissemination, such as using LED screens at entrances to display the number of available parking spaces in each area, or providing static parking lot location information and pricing information on mobile apps. With the development of the Internet of Things (IoT), more advanced solutions have begun to utilize sensors such as ultrasonic, geomagnetic, or video to detect the occupancy status of each parking space in real time. Based on this real-time data, some guidance systems can recommend available parking spaces to users. Recommendation logic is typically simple, for example, using the proximity principle to recommend the available space closest to the parking lot entrance. To improve the rationality of recommendations, some research has begun to incorporate artificial intelligence methods such as fuzzy reasoning. By presetting a fixed set of if-then rules (for example, if the distance is close AND the parking space is wide, then the recommendation is high), a weighted comprehensive evaluation of multiple parking space attributes, such as distance, width, and surveillance coverage, is performed to recommend the highest-scoring parking space. Furthermore, to model user preferences, some systems allow users to manually set preference tags (e.g., prefer charging stations, prefer locations near elevators), and then filter recommendations based on these static tags.

[0004] However, existing technologies still face the following specific technical issues when dealing with complex and dynamic real-world parking scenarios: First, the decision-making logic is static and rigid. Whether it is a simple proximity principle or fixed fuzzy rules, it cannot adaptively learn and respond to dynamic changes in the parking environment (such as changes in traffic patterns caused by the opening of new shops). As a result, its recommendation strategy may become mismatched with the real environment over time. Second, existing technologies generally adopt an individual optimal recommendation paradigm. When finding the optimal parking space for each user, they ignore the interactions between users. This can easily lead to group conflicts during peak hours when multiple users are directed to the same popular area, causing local congestion and wasted time searching for parking spaces, resulting in a dilemma of local optimality but global inefficiency. Summary of the Invention

[0005] The purpose of the invention is to provide a parking space recommendation and reservation method to solve the problems in the existing technology that the guidance strategy is single, cannot meet personalized needs and is prone to cause car-hunting conflicts.

[0006] Technical solution: a parking space recommendation and reservation method, including:

[0007] Process the acquired multi-source heterogeneous data to generate parking lot static attribute data, real-time parking space status data and dynamic user request data;

[0008] Based on the static attribute data of the parking lot and the real-time parking space status data, fuzzy reasoning is used to construct the objective basic potential field;

[0009] Based on dynamic user request data and historical behavior data, Bayesian reasoning is used to generate user subjective preference modulators;

[0010] The total potential energy function of the system is constructed by integrating the objective basic potential field, the user's subjective preference modulator, and the dynamic repulsive potential field calculated based on the status of all active users;

[0011] A global optimization is performed based on the total potential energy function of the system to generate a collaborative guidance strategy for parking spaces.

[0012] Preferably, when in use, the rule set based on the fuzzy reasoning is adaptively evolved, specifically:

[0013] Obtain the total potential energy index of the system output by the previous round of global optimization process, which describes the overall performance of the system;

[0014] Calculate the performance deviation that represents the difference between the system prediction and the user's actual choice;

[0015] When the system's total potential energy index continues to deteriorate or the accumulation of performance deviation exceeds a preset threshold, the topology of the rule set evolves to generate a new generation of optimal fuzzy rule sets.

[0016] Preferably, the topological evolution of the rule set adopts a reaction-diffusion based self-organization process, including:

[0017] Initialize the chemical concentration field representing the importance of each node and connection on the preset fully connected knowledge graph;

[0018] Based on the user's actual selection data, the potential rule paths on the knowledge graph are verified, and the weighted summation of multiple evolution results with different diffusion coefficients is performed to update the chemical concentration field; the concentration of the confirmed path is enhanced, and the concentration of the falsified path is suppressed;

[0019] The updated chemical concentration field is evolved and calculated, and the evolution process is guided by a topological constraint matrix constructed based on the logical relationship between rules.

[0020] After the concentration field is continuously iterated and a stable first concentration region emerges, the subgraph corresponding to the first concentration region is extracted and decoded to generate a new generation of optimal fuzzy rule sets.

[0021] Preferably, the topological evolution of the rule set is implemented based on a genetic algorithm, including:

[0022] Encode multiple sets of fuzzy rules into rule individuals with a tree structure to form an initial population;

[0023] Evaluate the fitness score of each regular individual in the population based on the system's total potential energy index;

[0024] Based on the fitness score, genetic operators including selection, crossover and mutation are performed on the regular individuals to reproduce a new generation of regular individuals;

[0025] The optimal individual is determined from the new generation of rule individuals and decoded to generate the new generation of optimal fuzzy rule sets.

[0026] Preferably, the total potential energy function of the system includes:

[0027] The basic allocation potential energy term, whose value is the sum of the global collaborative dynamic potential energy of all allocated users, is used to measure the sum of individual satisfaction with the allocation plan;

[0028] The path intersection penalty potential energy term, whose value is the sum of the path intersection penalties between all users' planned paths, is used to quantify the potential traffic congestion risk caused by the allocation plan;

[0029] Weight coefficient, used to weight the basic allocation potential energy term and the path crossing penalty potential energy term.

[0030] Preferably, the calculation process of the path crossing penalty potential energy term includes:

[0031] Each guided path is expressed as a sequence of discretized grid cells arranged in time sequence;

[0032] For any two paths, detect and record spatiotemporal conflict events that occupy the same grid cell at the same time step;

[0033] The sum of the spatiotemporal conflict events occurring between all path pairs is calculated to obtain the value of the path crossing penalty potential energy term.

[0034] Preferably, the process of generating the user subjective preference modulator includes:

[0035] Based on the user identity in the dynamic user request data, the user's decision features are extracted from the historical behavior database and quantified into the user's historical behavior feature vector;

[0036] The user's historical behavior feature vector is used as evidence and input into a pre-built dynamic Bayesian network model used to describe the temporal dependency relationship between user behavior and implicit preferences. Probabilistic reasoning is performed to solve the user preference probability distribution that represents the possibility of user implicit preferences, and based on this, a user subjective preference modulator is constructed.

[0037] Preferably, the process of constructing a user subjective preference modulator based on the user preference probability distribution includes:

[0038] Convert the user preference probability distribution into a dimension scaling factor vector through an exponential response function;

[0039] The components of the dimensional scaling factor vector are arranged on the diagonal of the matrix to construct an affine transformation matrix that serves as a modulator of the user's subjective preference.

[0040] Preferably, the dynamic repulsive potential field comprises:

[0041] The influence amplitude of the repulsive potential field A(ρ) is given by the linear growth formula A(ρ)=A base +k A ρ is determined; where ρ is the real-time traffic density, A base is the basic repulsion amplitude, k A is the amplitude growth coefficient;

[0042] The influence radius of the repulsive potential field σ(ρ) is determined by the inverse proportional attenuation formula σ(ρ)=σ base / (1+k σ ρ) is determined; where σ base is the basic repulsion radius, k σ is the radius attenuation coefficient;

[0043] The real-time traffic density ρ is the ratio of the number of active users looking for cars in the parking lot to the total number of currently available parking spaces.

[0044] Preferably, the process of generating the user subjective preference modulator specifically includes:

[0045] A hierarchical Bayesian model is used to model historical user behavior data to infer long-term stable preferences that represent users' inherent habits, and short-term dynamic preferences that represent users' recent intentions.

[0046] Generating a dynamic weight for integrating the long-term and short-term preferences according to the current dynamic user request data or real-time context;

[0047] Based on the dynamic weights, the long-term stable preference and the short-term dynamic preference are weightedly combined to obtain the final user preference distribution, and the user subjective preference modulator is constructed accordingly.

[0048] Preferably, the collaborative guidance strategy is generated by optimizing a global objective function including the states of at least two users and a potential field term or a penalty term used to characterize the interaction between the users.

