Intelligent parking charging pile positioning method and system combined with solar energy

Abstract

The present invention provides a method and system for positioning a smart parking charging pile in combination with solar energy, which relates to the technical field of data processing, including: drawing a three-dimensional electronic map of a parking lot through a sensor, marking parking spaces and sensing changes in occupancy status, building a multivariate power generation prediction model, a vehicle information knowledge map, determining the association law between vehicle information and building a charging demand prediction model, and generating a vehicle prediction report; obtaining vehicle attribute information, calculating a fitness score and predicting the optimal parking area for the current vehicle, generating a recommended parking area list, solving the optimal parking path and parking space selection strategy, and solving a vehicle charging scheduling plan with the goal of maximizing charging efficiency and minimizing waiting time; solving a charging power scheduling curve and updating a multivariate power generation prediction model, and if a deviation occurs and the deviation is greater than a preset deviation threshold, updating the vehicle charging scheduling plan, charging, and generating a charging report according to the charging process.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for positioning an intelligent parking charging pile combined with solar energy. Background Art

[0002] As global warming and environmental pollution become increasingly prominent, the development of clean energy and green transportation has become a consensus among countries around the world. Electric vehicles, with their advantages of zero emissions, low noise and high efficiency, have become an important direction for the transformation and upgrading of the automotive industry, showing a rapid development trend.

[0003] The large-scale promotion of electric vehicles still faces many challenges. Among them, the lagging charging infrastructure is one of the main bottlenecks restricting its development. The installed capacity of renewable energy power generation represented by photovoltaic power generation continues to expand, but due to the intermittent fluctuations of natural conditions and the uncertainty of load demand, the phenomenon of "abandonment of photovoltaic power generation and power rationing" occurs from time to time, and the resource utilization efficiency is not high.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for positioning an intelligent parking charging pile in combination with solar energy, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for positioning a smart parking charging pile in combination with solar energy, comprising:

[0007] The three-dimensional spatial information of the parking lot is obtained through pre-deployed sensors, and the image data obtained by the panoramic camera is combined to draw a three-dimensional electronic map of the parking lot through a synchronous positioning and map construction algorithm. The parking space coordinates are marked and the occupancy status changes are sensed through magnetic sensors pre-installed on the parking spaces. According to the environmental information collected by the environmental sensors, combined with historical environmental parameters and historical power generation data of solar panels, a multivariate power generation prediction model is built in combination with a long short-term memory network. According to the traffic video data, parked vehicles are identified through target detection and tracking algorithms, and parking space occupancy data is obtained and uploaded to the edge computing server. A structured vehicle information knowledge graph is built based on vehicle attribute information and vehicle preference information. The association between vehicle information is determined through a graph neural network algorithm and a charging demand prediction model is built to generate a vehicle prediction report.

[0008] The unique identification number of the vehicle entering the parking lot is obtained through the information collection device, and the real-time location coordinates and battery capacity are obtained. The vehicle attribute information corresponding to the current vehicle is obtained in combination with the vehicle information knowledge graph. According to the vehicle attribute information and the remaining resource information of the parking lot, the fitness score is calculated and the optimal parking area of ​​the current vehicle is predicted through the collaborative filtering recommendation algorithm based on graph network embedding, and a parking area recommendation list is generated. The optimal parking path and parking space selection strategy are solved by a multi-objective reinforcement learning algorithm with the parking area fitness and parking time as the reward function, the vehicle is parked and the remaining resource information of the parking lot is updated. The data center solves the vehicle charging scheduling plan based on the vehicle battery information with the goal of maximizing charging efficiency and minimizing waiting time;

[0009] The short-term power generation corresponding to the solar panel is obtained according to the multivariable power generation prediction model, the charging power scheduling curve is solved in combination with the vehicle charging scheduling plan, and the multivariable power generation prediction model is dynamically updated through incremental learning. If the charging power scheduling curve deviates and the deviation is greater than a preset deviation threshold, the vehicle charging scheduling plan is updated through a rolling optimization strategy, and charging is performed according to the updated charging scheduling plan and the fuzzy logic control algorithm. During the charging process, the battery status parameters are obtained in real time and the noise is filtered out through the Kalman filtering algorithm. The charging strategy is dynamically adjusted in combination with the adaptive neural fuzzy inference algorithm and the solar power generation status parameters until charging is completed. A charging report is generated according to the charging process and the vehicle information knowledge graph is updated.

[0010] In an optional embodiment,

[0011] The three-dimensional spatial information of the parking lot is obtained through pre-deployed sensors, combined with the image data obtained by the panoramic camera, and a three-dimensional electronic map of the parking lot is drawn through a synchronous positioning and map construction algorithm. The parking space coordinates are marked and the occupancy status changes are sensed through magnetic sensors pre-installed on the parking spaces. According to the environmental information collected by the environmental sensors, combined with historical environmental parameters and historical power generation data of solar panels, a multivariate power generation prediction model is built in combination with a long short-term memory network. According to the traffic video data, parked vehicles are identified through target detection and tracking algorithms, parking space occupancy data is obtained and uploaded to the edge computing server, and a structured vehicle information knowledge graph is built based on vehicle attribute information and vehicle preference information. The association between vehicle information is determined through a graph neural network algorithm and a charging demand prediction model is built. The vehicle prediction report generated includes:

[0012] The parking lot space is fully scanned by a laser radar sensor pre-deployed in the parking lot to obtain the three-dimensional spatial information of the parking lot in the form of a point cloud. Combined with a pre-set panoramic camera, a panoramic color image corresponding to the parking lot is obtained. The three-dimensional spatial information of the parking lot and the panoramic color image are synchronously aligned in time and space. The point cloud data is pre-processed based on a synchronous positioning and map building algorithm, feature points are selected and the initial pose estimation corresponding to the sensor is obtained by tracking the feature points between different frames, a constraint graph between the sensor pose and the feature points is constructed, the pose is optimized by a graph optimization library, and the laser point cloud is inserted into a pre-set octree map to obtain the three-dimensional electronic map of the parking lot.

