Intelligent property parking lot intelligent dispatching system and method based on artificial intelligence

Abstract

The application relates to the technical field of intelligent control and information system integration, and particularly discloses a smart property parking lot intelligent scheduling system and method based on artificial intelligence. The application solves the problems of extensive management of traditional parking lots relying on fixed parking space allocation and manual on-site scheduling, structural contradictions of parking guidance, security monitoring and data interaction blockage of a property service platform, and inability of collaborative optimization by means of multi-source heterogeneous data fusion, situation prediction based on a convolutional neural network and a gated recurrent unit, deep reinforcement learning strategy generation and multi-objective evolutionary algorithm collaborative optimization, and improves the parking lot resource optimal allocation capability and real-time response level.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and information system integration technology, and in particular to an intelligent dispatching system and method for smart property parking lots based on artificial intelligence. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of motor vehicle ownership, the contradiction between the supply and demand of parking resources is becoming increasingly prominent. Parking scheduling efficiency in smart property scenarios has become a key indicator for measuring the level of property management. Existing parking management systems mostly rely on fixed parking space allocation and manual on-site scheduling, or only achieve basic information recording functions, lacking the ability to intelligently perceive and predict vehicle entry and exit patterns, real-time parking space status, and dynamic environmental changes. This extensive management model results in low parking space turnover rates and long search times during peak hours, and makes it difficult to cope with sudden traffic surges, creating a structural contradiction of resource idleness and congestion. Furthermore, the traditional system suffers from blocked data interaction between functional modules; parking guidance, security monitoring, and the property service platform operate independently, failing to achieve collaborative optimization of multi-source information and real-time adjustment of dynamic scheduling strategies, thus hindering the improvement of overall parking lot operational efficiency. Therefore, there is an urgent need for an intelligent scheduling solution integrating artificial intelligence technology to overcome the bottlenecks of existing management models in resource optimization and real-time response capabilities. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an intelligent dispatching system and method for smart property parking lots based on artificial intelligence.

[0004] In a first aspect, the present invention provides an intelligent dispatching system for smart property parking lots based on artificial intelligence, the technical solution of which is as follows: The data acquisition module is used to collect multi-source heterogeneous data from the parking lot. The multi-source heterogeneous data includes historical operation data, real-time status data, and external environment data. The historical operation data includes vehicle entry and exit records, parking space occupancy time series, and user behavior logs. The real-time status data includes the occupancy status of each parking space, vehicle queue length at entrances and exits, vehicle type identification results, and in-park video streams. The external environment data includes meteorological information and surrounding traffic conditions. The multimodal data fusion module is used to perform spatiotemporal registration and feature-level fusion on the multi-source heterogeneous data to generate aligned multimodal fused data, and to perform feature enhancement on the aligned multimodal fused data using a feature weighting method based on an attention mechanism to generate a fused feature tensor. The situation prediction module is used to input the fused feature tensor into a pre-trained artificial intelligence prediction model. The artificial intelligence prediction model is a combination model of convolutional neural network and gated recurrent unit, and outputs multi-dimensional prediction data of parking lot operation situation for a future preset period. The multi-dimensional prediction data includes vehicle arrival rate time series, parking space occupancy rate distribution and turnover rate trend. The dynamic scheduling strategy generation module is used to call a deep reinforcement learning decision network to generate an initial dynamic scheduling strategy based on the multidimensional prediction data and the real-time status data. The initial dynamic scheduling strategy includes a dynamic parking space zoning scheme, a multi-path guidance scheme, and an entrance and exit gate timing control scheme. The multi-objective collaborative optimization module is used to access the property service platform to obtain security monitoring data and property service data. Taking the initial dynamic scheduling strategy as input, it constructs a multi-objective optimization model that includes traffic efficiency, energy consumption balance and user satisfaction, and uses a multi-objective evolutionary algorithm to solve for Pareto optimal solution set, and selects optimized scheduling instructions from the Pareto optimal solution set. The instruction execution and feedback module is used to send the optimized scheduling instruction to the parking lot terminal execution device and collect execution feedback data. The execution feedback data is used to update the artificial intelligence prediction model and the deep reinforcement learning decision network.

[0005] Secondly, this invention provides an intelligent scheduling method for smart property parking lots based on artificial intelligence. The technical solution of this method is as follows: Collect multi-source heterogeneous data from the parking lot. The multi-source heterogeneous data includes historical operation data, real-time status data, and external environment data. The historical operation data includes vehicle entry and exit records, parking space occupancy time series, and user behavior logs. The real-time status data includes the occupancy status of each parking space, vehicle queue length at entrances and exits, vehicle type identification results, and in-park video streams. The external environment data includes meteorological information and surrounding traffic conditions. The multi-source heterogeneous data is spatiotemporally registered and feature-level fused to generate aligned multimodal fused data. The aligned multimodal fused data is then enhanced using a feature weighting method based on an attention mechanism to generate a fused feature tensor. The fused feature tensor is input into a pre-trained artificial intelligence prediction model, which is a combination model of convolutional neural network and gated recurrent unit, and outputs multi-dimensional prediction data of parking lot operation status for a future preset time period. The multi-dimensional prediction data includes vehicle arrival rate time series, parking space occupancy rate distribution and turnover rate trend. Based on the multidimensional prediction data and the real-time status data, a deep reinforcement learning decision network is invoked to generate an initial dynamic scheduling strategy. The initial dynamic scheduling strategy includes a dynamic parking space zoning scheme, a multi-path guidance scheme, and an entrance and exit gate timing control scheme. The system accesses the property service platform to obtain security monitoring data and property service data. Using the initial dynamic scheduling strategy as input, it constructs a multi-objective optimization model that includes traffic efficiency, energy consumption balance, and user satisfaction. The system then uses a multi-objective evolutionary algorithm to solve for Pareto optimal solutions and selects optimized scheduling instructions from the Pareto optimal solution set. The optimized scheduling command is sent to the parking lot terminal execution device, and execution feedback data is collected. The execution feedback data is used to update the artificial intelligence prediction model and the deep reinforcement learning decision network.

