Parking lot pricing prediction system and method based on artificial intelligence and big data

Abstract

The invention relates to a parking lot pricing prediction system and method based on artificial intelligence and big data, and belongs to the technical field of big data analysis. The system comprises a three-layer state sensing and coding module, wherein a graph loop fusion network, an attention scoring model and a user clustering model are arranged in the three-layer state sensing and coding module; the graph loop fusion network is used for encoding the network game layer data into graph embedding vectors; the attention scoring model is used for carrying out dynamic weight distribution and fusion according to the importance of the influence of the current moment on the parking demand; the user clustering model is used for clustering individual heterogeneous layer data into a plurality of typical user groups, and extracting group features to obtain user heterogeneous state vectors. According to the method, a'three-layer perception-collaborative game 'framework is constructed, a regional Pareto improved pricing strategy is generated by utilizing a multi-agent dynamic game, meanwhile, the method has real-time response and decision interpretation capabilities, and social benefits and collaborative commercial and public benefits are embedded in a reward function of the method.

Description

Parking Pricing Prediction System and Method Based on Artificial Intelligence and Big Data Technical Field

[0001] This invention belongs to the field of big data analysis technology, specifically relating to a parking lot pricing prediction system and method based on artificial intelligence and big data. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of motor vehicle ownership, reasonable parking lot pricing can not only regulate supply and demand and improve parking lot operating revenue, but also effectively guide traffic flow and alleviate regional congestion. Traditional parking lot pricing often relies on experience-based judgment, static cost-plus pricing, or simple tiered pricing based on real-time occupancy rates, lacking the ability to systematically perceive and dynamically respond to complex external environments (such as traffic conditions, competitive landscape, and user preferences).

[0003] With the development of big data and artificial intelligence technologies, some intelligent pricing methods based on predictive models have emerged. However, these methods typically treat each parking lot as an independent entity, ignoring the competition and substitution relationships (i.e., network game effects) among multiple parking lots within a region. When one parking lot adjusts its price, its demand changes affect other parking lots, creating complex dynamic interactions. Simple demand forecasting cannot characterize this game equilibrium, potentially leading to strategy failure. Furthermore, existing methods often lack interpretability, making it difficult for managers to understand the logic behind pricing recommendations, thus reducing trust in decision-making. Simultaneously, most solutions focus on maximizing the operator's economic benefits, failing to internalize the social costs (such as congestion) of parking behavior on the surrounding road network. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a parking lot pricing prediction system and method based on artificial intelligence and big data. By constructing a "three-layer perception-cooperative game" framework, it utilizes multi-agent dynamic game to generate regional Pareto improved pricing strategies, while possessing real-time response and decision interpretation capabilities. Its reward function embeds social benefits, synergizing commercial and public interests.

[0005] The objective of this invention can be achieved through the following technical solution: The first aspect of this disclosure provides a parking lot pricing prediction system based on artificial intelligence and big data, including a multi-source data acquisition and fusion module, a three-layer state perception and encoding module, a collaborative pricing strategy generator, an interpretive decision support interface, and a strategy execution and feedback closed-loop module; the three-layer state perception and encoding module... The system incorporates a graph cyclic fusion network, an attention scoring model, and a user clustering model to fuse multi-source data and obtain the fusion state. The graph cyclic fusion network models the spatiotemporal dependencies and competitive relationships of the parking network, encoding the network game layer data into graph embedding vectors. The attention scoring model evaluates the importance of urban dynamic layer data to parking demand at the current moment and dynamically allocates weights for fusion, outputting an enhanced context vector. The user clustering model clusters individual heterogeneous layer data into several typical user groups, extracting group features to obtain user heterogeneous state vectors. The collaborative pricing strategy generator incorporates a partially observable stochastic game model based on a multi-agent deep deterministic policy gradient framework. This partially observable stochastic game model takes the fusion state as input, treats each parking lot as an agent, and outputs the collaborative pricing strategy for each parking lot in the network through a centralized training and decentralized execution mechanism.

[0006] As a preferred technical solution of the present invention, the multi-source data acquisition and fusion module is used to acquire and structure heterogeneous data in real time from the Internet and IoT platforms, wherein the heterogeneous data includes urban dynamic layer data, network game layer data and individual heterogeneous layer data.