[0013] The coordinates of the parking spaces are marked in the three-dimensional electronic map of the parking lot, and the magnetic field strength data in the state of having a car and the state of not having a car are collected. The judgment threshold is solved by a linear discriminant analysis method, and the three-axis magnetic field strength data of the current parking space is obtained in combination with a magnetic sensor pre-installed on the parking space. The change in the occupancy state of the current parking space is judged according to the three-axis magnetic field strength data and the judgment threshold. The environmental parameters at the current moment are collected by a pre-set multi-parameter sensor, wherein the environmental parameters include temperature, humidity, light and wind speed. The historical power generation data of the solar panels in the parking lot are obtained, and a nonlinear mapping relationship between the current environmental parameters and the historical power generation data is constructed by combining a long short-term memory network with the collected environmental parameter time series data as input, and the multivariate power generation prediction model is constructed based on the nonlinear mapping relationship;

[0014] According to the traffic video data acquired by the panoramic camera, a multi-scale feature map is generated through the feature extraction network in the target detection model, and the target bounding box and category probability are predicted at each position of the multi-scale feature map in combination with the anchor frame mechanism, and the detection result is output. According to the detection result, a deep simple sorting multi-target tracking algorithm is used, and the vehicle bounding box in the detection result is used as an observation. The position and speed of each vehicle target in the next frame are predicted through a Kalman filter, and the vehicle bounding box detected in the current frame is matched with the tracking result of the previous frame by solving the minimum cost matrix to generate a tracking trajectory, and the parked vehicle and parking time are determined based on the spatial relationship between the tracking trajectory and the parking space coordinates, and the parking space occupancy data is obtained by combining and uploading to the edge computing server;

[0015] For each vehicle entering the current parking lot, the edge node obtains the vehicle attribute information and vehicle preference information corresponding to the current vehicle, wherein the vehicle attribute information includes vehicle category, model, brand, fast charging rate and driver information, and the vehicle preference information includes historical parking location, historical parking duration and charging preference. If the current vehicle enters the current parking lot for the first time, the vehicle preference information is marked as empty. The vehicle attribute information and the vehicle preference information are imported into the database in graph form to generate a structured vehicle information knowledge graph. The charging demand prediction model is constructed through a graph neural network algorithm. The vehicle information knowledge graph is divided into subgraphs centered on a single vehicle and containing multi-hop neighbors. For each subgraph, the node features are updated and the neighbor node features are aggregated through a graph convolutional neural network. The vehicle node feature representation is repeatedly aggregated. The association rules between vehicle information and the influence between vehicle information and charging demand are determined based on the vehicle node feature representation. The hyperparameters of the charging demand prediction model are updated and predicted in combination with the back propagation algorithm to generate the predicted charging time corresponding to the current vehicle. The vehicle prediction report is obtained in combination with the tracking trajectory and vehicle attribute information.

[0016] In an optional embodiment,

[0017] The following formula is used to update node features and aggregate neighbor node features through graph convolutional neural network:

[0018] ;

[0019] in, Indicates l + 1 Tier i The feature representation of a node, s represents the activation function, N ( i ) represents a node i The set of neighbor nodes of a represents the attention vector, T represents transpose, W Indicates l The weight matrix of the layer, Indicates l Tier i The feature representation of a node, Indicates l Tier j The feature representation of a node.

[0020] In an optional embodiment,

[0021] The unique identification number of the vehicle entering the parking lot is obtained through the information collection device, and the real-time location coordinates and battery capacity are obtained. The vehicle attribute information corresponding to the current vehicle is obtained in combination with the vehicle information knowledge graph. According to the vehicle attribute information and the remaining resource information of the parking lot, the fitness score is calculated and the optimal parking area of ​​the current vehicle is predicted through the collaborative filtering recommendation algorithm based on graph network embedding, and a parking area recommendation list is generated. The optimal parking path and parking space selection strategy are solved by a multi-objective reinforcement learning algorithm with the parking area fitness and parking time as the reward function, the vehicle is parked and the remaining resource information of the parking lot is updated. The data center solves the vehicle charging scheduling scheme based on the vehicle battery information with the goal of maximizing charging efficiency and minimizing waiting time, including:

[0022] The unique identification number corresponding to the vehicle entering the parking lot is obtained through a pre-set information collection device, and the vehicle model parameters and battery capacity are obtained in combination with the vehicle model library in the network. At the same time, the real-time position coordinates of the current vehicle are obtained through the panoramic camera. Based on the vehicle information knowledge graph, the vehicle attribute information corresponding to the current vehicle is queried, and the vehicle attribute information and associated entities corresponding to the current vehicle are obtained according to pattern matching;

[0023] Based on the vehicle attribute information and parking lot remaining resource information, combined with the network embedding representation learning algorithm, the vehicle information knowledge graph is used as a heterogeneous information network, a heterogeneous node set and a semantic relationship edge set are defined, and the link topology and node attribute information of the heterogeneous information network are learned to obtain the comprehensive feature representation corresponding to the current vehicle. Based on the comprehensive feature representation, the low-dimensional dense vector of each vehicle node is learned through the graph convolution network model, and the latent vector matrix is ​​generated and the node embedding matrix is ​​obtained by stacking multiple layers of graph convolution. The aggregation weights of the neighbor nodes corresponding to the current vehicle node are dynamically adjusted in combination with the attention mechanism, and the neighbor nodes are aggregated in combination with the attribute feature projection matrix. According to the aggregation results, irrelevant vehicle-vehicle edges are judged and shielded. Through the multi-task learning algorithm, as well as node classification and relationship prediction as the goal, the graph convolution network model is trained in combination with the cross entropy loss function to obtain a trained graph convolution network;