[0049] Preferably, in the mutation operation of the genetic operator, when a new environmental variable is introduced, the step of regularizing the data will be automatically triggered to obtain and integrate real-time data related to the new environmental variable.

[0050] The present invention can provide efficient, personalized parking guidance with macro-coordination capabilities, significantly improving user experience and parking lot operation efficiency. The relevant technical effects will be described in detail below in conjunction with specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the present invention.

[0052] Figure 2 It is a flow chart of the adaptive evolution of the rule set based on the fuzzy reasoning of the present invention.

[0053] Figure 3 This is a flow chart of the topological evolution of a rule set based on a reaction-diffusion self-organizing process of the present invention.

[0054] Figure 4 It is a flow chart of the present invention for calculating the path crossing penalty potential energy term. DETAILED DESCRIPTION

[0055] In order to better illustrate the technical solution of the present invention, Figures 1 to 4 It should be noted that, unless there is any conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0056] The applicant has conducted an in-depth study of the existing technologies and analyzed the technical problems existing in the existing technologies as follows:

[0057] First, there's the issue of static decision-making logic and mismatch with the environment. Whether employing a simple proximity principle or a fuzzy inference system based on fixed rules, the decision-making logic is pre-set and unchanging, making it unable to adaptively respond to dynamic changes in the parking lot's internal and external environments. For example, when a popular new store opens in a certain area of ​​the parking lot and traffic surges, nearby parking spaces previously considered good may actually become unsafe or inconvenient due to the increased pedestrian traffic. However, static rule-based systems are unable to perceive and learn from this changing pattern and will continue to recommend these unsuitable spaces to users, resulting in a poor user experience. In other words, existing technologies generally lack a mechanism for autonomous rule evolution that enables their decision-making models to continuously learn from the environment, self-optimize, and iterate after deployment.

[0058] Secondly, there are the issues of group conflicts and local optimal traps caused by individual optimal recommendations. When making parking space recommendations, the basic idea of ​​existing technologies is to find the optimal parking space for each individual user, completely ignoring the interactions between users. This model may work during off-peak hours, but it can cause serious problems during peak hours. For example, when multiple users enter the parking lot at the same time, if the system guides them all to the same global optimal area (such as the area closest to the main entrance), it will inevitably cause serious traffic congestion and jams at the entrances and passages of this area, and users will conflict over the same or adjacent parking spaces. Ultimately, although everyone's goal is optimal, the lack of coordination causes the overall efficiency of the group's car search (i.e., the total potential energy of the system) to drop sharply. Existing technologies lack a negative entropy guidance mechanism that treats all users as a whole and performs global collaborative optimization, and are unable to fundamentally avoid group conflicts caused by individual optimality. To this end, the following solutions are provided:

[0059] Example 1: This example provides a parking space recommendation and reservation method. By constructing an adaptively evolving decision model and performing global collaborative optimization, it aims to address the technical issues of existing parking guidance systems, such as low efficiency, inability to meet user personalized needs, and inability to resolve traffic conflicts on a macro level. This example uses a reaction-diffusion scheme as the core of rule evolution to demonstrate the complete technical process. Specifically, the method may include the following steps:

[0060] Step S1: Process / regulate the acquired multi-source heterogeneous data to generate parking lot static attribute data, real-time parking space status data, and dynamic user request data.

[0061] Static parking lot attribute data is a collection of data describing the fixed physical characteristics of a parking lot, such as the geographic coordinates, dimensions, and whether each parking space is a special parking space (charging, disabled access), etc. Real-time parking space status data is dynamic, time-stamped data describing whether each parking space is currently occupied. Dynamic user request data refers to data packets initiated in real time by users through the client app, containing their identity, location, vehicle information, and potential intent.

[0062] In some embodiments, specifically:

[0063] During system initialization, the system loads the parking lot's high-precision map data. Specifically, it uses geometric parsing and attribute extraction algorithms to identify and encode each parking space's unique ID, geographic coordinates, dimensions, and whether it's a charging or disabled parking space. Ultimately, it outputs structured static attribute data for the parking lot.

[0064] The system continuously receives raw parking lot sensor data from geomagnetic sensors or surveillance cameras deployed under each parking space. For example, for geomagnetic data, a threshold for magnetic field change can be set to determine vehicle entry and exit. For video stream data, a YOLOv8-based visual recognition model can be applied to select and identify vehicles in parking spaces in real time, thereby determining the occupancy status of each space. This module ultimately outputs real-time parking space status data with a timestamp.

[0065] Specifically, the YOLOv8 model was trained on a dataset of approximately 100,000 parking lot images captured under varying lighting conditions, weather conditions, and congestion levels, encompassing over 5,000 different parking spaces. During deployment, the model was quantized and run on the parking lot's edge computing gateway, enabling real-time detection and analysis of a single 1080p video stream in under 50 milliseconds, ensuring the high timeliness of parking space status data.

[0066] The server interface generates dynamic user request data by monitoring and parsing real-time user app requests, extracting information such as user ID, current GPS location, vehicle model, and possible on-site destination descriptions (such as near the elevator in Area B).

[0067] By standardizing multi-source heterogeneous data, unified, clean, and real-time data input is provided for subsequent potential field construction, preference reasoning, and global optimization, solving the technical problems of diverse sources, different formats, and difficulty in direct use of original data.

[0068] In some embodiments, data acquisition can also include the collection of environmental data. For example, when a subsequent rule evolution step (such as a genetic algorithm mutation operation) generates a rule containing a new variable (such as weather), it triggers a request for a new data dimension. At this point, the dynamic data demand response function in this step is activated: weather data is acquired through external APIs and integrated into the data structure of the parking lot's static attribute data for use in the new rule, achieving dynamic data demand response.

[0069] Step S2, adaptive fuzzy reasoning and potential field construction based on topological evolution, specifically includes: using fuzzy reasoning to construct an objective basic potential field based on parking lot static attribute data and real-time parking space status data.

[0070] Among them, the potential field P is used to map any parking space s_i in the parking space set to a function of real potential energy value, that is, P:s_i→E_i.

[0071] Furthermore, in this embodiment, the rule set underlying fuzzy reasoning is adaptively evolving. To achieve this evolution, the method also includes: obtaining a system total potential energy indicator, output by the global optimization process, that quantifies the overall system performance; calculating a performance deviation that characterizes the difference between the system's prediction and the user's actual selection; and initiating topological evolution of the rule set when the system total potential energy indicator continues to degrade or the cumulative performance deviation exceeds a preset threshold. This topological evolution uses the system total potential energy indicator as the core fitness function to generate a new generation of optimal fuzzy rule sets, thereby dynamically updating the inference logic underlying the subsequent construction of the objective foundation potential field.

[0072] Furthermore, in this embodiment, the evolution process is specifically as follows:

[0073] Initiating the topological evolution of the rule set specifically includes a reaction-diffusion-based self-organizing process, which is as follows: on a preset fully connected knowledge graph, a chemical concentration field that characterizes the importance of each node and connection is initialized; based on the user's actual selection data, the potential rule paths on the knowledge graph are verified to update the chemical concentration field, wherein the concentration of the verified path is nonlinearly enhanced, while the concentration of the falsified path is suppressed; applying the reaction-diffusion equation based on the graph Laplace operator, the updated chemical concentration field is evolved and calculated to simulate the knowledge propagation between logically associated nodes; after the concentration field is continuously iterated and spontaneously organized to produce stable high-concentration areas, the subgraphs corresponding to these areas are extracted and decoded to generate a new generation of optimal fuzzy rule sets.