[0006] The technical solution of this invention solves the problem of extensive management of traditional parking lots that rely on fixed parking space allocation and manual on-site scheduling, as well as the structural contradiction of blocked data interaction and inability to coordinate optimization between parking guidance, security monitoring and property service platforms, through multi-source heterogeneous data fusion, situation prediction based on convolutional neural networks and gated recurrent units, deep reinforcement learning strategy generation and multi-objective evolutionary algorithm collaborative optimization. It achieves the improvement of parking lot resource optimization and real-time response level.

[0007] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0009] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of an embodiment of an intelligent parking lot dispatching system based on artificial intelligence according to the present invention. Figure 2 This is a flowchart illustrating an embodiment of an intelligent parking lot scheduling method based on artificial intelligence according to the present invention. Detailed Implementation

[0010] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0011] Figure 1 This diagram illustrates the structure of an embodiment of an intelligent parking lot dispatching system based on artificial intelligence provided by the present invention. Figure 1 As shown, the AI-based smart property parking intelligent dispatch system includes: The data acquisition module 110 is used to collect multi-source heterogeneous data of the parking lot. The multi-source heterogeneous data includes historical operation data, real-time status data and external environment data. The historical operation data includes vehicle entry and exit records, parking space occupancy time series and user behavior logs. The real-time status data includes the occupancy status of each parking space, the queue length of vehicles at the entrance and exit, vehicle type identification results and in-park video stream. The external environment data includes meteorological information and surrounding traffic conditions.

[0012] The multimodal data fusion module 120 is used to perform spatiotemporal registration and feature-level fusion on the multi-source heterogeneous data to generate aligned multimodal fused data, and to perform feature enhancement on the aligned multimodal fused data using a feature weighting method based on an attention mechanism to generate a fused feature tensor.

[0013] The situation prediction module 130 is used to input the fused feature tensor into a pre-trained artificial intelligence prediction model. The artificial intelligence prediction model is a combination model of convolutional neural network and gated recurrent unit, and outputs multi-dimensional prediction data of parking lot operation situation for a future preset period. The multi-dimensional prediction data includes vehicle arrival rate time series, parking space occupancy rate distribution and turnover rate trend.

[0014] The dynamic scheduling strategy generation module 140 is used to generate an initial dynamic scheduling strategy by calling a deep reinforcement learning decision network based on the multidimensional prediction data and the real-time status data. The initial dynamic scheduling strategy includes a dynamic parking space zoning scheme, a multi-path guidance scheme, and an entrance and exit gate timing control scheme.

[0015] The multi-objective collaborative optimization module 150 is used to access the property service platform to obtain security monitoring data and property service data. Taking the initial dynamic scheduling strategy as input, it constructs a multi-objective optimization model that includes traffic efficiency, energy consumption balance and user satisfaction, and uses a multi-objective evolutionary algorithm to solve for Pareto optimal solution set, and selects optimized scheduling instructions from the Pareto optimal solution set.

[0016] The instruction execution and feedback module 160 is used to send the optimized scheduling instruction to the parking lot terminal execution device and collect execution feedback data. The execution feedback data is used to update the artificial intelligence prediction model and the deep reinforcement learning decision network.

[0017] The technical solution of this embodiment solves the problem of extensive management of traditional parking lots that rely on fixed parking space allocation and manual on-site scheduling, as well as the structural contradiction of blocked data interaction and inability to coordinate optimization between parking guidance, security monitoring and property service platforms, by multi-source heterogeneous data fusion, situation prediction based on convolutional neural networks and gated recurrent units, deep reinforcement learning strategy generation and multi-objective evolutionary algorithm collaborative optimization. It realizes the improvement of parking lot resource optimization and real-time response level.

[0018] In one alternative embodiment, the data acquisition module 110 is specifically used for: The historical operation data is read from the historical database of the parking management system. The historical operation data includes the vehicle entry and exit records, the parking space occupancy time series, and the user behavior log. The occupancy status of each parking space is collected by parking space sensors deployed in each parking space, the vehicle type identification results and the video stream inside the site are collected by cameras deployed at the entrance and exit, and the vehicle queue length at the entrance and exit is collected by vehicle detectors deployed at the entrance and exit gates. The meteorological information is obtained from the meteorological service platform through a network interface, and the surrounding traffic conditions are obtained from the traffic information service platform.

[0019] In the above-mentioned optional methods, parking space sensors, cameras and network interfaces deployed in various parts of the parking lot are used to collect historical operation data, real-time status data and external environment data from historical databases, meteorological service platforms and traffic information service platforms to form a comprehensive dataset covering vehicle entry and exit records, parking space occupancy status, meteorological information and traffic conditions, providing a complete information foundation for subsequent data fusion.