[0024] For the current vehicle, the vehicle similarity is calculated by cosine similarity, and the vehicle with the highest similarity to the current vehicle in the parking lot is selected as the comparison vehicle according to the vehicle similarity, and the vehicle-parking space similarity between the parking spaces corresponding to the current vehicle and the comparison vehicle is calculated. The regional scoring model is constructed by combining the matrix decomposition technology. For each area, the scoring vector of each parking space is calculated in combination with the current vehicle preference. According to the scoring vector, the parking spaces are sorted in descending order and the parking area recommendation list is generated;

[0025] According to the pre-acquired three-dimensional electronic map of the parking lot, the current parking lot road is modeled to obtain an undirected weighted graph. According to the reinforcement learning algorithm, the state space is constructed in combination with the vehicle position, road occupancy and target parking space conditions. The vehicle driving is used as an action and the action space is constructed. According to the parking time and parking difficulty, a reward function is set and a path search is performed to obtain the optimal parking path and parking space selection strategy. The vehicle is parked and the remaining resource information of the parking lot is updated. According to the vehicle battery information, the data center solves the vehicle charging scheduling plan with the goal of maximizing the charging efficiency and minimizing the waiting time.

[0026] In an optional embodiment,

[0027] The loss function corresponding to path search based on reinforcement learning is shown in the following formula:

[0028] ;

[0029] in, L ( i ) represents the loss value, Indicates in status s Take action a Rewards received after r and the next state s' The expected value of r Indicates the current time step t Rewards received, c represents the discount factor, Indicates that in the next state s' Next, all actions a' The maximum Q value, Indicates the current status s and actions a The Q value of the network is i Calculated.

[0030] In an optional embodiment,

[0031] According to the multivariable power generation prediction model, the short-term power generation corresponding to the solar panel is obtained, and the charging power scheduling curve is solved in combination with the vehicle charging scheduling plan, and the multivariable power generation prediction model is dynamically updated through incremental learning. If the charging power scheduling curve deviates and the deviation is greater than the preset deviation threshold, the vehicle charging scheduling plan is updated through a rolling optimization strategy, and charging is performed according to the updated charging scheduling plan and the fuzzy logic control algorithm. During the charging process, the battery state parameters are obtained in real time and the noise is filtered through the Kalman filter algorithm. The charging strategy is dynamically adjusted in combination with the adaptive neural fuzzy reasoning algorithm and the solar power generation state parameters until charging is completed. A charging report is generated according to the charging process and the vehicle information knowledge graph is updated, including:

[0032] According to the preset multivariable power generation prediction model, the short-term power generation corresponding to the solar panel is obtained, and the standard charging power scheduling curve is calculated in combination with the vehicle charging scheduling plan. The incremental samples are added to the multivariable power generation prediction model through the incremental learning algorithm, and the gradient of the loss function of each layer is calculated. The parameters are updated in combination with the stochastic gradient descent algorithm to determine the influence weight of the incremental sample on the original loss function. If the influence weight is greater than the preset influence threshold, it is considered that the incremental sample has information gain, and it is added to the training set and incremental learning is performed. The updated prediction mean and variance are determined in combination with the Bayesian theory to obtain the updated multivariable power generation prediction model. According to the updated multivariable power generation prediction model, the actual charging power scheduling curve is obtained by solving the vehicle charging constraints and the power grid constraints.

[0033] The actual charging power dispatch curve is compared with the standard charging power dispatch curve, and the deviation value is calculated and compared with the preset deviation threshold. If the deviation value is greater than the deviation threshold, a hierarchical collaborative optimization model is set through a rolling optimization strategy, wherein the hierarchical collaborative optimization model includes a day-ahead optimization layer, an intraday correction layer, and a real-time control layer. According to the hierarchical collaborative optimization model, the goal is to maximize the solar energy absorption and minimize the charging cost and the peak-to-valley difference, and the battery physical constraints and the grid technical constraints are used as constraints. The vehicle entry time and departure time and the vehicle battery capacity are initialized according to the vehicle information knowledge graph, and the vehicle charging scheduling plan is updated according to the electricity price information and the photovoltaic output information in combination with the hardware utility function;

[0034] Charging is performed based on the updated vehicle charging scheduling plan, and battery status parameters during the charging process are obtained in real time. Noise is filtered out through the Kalman filtering algorithm, and a fuzzy control rule base is constructed in combination with the fuzzy logic control algorithm. The clear value of the control quantity is determined according to the control rules and the center of gravity defuzzification method, and the optimal charging power is obtained by inferring the clear value. At the same time, according to the uncertainty of battery characteristics and photovoltaic fluctuations, the rules in the fuzzy control rule base are updated through an adaptive neural fuzzy inference algorithm, and the update is repeated until convergence. The updated vehicle charging scheduling plan is dynamically adjusted according to the updated rules until charging is completed. After charging is completed, a charging report is generated according to the charging process and the vehicle information knowledge graph is updated.

[0035] In an optional embodiment,

[0036] The hierarchical collaborative optimization model updates the vehicle charging scheduling scheme including:

[0037] The day-ahead optimization layer obtains the prediction information and combines it with the vehicle information knowledge graph to obtain the power dispatch plan of the current vehicle in the next 24 hours. Based on the dispatch plan, the intraday correction layer obtains the predicted power generation power in the next 4 hours in a rolling manner, and corrects the charging power in combination with the current vehicle data in the parking lot. The field-level charging plan in the next four hours is updated and sent to the charging piles in each parking space. The solar power generation power is tracked on the premise of meeting the charging demand through the rolling time domain optimization strategy.

[0038] The real-time control layer determines the charging power for the current period through fuzzy logic control based on the real-time output of solar power generation, combined with the battery charge state and temperature, and adjusts the vehicle charging scheduling plan based on the generated charging power decision.