[0074] The objective base potential field is a function or data structure that maps each unoccupied parking space in the parking lot to a scalar potential energy value. Lower potential energy values ​​represent higher overall evaluations of the parking space within the objective physical environment. The chemical substance concentration field is a mathematical tool used in this embodiment to represent the importance or confidence of rules. High-concentration regions correspond to more effective and reliable rule combinations within the current environment.

[0075] In this embodiment, the process mainly includes:

[0076] First, for each unoccupied parking space, extract its multi-dimensional attributes from the static attribute data of the parking lot and the real-time parking space status data generated in step S1 to form a parking space multi-dimensional numerical vector V i .

[0077] Secondly, the current optimal fuzzy rule set is loaded from the rule gene library, and the parking space basic potential energy scalar is calculated for each parking space i through standard processes such as fuzzification, rule activation, reasoning, aggregation and defuzzification. The collection of all these scalars constitutes the objective basic potential field.

[0078] Specifically, in some embodiments, this step may be:

[0079] Load the fuzzy rule set currently marked as optimal from the rule gene library.

[0080] The multi-dimensional numerical vector of parking spaces is read and fuzzified, and then the rule activation degree set of each rule in the fuzzy rule set is calculated.

[0081] The rule activation degree set is read, and the output fuzzy set of each rule is truncated using the Mamdani implication method to obtain the rule consequent fuzzy set.

[0082] The rule consequent fuzzy set is received, and the max aggregation operator is used to superimpose all consequent fuzzy graphs to obtain the aggregated total output fuzzy set.

[0083] The center of gravity method is applied to the aggregated total output fuzzy set for defuzzification, the basic potential energy scalar of the parking space is calculated, and the potential energy scalars of all parking spaces are summarized to finally construct the objective basic potential field.

[0084] Finally, the system continuously monitors two performance indicators: E_best, the system's total potential energy indicator (feedback from the previous round), and the prediction-to-reality deviation δ. If E_best continues to deteriorate, or if the moving average of δ exceeds a preset threshold θ (for example, θ = 0.3), the system triggers the following rule-based self-organization process.

[0085] Performance deviation δ=(1 / N)∑|Recommended _Score (i)-Actual _Choice_Score(i)|, where Recommended _Score is the system's rating of parking spaces. _Choice_Scor e is a binary variable (1 if selected by the user, 0 otherwise).

[0086] The self-organizing topological evolution based on reaction-diffusion rules includes the following steps:

[0087] Load a pre-set, fully connected knowledge graph G containing all potential fuzzy variables, fuzzy sets, and logical operators. On this graph, assign an initial rule catalyst concentration C(0) to each node and edge, for example, uniformly set to 0.1. The entire graph and its concentration distribution together constitute the rule topology graph and the initial concentration field. Concentration represents the importance or confidence of a concept or connection in the current rule system.

[0088] After receiving the user's actual selection data, if the user's behavior confirms a potential rule path, the concentration C(t) of all nodes and edges on the path is nonlinearly enhanced (catalytic reaction). Otherwise, its concentration is suppressed.

[0089] For example, when a user's actual selection behavior confirms a potential rule path (for example, the user chooses a nearby and safe parking space, and this path corresponds to a high potential energy prediction on the graph), the catalyst concentration of all nodes and edges along this path is nonlinearly enhanced (catalytic reaction). Conversely, if the user's behavior falsifies a rule path, its concentration is suppressed. This process updates the chemical concentration field C(t) at the previous moment to the chemical concentration field C'(t) after the reaction.

[0090] The concentration field evolution of the entire graph is calculated using a set of reaction-diffusion partial differential equations based on the graph Laplace operator. This equation simulates the propagation and mutual influence of confidence between logically connected nodes.

[0091] Specifically, the formula is: dCk / dt =R k (C)+∑ j D kj grad 2 C j . Among them, Rk(C) is the reaction term, D kj is the diffusion coefficient matrix between different chemical substances, grad 2 is the graph Laplacian. This step simulates the propagation and mutual influence of knowledge or confidence between logically connected nodes (for example, from safe nodes to monitored nodes). After numerical integration over one time step, the evolved chemical concentration field C(t+1) is obtained.

[0092] After a sufficiently long iteration period, the concentration field will spontaneously organize into multiple stable high-concentration regions (Turing patterns). These high-concentration subgraphs are extracted using graph theory algorithms and decoded into human-readable fuzzy rules, forming a new generation of optimal fuzzy rule sets. These are then updated back into the rule gene library for use in the next round of routine calculations. The subgraphs circled on the topological map by the stable pattern represent the most stable and collaborative fuzzy rules that have emerged in the current environment.

[0093] In this embodiment, for example, R k (C) can be a nonlinear function (such as a quadratic or cubic polynomial) related to the concentration C, D kj It can be a diagonal matrix, where the diagonal elements represent the diffusion capacity of each chemical substance. In some embodiments, the model that can be used is as follows:

[0094] R k (C)=r k *C k *(1-C k / K k ); where r k is the growth rate, K k is the carrying capacity;

[0095] R k (C)=α k * C k -β k * C k 2 -Σ(γ kj * C k *C j ); where α k is the catalytic coefficient, β k is the self-inhibition coefficient, γ kj is the competition coefficient;

[0096] R k (C)=f(feedback) * (α* C k -β* C k 2 ); f(feedback) is adjusted according to the user's actual selection data.

[0097] This step addresses the technical issues of traditional systems, which suffer from rigid logic and an inability to adapt to changing environments. By simulating chemical reactions and diffusion processes, the system allows the optimal combination of rules to emerge spontaneously from a vast potential space of rules, rather than relying on traditional brute-force search. This is more efficient and intelligent.

[0098] Step S3: Constructing user subjective preference modulator based on Bayesian reasoning

[0099] Furthermore, the process of generating a user subjective preference modulator includes: based on the user identity in the dynamic user request data, extracting the user's decision characteristics from the historical behavior database and quantifying them into a user historical behavior feature vector; using the user historical behavior feature vector as evidence, instantiating it into a pre-defined dynamic Bayesian network model that describes the temporal dependency relationship between user behavior and implicit preferences; performing probabilistic reasoning on the instantiated dynamic Bayesian network to solve the user preference probability distribution that represents the possibility of the user's implicit preference, and constructing a user subjective preference modulator based on the distribution.

[0100] Furthermore, the construction process is specifically as follows: the user preference probability distribution is converted into a dimensional scaling factor vector through an exponential response function; the components of the dimensional scaling factor vector are arranged on the diagonal of the matrix, thereby constructing an affine transformation matrix as a modulator of the user's subjective preference.

[0101] In this embodiment, this step can construct a corresponding personalized filter for each user, namely a user subjective preference modulator. This modulator can adjust the objective and universal potential field to a subjective potential field that conforms to the user's personal habits and preferences.

[0102] In this paper, the user's subjective preference modulator is strictly defined as a mathematical operator, specifically an affine transformation matrix. Its function is to operate on the parking space attribute vector representing the objective environment, thereby mathematically modulating or distorting the objective underlying potential field. A dynamic Bayesian network (DBN) is a probabilistic graphical model for processing time series data. It can infer the probability distribution of hidden states (such as user preferences) based on a series of observed behaviors (evidence).

[0103] In some embodiments, the process is as follows:

[0104] When dynamic user request data is generated, the system queries the backend database of historical user behavior based on the user ID. Specifically, the system collects and quantifies the user's parking selection characteristics over the past N times (e.g., N = 50), such as the average distance between the selected parking space and the elevator, the frequency of selecting charging pile parking spaces, the average parking time, and the preference for spacious parking spaces (derived by comparing the vehicle model with the selected parking space size). These quantitative characteristics are then combined into a user historical behavior feature vector.