[0020] In one alternative embodiment, the multimodal data fusion module 120 is specifically used for: Extract the timestamps of each data source in the multi-source heterogeneous data, align the historical running data, the real-time status data and the external environment data with different sampling frequencies in the time dimension according to the timestamps, and fill in the missing time points with interpolation methods to generate time-aligned multi-source data; The spatial coordinates and spatial reference system of each sensor in the multi-source heterogeneous data are obtained. The spatial coordinates of each data point in the time-aligned multi-source data are transformed to the preset parking lot global coordinate system through the coordinate transformation matrix, thereby eliminating the spatial position deviation between different sensors and generating spatiotemporally registered multimodal data. The structured data from the spatiotemporally registered multimodal data is input into a fully connected neural network. This structured data includes vehicle entry / exit records, parking space occupancy time series, user behavior logs, parking space occupancy status, entrance / exit vehicle queue length, weather information, and surrounding traffic conditions. The output is a structured feature vector. The image data from the spatiotemporally registered multimodal data is input into a convolutional neural network. This image data includes vehicle type recognition results and the on-site video stream. The output is an image feature map. The text data from the spatiotemporally registered multimodal data is input into a natural language processing model. This text data includes the text description portion of the user behavior logs. The output is a text feature vector. The structured feature vector, the image feature map, and the text feature vector are concatenated along the feature dimension, and a linear transformation is performed on the concatenated features to unify the feature scale, thereby generating aligned multimodal fusion features. The aligned multimodal fusion features are input into an attention mechanism network. The attention mechanism network calculates the weight coefficients of each feature channel through self-attention, and weights the aligned multimodal fusion features according to the weight coefficients. The weighted features are then residually connected with the original features to obtain the enhanced multimodal fusion features. The enhanced multimodal fusion features are reshaped according to preset batch dimension, time dimension and feature channel dimension to generate the fusion feature tensor.

[0021] In the above-mentioned optional methods, time alignment and spatial coordinate transformation are further performed on multi-source heterogeneous data to eliminate spatiotemporal bias. Structured data, image data and text data are processed by fully connected neural networks, convolutional neural networks and natural language processing models respectively. After weighting and enhancement by attention network, a fusion feature tensor is generated to improve the quality and consistency of input data.

[0022] In one alternative embodiment, the situation prediction module 130 is specifically used for: The fused feature tensor is divided into multiple consecutive time slices along the time dimension, with each time slice corresponding to the input data of one time step, denoted as . ,in Indicates the time step index; Each time slice The input is a convolutional neural network, which extracts spatial features through multiple convolutional and pooling layers, and outputs a spatial feature vector at each time step. ; Time slices at each time step and the corresponding spatial feature vector A common input spatially augmented gated recurrent unit network is used, wherein each cell in the spatially augmented gated recurrent unit network calculates the hidden state at the current time step according to the following formula: in, Indicates the first The time slice of each time step Indicates the first The spatial feature vector at each time step, This indicates the hidden state of the previous time step. This indicates the hidden state at the current time step. This indicates an update to the gate output. Indicates resetting the gate output. Indicates the context gate output, This represents a context vector that integrates spatial features and historical information. Indicates the candidate hidden state. The weight matrix is ​​a learnable matrix. This is a learnable bias term.

[0023] Wherein, the context gate The output range is [-1, 1], when When it is close to 1, the context vector Mainly composed of historical information The contribution after linear transformation, when When it approaches -1, Mainly due to current spatial characteristics Contribution, when When approaching 0, for and The mechanism employs a balanced weighting system. This system enables the network to adaptively select a fusion mode that prioritizes spatial features, historical information, or a balance of both, to address the dynamic spatial changes in different areas of the parking lot.

[0024] It should be noted that during the training process, the spatially augmented gated recurrent unit network uses the mean square error between the multidimensional predicted data and the historical actual operating data as the loss function, and updates the learnable weight matrix and bias terms through the time backpropagation algorithm.

[0025] Hidden state of the last time step Three parallel fully connected layers are input respectively. The first fully connected layer outputs a time-series prediction of vehicle arrival rate, which represents the sequence of the number of vehicles arriving at each time point within the future preset time period. The second fully connected layer outputs a parking space occupancy rate prediction, which represents the probability distribution matrix of each parking space being occupied within the future preset time period. The third fully connected layer outputs a turnover rate trend prediction, which represents the curve of change in parking space turnover rate within the future preset time period.

[0026] In the above-mentioned optional approach, an artificial intelligence prediction model that combines feature tensor input with convolutional neural network and gated recurrent unit is further used. Through three parallel fully connected layers, the model outputs the time series of vehicle arrival rate, parking space occupancy rate distribution and turnover rate trend for future time periods, so as to achieve accurate prediction of the parking lot operation status and provide forward-looking data support for the generation of dynamic scheduling strategy.

[0027] In an alternative embodiment, the dynamic scheduling strategy generation module 140 is specifically used for: The multidimensional prediction data and the real-time status data are fused according to a preset feature splicing rule to generate an environmental status vector at the current moment. The multidimensional prediction data includes the vehicle arrival rate time series, the parking space occupancy rate distribution and the turnover rate trend. The real-time status data includes the occupancy status of each parking space, the vehicle queue length at the entrance and exit and the vehicle type identification result. The environment state vector is input into a deep reinforcement learning decision network, which includes an online policy network and a target policy network. The online policy network outputs an initial action probability distribution based on the environment state vector, and the target policy network is used to provide a stable target value during training to update the online policy network. An initial action vector is obtained by sampling the initial action probability distribution output by the online policy network. The initial action vector includes a partition action component, an induced action component, and a temporal action component. The partition action component is input into the first decoding network, which consists of multiple fully connected layers. The partition action component is converted into a dynamic parking space partitioning scheme through nonlinear mapping. The dynamic parking space partitioning scheme includes the boundary coordinates of each partition, the number of open parking spaces in each partition within a preset time period, and the types of vehicles allowed to be parked in each partition. The induced action component is input into the second decoding network. The second decoding network combines the real-time congestion index of each lane extracted from the real-time status data and generates a multi-path guidance scheme through a path search algorithm. The multi-path guidance scheme includes the preferred driving path sequence for entering vehicles, the exit guidance sign for leaving vehicles, and the path switching triggering condition. The timing action components are input into the third decoding network. The third decoding network combines the real-time queue length of each entrance and exit extracted from the real-time status data and generates an entrance and exit gate timing control scheme through a timing planning algorithm. The entrance and exit gate timing control scheme includes the opening time, closing time of each entrance and exit gate in a continuous time window, as well as the linkage and locking relationship between adjacent gates. The dynamic parking space zoning scheme, the multi-path guidance scheme, and the entrance / exit gate timing control scheme are combined to generate the initial dynamic scheduling strategy.