[0007] As a preferred technical solution of the present invention, the urban dynamic layer data includes real-time traffic flow, road congestion index, weather conditions, schedules of major public events, and weekday / holiday markers for the target area and its surroundings; the network game layer data includes historical and real-time prices, parking space occupancy rates, total number of parking spaces, and geographical locations of each parking lot within the area; and the individual heterogeneous layer data includes user parking selection preferences, willingness to pay, and differences in sensitivity to price / distance / waiting time, all collected anonymously through mobile applications.

[0008] As a preferred technical solution of the present invention, the method of encoding network game layer data into graph embedding vectors includes the following steps: Constructing a dynamic attribute graph: abstracting each parking lot in the area as a graph node, assigning static and dynamic attributes to each node, defining the connection edges between nodes by calculating the actual road distance between parking lots and the similarity of historical price behavior, constructing a composite competitive relationship graph, and quantifying the mutual influence potential between nodes; Spatial aggregation encoding: using a graph attention network to propagate and aggregate information in the constructed competitive relationship graph, each parking lot node collects information based on directly connected competitor nodes; through a learnable attention mechanism, dynamically calculating and allocating attention weights for different neighbor information; after multiple rounds of aggregation, each node generates a spatial feature vector that integrates its own state and competitive environment context information; Temporal fusion encoding: inputting the spatial feature vector along the time axis into a gated loop unit, adaptively controlling the retention and forgetting of historical information through update gate and reset gate mechanisms, capturing evolution patterns, and outputting a spatiotemporal fusion feature vector for each parking lot node that integrates historical patterns and current state; Outputting graph embedding vectors: outputting the spatiotemporal fusion feature vectors of all nodes as graph embedding vectors.

[0009] As a preferred technical solution of the present invention, the attention scoring model outputs an enhanced context vector, comprising the following steps: Feature input and standardization: The urban dynamic layer data is standardized according to feature type, and all features are mapped to similar numerical ranges to generate a standardized dynamic feature matrix; Feature mapping and interactive computation: The model uses the parameter matrix obtained through training with historical data to map the input features into three sets of vectors, including query vector, key vector, and value vector; The interactive attention score matrix is ​​obtained by calculating the scaled dot product of the query vector and all key vectors; Weight normalization and dynamic allocation: The interactive attention score matrix is ​​normalized by applying the Softmax function to each row, and the initial attention score is converted into a probability distribution to obtain a probabilistic dynamic weight vector; Weighted fusion and feature enhancement: The value vectors of the urban dynamic layer data features are weighted and summed using the probabilistic dynamic weight vector, thereby selectively focusing and filtering the original urban dynamic feature information to generate an enhanced context vector.

[0010] As a preferred technical solution of the present invention, the user clustering model extracts group features, including the following steps: Data preparation and feature engineering: cleaning, denoising, and consistency checks are performed on individual heterogeneous layer data, user behavior is converted into numerical feature vectors, and then user feature matrix is ​​obtained through feature standardization; Clustering algorithm training and group division: an unsupervised clustering algorithm is used to analyze the user feature matrix, the model is trained by optimizing the objective function, and a group label is assigned to each user sample; Group feature profile extraction: based on the group label, the features of all users belonging to the same group are statistically aggregated and analyzed, and the central feature vector and the dispersion of features of each group are calculated; at the same time, the groups are semantically interpreted in combination with business logic, and finally a typical user group profile library is output, wherein each profile includes the group size proportion and the group central feature vector; Real-time data encoding and output: using the trained clustering model, the group affiliation of the feature vector of new users is predicted, and the estimated proportion distribution of each typical group among potential users who visit the target area or are affected by the strategy is statistically analyzed in real time during the current decision period, and the user heterogeneous state vector is output, wherein the user heterogeneous state vector is usually composed of the weighted sum or concatenation of the estimated proportion of each group and the central feature vector.