[0105] The system loads a predefined 2-TBN (Dual Time Slicing Bayesian Network) network structure. This network structure describes the causal relationship and temporal dependencies between users' implicit preferences (such as convenience, safety, and cost sensitivity) and their observable behavioral characteristics (such as choosing to be near an elevator, in a monitored area, or in an affordable area).

[0106] The user historical behavior feature vector generated in the previous step is used as evidence and instantiated into the DBN model.

[0107] On this instantiated DBN, we execute an exact inference algorithm, such as the junction tree algorithm, to calculate the posterior probability of each implicit preference node given the observed behavior. Ultimately, we solve for the user's preference probability distribution. For example, the preference distribution for user A is {P (convenience-seeking) = 0.8, P (security-conscious) = 0.5, P (cost-sensitive) = 0.2}.

[0108] The above probability distribution is aligned with the specific physical attribute dimensions of the parking space through a preset mapping rule. For example, the preference for convenience can be aligned with the distance from the elevator / exit, a parking space attribute dimension.

[0109] Using an exponential response function, we convert the probability distribution of user preferences into a vector of dimensional scaling factors. For example, scaling factor = exp(k(probability - 0.5)), where k is the gain factor. For user A, the scaling factor on the convenience dimension will be significantly greater than 1, while the scaling factor on the cost-sensitive dimension will be less than 1.

[0110] The components of the dimensional scaling factor vector are arranged on the diagonal of the square matrix, and the remaining positions are 0, thereby constructing a diagonal affine transformation matrix M. This matrix M is the final generated user subjective preference modulator.

[0111] This step addresses the pain point of traditional recommendation systems, which lack a deep understanding of users' personalized and implicit needs. By introducing DBN, the system can dynamically learn from users' evolving preferences through their behavioral sequences, rather than relying on static user labels. Even more innovative, it cleverly transforms these fuzzy, probabilistic preferences into a precise, operational linear algebraic operator (the modulator matrix), providing a solid foundation for the subsequent seamless mathematical fusion of subjective preferences with the objective environment.

[0112] For example, suppose there are two empty parking spaces, P1 and P2, in a parking lot. In the objective underlying potential field, P1 has a lower potential energy (more favorable) because it is closer to the entrance, while P2 has a higher potential energy because it is farther away. Their potential energies are represented by vectors V_P1 and V_P2.

[0113] After analysis, user A generates the user subjective preference modulator M A There is a strong scaling (enhancement) effect in the dimension close to the elevator.

[0114] In the following steps, the system will calculate the subjective potential energy: V _subjective_P1 =M _A V _P1 and V _subjective_P2 =M_A V _P2 .

[0115] If parking space P2 is right next to the elevator in area B, even if its original objective potential energy is higher, after passing through the modulator M A After the effect of , the advantage of being close to the elevator is greatly magnified, and the final subjective potential energy V_subjective_P2 may actually be lower than V_subjective_P1. Therefore, the system will prioritize recommending P2 to user A, achieving truly personalized recommendations.

[0116] Step S4, global collaborative dynamic potential field generation and negative entropy guidance strategy calculation, including: integrating the objective basic potential field, the user subjective preference modulators of all active users, and the dynamic repulsion effect between users, and generating and outputting a collaborative guidance strategy for all users by solving a global optimization problem aimed at minimizing the total potential energy of the system.

[0117] Furthermore, a global optimization is performed based on the total potential energy function of the system. Specifically, the final collaborative guidance strategy is determined by solving the minimum value of the total potential energy function of the system. The total potential energy function of the system includes:

[0118] The basic allocation potential energy term, whose value is the sum of the global collaborative dynamic potential energy of all allocated users, is used to measure the sum of individual satisfaction with the allocation plan;

[0119] The path intersection penalty potential energy term, whose value is the sum of the path intersection penalties between all users' planned paths, is used to quantify the potential traffic congestion risk caused by the allocation plan;

[0120] Weight coefficient, used to weight the basic allocation potential energy term and the path crossing penalty potential energy term.

[0121] Furthermore, the calculation process of the path crossing penalty potential energy term includes: expressing each guided path as a sequence of discretized grid cells arranged in time sequence; for any two paths, detecting and recording the spatiotemporal conflict events occupying the same grid cell at the same time step; and the value of the path crossing penalty potential energy term is determined as the sum of the spatiotemporal conflict events occurring between all path pairs.

[0122] Moreover, the dynamic repulsion effect is achieved by constructing a multi-agent repulsion potential field whose key parameters are dynamically adjusted according to the real-time traffic density, and the influence amplitude of the repulsion potential field increases with the increase of traffic density, so as to numerically increase the peak value of the repulsion potential energy; the influence radius of the repulsion potential field decreases with the increase of traffic density, so as to spatially shorten the effective range of the repulsion potential energy.

[0123] In addition, the real-time traffic density is calculated as follows: the real-time traffic density ρ is quantified into a normalized dimensionless value, which is obtained by calculating the ratio of the number of active car-seeking users in the parking lot to the total number of currently available parking spaces.

[0124] In this embodiment, this step is no longer about finding the optimal parking space for a single user, but rather considering all users who are looking for a car as a whole system. By solving the global optimization problem, a collaborative allocation solution with the lowest total cost, the least conflict, and the highest overall satisfaction is found for everyone.

[0125] The total system potential energy function, E(S), quantifies the total system cost or disorder of a given parking allocation solution, S, encompassing all users. High E(S) values ​​correspond to high entropy states (e.g., a large number of users are directed to congested areas, resulting in significant path intersection). Negative entropy guidance, as defined in this paper, is an optimization process that seeks to minimize the total system potential energy function, E(S). A collaborative guidance strategy is a globally optimized and conflict-reconciled set of navigation instructions for all active users, as distinguished from independent individual optimal navigation.

[0126] In this step, the main processes are as follows:

[0127] For each active user u, the system generates the user subjective preference modulator M u (a matrix) acts on the generated objective basic potential field. Specifically, through matrix transformation, a unique individual subjective potential field is generated for the user.

[0128] The system calculates the real-time traffic density ρ based on the current state. The calculation formula is: ρ = (number of active users looking for a car) / (total number of currently available parking spaces). ρ is a normalized dimensionless value that intuitively reflects the degree of parking resource scarcity.

[0129] Then, the ρ value is substituted into the preset parameter adjustment formula to dynamically calculate the influence amplitude A(ρ) and influence radius σ(ρ) of the repulsive potential function. In this embodiment, the formula is: A(ρ)=A base +k A× ρ and σ(ρ)=σ base / (1+k σ× ρ), where A base ,k A ,σ base ,k σ All are preset coefficients.

[0130] The Gaussian function is used as the single-agent repulsion potential function, and the repulsion potential fields of all users are linearly superimposed to finally obtain the multi-agent repulsion potential field.

[0131] For each user u and each parking space i, the system uses a weighted sum formula P_global(u,i) =w _ s P _subjective(u,i) +w _r P _repulsion (i) The individual subjective potential field and the multi-agent repulsive potential field are synthesized point by point to obtain the final global collaborative dynamic potential field.

[0132] The system constructs the final system total potential energy function E(S), whose mathematical expression is: E(S)=∑ ui∈U P global (u i ,s i )+λ∑ i The first term of this function is the sum of the potential energy of all users in their allocated parking spaces (basic allocation potential energy term), the second term is the sum of the path crossing penalties between all pairs of users (path crossing penalty potential energy term), and λ is the weight coefficient that balances the two.

[0133] For any two planned paths, route_u and route_v, the system first divides the parking map into a uniform grid and represents the path as a sequence of grid cells arranged in time. It then calculates the penalty term C(route_u, route_v) by detecting the number of times the two paths occupy the same grid cell at the same time step.