[0028] It should be noted that the path search algorithm is the dynamic A* algorithm. This algorithm uses the real-time congestion index of each lane as the cost function to dynamically update the passage cost of each road segment and generate the top K paths with the minimum cumulative cost from the entrance to each zone as the preferred driving path sequence.

[0029] In the above-mentioned optional methods, based on multidimensional prediction data and real-time status data, a deep reinforcement learning decision network containing an online policy network and a target policy network is invoked to generate initial action vectors. Through three decoding networks, dynamic parking space zoning schemes, multi-path guidance schemes, and entrance / exit gate timing control schemes are output respectively, so as to realize intelligent allocation and optimized regulation of parking resources.

[0030] In an alternative embodiment, the multi-objective collaborative optimization module 150 is specifically used for: Access the property service platform via application programming interface (API) to obtain security monitoring data and property service data from the property service platform. The security monitoring data includes security event records of various areas in the parking lot and abnormal behavior detection results obtained from real-time video analysis. The property service data includes users' historical complaint records, user level information, and users' preference settings for parking services. Taking the initial dynamic scheduling strategy as the individual to be optimized, a multi-objective optimization model is constructed. The multi-objective optimization model includes three objective functions. The first objective function is the traffic efficiency function, which calculates the average vehicle entry and exit time based on the dynamic parking space zoning scheme and multi-path guidance scheme in the initial dynamic scheduling strategy. The second objective function is the energy consumption balance function, which calculates the number of start-stop cycles and the running time of each gate device based on the entrance and exit gate timing control scheme in the initial dynamic scheduling strategy. The third objective function is the user satisfaction function, which calculates the comprehensive user satisfaction score based on the dynamic parking space zoning scheme and multi-path guidance scheme in the initial dynamic scheduling strategy, combined with the user preference settings in the property service data. A multi-objective evolutionary algorithm is used to solve the multi-objective optimization model. The multi-objective evolutionary algorithm generates a Pareto optimal solution set iteratively through population initialization, crossover mutation, and Pareto dominance comparison. Each solution in the Pareto optimal solution set corresponds to a candidate scheduling instruction.

[0031] It should be noted that the multi-objective evolutionary algorithm is a non-dominated sorting genetic algorithm based on reference points. This algorithm guides the population to be evenly distributed towards the Pareto front by generating reference points, and adopts an elite retention strategy during the iteration process to ensure the diversity of the Pareto optimal solution set.

[0032] Based on the security monitoring data and the property service data, the preference score for each candidate scheduling instruction in the Pareto optimal solution set is calculated according to the following formula: in, This represents the candidate scheduling instructions in the Pareto optimal solution set. Indicates candidate scheduling instructions The corresponding number The objective function value, These correspond to the traffic efficiency function, the energy consumption balance function, and the user satisfaction function, respectively. and They represent the current Pareto optimal solution set, i.e., the first... The minimum and maximum values ​​of the objective function. This represents the feature vector of the property service data. This represents the feature vector of the security monitoring data. This indicates the extraction of the first [item] from the property service data. The intensity of preference for each objective This represents the security risk coefficient extracted from the security monitoring data. For the first The preference weight coefficients for each objective. For the first The security modulation coefficient of each target.

[0033] Wherein, the security modulation coefficient The value varies depending on the objective function; for the aforementioned traffic efficiency function... A smaller value is chosen to reduce the impact of safety factors on traffic efficiency; for the user satisfaction function... A larger value is selected to prioritize user security perception when security risks are high. This differentiated modulation mechanism achieves asymmetric influence of security factors on different optimization objectives.

[0034] Select the preference score The largest candidate scheduling instruction is selected as the optimized scheduling instruction.

[0035] In the above-mentioned optional methods, security monitoring data and property service data can be obtained by accessing the property service platform through the application programming interface, a multi-objective optimization model of traffic efficiency, energy consumption balance and user satisfaction can be constructed, and Pareto optimal solution set can be solved by multi-objective evolutionary algorithm and optimized scheduling instructions can be selected to achieve the coordinated balance of multi-dimensional objectives of parking lot operation.

[0036] In an alternative embodiment, the instruction execution and feedback module 160 is specifically used for: The optimized scheduling instructions are parsed into control instructions for each parking lot terminal execution device, which includes a barrier gate controller, a parking space lock controller, a guidance screen controller, and a voice broadcast device. The control commands are sent to the corresponding parking lot terminal execution devices through a preset communication protocol, and the command execution status information returned by the parking lot terminal execution devices is received. The command execution status information includes a command reception confirmation signal, a device action completion signal, and a device fault alarm signal. The execution feedback data during the execution of the optimized scheduling instructions is collected. The execution feedback data includes the actual occupancy status change sequence of each parking space, the actual vehicle queue length change sequence of each entrance and exit, the actual opening and closing time of each barrier gate, the display content switching time of each guidance screen, and the location search time and satisfaction rating reported by users through mobile terminals. The execution feedback data is organized according to a preset data format to generate a feedback sample set. Each sample in the feedback sample set contains an environment state vector when the optimized scheduling instruction is executed, the optimized scheduling instruction, and the effect index after execution. The samples in the feedback sample set are stored in the experience replay buffer in chronological order. When the number of samples in the experience replay buffer reaches a preset threshold, a batch of samples are randomly selected from the experience replay buffer, and the artificial intelligence prediction model is fine-tuned using the gradient descent method to update the network parameters of the artificial intelligence prediction model. The samples in the feedback sample set are stored in the policy experience buffer in the same time order. When the number of samples in the policy experience buffer reaches a preset threshold, a batch of samples are randomly drawn from the policy experience buffer. The policy gradient method is used to update the deep reinforcement learning decision network and adjust the network parameters of the online policy network and the target policy network in the deep reinforcement learning decision network.