[0039] A second aspect of an embodiment of the present invention provides a smart parking charging pile positioning system combined with solar energy, comprising:

[0040] The first unit is used to obtain the three-dimensional spatial information of the parking lot through pre-deployed sensors, combine the image data obtained by the panoramic camera, draw a three-dimensional electronic map of the parking lot through a synchronous positioning and map construction algorithm, mark the coordinates of the parking spaces, and sense the changes in occupancy status through magnetic sensors pre-installed on the parking spaces. According to the environmental information collected by the environmental sensors, combined with historical environmental parameters and historical power generation data of solar panels, a multivariate power generation prediction model is constructed in combination with a long short-term memory network. According to the traffic video data, parked vehicles are identified through target detection and tracking algorithms, parking space occupancy data is obtained and uploaded to the edge computing server, a structured vehicle information knowledge graph is constructed based on vehicle attribute information and vehicle preference information, the association rules between vehicle information are determined through a graph neural network algorithm, and a charging demand prediction model is constructed to generate a vehicle prediction report;

[0041] The second unit is used to obtain the unique identification number of the vehicle entering the parking lot through the information collection device, obtain the real-time location coordinates and battery capacity, and obtain the vehicle attribute information corresponding to the current vehicle in combination with the vehicle information knowledge graph. According to the vehicle attribute information and the remaining resource information of the parking lot, the fitness score is calculated and the optimal parking area of ​​the current vehicle is predicted through the collaborative filtering recommendation algorithm based on graph network embedding, and a parking area recommendation list is generated. Through the multi-objective reinforcement learning algorithm, the optimal parking path and parking space selection strategy are solved with the parking area fitness and parking time as the reward function, the vehicle is parked and the remaining resource information of the parking lot is updated. The data center solves the vehicle charging scheduling plan based on the vehicle battery information with the goal of maximizing the charging efficiency and minimizing the waiting time;

[0042] The third unit is used to obtain the short-term power generation corresponding to the solar panel according to the multivariable power generation prediction model, solve the charging power scheduling curve in combination with the vehicle charging scheduling plan, and dynamically update the multivariable power generation prediction model through incremental learning. If the charging power scheduling curve deviates and the deviation is greater than a preset deviation threshold, the vehicle charging scheduling plan is updated through a rolling optimization strategy, and charging is performed according to the updated charging scheduling plan and fuzzy logic control algorithm. During the charging process, battery status parameters are obtained in real time and noise is filtered through the Kalman filtering algorithm. The charging strategy is dynamically adjusted in combination with the adaptive neural fuzzy inference algorithm and the solar power generation status parameters until charging is completed, and a charging report is generated according to the charging process and the vehicle information knowledge graph is updated.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the embodiments of the present invention is:

[0049] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0050] In the present invention, a multivariate power generation prediction model constructed by environmental sensors and historical data is combined with incremental learning to continuously optimize the prediction accuracy, so that the charging power scheduling is highly matched with the photovoltaic power generation characteristics, the local consumption of solar energy is maximized, and the power curtailment is reduced. The optimal parking path and parking space selection are obtained through reinforcement learning, and the parking time is shortened. The charging scheduling scheme takes into account both charging efficiency and waiting time. The adaptive control algorithm ensures the safety and efficiency of the charging process and improves the user experience. The vehicle information knowledge graph mines vehicle and user characteristics. The personalized parking and charging service is conducive to cultivating user habits. Intelligent operation also creates a better infrastructure environment for the promotion and popularization of new energy vehicles. A rolling optimization mechanism is introduced to perform online correction of the charging power deviation, maximize the matching of solar power generation characteristics, and improve the local consumption level of photovoltaic power generation. In summary, the present invention has achieved remarkable results in improving the utilization efficiency of solar energy, optimizing the user charging experience, and accelerating the popularization of new energy vehicles. At the same time, it provides a technical solution that can be used as a reference for promoting the energy revolution and the construction of smart cities, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a method for positioning a smart parking charging pile using solar energy according to an embodiment of the present invention;

[0052] Figure 2 It is a structural schematic diagram of an intelligent parking charging pile positioning system combined with solar energy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0055] Figure 1 FIG. 1 is a flow chart of a method for positioning a smart parking charging pile using solar energy according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] S1. Obtain the three-dimensional spatial information of the parking lot through pre-deployed sensors, combine the image data obtained by the panoramic camera, draw a three-dimensional electronic map of the parking lot through the synchronous positioning and map construction algorithm, mark the coordinates of the parking spaces, and sense the changes in occupancy status through the magnetic sensors pre-installed on the parking spaces. According to the environmental information collected by the environmental sensors, combined with the historical environmental parameters and the historical power generation data of the solar panels, a multivariate power generation prediction model is constructed in combination with the long short-term memory network. According to the traffic video data, the parked vehicles are identified through the target detection and tracking algorithm, the parking space occupancy data is obtained and uploaded to the edge computing server, and a structured vehicle information knowledge graph is constructed based on the vehicle attribute information and vehicle preference information. The association rules between vehicle information are determined through the graph neural network algorithm and a charging demand prediction model is constructed to generate a vehicle prediction report;

[0057] The three-dimensional spatial information of the parking lot involves building a stereo model of the parking lot, using a laser radar or a depth camera to obtain the height and structural information of the site to optimize parking management and navigation. The panoramic camera provides a 360-degree view and is suitable for monitoring and navigation. The synchronous positioning and map building algorithm combines sensor data (such as a laser radar, camera) and motion information to build an environmental map in real time and locate its own position. The magnetic sensor is used to detect the location information of the vehicle and provide accurate positioning data by measuring changes in the earth's magnetic field. The long short-term memory network is a special recursive neural network that is suitable for processing time series data and can capture long-term dependencies. The multivariate power generation prediction model uses multiple influencing factors (such as weather and historical power generation data) to predict power generation, and usually uses machine learning algorithms (such as regression and neural networks) to improve prediction accuracy. The target detection algorithm identifies specific objects in the image, and the tracking algorithm continuously tracks these objects in the video. The graph neural network processes graph structure data and realizes information dissemination and aggregation through feature learning of nodes and edges. The charging demand prediction model predicts the demand for electric vehicle charging, usually combining historical charging data, user behavior and environmental factors, and using time series analysis or machine learning methods to optimize the layout and resource allocation of charging facilities.