[0134] A simulated annealing algorithm is used to minimize the total potential energy function of the system. Specifically, a complete user-parking space assignment is defined as a state of the algorithm, and randomly swapping two users' assignments is defined as a neighborhood perturbation operation. In other words, randomly swapping two users' assignments is defined as a neighborhood function. Starting from a random state, the algorithm iteratively searches and probabilistically accepts worse solutions according to the Metropolis criterion, gradually annealing until convergence criteria are met (e.g., temperature below a threshold, energy continuously stable, or the maximum number of iterations is reached). Ultimately, the optimal parking space assignment S_best with the lowest total energy is found.

[0135] The optimal parking space allocation plan S_best is converted into specific navigation path instructions for each user to form the final output collaborative guidance strategy.

[0136] The system's lowest total potential energy value E_best is used as a quantitative global performance indicator and fed back to the evolution trigger and fitness evaluation module to drive and evaluate rule evolution.

[0137] This step creatively transforms the vehicle-finding problem from multiple independent, single-objective optimization problems into a multi-objective, coupled, global combinatorial optimization problem. By minimizing the total energy of four factors, including individual preferences, the physical environment, group exclusion, and path conflicts, the system is able to find a collaborative solution that avoids the prisoner's dilemma and is optimal for all participants. This overcomes the technical challenges of traditional guidance methods, which are prone to local congestion and path conflicts.

[0138] For example, in a certain scenario: during peak hours, users A and B enter the parking lot at the same time. The system identifies two empty parking spaces: P1 (close, but with a narrow path) and P2 (slightly farther, but with an open path).

[0139] In a candidate allocation scheme S1={A→P1,B→P2}, the basic allocation potential energy term may be low, but due to the overlap of paths A and B, the path crossing penalty term is high, resulting in a high total potential energy E(S1).

[0140] In another candidate solution S2={A→P2,B→P1}, the basic allocation potential energy term may be slightly higher, but due to the staggered paths, the path crossing penalty term is zero, resulting in the total potential energy E(S2) being lower.

[0141] At the same time, due to the large number of users in the parking lot and the high traffic density ρ, the influence amplitude and radius of the repulsive potential field have changed, making the system tend to guide users to areas farther away.

[0142] Through exploration, the simulated annealing algorithm ultimately found S2 to be the global optimal solution and output the corresponding collaborative guidance strategy: User A goes to P2, and User B goes to P1. Test data shows that compared to traditional guidance methods, this method can reduce the conflict rate of car search paths during peak hours by approximately 40% and the average car search time by approximately 25%.

[0143] Through the above steps, this embodiment fully describes an intelligent parking guidance method that can achieve personalization, self-adaptation and global coordination.

[0144] In another embodiment of the present application, another expression of the system total potential energy function E(S) is:

[0145] The system state S is defined as assigning all active users U={u1,u2,...} to a set of unique parking spaces P S ={s1,s2,...}. The total potential energy function E(S) of the system consists of two parts:

[0146] Base allocation potential and path crossing penalty potential.

[0147] E(S)=∑ ui∈U P global (u i ,si )+λ i <∑ j C(route(u i ,s i ),route(u j ,s j ));

[0148] Where: P global (u i ,s i ) is user u i In its assigned parking space i The global cooperative dynamic potential energy value on C(route u ,route v ) is the cross penalty term between the planned paths of user u and user v. λ is a weight coefficient used to balance the quality of the parking space itself and the severity of the path conflict.

[0149] In another embodiment of the present application, the termination (convergence) condition of the simulated annealing algorithm is composed of the following three conditions, and the iteration is terminated when any one of them is satisfied:

[0150] ① The current temperature T drops to the preset minimum temperature T_min (e.g. 1e-5). This is the most important termination condition.

[0151] ② In k consecutive iterations (k is a preset value, such as 100), the change in the system's optimal energy value E(S_best) is less than a very small threshold ε (such as 1e-6). This indicates that the system has entered a stable energy basin.

[0152] ③ The total number of iterations reaches the preset upper limit Iter_max (for example, 1,000,000) to prevent the algorithm from not converging for a long time in extreme cases.

[0153] In another embodiment of the present application, the path intersection penalty term C(route_u, route_v) is used to quantify the degree of conflict between two guided paths in time and space, and is calculated as follows:

[0154] The parking map is divided into a uniform grid. For any guided route, it is represented as a sequence of grid cells arranged in time order: route = {cell_1(t_1), cell_2(t_2), ...}.

[0155] For two paths route_u and route_v, if there is a time step t where they are located in the same grid cell, that is, cell_u(t)=cell_v(t), then it is recorded as a spatiotemporal conflict. The penalty term C is defined as the sum of the number of spatiotemporal conflicts between the two paths during the entire guidance process.

[0156] It should be noted that the conflict resolution mechanism is inherent in the optimization process, rather than an independent post-processing step.

[0157] The repulsive potential field creates a "repulsive zone" for each user in space, physically preventing the intention of directing multiple users into the same small area and avoiding conflicts at the source. Global optimization treats the allocation of all users as a holistic problem. By minimizing the system's total potential energy function E(S), which incorporates all user potentials and path conflicts, the algorithm seeks the globally optimal, conflict-minimizing allocation solution. Therefore, the optimal solution S_best already resolves competing conflicts. Conflict resolution is achieved by selecting the path with lower total energy (i.e., less conflict), rather than reconciling conflicts after the fact.

[0158] In another embodiment of the present application, in the multi-agent repulsive potential field, the parameters A (amplitude) and σ (radius) of the repulsive potential function are dynamically adjusted according to the traffic density ρ to achieve better adaptability. As mentioned above, the specific functional relationship can be set as: A(ρ)=A base +k A× ρwhere A base is the basic repulsion amplitude, k A is the amplitude growth coefficient. σ(ρ)=σ base / (1+k σ× ρ); where σ base is the base repulsion radius, k σ is the radius attenuation coefficient.

[0159] Therefore, in this embodiment, the amplitude A(ρ) adopts a linear growth relationship. As the density increases, the repulsion strength should also increase linearly to exert a greater avoidance thrust. The radius σ(ρ) adopts an inversely proportional decay relationship. As the density increases, each user's "personal space" should be appropriately reduced to accommodate more people and avoid an excessively large impact area that would result in an unsolvable problem.

[0160] According to one aspect of this application, multiple potential fields (individual subjective potential field, multi-agent repulsive potential field) are linearly superimposed. To ensure numerical stability, the following processing method is adopted:

[0161] All potential energy calculations use double-precision floating-point numbers (double) to ensure sufficient accuracy and prevent precision loss in large amounts of addition operations.

[0162] Before superposition, various potential fields can be normalized (for example, their value ranges can be scaled to [0,1]) so that the potential energies from different sources are at the same numerical level, thus preventing a certain type of potential energy from dominating the entire superposition result due to its excessively large value.

[0163] Due to the characteristics of the Gaussian repulsive potential, at any point i, the total repulsive potential energy is primarily contributed by the few closest users, while the potential energy contributed by more distant users decays exponentially to near zero. This naturally avoids catastrophic floating-point cancellation or overflow issues that can occur when summing the values ​​of all users in the field.

[0164] In this scheme, negative entropy is used to describe the process of a system evolving from "disorder" to "order".

[0165] In this context, entropy is defined as the degree of disorder or inefficiency in the parking system, and its quantifiable metric is the system's total potential energy function, E(S). High E(S) values ​​correspond to high entropy states (for example, a large number of users are directed to congested areas, resulting in significant path intersections and low overall parking efficiency). The negative entropy guidance process is defined as an optimization process aimed at minimizing the system's total potential energy function, E(S).