[0011] As a preferred embodiment of the present invention, the collaborative pricing strategy of each parking lot in the partially observable stochastic game model output network includes the following steps: fusion state reception and agent state distribution: receiving the fusion state and distributing it to N agents representing N parking lots in the region in the partially observable stochastic game model, each agent... Obtain the complete fusion state; where the fusion state is composed of a graph embedding vector, an augmentation context vector, and a user heterogeneous state vector; Distributed price decision based on a policy network: each agent... Dedicated policy network The system independently processes the fusion state. First, it extracts features relevant to itself from the graph embedding vector, augmented context vector, and user heterogeneous state vector through a fully connected layer, and then performs feature fusion to output a deterministic optimal price action value. The action values ​​output by each of the N agents are combined to obtain the optimal price action set. For policy posterior analysis and price range generation, a posterior analysis is performed on each optimal price action value. A lightweight demand-price response simulator is used to conduct small-scale price perturbation tests around the optimal price action value to predict the changes in occupancy rate and revenue. Simultaneously, a recommended price range is determined based on preset business rules. Finally, the recommended price range and the optimal recommended price are output, where the optimal recommended price is the optimal price action value.

[0012] As a preferred technical solution of the present invention, the explanatory decision support interface is used to provide parking lot planners or administrators with a visual interactive interface, displaying content including pricing basis, predicted effects and alternative plans.

[0013] As a preferred technical solution of the present invention, the strategy execution and feedback closed-loop module continuously collects actual data after the pricing strategy is adopted and applied to a real scenario, and uses the actual data as new samples to flow back to the model training pool, driving the model to perform online fine-tuning and continuous learning, thereby realizing the system's self-evolution.

[0014] The second aspect of this disclosure provides a parking pricing prediction method based on artificial intelligence and big data, applied to a parking pricing prediction system based on artificial intelligence and big data as described above, comprising the following steps: S1, acquiring and preprocessing urban dynamic layer, network game layer, and individual heterogeneous layer data in real time through a multi-source data acquisition and fusion module; S2, extracting parking network structure features using GRFN through a three-layer state perception and encoding module, dynamically weighting macro-dynamic information using an attention mechanism, refining user group features using a clustering model, and finally fusing to generate a fusion state at time t. S3, Collaborative Pricing Strategy Generator receives The system uses the policy networks of each agent in the trained POSG model to calculate the optimal recommended price for each parking lot in the area in parallel and map it to an executable price range. S4: The interpretive decision support interface receives the output and interpretive data from the collaborative pricing strategy generator, generates a visual report, and assists the administrator in understanding, evaluating, and selecting the final pricing scheme. S5: The strategy execution and feedback closed-loop module monitors the actual feedback data after the scheme is executed, compares it with the predicted value, calculates the prediction deviation, and uses reinforcement learning algorithms to update the model parameters to achieve closed-loop optimization.

[0015] The beneficial effects of this invention are as follows: Compared with the prior art, this invention has the following significant advantages: High theoretical originality: For the first time, the parking pricing problem is systematically modeled as a "dynamic network game", and a complete "three-layer perception (city, network, individual) - collaborative game" mathematical framework and modeling method are proposed, which breaks through the limitations of traditional prediction models.

[0016] The strategy has strong global optimization capabilities: the generated pricing strategy is not a local optimal solution for a single parking lot, but rather seeks an approximate Pareto optimal solution for the regional parking network under multiple objectives such as revenue, utilization efficiency, and traffic management through multi-agent collaborative game, thereby improving the overall allocation efficiency of urban transportation resources.

[0017] The system is highly adaptable: it can keenly perceive urban macro-dynamics (such as traffic and events) and respond quickly through pricing levers, effectively smoothing demand shocks caused by emergencies and enhancing system robustness.

[0018] Highly practical and interpretable: It provides intuitive decision support explanations, transforming the decision-making logic of the "black box" model into business language that administrators can understand, greatly enhancing the credibility and acceptability of artificial intelligence systems in key business decisions.

[0019] Social benefit orientation: By embedding social benefit indicators (such as congestion relief) into the model's reward function, commercial pricing behavior is guided to spontaneously align with the overall public interest of the smart city, thus achieving a unity of commercial and social value. Attached Figure Description

[0020] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0021] Figure 1 is a schematic diagram of the structure of a parking lot pricing prediction system based on artificial intelligence and big data provided in an embodiment of the present invention; Figure 2 is a schematic diagram of the steps of the collaborative pricing strategy of each parking lot in the output network of a partially observable random game model provided in an embodiment of the present invention; Figure 3 is a schematic diagram of the steps of a parking lot pricing prediction method based on artificial intelligence and big data provided in an embodiment of the present invention. Detailed Implementation

[0022] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0023] Example 1 provides a parking lot pricing prediction system based on artificial intelligence and big data, as shown in Figure 1. It includes a multi-source data acquisition and fusion module, a three-layer state perception and encoding module, a collaborative pricing strategy generator, an interpretive decision support interface, and a strategy execution and feedback closed-loop module.