[0037] In the above-mentioned optional methods, the optimized scheduling instructions are further parsed into control instructions for the gate controller, parking lock controller, guidance screen controller and voice broadcasting equipment and issued for execution. The execution feedback data is collected and stored in the experience playback buffer and the strategy experience buffer. The sample set is used to update the parameters of the artificial intelligence prediction model and the deep reinforcement learning decision network.

[0038] In an alternative embodiment, the system further includes: The model pre-training module is used to extract sample data from multiple historical time periods from the historical database before the system is deployed. The sample data includes historical fusion feature tensors, historical multidimensional prediction data, and corresponding historical actual operation data. The historical fusion feature tensors are used as input, and the historical multidimensional prediction data are used as supervision signals. The AI ​​prediction model is pre-trained using supervised learning. The network parameters of the AI ​​prediction model are adjusted using the mean squared error loss function. The historical actual operation data are used as environmental interaction samples. The deep reinforcement learning decision network is pre-trained using offline reinforcement learning. The network parameters of the online policy network and the target policy network are initialized by combining behavior cloning and policy constraints.

[0039] It should be noted that the behavior cloning is achieved by minimizing the cross-entropy loss between the action probability distribution output by the online policy network and the historical best action. The policy constraint limits the deviation between the online policy network and the behavior cloning network by introducing a KL divergence penalty term. The total loss function is... ,in, This represents the total loss function value. This represents the value of the behavior cloning loss function, which is the cross-entropy loss function. Indicates the trade-off coefficient. This represents the action probability distribution output by the online policy network. Action probability distribution output by the behavior cloning network KL divergence between them This represents the action probability distribution output by the online policy network. This represents the action probability distribution output by the behavior cloning network, which is a policy network obtained through behavior cloning pre-training.

[0040] In the above-mentioned optional methods, before system deployment, historical fusion feature tensors, historical multidimensional prediction data and historical actual operation data are further extracted from the historical database. The artificial intelligence prediction model is pre-trained using supervised learning and the deep reinforcement learning decision network is pre-trained using offline reinforcement learning to complete the network parameter initialization.

[0041] In an alternative embodiment, the system further includes: The user interaction and personalized strategy generation module receives parking and departure requests from users via a mobile terminal application. It extracts user preference features from the user's historical behavior logs, including preferred parking areas, walking distance tolerance, and willingness to pay. These user preference features are then fused with the current multidimensional prediction data and real-time status data to generate a personalized user state vector. This vector is input into a lightweight decision network, which extracts information from the deep reinforcement learning decision network through knowledge distillation. The lightweight decision network outputs personalized guidance information for the user, including a recommended parking space number, a customized route to the recommended parking space, and estimated walking time. This personalized guidance information is then pushed to the mobile terminal application for display.

[0042] In the above-mentioned optional methods, user parking requests and departure requests are further received through a mobile terminal application. User preference features are extracted based on historical behavior logs and a personalized user state vector is generated. The lightweight decision network obtained through knowledge distillation is then called to output personalized guidance information such as recommended parking space number, customized route, and estimated walking time.

[0043] Figure 2 This diagram illustrates a flowchart of an embodiment of an intelligent parking lot scheduling method based on artificial intelligence provided by the present invention. Figure 2 As shown, it includes the following steps: S1. Collect multi-source heterogeneous data of the parking lot. The multi-source heterogeneous data includes historical operation data, real-time status data and external environment data. The historical operation data includes vehicle entry and exit records, parking space occupancy time series and user behavior logs. The real-time status data includes the occupancy status of each parking space, the queue length of vehicles at the entrance and exit, vehicle type identification results and in-park video stream. The external environment data includes meteorological information and surrounding traffic conditions. S2. Perform spatiotemporal registration and feature-level fusion on the multi-source heterogeneous data to generate aligned multimodal fusion data, and perform feature enhancement on the aligned multimodal fusion data using a feature weighting method based on an attention mechanism to generate a fusion feature tensor. S3. Input the fused feature tensor into a pre-trained artificial intelligence prediction model. The artificial intelligence prediction model is a combination model of convolutional neural network and gated recurrent unit. It outputs multi-dimensional prediction data of parking lot operation status for a future preset time period. The multi-dimensional prediction data includes vehicle arrival rate time series, parking space occupancy rate distribution and turnover rate trend. S4. Based on the multidimensional prediction data and the real-time status data, call the deep reinforcement learning decision network to generate an initial dynamic scheduling strategy. The initial dynamic scheduling strategy includes a dynamic parking space zoning scheme, a multi-path guidance scheme, and an entrance and exit gate timing control scheme. S5. Access the property service platform to obtain security monitoring data and property service data. Using the initial dynamic scheduling strategy as input, construct a multi-objective optimization model that includes traffic efficiency, energy consumption balance and user satisfaction. Use a multi-objective evolutionary algorithm to solve for Pareto optimal solution set. Select optimized scheduling instructions from the Pareto optimal solution set. S6. The optimized scheduling instruction is sent to the parking lot terminal execution device, and execution feedback data is collected. The execution feedback data is used to update the artificial intelligence prediction model and the deep reinforcement learning decision network.