[0058] In an optional embodiment,

[0059] The three-dimensional spatial information of the parking lot is obtained through pre-deployed sensors, combined with the image data obtained by the panoramic camera, and a three-dimensional electronic map of the parking lot is drawn through a synchronous positioning and map construction algorithm. The parking space coordinates are marked and the occupancy status changes are sensed through magnetic sensors pre-installed on the parking spaces. According to the environmental information collected by the environmental sensors, combined with historical environmental parameters and historical power generation data of solar panels, a multivariate power generation prediction model is built in combination with a long short-term memory network. According to the traffic video data, parked vehicles are identified through target detection and tracking algorithms, parking space occupancy data is obtained and uploaded to the edge computing server, and a structured vehicle information knowledge graph is built based on vehicle attribute information and vehicle preference information. The association between vehicle information is determined through a graph neural network algorithm and a charging demand prediction model is built. The vehicle prediction report generated includes:

[0060] The parking lot space is fully scanned by a laser radar sensor pre-deployed in the parking lot to obtain the three-dimensional spatial information of the parking lot in the form of a point cloud. Combined with a pre-set panoramic camera, a panoramic color image corresponding to the parking lot is obtained. The three-dimensional spatial information of the parking lot and the panoramic color image are synchronously aligned in time and space. The point cloud data is pre-processed based on a synchronous positioning and map building algorithm, feature points are selected and the initial pose estimation corresponding to the sensor is obtained by tracking the feature points between different frames, a constraint graph between the sensor pose and the feature points is constructed, the pose is optimized by a graph optimization library, and the laser point cloud is inserted into a pre-set octree map to obtain the three-dimensional electronic map of the parking lot.

[0061] The coordinates of the parking spaces are marked in the three-dimensional electronic map of the parking lot, and the magnetic field strength data in the state of having a car and the state of not having a car are collected. The judgment threshold is solved by a linear discriminant analysis method, and the three-axis magnetic field strength data of the current parking space is obtained in combination with a magnetic sensor pre-installed on the parking space. The change in the occupancy state of the current parking space is judged according to the three-axis magnetic field strength data and the judgment threshold. The environmental parameters at the current moment are collected by a pre-set multi-parameter sensor, wherein the environmental parameters include temperature, humidity, light and wind speed. The historical power generation data of the solar panels in the parking lot are obtained, and a nonlinear mapping relationship between the current environmental parameters and the historical power generation data is constructed by combining a long short-term memory network with the collected environmental parameter time series data as input, and the multivariate power generation prediction model is constructed based on the nonlinear mapping relationship;

[0062] According to the traffic video data acquired by the panoramic camera, a multi-scale feature map is generated through the feature extraction network in the target detection model, and the target bounding box and category probability are predicted at each position of the multi-scale feature map in combination with the anchor frame mechanism, and the detection result is output. According to the detection result, a deep simple sorting multi-target tracking algorithm is used, and the vehicle bounding box in the detection result is used as an observation. The position and speed of each vehicle target in the next frame are predicted through a Kalman filter, and the vehicle bounding box detected in the current frame is matched with the tracking result of the previous frame by solving the minimum cost matrix to generate a tracking trajectory, and the parked vehicle and parking time are determined based on the spatial relationship between the tracking trajectory and the parking space coordinates, and the parking space occupancy data is obtained by combining and uploading to the edge computing server;

[0063] For each vehicle entering the current parking lot, the edge node obtains the vehicle attribute information and vehicle preference information corresponding to the current vehicle, wherein the vehicle attribute information includes vehicle category, model, brand, fast charging rate and driver information, and the vehicle preference information includes historical parking location, historical parking duration and charging preference. If the current vehicle enters the current parking lot for the first time, the vehicle preference information is marked as empty. The vehicle attribute information and the vehicle preference information are imported into the database in graph form to generate a structured vehicle information knowledge graph. The charging demand prediction model is constructed through a graph neural network algorithm. The vehicle information knowledge graph is divided into subgraphs centered on a single vehicle and containing multi-hop neighbors. For each subgraph, the node features are updated and the neighbor node features are aggregated through a graph convolutional neural network. The vehicle node feature representation is repeatedly aggregated. The association rules between vehicle information and the influence between vehicle information and charging demand are determined based on the vehicle node feature representation. The hyperparameters of the charging demand prediction model are updated and predicted in combination with the back propagation algorithm to generate the predicted charging time corresponding to the current vehicle. The vehicle prediction report is obtained in combination with the tracking trajectory and vehicle attribute information.