[0024] The multi-source data acquisition and fusion module is used to acquire and structure heterogeneous data in real time from the Internet and IoT platforms. The heterogeneous data includes urban dynamic layer data, network game layer data, and individual heterogeneous layer data.

[0025] Specifically, the urban dynamic layer data includes real-time traffic flow, road congestion index, weather conditions, schedules of major public events, and weekday / holiday markers for the target area and its surroundings. The network game layer data includes historical and real-time prices, parking space occupancy rates, total number of parking spaces, and geographical locations (network topology) for each parking lot within the area. The individual heterogeneous layer data includes user parking selection preferences, willingness to pay, and differences in sensitivity to price / distance / waiting time, collected anonymously through mobile applications.

[0026] It should be noted that the collection and application of individual heterogeneous layer data comply with user privacy regulations.

[0027] The three-layer state perception and encoding module incorporates a Graph Recurrent Fusion Network (GRFN), an attention scoring model, and a user clustering model to fuse multi-source data and obtain the fusion state. .

[0028] The graph cyclic fusion network is used to model the spatiotemporal dependencies and competitive relationships of the parking lot network, encoding the network game layer data into graph embedding vectors. The attention scoring model is used to evaluate the importance of the urban dynamic layer data to parking demand at the current moment and dynamically allocate weights for fusion, outputting an enhanced context vector. The user clustering model is used to cluster individual heterogeneous layer data into several typical user groups, extracting group characteristics (such as "price-sensitive" and "convenience-first") to obtain heterogeneous user state vectors.

[0029] The process of encoding network game layer data into graph embedding vectors includes the following steps: Constructing a dynamic attribute graph: Abstracting each parking lot in the area into a graph node, assigning static attributes (such as parking space capacity) and dynamic attributes (such as real-time price and number of empty parking spaces) to each node, defining the connection edges between nodes by calculating the actual road distance between parking lots and the similarity of historical price behavior, constructing a composite competitive relationship graph, and quantifying the potential for mutual influence between them.

[0030] Spatial aggregation encoding: Graph Attention Network (GAT) is used to propagate and aggregate information in the constructed competitive relationship graph. Each parking lot node collects information (such as price and occupancy status) from directly connected competitor nodes. Through a learnable attention mechanism, attention weights for different neighbor information are dynamically calculated and assigned (the weight depends on the relative importance of the neighbor node's state). After multiple rounds of aggregation, each node generates a spatial feature vector that integrates its own state and competitive environment context information.

[0031] Temporal fusion encoding: Spatial feature vectors are input into a gated recurrent unit (GRU) along the time axis. The GRU adaptively controls the retention and forgetting of historical information through the gate update and gate reset mechanisms, captures evolution patterns, and outputs a spatiotemporal fusion feature vector for each parking lot node that integrates historical patterns and the current state.

[0032] Output graph embedding vectors: The spatiotemporal fusion feature vectors of all nodes are output as graph embedding vectors. The final output of the graph cyclic fusion network is a set of fixed-length graph embedding vectors corresponding to all parking lot nodes. This set is a complete encoding of the network game layer data and is directly used as the input state for the downstream collaborative pricing strategy model.

[0033] The attention scoring model outputs an enhanced context vector, which includes the following steps: Feature input and standardization: The urban dynamic layer data is standardized according to feature type (such as time series features, event features, traffic features and environmental features), and all features are mapped to similar numerical ranges to generate a standardized dynamic feature matrix.

[0034] Feature mapping and interactive computation: The model uses the parameter matrix (query weight matrix, key weight matrix and value weight matrix) obtained by training with historical data to map the input features into three sets of vectors, including query vector, key vector and value vector.

[0035] Understandably, the query vector encodes the abstract question, "At the current moment, what feature information does the system need to understand parking demand?" The key vector is the "identifier" or "index label" for each feature, used to respond to the query. It represents "what information I (this feature) can provide about the impact of demand." The value vector carries the original, detailed information of the feature to be aggregated.

[0036] The interaction attention score matrix is ​​obtained by calculating the scaled dot product of the query vector and all key vectors. Understandably, this score reflects the relevance of each feature to the current demand scenario. For example, in a rainstorm scenario, the interaction calculation of the "rainfall intensity" feature with the query vector will produce a high initial score.