[0044] The technical solution of this embodiment solves the problem of extensive management of traditional parking lots that rely on fixed parking space allocation and manual on-site scheduling, as well as the structural contradiction of blocked data interaction and inability to coordinate optimization between parking guidance, security monitoring and property service platforms, by multi-source heterogeneous data fusion, situation prediction based on convolutional neural networks and gated recurrent units, deep reinforcement learning strategy generation and multi-objective evolutionary algorithm collaborative optimization. It realizes the improvement of parking lot resource optimization and real-time response level.

[0045] Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0046] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0047] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and do not imply a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0048] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An intelligent dispatching system for smart property parking lots based on artificial intelligence, characterized in that, The system includes: The data acquisition module is used to collect multi-source heterogeneous data from the parking lot. The multi-source heterogeneous data includes historical operation data, real-time status data, and external environment data. The historical operation data includes vehicle entry and exit records, parking space occupancy time series, and user behavior logs. The real-time status data includes the occupancy status of each parking space, vehicle queue length at entrances and exits, vehicle type identification results, and in-park video streams. The external environment data includes meteorological information and surrounding traffic conditions. The multimodal data fusion module is used to perform spatiotemporal registration and feature-level fusion on the multi-source heterogeneous data to generate aligned multimodal fused data, and to perform feature enhancement on the aligned multimodal fused data using a feature weighting method based on an attention mechanism to generate a fused feature tensor. The situation prediction module is used to input the fused feature tensor into a pre-trained artificial intelligence prediction model. The artificial intelligence prediction model is a combination model of convolutional neural network and gated recurrent unit, and outputs multi-dimensional prediction data of parking lot operation situation for a future preset period. The multi-dimensional prediction data includes vehicle arrival rate time series, parking space occupancy rate distribution and turnover rate trend. The dynamic scheduling strategy generation module is used to call a deep reinforcement learning decision network to generate an initial dynamic scheduling strategy based on the multidimensional prediction data and the real-time status data. The initial dynamic scheduling strategy includes a dynamic parking space zoning scheme, a multi-path guidance scheme, and an entrance and exit gate timing control scheme. The multi-objective collaborative optimization module is used to access the property service platform to obtain security monitoring data and property service data. Taking the initial dynamic scheduling strategy as input, it constructs a multi-objective optimization model that includes traffic efficiency, energy consumption balance and user satisfaction, and uses a multi-objective evolutionary algorithm to solve for Pareto optimal solution set, and selects optimized scheduling instructions from the Pareto optimal solution set. The instruction execution and feedback module is used to send the optimized scheduling instruction to the parking lot terminal execution device and collect execution feedback data. The execution feedback data is used to update the artificial intelligence prediction model and the deep reinforcement learning decision network.

2. The intelligent dispatching system for smart property parking lots based on artificial intelligence according to claim 1, characterized in that, The data acquisition module is specifically used for: The historical operation data is read from the historical database of the parking management system. The historical operation data includes the vehicle entry and exit records, the parking space occupancy time series, and the user behavior log. The occupancy status of each parking space is collected by parking space sensors deployed in each parking space, the vehicle type identification results and the video stream inside the site are collected by cameras deployed at the entrance and exit, and the vehicle queue length at the entrance and exit is collected by vehicle detectors deployed at the entrance and exit gates. The meteorological information is obtained from the meteorological service platform through a network interface, and the surrounding traffic conditions are obtained from the traffic information service platform.

3. The intelligent dispatching system for smart property parking lots based on artificial intelligence according to claim 1, characterized in that, The multimodal data fusion module is specifically used for: Extract the timestamps of each data source in the multi-source heterogeneous data, align the historical running data, the real-time status data and the external environment data with different sampling frequencies in the time dimension according to the timestamps, and fill in the missing time points with interpolation methods to generate time-aligned multi-source data; The spatial coordinates and spatial reference system of each sensor in the multi-source heterogeneous data are obtained. The spatial coordinates of each data point in the time-aligned multi-source data are transformed to the preset parking lot global coordinate system through the coordinate transformation matrix, thereby eliminating the spatial position deviation between different sensors and generating spatiotemporally registered multimodal data. The structured data from the spatiotemporally registered multimodal data is input into a fully connected neural network. This structured data includes vehicle entry / exit records, parking space occupancy time series, user behavior logs, parking space occupancy status, entrance / exit vehicle queue length, weather information, and surrounding traffic conditions. The output is a structured feature vector. The image data from the spatiotemporally registered multimodal data is input into a convolutional neural network. This image data includes vehicle type recognition results and the on-site video stream. The output is an image feature map. The text data from the spatiotemporally registered multimodal data is input into a natural language processing model. This text data includes the text description portion of the user behavior logs. The output is a text feature vector. The structured feature vector, the image feature map, and the text feature vector are concatenated along the feature dimension, and a linear transformation is performed on the concatenated features to unify the feature scale, thereby generating aligned multimodal fusion features. The aligned multimodal fusion features are input into an attention mechanism network. The attention mechanism network calculates the weight coefficients of each feature channel through self-attention, and weights the aligned multimodal fusion features according to the weight coefficients. The weighted features are then residually connected with the original features to obtain the enhanced multimodal fusion features. The enhanced multimodal fusion features are reshaped according to preset batch dimension, time dimension and feature channel dimension to generate the fusion feature tensor.