[0166] In another embodiment of the present application, in collaborative guidance, the guidance strategy is generated based on global information rather than local information. A guidance strategy is collaborative if and only if it is generated by optimizing a global objective function (i.e., E(S)) that incorporates the states and interactions of at least two or more users.

[0167] The criterion for determining whether a guidance system is collaborative is whether changing the goals or preferences of one user, u_A, could potentially affect the guidance strategy of another, unrelated user, u_B. In this solution, the answer is yes, because any change in any user affects the global optimal solution of E(S), potentially changing the optimal strategy for everyone. In an individual guidance system, users are isolated from each other.

[0168] In this embodiment, most steps of the method, including S1 data acquisition and normalization, S3 user preference modulator construction, and S4 global collaborative optimization and strategy calculation, are the same or substantially the same as those in the first embodiment. The main difference in this embodiment is that the rule evolution process in step S2 is implemented using a different technical path.

[0169] In this embodiment, when the evolution trigger is activated, the system will execute the optimization process based on the genetic algorithm. The process is:

[0170] Multiple sets of fuzzy rule sets are encoded into a group of rule individuals with a tree structure to form an initial population; the fitness score of each rule individual in the population is evaluated based on the system's total potential energy index; based on the fitness score, genetic operator operations including selection, crossover and mutation are performed on the rule individuals to reproduce a new generation of rule individuals; the optimal individual is determined from the new generation of rule individuals and decoded to generate a new generation of optimal fuzzy rule sets.

[0171] In this embodiment, the system simulates the principle of survival of the fittest in biological evolution, and ultimately selects the rule set with the best performance from a population consisting of a large number of candidate rule sets through multiple generations of reproduction and selection.

[0172] Genetic algorithms are used to search for optimal solutions in complex spaces. A rule population is a collection of multiple candidate fuzzy rule sets, each of which participates in the evolutionary process as an individual. The fitness score is a quantitative value used to evaluate the quality of each individual rule (i.e., rule set). Individuals with higher scores are more likely to be selected during reproduction.

[0173] Specifically, the system extracts multiple different fuzzy rule sets from a rule gene library. Each rule set is encoded as a rule individual with a tree structure. For example, the rule IFaisXANDbisYTHENcisZ can be encoded as a tree with the logical operator AND as the root node, and the child nodes being the conditions aisX and bisY. Combining multiple such individuals forms the algorithm's initial population.

[0174] For each individual rule in the population, the system needs to evaluate its performance. Specifically, the system uses the system's total potential energy indicator, E_best, as the core evaluation criterion, and calculates each individual's fitness score, Fitness, using the following formula: Fitness = -E_best. It should be noted that lower E_best values ​​represent better system performance. Therefore, taking a negative value ensures that the better-performing rule sets receive higher fitness scores, which conforms to the principle of prioritizing high-fitness individuals in genetic algorithms.

[0175] The operation process of the genetic operator is as follows:

[0176] Based on the fitness score calculated in the previous step, a roulette wheel selection algorithm is used to select the parent individuals from the current population. Individuals with higher fitness scores have a greater probability of being selected.

[0177] Two parent individuals are randomly selected and a subtree crossover operation is performed on them. Specifically, a crossover point is randomly selected in the tree structure of the two parent individuals, and then the entire subtree below these two points is exchanged to generate two new offspring individuals.

[0178] In order to maintain the diversity of the population and explore new possibilities, the newly generated offspring individuals are subjected to small-probability node mutations, for example, randomly changing the operator of a node (such as AND to OR), or changing a fuzzy set (such as close to very close).

[0179] Optionally, as a preferred approach, the mutation operation can also include the introduction of a new environment variable. For example, a leaf node might be replaced with weather instead of distance. When this mutation occurs, the system recognizes this as a new variable that doesn't exist in the current data structure and immediately triggers the dynamic data demand response function to obtain weather data through external APIs and integrate it into the data stream for use by this new rule.

[0180] New population generation and optimal rule updating: New individuals, after undergoing crossover and mutation operations, replace individuals with lower fitness in the population, thus forming a new generation of rule populations. The algorithm terminates after a preset number of evolutionary generations (e.g., 200 generations) or when the population's highest fitness score stops improving significantly over multiple generations. Finally, the system selects the rule individual with the highest fitness from the final generation population, decodes it from the tree structure into human-readable fuzzy rule text, and updates it back into the rule gene library as the new generation's optimal fuzzy rule set.

[0181] This embodiment provides another powerful mechanism for achieving adaptive evolution of rules. Compared with the self-organizing pattern based on chemical diffusion in Example 1, the genetic algorithm is a widely verified and robust global search heuristic algorithm. By exploring multiple points in the rule space in parallel and using operations such as crossover and mutation, it can effectively jump out of the local optimal solution and efficiently converge to a rule set with better global performance in the huge and complex potential rule combination space. In this way, the technical problem of the traditional system's solidified logic and inability to continuously optimize is also solved.

[0182] According to one aspect of the present application, a dynamic Bayesian network (DBN) structure example is as follows:

[0183] In this embodiment, the 2-TBN network structure for user preference inference can be defined as follows:

[0184] Hidden Nodes: represent the user’s implicit preferences, for example:

[0185] H1: Convenience_Preference (convenience preference);

[0186] H2: Safety_Preference (safety preference);

[0187] H3: Cost_Preference (cost preference);

[0188] Observed Nodes: Represents the historical behavior characteristics of users that can be directly observed, for example:

[0189] O1:Chose_Near_Elevator (selected a parking space near the elevator);

[0190] O2:Chose_Well_Lit_Area (selected the well-lit area);

[0191] O3:Chose_Under_Camera (selected a parking space covered by surveillance);

[0192] O4:Chose_Low_Price_Zone (Chose a parking space in the low-price zone).

[0193] Network topology (edges): defines the causal dependencies between nodes.

[0194] H1 points to O1 (convenience preference influences whether to choose to be close to the elevator);

[0195] H2 points to O2 and O3 (safety preference influences whether to choose areas with good lighting and surveillance);

[0196] H3 points to O4 (cost preference affects whether to choose a low-price zone);

[0197] In addition, the state of each hidden layer node at time step t also depends on its state at time step t-1 to reflect the temporal continuity of preferences.

[0198] According to one aspect of the present application, an example of a fuzzy rule set is as follows:

[0199] In this embodiment, the rule gene library may include but is not limited to the following specific fuzzy rules:

[0200] R1: IF distance to entrance is close AND monitoring coverage is high THEN potential energy is very low;

[0201] R2: IF distance to elevator is far OR light intensity is weak THEN potential energy is high;

[0202] R3:IF parking space width is wide AND path complexity is low THEN potential energy is low;

[0203] R4: IF charging pile is yes AND user request is charging THEN potential energy is very low;

[0204] R5: IF historical congestion is high AND current traffic density is high THEN potential energy is very high;

[0205] Preferably, in order to further improve the convergence speed of the rule evolution process and the rationality of the results, the following method can be used for the evolution calculation step:

[0206] That is, the steps of applying the reaction-diffusion equation for evolutionary calculation are further defined as: using a multi-scale diffusion mechanism to update the chemical concentration field by weighted summing of multiple evolutionary results with different diffusion coefficients; and the evolutionary calculation process is guided by a topological constraint matrix constructed based on the logical relationship between rules.

[0207] In this preferred embodiment, two key improvements are made to the reaction-diffusion process to solve the problems of slow convergence or diffusion to logically irrelevant areas that may occur on high-dimensional complex knowledge graphs.

[0208] A multiscale diffusion mechanism is a strategy that uses a set of diffusion coefficients rather than a single one to simultaneously simulate the spread of knowledge at different distance scales. A topological constraint matrix is ​​a predefined matrix that encodes the strength of the logical connections between nodes in a knowledge graph and is used to guide the flow of chemical concentrations in diffusion calculations.