[0037] Weight normalization and dynamic allocation: The interaction attention score matrix is ​​normalized by applying the Softmax function to each row (usually the feature dimension), which converts the initial attention scores into a probability distribution and obtains a probabilistic dynamic weight vector.

[0038] Weighted fusion and feature enhancement: The value vectors of urban dynamic layer data features are weighted and summed using probabilistic dynamic weight vectors, thereby selectively focusing and filtering the original urban dynamic feature information and generating an enhanced context vector.

[0039] The user clustering model extracts group features, including the following steps: data preparation and feature engineering: cleaning, denoising and consistency checks are performed on individual heterogeneous layer data, user behavior is converted into numerical feature vectors, and then user feature matrix is ​​obtained through feature standardization.

[0040] Clustering Algorithm Training and Group Partitioning: Unsupervised clustering algorithms (such as K-means++, Gaussian Mixture Model, or spectral clustering) are used to analyze the user feature matrix. The model is trained by optimizing the objective function (such as minimizing intra-class distance and maximizing inter-class distance), and a group label is assigned to each user sample.

[0041] Group Feature Profiling: Based on group tags, the features of all users belonging to the same group are statistically aggregated and analyzed to calculate the central feature vector (usually the mean or median of features within the group) and the dispersion of features (such as variance) for each group. Simultaneously, semantic interpretation of groups is performed in conjunction with business logic, such as identifying "weekday commuting groups with rigid demand," "weekend leisure groups with high price tolerance," and "price-sensitive random visitor groups." Finally, a typical user group profile library is output, where each profile includes the group size proportion (the proportion of users in this group to the total sample) and the group central feature vector (representing the typical behavioral patterns of this group).

[0042] Real-time data encoding and output: Using a trained clustering model, the feature vectors of new users are used to predict their group affiliation. At the same time, the estimated proportion distribution of each typical group among potential users who visit the target area or are affected by the strategy is statistically analyzed in real time during the current decision-making period. The output is structured data describing the user composition in the current context—the user heterogeneous state vector. The user heterogeneous state vector is usually composed of the weighted sum or concatenation of the estimated proportion of each group and the central feature vector.

[0043] It should be noted that the three-layer state perception and encoding module ultimately fuses the outputs of the graph cyclic fusion network, the attention scoring model, and the user clustering model into a unified, high-dimensional global fusion state representation. It comprehensively depicts the current and near-future state of the regional parking ecosystem.

[0044] The collaborative pricing strategy generator incorporates a partially observable stochastic game (POSG) model based on a multi-agent deep deterministic policy gradient (MADDPG) framework. The partially observable stochastic game model uses a fusion state... Using this as input, each parking lot is treated as an intelligent agent. Through a mechanism of centralized training and decentralized execution, the output is a collaborative pricing strategy for each parking lot in the network. Its reward function design integrates parking lot revenue, overall network occupancy balance, and the negative impact on congestion of surrounding main roads to achieve multi-objective collaborative optimization.

[0045] Specifically, the partially observable stochastic game model outputs the collaborative pricing strategy of each parking lot in the network, as shown in Figure 2, including the following steps: fusion state reception and agent state distribution: receiving the fusion state. Distribute it to N agents representing N parking lots in a partially observable stochastic game model, each agent... Obtain the complete fusion state Among them, the fusion state It is composed of graph embedding vector, augmented context vector and user heterogeneous state vector.

[0046] Distributed price decision-making based on policy networks: each agent Dedicated policy network (Actor network) independently of the fusion state The process involves first extracting features relevant to the user from the graph embedding vector, augmented context vector, and user heterogeneous state vector through a fully connected layer (e.g., focusing on neighboring nodes with direct edges in the graph from the graph embedding vector). After feature fusion, a deterministic optimal price action value is output. The action values ​​output by each of the N agents are combined to obtain the optimal price action set.

[0047] It should be noted that the decision function for each agent... It was learned through a global optimization objective (guided by a central commentator network) during the initial "intensive training" phase. Therefore, its output optimal price action value is not a local optimum, but rather an approximate Nash equilibrium solution that takes into account network game effects. The action value is a continuous scalar representing the agent. (i.e., parking lot) Recommended parking rate per unit time for the next period.