4. The intelligent dispatching system for smart property parking lots based on artificial intelligence according to claim 3, characterized in that, The situation prediction module is specifically used for: The fused feature tensor is divided into multiple consecutive time slices along the time dimension, with each time slice corresponding to the input data of one time step, denoted as . ,in Indicates the time step index; Each time slice The input is a convolutional neural network, which extracts spatial features through multiple convolutional and pooling layers, and outputs a spatial feature vector at each time step. ; Time slices at each time step and the corresponding spatial feature vector A common input spatially augmented gated recurrent unit network is used, wherein each cell in the spatially augmented gated recurrent unit network calculates the hidden state at the current time step according to the following formula: in, Indicates the first The time slice of each time step Indicates the first The spatial feature vector at each time step, This indicates the hidden state of the previous time step. This indicates the hidden state at the current time step. This indicates an update to the gate output. Indicates resetting the gate output. Indicates the context gate output, This represents a context vector that integrates spatial features and historical information. Indicates the candidate hidden state. The weight matrix is ​​a learnable matrix. For learnable bias terms, This represents the sigmoid activation function. This represents the hyperbolic tangent activation function. This represents element-wise multiplication; Hidden state of the last time step Three parallel fully connected layers are input respectively. The first fully connected layer outputs a time-series prediction of vehicle arrival rate, which represents the sequence of the number of vehicles arriving at each time point within the future preset time period. The second fully connected layer outputs a parking space occupancy rate prediction, which represents the probability distribution matrix of each parking space being occupied within the future preset time period. The third fully connected layer outputs a turnover rate trend prediction, which represents the curve of change in parking space turnover rate within the future preset time period.

5. The intelligent dispatching system for smart property parking lots based on artificial intelligence according to claim 4, characterized in that, The dynamic scheduling strategy generation module is specifically used for: The multidimensional prediction data and the real-time status data are fused according to a preset feature splicing rule to generate an environmental status vector at the current moment. The multidimensional prediction data includes the vehicle arrival rate time series, the parking space occupancy rate distribution and the turnover rate trend. The real-time status data includes the occupancy status of each parking space, the vehicle queue length at the entrance and exit and the vehicle type identification result. The environment state vector is input into a deep reinforcement learning decision network, which includes an online policy network and a target policy network. The online policy network outputs an initial action probability distribution based on the environment state vector, and the target policy network is used to provide a stable target value during training to update the online policy network. An initial action vector is obtained by sampling the initial action probability distribution output by the online policy network. The initial action vector includes a partition action component, an induced action component, and a temporal action component. The partition action component is input into the first decoding network, which consists of multiple fully connected layers. The partition action component is converted into a dynamic parking space partitioning scheme through nonlinear mapping. The dynamic parking space partitioning scheme includes the boundary coordinates of each partition, the number of open parking spaces in each partition within a preset time period, and the types of vehicles allowed to be parked in each partition. The induced action component is input into the second decoding network. The second decoding network combines the real-time congestion index of each lane extracted from the real-time status data and generates a multi-path guidance scheme through a path search algorithm. The multi-path guidance scheme includes the preferred driving path sequence for entering vehicles, the exit guidance sign for leaving vehicles, and the path switching triggering condition. The timing action components are input into the third decoding network. The third decoding network combines the real-time queue length of each entrance and exit extracted from the real-time status data and generates an entrance and exit gate timing control scheme through a timing planning algorithm. The entrance and exit gate timing control scheme includes the opening time, closing time of each entrance and exit gate in a continuous time window, as well as the linkage and locking relationship between adjacent gates. The dynamic parking space zoning scheme, the multi-path guidance scheme, and the entrance / exit gate timing control scheme are combined to generate the initial dynamic scheduling strategy.

6. The intelligent dispatching system for smart property parking lots based on artificial intelligence according to claim 5, characterized in that, The multi-objective collaborative optimization module is specifically used for: Access the property service platform via application programming interface (API) to obtain security monitoring data and property service data from the property service platform. The security monitoring data includes security event records of various areas in the parking lot and abnormal behavior detection results obtained from real-time video analysis. The property service data includes users' historical complaint records, user level information, and users' preference settings for parking services. Taking the initial dynamic scheduling strategy as the individual to be optimized, a multi-objective optimization model is constructed. The multi-objective optimization model includes three objective functions. The first objective function is the traffic efficiency function, which calculates the average vehicle entry and exit time based on the dynamic parking space zoning scheme and multi-path guidance scheme in the initial dynamic scheduling strategy. The second objective function is the energy consumption balance function, which calculates the number of start-stop cycles and the running time of each gate device based on the entrance and exit gate timing control scheme in the initial dynamic scheduling strategy. The third objective function is the user satisfaction function, which calculates the comprehensive user satisfaction score based on the dynamic parking space zoning scheme and multi-path guidance scheme in the initial dynamic scheduling strategy, combined with the user preference settings in the property service data. A multi-objective evolutionary algorithm is used to solve the multi-objective optimization model. The multi-objective evolutionary algorithm generates a Pareto optimal solution set iteratively through population initialization, crossover mutation, and Pareto dominance comparison. Each solution in the Pareto optimal solution set corresponds to a candidate scheduling instruction. Based on the security monitoring data and the property service data, the preference score for each candidate scheduling instruction in the Pareto optimal solution set is calculated according to the following formula: in, This represents the candidate scheduling instructions in the Pareto optimal solution set. Indicates candidate scheduling instructions The corresponding number The objective function value, These correspond to the traffic efficiency function, the energy consumption balance function, and the user satisfaction function, respectively. and They represent the current Pareto optimal solution set, i.e., the first... The minimum and maximum values ​​of the objective function. This represents the feature vector of the property service data. This represents the feature vector of the security monitoring data. This indicates the extraction of the first [item] from the property service data. The intensity of preference for each objective This represents the security risk coefficient extracted from the security monitoring data. For the first The preference weight coefficients for each objective. For the first The security modulation coefficient of each target; Select the preference score The largest candidate scheduling instruction is selected as the optimized scheduling instruction.