[0209] Specifically, the system no longer uses a single diffusion coefficient D, but presets a set of diffusion coefficients, such as {D_local, D_regional, D_global}, which represent the diffusion intensities of small range (local), medium range (regional) and long distance (global), respectively.

[0210] In each evolution time step, the system performs three independent diffusion calculations on the concentration field C'(t) after the reaction based on the three diffusion coefficients, and obtains three different candidate concentration fields: C _local (t+1),C _regional (t+1),C _global (t+1).

[0211] Finally, through a weighted summation formula

[0212] C(t+1)=w _local C _local (t+1)+w _regional C _regional (t+1)+w _global C _global (t+1) to generate the final updated concentration field. The weight coefficient w can be preset (for example, w _local =0.5,w _regional =0.3,w _global=0.2), and can also be adjusted adaptively based on indicators such as the system entropy change rate.

[0213] The mechanism of this embodiment enables knowledge closely related to a node to be quickly deepened locally (via D_local), while some knowledge of universal significance can also be efficiently propagated to distant areas of the graph (via D_global), thereby improving evolution efficiency.

[0214] Introduction and application of topological constraints:

[0215] During system initialization, domain experts can construct a topological constraint matrix T based on prior knowledge. The dimensions of this matrix are the same as the adjacency matrix of the knowledge graph. If nodes i and j are logically strongly related (for example, i is monitored and j has a high safety rating), the value of the matrix element T(i, j) is close to 1. If the two are logically unrelated or mutually exclusive (for example, i is close to the entrance and j has a high charging power), the value of T(i, j) is close to 0.

[0216] When performing diffusion calculations, the original reaction-diffusion equation is modified to include this constraint matrix. For example, the graph Laplacian L can be replaced by a constrained Laplacian L'=TL (where . represents element-by-element multiplication).

[0217] The constraint matrix is ​​like a guiding template, which forces the diffusion of rule concentration to follow a logically reasonable path, effectively preventing the generation of meaningless rule combinations, ensuring the professionalism and usability of the evolution results, and fundamentally accelerating the convergence process.

[0218] In another embodiment of the present application, in another preferred embodiment of the present invention, step S3 for generating a user subjective preference modulator can be implemented by using a model that can distinguish between the user's long-term and short-term preferences.

[0219] Specifically, the process of generating a user subjective preference modulator includes: using a hierarchical Bayesian model to model historical user behavior data to infer the long-term stable preferences that represent the user's inherent habits, and the short-term dynamic preferences that represent the user's recent intentions; applying the attention mechanism to generate a dynamic weight for fusing long-term and short-term preferences based on the current dynamic user request data or real-time context; performing a weighted combination of the long-term stable preference and the short-term dynamic preference based on the dynamic weight to obtain the final user preference distribution, and constructing a user subjective preference modulator based on the final distribution.

[0220] In this preferred embodiment, the system no longer uses a single DBN model, but adopts a more sophisticated hierarchical model and introduces an attention mechanism to more accurately characterize complex user preferences.

[0221] The hierarchical Bayesian model is a probabilistic model that incorporates multiple levels of latent variables, enabling deeper, structured modeling of problems. The attention mechanism is a neural network technology that mimics human cognitive attention, dynamically calculating the degree of attention (weight) given to different information sources based on the current input.

[0222] Hierarchical Preference Modeling:

[0223] Upper layer: Long-term stable preference modeling: The system uses the user's entire historical behavior data to train a Gaussian mixture model (GMM) or a simple Bayesian network. This model aims to extract relatively fixed, unchanging personality traits or habitual preferences from a user's long-term, extensive behavior. For example, if a user's parking choices over the past two years were 90% in Zone B, the model will output a long-term stable preference vector with a very high confidence level for Zone B.

[0224] Lower layer: Short-term dynamic preference modeling: The system uses only the user's recent historical behavior data (for example, the past month or the last five times) to train a time series model (such as an RNN or a simple DBN). This model aims to capture the user's temporary, near-term intentions, which may not align with long-term habits. For example, if a user has chosen a charging station parking space for the last three times (and never in the past), the model will output a short-term dynamic preference vector that places a high weight on charging needs.

[0225] Dynamic fusion based on the attention mechanism: After obtaining the long-term stable preference vector and the short-term dynamic preference vector mentioned above, the system does not simply add them together. Instead, it uses the current user request (for example, the user app interface stays on the page looking for charging stations) as the query (Query), and the long-term and short-term preferences as the key (Key) and value (Value), which are input into the attention module. This module calculates the similarity between the query and different preferences and outputs a dynamic attention weight α(t). For example, if the current request is to charge, α(t) will assign a very high weight to the short-term preference (because recent behavior indicates charging intention).

[0226] The final user preference distribution is calculated using the weighted summation formula: _Preference =(1-α(t))Long _term_Preference +α(t)Short _term_Preference Calculated.

[0227] The Final _Preference (probability distribution or vector), the final user subjective preference modulator is generated through the exponential response function and diagonal matrix construction method.

[0228] This embodiment improves the intelligence level of user preference modeling. It solves the contradiction that a single model cannot take into account the user's long-term habits and short-term intentions. Through hierarchical modeling and dynamic weighting of the attention mechanism, the system can understand the user's real needs in a specific situation like a real human assistant (for example, I know you usually like to park in Zone B, but you have been looking for charging piles recently and now you have turned on the charging pile filter, so I guess what you need most now is a charging pile parking space), thereby achieving truly deep personalization and context-aware recommendations.

[0229] In summary, in this application, by mapping the confidence of the rules to the concentration of chemical substances on the knowledge graph and applying the reaction-diffusion equation based on the graph Laplace operator for evolution, an intelligent mechanism that can spontaneously emerge the optimal rules is realized. When the user's actual selection behavior is input into the system as a catalyst or inhibitor, the high-concentration area representing the effective logic will spontaneously propagate, enhance and form a stable pattern on the graph, while the invalid logic will decay. It solves the technical difficulties of the traditional expert system rule solidification. In the parking scenario, it enables the recommendation system to autonomously learn and adapt to the dynamic changes of the environment. For example, when an exit is temporarily closed due to construction, the system can automatically reduce the recommendation weight of the parking space on the path to the exit through feedback from user behavior without any manual intervention, thereby ensuring the long-term effectiveness and environmental adaptability of the recommendation logic, and its recommendation accuracy can still be maintained at a high level after environmental changes.

[0230] By constructing a hierarchical Bayesian model, the present invention successfully decouples the user's historical behavior data into two levels: long-term stable preferences and short-term dynamic preferences. It also innovatively introduces an attention mechanism to dynamically calculate the fusion weight based on the user's current real-time request or context. This enables the system to achieve a deep and accurate understanding of the user's intentions. In the parking scenario, this means that the system not only knows that a user always likes to park in Zone B (long-term preference), but also understands that he is looking for a charging station this week because he opened a new electric car (short-term preference). When the user opens the App, the attention mechanism can determine that charging is the most important intention at the moment, and thus prioritize the recommendation of charging parking spaces. This solves the pain point that traditional user portraits cannot distinguish between long-term and short-term intentions, and that recommendation results are often inappropriate. By providing personalized services that truly understand customers, user satisfaction and trust in the system are improved.