[0048] Posterior analysis and price range generation: Posterior analysis is performed on each optimal price action value using a lightweight demand-price response simulator (this simulator is trained on historical data). Using a set of tentative prices as input, the system predicts occupancy rates. It then conducts small-scale price fluctuation tests around the optimal price action value to predict the changes in occupancy rates and revenue. Simultaneously, based on preset business rules (such as "occupancy rate not lower than X%" or "revenue fluctuation less than Y%), it determines a recommended price range. Finally, it outputs the recommended price range and the optimal recommended price, where the optimal recommended price is the optimal price action value.

[0049] Understandably, firstly Distributed to all agents, this establishes the information foundation for collaborative decision-making. Subsequently, each agent, based on its centrally trained policy network, [makes decisions / adjustments]. Parallel processing is performed to independently output the original optimal price action, achieving decentralized game theory decision-making. Finally, the system generates a final output—a recommended price range and the optimal price—that combines optimality and operational flexibility by conducting simulation-based posterior analysis on these original actions. The entire process realizes an end-to-end transformation from "environmental state perception" to "cooperative game theory decision-making" and then to "executable strategy generation."

[0050] The explanatory decision support interface is used to provide parking lot planners or administrators with a visual interactive interface, displaying content including pricing basis, predicted effects, and contingency plans.

[0051] The pricing rationale is presented using attribution diagrams and other methods to illustrate the key factors influencing the pricing decision (e.g., "Due to an upcoming concert at the stadium 500 meters away, demand is expected to increase by 30%, therefore a 15% price increase is recommended"). The predicted effect simulation displays the expected changes in occupancy rates, revenue, and average vehicle speeds of the surrounding road network for each parking lot after adopting the recommended pricing. The contingency plan provides 2-3 pricing strategies with different focuses (e.g., "revenue priority," "traffic diversion priority," "balanced solution") and their predicted implementation consequences for administrators' decision-making reference.

[0052] The strategy execution and feedback closed-loop module continuously collects actual data (actual occupancy rate, revenue, user selection distribution) after the pricing strategy is adopted and applied to real-world scenarios. This actual data is then fed back into the model training pool as new samples, driving the model to perform online fine-tuning and continuous learning, thereby enabling the system to self-evolve.

[0053] Example 2 This example also provides a parking lot pricing prediction method based on artificial intelligence and big data, as shown in Figure 3, including the following steps: S1, real-time acquisition and preprocessing of urban dynamic layer, network game layer, and individual heterogeneous layer data through a multi-source data acquisition and fusion module; S2, extraction of parking lot network structure features using GRFN through a three-layer state perception and coding module, dynamic weighting of macro-dynamic information using an attention mechanism, extraction of user group features using a clustering model, and finally fusion to generate the fusion state at time t. S3, Collaborative Pricing Strategy Generator receives The system uses the policy networks of each agent in the trained POSG model to calculate the optimal recommended price for each parking lot in the area in parallel and map it to an executable price range. S4: The interpretive decision support interface receives the output and interpretive data from the collaborative pricing strategy generator, generates a visual report, and assists the administrator in understanding, evaluating, and selecting the final pricing scheme. S5: The strategy execution and feedback closed-loop module monitors the actual feedback data after the scheme is executed, compares it with the predicted value, calculates the prediction deviation, and uses reinforcement learning algorithms to update the model parameters to achieve closed-loop optimization.