[0064] The panoramic color image is a color image with a 360-degree field of view captured by a panoramic camera, and is commonly used in environmental monitoring and navigation applications to provide rich visual information. The spatiotemporal synchronization alignment is to align data from different times and spaces so that they can be analyzed in the same coordinate system, and is commonly used in multi-sensor data fusion and time series analysis. The initial pose estimation is the process of determining the initial position and orientation of an object in space, and usually relies on sensor data such as IMU or visual information to lay the foundation for subsequent positioning and map construction. The constraint graph is a graph structure that represents the constraint relationship between states, where nodes represent states and edges represent constraints. The graph optimization library provides efficient algorithms and tools for solving graph structure optimization problems. The octree map is a spatial data structure for storing three-dimensional spatial information. The linear discriminant analysis is a supervised learning algorithm for dimensionality reduction and classification. The three-axis magnetic field strength data comes from a magnetic sensor, representing the magnetic field strength in the X, Y and Z directions, and is commonly used for position estimation and direction perception, especially in GPS signals. In an unstable environment, the anchor frame mechanism is used for target detection, defining a set of preset frames (anchor frames) to predict the target position and size. The deep simple sorting multi-target tracking algorithm combines deep learning and sorting strategies to track multiple targets in the video, and uses deep features to improve the accuracy and robustness of tracking. The vehicle bounding box is a rectangular box that identifies the vehicle in target detection, which defines the position and size of the vehicle in the image and is commonly used in autonomous driving and monitoring systems. The Kalman filter is a recursive filtering algorithm used to estimate the system state and can process noisy measurement data. The process of minimizing the cost matrix involves optimizing a certain cost function under given constraints to achieve the optimal solution to problems such as resource allocation or path planning. The tracking trajectory is a path that describes the position change of the target in a time series, which is usually obtained through a positioning algorithm or sensor data. The subgraph of the multi-hop neighbor refers to a set of neighbor nodes in the graph structure that is obtained through multi-hop connections starting from a certain node, and is commonly used for information propagation and feature learning in graph neural networks.

[0065] Pre-deploy a laser radar sensor in the parking lot, perform a full coverage scan of the parking lot space, obtain the three-dimensional spatial information of the parking lot in the form of a point cloud, pre-set a panoramic camera in the parking lot, obtain a panoramic color image corresponding to the parking lot, synchronize the acquired three-dimensional spatial information of the parking lot and the panoramic color image in time and space, pre-process the point cloud data based on the simultaneous localization and mapping algorithm (SLAM), select feature points and obtain the initial pose estimation corresponding to the sensor by tracking the feature points between different frames, construct a constraint graph between the sensor pose and the feature points, optimize the pose through the graph optimization library, insert the optimized laser point cloud into the pre-set octree map, and obtain a three-dimensional electronic map of the parking lot. For example, an underground parking lot of a large shopping mall covers an area of ​​about 5,000 square meters. By arranging 4 laser radar sensors and 6 panoramic cameras in the parking lot, a three-dimensional electronic map of the parking lot is constructed through the above steps, and the specific coordinates of the parking spaces are marked on the map;

[0066] The coordinates of each parking space are marked on a three-dimensional electronic map of the parking lot. The magnetic field strength data of each parking space is collected in both the states of having a car and not having a car. The decision threshold is solved by a linear discriminant analysis method. A magnetic sensor is pre-installed on each parking space to obtain the three-axis magnetic field strength data of the current parking space in real time. According to the obtained three-axis magnetic field strength data and the calculated decision threshold, it is determined whether the occupancy status of the current parking space has changed. For example, a magnetic sensor is installed on each parking space in the above parking lot. By collecting 100 groups of data in the states of having a car and not having a car, the decision threshold is calculated to be [X1, X2, X3]. When the current three-axis magnetic strength data of a parking space is [x1, x2, x3], and x1>X1, x2>X2, x3>X3 is satisfied, it is determined that the parking space has changed from having a car to having a car.

[0067] The current environmental parameters, including temperature, humidity, light and wind speed, are collected through pre-set multi-parameter sensors, and the historical power generation data of the solar panels in the parking lot are obtained. Combined with the long short-term memory (LSTM) neural network, the collected environmental parameter time series data is used as input and the historical power generation data is used as output to construct a nonlinear mapping relationship between environmental parameters and power generation data. A multivariate power generation prediction model is constructed based on the above nonlinear mapping relationship. For example, 10 solar panels are installed in the parking lot, and the power generation is recorded every 5 minutes while the environmental parameters are collected. Take the data from the most recent week, and the data for every 30 minutes constitutes a sample, for a total of 336 samples. 80% are randomly selected as the training set, and the rest are used as the test set. The multivariate power generation prediction model is trained to predict the power generation in the next 3 hours every 1 hour on the same day;

[0068] Obtain traffic video data in the parking lot through a panoramic camera, generate a multi-scale feature map using the feature extraction network of the target detection model, predict the target bounding box and category probability at each position of the multi-scale feature map in combination with the anchor frame mechanism, and output the vehicle detection result. According to the vehicle detection result, a deep simple online real-time tracking algorithm is used, and the detected vehicle bounding box is used as an observation. The position and speed of each vehicle target in the next frame are predicted through a Kalman filter. The minimum cost matrix is ​​solved by the Hungarian algorithm, and the vehicle bounding box detected in the current frame is matched with the tracking result of the previous frame to generate a vehicle tracking trajectory. Based on the spatial relationship between the vehicle tracking trajectory and the parking space coordinates, the parked vehicle and the parking time are determined, and the parking space occupancy data is obtained by combining them. The parking space occupancy data is uploaded to the edge computing server. For example, 6 panoramic cameras are arranged at the entrance and exit and the passage of the parking lot to detect and track vehicles in real time. When a vehicle is detected to enter the bounding box of a parking space and remain stationary for more than 5 minutes, it is determined that the vehicle is parked in the parking space, and the vehicle entry time is recorded. When the vehicle leaves the parking space bounding box, the departure time is recorded, the parking time is calculated, and the parking space occupancy record is generated and uploaded to the server;