[0231] By constructing a system-wide total potential energy function E(S) that incorporates the individual potential energy of all users and a penalty term for path intersection between each user, and transforming the problem of finding the optimal allocation solution into the mathematical problem of minimizing E(S), collaborative guidance is achieved at a macro level. Instead of planning the optimal path for each user in isolation, a globally optimal allocation and routing solution with the lowest total cost is designed for the entire user group. In parking scenarios, this negative entropy guidance mechanism can proactively guide users entering the parking lot simultaneously to spaces that are spatially staggered and do not intersect with each other, even if these spaces are not necessarily the closest to each user. During peak hours in the parking lot, the number of internal road intersections and conflict points caused by searching for a car can be reduced by over 45%, effectively avoiding group conflicts, significantly reducing congestion and waiting times, and improving the overall vehicle turnover efficiency of the parking lot.

[0232] By functionally linking the key parameters of the repulsive potential field (the influence amplitude and radius) to the real-time calculated traffic density ρ, an adaptive social safety distance mechanism is introduced for the multi-user system, enabling the system to implement different guidance strategies under different levels of congestion. When the parking lot is empty and the traffic density ρ is low, the repulsive potential field has a weak effect, allowing users to be guided to relatively concentrated, high-quality areas. However, as the parking lot approaches saturation and the traffic density ρ increases, the influence amplitude A(ρ) automatically increases and the influence radius σ(ρ) automatically decreases, resulting in a stronger but more concentrated repulsive force around each user. This embodiment effectively addresses the problem of congestion in localized areas. For example, when multiple users simultaneously enter the same parking lane, the dynamically enhanced repulsive force prompts the optimization algorithm to assign them to spaces farther apart within the lane, avoiding the need for slow maneuvering and waiting for one or two adjacent spaces in a confined space, thereby improving the user experience during the final stage of the parking search.

[0233] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A parking space recommendation and reservation method, characterized in that: include: Process the acquired multi-source heterogeneous data to generate parking lot static attribute data, real-time parking space status data and dynamic user request data; Based on the static attribute data of the parking lot and the real-time parking space status data, fuzzy reasoning is used to construct an objective basic potential field. This includes: for each unoccupied parking space, its multi-dimensional attributes are extracted from the static attribute data and the real-time parking space status data to form a multi-dimensional numerical vector of the parking space; the current optimal fuzzy rule set is loaded from the rule gene library, and the basic potential energy scalar of each parking space is calculated through fuzzification, rule activation, reasoning, aggregation and defuzzification. The collection of all these scalars constitutes the objective basic potential field; Based on dynamic user request data and historical behavior data, Bayesian reasoning is used to generate user subjective preference modulators; The total potential energy function of the system is constructed by integrating the objective basic potential field, the user's subjective preference modulator, and the dynamic repulsive potential field calculated based on the status of all active users; Perform global optimization based on the system's total potential energy function to find the minimum value of the system's total potential energy function, find the optimal parking space allocation plan with the lowest total energy, convert it into specific navigation path instructions for each user, and generate a collaborative parking guidance strategy; Among them, when in use, the rule set based on fuzzy reasoning is adaptively evolved, specifically: Obtain the total potential energy index of the system output by the previous round of global optimization process, which describes the overall performance of the system; Calculate the performance deviation that represents the difference between the system prediction and the user's actual choice; When the system's total potential energy index continues to deteriorate or the accumulation of performance deviation exceeds the preset threshold, the topology of the rule set evolves to generate a new generation of optimal fuzzy rule sets; Among them, the total potential energy function of the system includes: The basic allocation potential energy term, whose value is the sum of the global collaborative dynamic potential energy of all assigned users, is used to measure the total individual satisfaction with the allocation scheme. The calculation process of the global collaborative dynamic potential energy includes: for each active user, the user subjective preference modulator generated by the user acts on the objective basic potential field, and through matrix transformation, an individual subjective potential field is generated for the user; the dynamic repulsion potential fields of all users are linearly superimposed to obtain a multi-agent repulsion potential field; and the individual subjective potential field and the multi-agent repulsion potential field are synthesized point by point to obtain the global collaborative dynamic potential energy. The path intersection penalty potential energy term, whose value is the sum of the path intersection penalties between all users' planned paths, is used to quantify the potential traffic congestion risk caused by the allocation plan; Weight coefficient, used to weight the basic allocation potential energy term and the path crossing penalty potential energy term; The process of generating a user subjective preference modulator includes: using Bayesian inference to obtain a user preference probability distribution; converting the user preference probability distribution into a dimensional scaling factor vector through an exponential response function; arranging the components of the dimensional scaling factor vector on the diagonal of the matrix to construct an affine transformation matrix as the user subjective preference modulator; Among them, the dynamic repulsive potential field includes: The influence amplitude of the repulsive potential field A(ρ) is given by the linear growth formula A(ρ)=A base +k A ρ is determined; where ρ is the real-time traffic density, A base is the basic repulsion amplitude, k A is the amplitude growth coefficient; The influence radius of the repulsive potential field σ(ρ) is determined by the inverse proportional attenuation formula σ(ρ)=σ base / (1+k σ ρ) is determined; where σ base is the basic repulsion radius, k σ is the radius attenuation coefficient; The real-time traffic density ρ is the ratio of the number of active users looking for cars in the parking lot to the total number of currently available parking spaces.

2. The method according to claim 1, characterized in that The topological evolution of the rule set adopts a reaction-diffusion-based self-organization process, including: Initialize the chemical concentration field representing the importance of each node and connection on the preset fully connected knowledge graph; Based on the user's actual selection data, the potential rule paths on the knowledge graph are verified, and the weighted summation of multiple evolution results with different diffusion coefficients is performed to update the chemical concentration field; the concentration of the confirmed path is enhanced, and the concentration of the falsified path is suppressed; The updated chemical concentration field is evolved and calculated, and the evolution process is guided by a topological constraint matrix constructed based on the logical relationship between rules. After the concentration field is continuously iterated and a stable first concentration region emerges, the subgraph corresponding to the first concentration region is extracted and decoded to generate a new generation of optimal fuzzy rule sets.

3. The method according to claim 1, characterized in that The topological evolution of the rule set is implemented based on a genetic algorithm, including: Encode multiple sets of fuzzy rules into rule individuals with a tree structure to form an initial population; Evaluate the fitness score of each regular individual in the population based on the system's total potential energy index; Based on the fitness score, genetic operators including selection, crossover and mutation are performed on the regular individuals to reproduce a new generation of regular individuals; The optimal individual is determined from the new generation of rule individuals and decoded to generate the new generation of optimal fuzzy rule sets.

4. The method according to claim 1, wherein The calculation process of the path crossing penalty potential energy term includes: Each guided path is expressed as a sequence of discretized grid cells arranged in time sequence; For any two paths, detect and record spatiotemporal conflict events that occupy the same grid cell at the same time step; The sum of the spatiotemporal conflict events occurring between all path pairs is calculated to obtain the value of the path crossing penalty potential energy term.

5. The method according to claim 1, wherein The process of generating a user's subjective preference modulator includes: Based on the user identity in the dynamic user request data, the user's decision features are extracted from the historical behavior database and quantified into the user's historical behavior feature vector; The user's historical behavior feature vector is used as evidence and input into a pre-built dynamic Bayesian network model used to describe the temporal dependency relationship between user behavior and implicit preferences. Probabilistic reasoning is performed to solve the user preference probability distribution that represents the possibility of user implicit preferences, and based on this, a user subjective preference modulator is constructed.

6. The method according to claim 1, characterized in that The process of generating a user's subjective preference modulator includes: A hierarchical Bayesian model is used to model historical user behavior data to infer long-term stable preferences that represent users' inherent habits, and short-term dynamic preferences that represent users' recent intentions. Generating a dynamic weight for integrating the long-term and short-term preferences according to the current dynamic user request data or real-time context; Based on the dynamic weights, the long-term stable preference and the short-term dynamic preference are weightedly combined to obtain the final user preference distribution, and the user subjective preference modulator is constructed accordingly.

Images