[0054] This invention is the first to systematically model the parking pricing problem as a "dynamic network game," proposing a complete "three-layer perception (city, network, individual) - collaborative game" mathematical framework and modeling method, breaking through the limitations of traditional prediction models. Secondly, the generated pricing strategy is not a local optimum for a single parking lot, but rather seeks an approximate Pareto optimum for the regional parking network under multiple objectives such as revenue, utilization efficiency, and traffic management through multi-agent collaborative game, improving the overall allocation efficiency of urban transportation resources. Simultaneously, this system can keenly perceive urban macro-dynamics (such as traffic and events) and quickly respond through pricing levers, effectively smoothing demand shocks caused by sudden events and enhancing system robustness. It provides intuitive decision support explanations, transforming the decision logic of the "black box" model into business language understandable to administrators, greatly enhancing the credibility and acceptability of the artificial intelligence system in key business decisions. Finally, by embedding social benefit indicators (such as congestion relief) into the model's reward function, it guides commercial pricing behavior to spontaneously align with the overall public interest of the smart city, achieving a unity of commercial and social value.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A parking lot pricing prediction system based on artificial intelligence and big data, characterized in that: It includes a multi-source data acquisition and fusion module, a three-layer state perception and encoding module, a collaborative pricing strategy generator, an interpretive decision support interface, and a strategy execution and feedback closed-loop module. The three-layer state perception and encoding module incorporates a graph cyclic fusion network, an attention scoring model, and a user clustering model to fuse multi-source data and obtain the fusion state. The graph cyclic fusion network is used to model the spatiotemporal dependencies and competitive relationships of the parking lot network, encoding the network game layer data into graph embedding vectors. The attention scoring model is used to evaluate the importance of urban dynamic layer data to parking demand at the current moment and dynamically allocate weights for fusion, outputting an enhanced context vector; the user clustering model is used to cluster individual heterogeneous layer data into several typical user groups, extract group features to obtain user heterogeneous state vectors; the collaborative pricing strategy generator has a built-in partially observable stochastic game model based on a multi-agent deep deterministic policy gradient framework. The partially observable stochastic game model takes the fusion state as input, treats each parking lot as an agent, and outputs the collaborative pricing strategy of each parking lot in the network through a centralized training and decentralized execution mechanism.

2. The parking lot pricing prediction system based on artificial intelligence and big data according to claim 1, characterized in that: The multi-source data acquisition and fusion module is used to acquire and structure heterogeneous data in real time from the Internet and IoT platforms, including urban dynamic layer data, network game layer data, and individual heterogeneous layer data.

3. The parking lot pricing prediction system based on artificial intelligence and big data according to claim 2, characterized in that: The urban dynamic layer data includes real-time traffic flow, road congestion index, weather conditions, schedules of major public events, and weekday / holiday markers for the target area and its surroundings; the network game layer data includes historical and real-time prices, parking space occupancy rates, total number of parking spaces, and geographical locations of each parking lot within the area; and the individual heterogeneous layer data includes user parking selection preferences, willingness to pay, and differences in sensitivity to price / distance / waiting time, all collected anonymously through mobile applications.

4. The parking lot pricing prediction system based on artificial intelligence and big data according to claim 1, characterized in that: The process of encoding network game layer data into graph embedding vectors includes the following steps: Constructing a dynamic attribute graph: Each parking lot within the region is abstracted as a graph node. Static and dynamic attributes are assigned to each node. Connection edges between nodes are defined by calculating the actual road distance between parking lots and the similarity of historical price behavior, thus constructing a composite competitive relationship graph and quantifying the mutual influence potential between nodes. Spatial aggregation encoding: Information propagation and aggregation are performed on the constructed competitive relationship graph using a graph attention network. Each parking lot node collects information based on its directly connected competitor nodes. Through a learnable attention mechanism, attention weights for different neighbor information are dynamically calculated and allocated. After multiple rounds of aggregation, each node generates a spatial feature vector that integrates its own state and the context of the competitive environment. Temporal fusion encoding: The spatial feature vector is input along the time axis into a gated recurrent unit. Through update and reset gate mechanisms, the retention and forgetting of historical information are adaptively controlled, capturing evolution patterns and outputting a spatiotemporal fusion feature vector for each parking lot node that integrates historical patterns and the current state. Outputting graph embedding vectors: The spatiotemporal fusion feature vectors of all nodes are output as graph embedding vectors.

5. The parking lot pricing prediction system based on artificial intelligence and big data according to claim 1, characterized in that: The attention scoring model outputs an enhanced context vector, comprising the following steps: Feature Input and Standardization: The urban dynamic layer data is standardized according to feature type, mapping all features to similar numerical ranges to generate a standardized dynamic feature matrix; Feature Mapping and Interactive Calculation: The model uses a parameter matrix trained on historical data to map the input features into three sets of vectors, including query vector, key vector, and value vector; the interactive attention score matrix is ​​obtained by calculating the scaled dot product of the query vector and all key vectors; Weight Normalization and Dynamic Allocation: The interactive attention score matrix is ​​normalized row-wise using the Softmax function, converting the initial attention scores into a probability distribution to obtain a probabilistic dynamic weight vector; Weighted Fusion and Feature Enhancement: The probabilistic dynamic weight vector is used to perform a weighted summation of the value vectors of the urban dynamic layer data features, thereby selectively focusing on and filtering the original urban dynamic feature information to generate an enhanced context vector.