[0069] For each vehicle entering the parking lot, the edge node first obtains the vehicle attribute information (vehicle category, model, brand, fast charging rate, driver information) and vehicle preference information (historical parking location, historical parking duration, charging preference) of the vehicle. If the vehicle enters for the first time, the vehicle preference information is marked as empty. The vehicle attribute information and preference information are imported into the database in the form of a graph to generate a structured vehicle information knowledge graph, and a charging demand prediction model is constructed through a graph neural network algorithm. The vehicle information knowledge graph is divided into a subgraph centered on a single vehicle and containing multi-hop neighbors. For each subgraph, the node features are updated through a graph convolutional neural network (GCN), and the neighbor node features are aggregated. The aggregation operation is repeated multiple times to obtain the feature representation of each vehicle node. Based on the feature representation of the vehicle node, the association rules between vehicle information and the impact of vehicle information on charging demand are mined. The back propagation algorithm is used to optimize the model hyperparameters and perform charging demand prediction to obtain the predicted charging time of the vehicle. The vehicle tracking trajectory, attribute information and predicted charging time are combined to generate a vehicle prediction report for the vehicle. For example, an electric car of model A with license plate number ABC123 enters the parking lot for the first time. The attribute information of the vehicle is obtained through the on-board terminal and imported into the graph database together with the preference information tag to obtain the feature representation vector of the vehicle. The feature representation vector is input into the charging demand prediction model, and the charging time of the vehicle is predicted to be 2.5 hours. The vehicle is tracked to enter the B3 parking space, and a report is generated: Vehicle ABC123 (Brand A) enters the B3 parking space at 14:20, and the estimated charging time is 2.5 hours.

[0070] In this embodiment, the three-dimensional spatial information and panoramic images of the parking lot are obtained through laser radar scanning and panoramic camera shooting. After point cloud processing and image matching, an accurate three-dimensional electronic map of the parking lot is constructed to provide a basis for subsequent positioning, monitoring and scheduling. By arranging magnetic sensors on the parking spaces, the magnetic field data of the parking spaces are collected in real time, and the data is classified by combining the linear discriminant analysis method to accurately judge the changes in the occupancy status of the parking spaces, and to achieve real-time monitoring of the use of parking spaces. Meteorological data is collected through environmental parameter sensors, combined with historical power generation data, and a multivariate power generation prediction model is constructed using a long short-term memory neural network, which can accurately predict the power generation in the future. The solar power generation provides a basis for the energy management of the parking lot, and the target detection model is used to realize the recognition and positioning of the vehicle. The deep simple online real-time tracking algorithm is used to track the vehicle to generate the vehicle trajectory. Combined with the parking space coordinate information, the parking space occupancy data is automatically generated to realize the automatic recording of the parking status of the vehicle. In summary, this embodiment improves the intelligent management level of the parking lot, realizes a significant improvement in the efficiency of parking space use, and alleviates the problem of parking difficulties. At the same time, through the accurate prediction of vehicle charging needs and the reasonable allocation of charging resources, it promotes the promotion and application of new energy vehicles, which is of great significance and value for the construction of smart cities and the development of smart transportation.

[0071] In an optional embodiment,

[0072] The following formula is used to update node features and aggregate neighbor node features through graph convolutional neural network:

[0073] ;

[0074] in, Indicates l + 1 Tier i The feature representation of a node, s represents the activation function, N ( i ) represents a node i The set of neighbor nodes of a represents the attention vector, T represents transpose, W Indicates l The weight matrix of the layer, Indicates l Tier i The feature representation of a node, Indicates l Tier j The feature representation of a node.

[0075] In this embodiment, by aggregating the features of neighbor nodes, it is possible to fully utilize the connection relationship between nodes, capture important information in the graph structure, and enable the node feature representation to contain more contextual information. The features of neighbor nodes are aggregated layer by layer to obtain a higher-level and more abstract node feature representation, which is helpful to explore high-order interaction patterns between nodes and improve the representation ability of node features. The importance of different neighbor nodes is distinguished by attention weights, so that important neighbor nodes occupy a larger weight in feature aggregation, thereby improving the effectiveness and representation ability of aggregated features. In summary, this embodiment fully utilizes the structural information in the vehicle information knowledge graph, adaptively learns the feature representation of vehicle nodes, and improves the accuracy and generalization ability of charging demand prediction, which is of great significance for realizing intelligent parking lot scheduling and improving energy utilization efficiency.

[0076] S2. Obtain the unique identification number of the vehicle entering the parking lot through the information collection device, obtain the real-time location coordinates and battery capacity, and obtain the vehicle attribute information corresponding to the current vehicle in combination with the vehicle information knowledge graph. According to the vehicle attribute information and the remaining resource information of the parking lot, calculate the fitness score and predict the optimal parking area of ​​the current vehicle through the collaborative filtering recommendation algorithm based on graph network embedding, generate a parking area recommendation list, and solve the optimal parking path and parking space selection strategy through the multi-objective reinforcement learning algorithm with the parking area fitness and parking time as the reward function, park the vehicle and update the remaining resource information of the parking lot. The data center solves the vehicle charging scheduling plan based on the vehicle battery information with the goal of maximizing charging efficiency and minimizing waiting time;

[0077] The unique identification number is a string or digital sequence used to uniquely identify a specific object. In vehicle management, each vehicle is usually assigned a unique identification number for easy tracking and management. The vehicle attribute information includes various characteristics of the vehicle, such as brand, model, color, year, license plate number, vehicle type, etc. The collaborative filtering recommendation algorithm based on graph network embedding uses graph neural network to capture the relationship between users and items, and improves the effect of the recommendation system by learning the embedded representation of the graph. The fitness score is an indicator to measure the degree of match between a certain plan or choice and specific needs or preferences. The parking area recommendation list is a parking space recommendation generated based on user needs and parking lot availability. The multi-objective reinforcement learning algorithm optimizes multiple objectives simultaneously, usually by setting different reward functions to balance the learning of multiple tasks.