7. The intelligent dispatching system for smart property parking lots based on artificial intelligence according to claim 6, characterized in that, The instruction execution and feedback module is specifically used for: The optimized scheduling instructions are parsed into control instructions for each parking lot terminal execution device, which includes a barrier gate controller, a parking space lock controller, a guidance screen controller, and a voice broadcast device. The control commands are sent to the corresponding parking lot terminal execution devices through a preset communication protocol, and the command execution status information returned by the parking lot terminal execution devices is received. The command execution status information includes a command reception confirmation signal, a device action completion signal, and a device fault alarm signal. The execution feedback data during the execution of the optimized scheduling instructions is collected. The execution feedback data includes the actual occupancy status change sequence of each parking space, the actual vehicle queue length change sequence of each entrance and exit, the actual opening and closing time of each barrier gate, the display content switching time of each guidance screen, and the location search time and satisfaction rating reported by users through mobile terminals. The execution feedback data is organized according to a preset data format to generate a feedback sample set. Each sample in the feedback sample set contains an environment state vector when the optimized scheduling instruction is executed, the optimized scheduling instruction, and the effect index after execution. The samples in the feedback sample set are stored in the experience replay buffer in chronological order. When the number of samples in the experience replay buffer reaches a preset threshold, a batch of samples are randomly selected from the experience replay buffer, and the artificial intelligence prediction model is fine-tuned using the gradient descent method to update the network parameters of the artificial intelligence prediction model. The samples in the feedback sample set are stored in the policy experience buffer in the same time order. When the number of samples in the policy experience buffer reaches a preset threshold, a batch of samples are randomly drawn from the policy experience buffer. The policy gradient method is used to update the deep reinforcement learning decision network and adjust the network parameters of the online policy network and the target policy network in the deep reinforcement learning decision network.

8. The intelligent dispatching system for smart property parking lots based on artificial intelligence according to claim 7, characterized in that, The system also includes: The model pre-training module is used to extract sample data from multiple historical time periods from the historical database before the system is deployed. The sample data includes historical fusion feature tensors, historical multidimensional prediction data, and corresponding historical actual operation data. The historical fusion feature tensors are used as input, and the historical multidimensional prediction data are used as supervision signals. The AI ​​prediction model is pre-trained using supervised learning. The network parameters of the AI ​​prediction model are adjusted using the mean squared error loss function. The historical actual operation data are used as environmental interaction samples. The deep reinforcement learning decision network is pre-trained using offline reinforcement learning. The network parameters of the online policy network and the target policy network are initialized by combining behavior cloning and policy constraints.

9. The intelligent dispatching system for smart property parking lots based on artificial intelligence according to claim 7, characterized in that, The system also includes: The user interaction and personalized strategy generation module receives parking and departure requests from users via a mobile terminal application. It extracts user preference features from the user's historical behavior logs, including preferred parking areas, walking distance tolerance, and willingness to pay. These user preference features are then fused with the current multidimensional prediction data and real-time status data to generate a personalized user state vector. This vector is input into a lightweight decision network, which extracts information from the deep reinforcement learning decision network through knowledge distillation. The lightweight decision network outputs personalized guidance information for the user, including a recommended parking space number, a customized route to the recommended parking space, and estimated walking time. This personalized guidance information is then pushed to the mobile terminal application for display.

10. A smart property parking lot intelligent scheduling method based on artificial intelligence, employing the smart property parking lot intelligent scheduling system based on artificial intelligence as described in any one of claims 1 to 9, characterized in that, The method includes: Collect multi-source heterogeneous data from the parking lot. The multi-source heterogeneous data includes historical operation data, real-time status data, and external environment data. The historical operation data includes vehicle entry and exit records, parking space occupancy time series, and user behavior logs. The real-time status data includes the occupancy status of each parking space, vehicle queue length at entrances and exits, vehicle type identification results, and in-park video streams. The external environment data includes meteorological information and surrounding traffic conditions. The multi-source heterogeneous data is spatiotemporally registered and feature-level fused to generate aligned multimodal fused data. The aligned multimodal fused data is then enhanced using a feature weighting method based on an attention mechanism to generate a fused feature tensor. The fused feature tensor is input into a pre-trained artificial intelligence prediction model, which is a combination model of convolutional neural network and gated recurrent unit, and outputs multi-dimensional prediction data of parking lot operation status for a future preset time period. The multi-dimensional prediction data includes vehicle arrival rate time series, parking space occupancy rate distribution and turnover rate trend. Based on the multidimensional prediction data and the real-time status data, a deep reinforcement learning decision network is invoked to generate an initial dynamic scheduling strategy. The initial dynamic scheduling strategy includes a dynamic parking space zoning scheme, a multi-path guidance scheme, and an entrance and exit gate timing control scheme. The system accesses the property service platform to obtain security monitoring data and property service data. Using the initial dynamic scheduling strategy as input, it constructs a multi-objective optimization model that includes traffic efficiency, energy consumption balance, and user satisfaction. The system then uses a multi-objective evolutionary algorithm to solve for Pareto optimal solutions and selects optimized scheduling instructions from the Pareto optimal solution set. The optimized scheduling command is sent to the parking lot terminal execution device, and execution feedback data is collected. The execution feedback data is used to update the artificial intelligence prediction model and the deep reinforcement learning decision network.

Images