6. The parking lot pricing prediction system based on artificial intelligence and big data according to claim 1, characterized in that: The user clustering model extracts group features, including the following steps: Data preparation and feature engineering: Cleaning, denoising, and consistency checks are performed on individual heterogeneous layer data. User behavior is converted into numerical feature vectors, and then the user feature matrix is ​​obtained through feature standardization; Clustering algorithm training and group partitioning: Unsupervised clustering algorithm is used to analyze the user feature matrix. The model is trained by optimizing the objective function, and a group label is assigned to each user sample; Group feature profile extraction: Based on the group label, the features of all users belonging to the same group are statistically aggregated and analyzed. The central feature vector and the dispersion of features for each group are calculated. At the same time, the groups are semantically interpreted in conjunction with business logic. Finally, a typical user group profile library is output, where each profile includes the group size proportion and the group central feature vector; Real-time data encoding and output: Using the trained clustering model, the group affiliation of the feature vectors of new users is predicted. At the same time, the estimated proportion distribution of each typical group among potential users who visit the target area or are affected by the strategy is statistically analyzed in real time during the current decision period. The user heterogeneous state vector is output, where the user heterogeneous state vector is usually composed of the weighted sum or concatenation of the estimated proportion of each group and the central feature vector.

7. The parking lot pricing prediction system based on artificial intelligence and big data according to claim 1, characterized in that: The partially observable stochastic game model outputs the collaborative pricing strategy for each parking lot in the network, including the following steps: Fusion state reception and agent state distribution: Receiving the fusion state and distributing it to N agents representing N parking lots in the region in the partially observable stochastic game model, each agent... Obtain the complete fusion state; where the fusion state is composed of a graph embedding vector, an augmentation context vector, and a user heterogeneous state vector; Distributed price decision based on a policy network: each agent... Dedicated policy network The system independently processes the fusion state. First, it extracts features relevant to itself from the graph embedding vector, augmented context vector, and user heterogeneous state vector through a fully connected layer, and then performs feature fusion to output a deterministic optimal price action value. The action values ​​output by each of the N agents are combined to obtain the optimal price action set. For policy posterior analysis and price range generation, a posterior analysis is performed on each optimal price action value. A lightweight demand-price response simulator is used to conduct small-scale price perturbation tests around the optimal price action value to predict the changes in occupancy rate and revenue. Simultaneously, a recommended price range is determined based on preset business rules. Finally, the recommended price range and the optimal recommended price are output, where the optimal recommended price is the optimal price action value.

8. The parking lot pricing prediction system based on artificial intelligence and big data according to claim 1, characterized in that: The explanatory decision support interface is used to provide parking lot planners or administrators with a visual interactive interface, displaying content including pricing basis, predicted effects, and contingency plans.

9. The parking lot pricing prediction system based on artificial intelligence and big data according to claim 1, characterized in that: The strategy execution and feedback closed-loop module continuously collects actual data after the pricing strategy is adopted and applied to real-world scenarios. This data is then fed back into the model training pool as new samples, driving the model to perform online fine-tuning and continuous learning, thereby enabling the system to self-evolve.

10. A parking lot pricing prediction method based on artificial intelligence and big data, applied to a parking lot pricing prediction system based on artificial intelligence and big data as described in any one of claims 1-9, characterized in that: Includes the following steps: S1. Through the multi-source data acquisition and fusion module, real-time acquisition and preprocessing of urban dynamic layer, network game layer and individual heterogeneous layer data are obtained; S2. Through the three-layer state perception and coding module, the parking lot network structure features are extracted using GRFN, macro dynamic information is dynamically weighted using attention mechanism, user group features are extracted using clustering model, and finally the fusion state at time t is generated. S3. The collaborative pricing strategy generator receives the fusion state and, through the policy networks of each agent in the trained POSG model, calculates the optimal recommended price for each parking lot in the area in parallel and maps it to an executable price range. S4. The interpretive decision support interface receives the output and interpretive data from the collaborative pricing strategy generator, generates a visual report, and assists administrators in understanding, evaluating, and selecting the final pricing scheme. S5. The strategy execution and feedback closed-loop module monitors the actual feedback data after the scheme is executed, compares it with the predicted value, calculates the prediction deviation, and uses reinforcement learning algorithms to update the model parameters to achieve closed-loop optimization.

Images