[0078] In an optional embodiment,

[0079] The unique identification number of the vehicle entering the parking lot is obtained through the information collection device, and the real-time location coordinates and battery capacity are obtained. The vehicle attribute information corresponding to the current vehicle is obtained in combination with the vehicle information knowledge graph. According to the vehicle attribute information and the remaining resource information of the parking lot, the fitness score is calculated and the optimal parking area of ​​the current vehicle is predicted through the collaborative filtering recommendation algorithm based on graph network embedding, and a parking area recommendation list is generated. The optimal parking path and parking space selection strategy are solved by a multi-objective reinforcement learning algorithm with the parking area fitness and parking time as the reward function, the vehicle is parked and the remaining resource information of the parking lot is updated. The data center solves the vehicle charging scheduling scheme based on the vehicle battery information with the goal of maximizing charging efficiency and minimizing waiting time, including:

[0080] The unique identification number corresponding to the vehicle entering the parking lot is obtained through a pre-set information collection device, and the vehicle model parameters and battery capacity are obtained in combination with the vehicle model library in the network. At the same time, the real-time position coordinates of the current vehicle are obtained through the panoramic camera. Based on the vehicle information knowledge graph, the vehicle attribute information corresponding to the current vehicle is queried, and the vehicle attribute information and associated entities corresponding to the current vehicle are obtained according to pattern matching;

[0081] Based on the vehicle attribute information and parking lot remaining resource information, combined with the network embedding representation learning algorithm, the vehicle information knowledge graph is used as a heterogeneous information network, a heterogeneous node set and a semantic relationship edge set are defined, and the link topology and node attribute information of the heterogeneous information network are learned to obtain the comprehensive feature representation corresponding to the current vehicle. Based on the comprehensive feature representation, the low-dimensional dense vector of each vehicle node is learned through the graph convolution network model, and the latent vector matrix is ​​generated and the node embedding matrix is ​​obtained by stacking multiple layers of graph convolution. The aggregation weights of the neighbor nodes corresponding to the current vehicle node are dynamically adjusted in combination with the attention mechanism, and the neighbor nodes are aggregated in combination with the attribute feature projection matrix. According to the aggregation results, irrelevant vehicle-vehicle edges are judged and shielded. Through the multi-task learning algorithm, as well as node classification and relationship prediction as the goal, the graph convolution network model is trained in combination with the cross entropy loss function to obtain a trained graph convolution network;

[0082] For the current vehicle, the vehicle similarity is calculated by cosine similarity, and the vehicle with the highest similarity to the current vehicle in the parking lot is selected as the comparison vehicle according to the vehicle similarity, and the vehicle-parking space similarity between the parking spaces corresponding to the current vehicle and the comparison vehicle is calculated. The regional scoring model is constructed by combining the matrix decomposition technology. For each area, the scoring vector of each parking space is calculated in combination with the current vehicle preference. According to the scoring vector, the parking spaces are sorted in descending order and the parking area recommendation list is generated;

[0083] According to the pre-acquired three-dimensional electronic map of the parking lot, the current parking lot road is modeled to obtain an undirected weighted graph. According to the reinforcement learning algorithm, the state space is constructed in combination with the vehicle position, road occupancy and target parking space conditions. The vehicle driving is used as an action and the action space is constructed. According to the parking time and parking difficulty, a reward function is set and a path search is performed to obtain the optimal parking path and parking space selection strategy. The vehicle is parked and the remaining resource information of the parking lot is updated. According to the vehicle battery information, the data center solves the vehicle charging scheduling plan with the goal of maximizing the charging efficiency and minimizing the waiting time.

[0084] The remaining parking resource information refers to the number and type of parking spaces currently available, including the number of vacant parking spaces, reservation status, and availability status in different parking areas. The network embedding representation learning algorithm aims to map nodes in the network into a low-dimensional space in order to capture the relationship and structural information between nodes. The link topology describes the connection mode between nodes in the network, including the relationship structure between nodes and edges. The node attribute information is the characteristic data about each node in the network, such as user personal information, item attributes, vehicle model, etc. The low-dimensional dense vector is a vector representation obtained by the embedding learning algorithm, which is usually used for the simplified representation of high-dimensional data and can retain the main features of the original data to facilitate subsequent analysis and calculation. The latent vector matrix is ​​a matrix representing node features, which is usually used in recommendation systems or graph embedding models. Each row in the matrix represents the implicit features of a node. The node embedding matrix is ​​a matrix generated by the network embedding algorithm. Each row corresponds to the representation of a node in a low-dimensional space, which is convenient for similarity calculation and downstream tasks. The aggregation weight is used to perform weighted aggregation on the information of neighbor nodes in the graph neural network, and adjust according to the importance or relevance of the neighbor nodes. The weight of information aggregation, the attribute feature projection matrix is ​​used to map high-dimensional attribute features to low-dimensional space, which is convenient for data dimensionality reduction and feature extraction, helps to reduce computational complexity and improve model performance, the vehicle-vehicle edge represents the connection relationship between vehicles, and is usually used to build an interaction map between vehicles, which is convenient for analyzing vehicle behavior, similarity, etc. The cross entropy loss function is a common loss function for classification tasks, which measures the difference between the predicted distribution and the true distribution. The smaller the value, the more accurate the model's prediction. The vehicle-parking space similarity is an indicator of the degree of match between a specific vehicle and an available parking space. The marker is usually calculated based on vehicle attributes and parking space characteristics to optimize parking recommendations. The matrix decomposition technology is used to decompose a large matrix into a product of low-rank matrices, which is often used in recommendation systems to capture the potential relationship between users and items through latent factor models. The rating vector is a user's rating representation of different items (such as parking spaces, vehicles, etc.), which is usually used in recommendation systems to help predict user preferences for unrated items. The undirected weighted graph is a graph structure in which the edges have no direction and each edge has a weight, which represents the strength of the association between nodes. It is often used in social network and transportation network analysis.

[0085] Vehicle information entering the parking lot is obtained through pre-set information collection devices (such as license plate recognition cameras, RFID readers, etc.). Static attributes such as license plate numbers, vehicle models, and unique identification numbers are acquired. Combining with the vehicle model library in the network, corresponding vehicle model parameters and battery capacities are obtained based on the recognized vehicle model information. The vehicle is tracked in real-time through a panoramic camera to obtain the position coordinates of the current vehicle within the parking lot. Based on the pre-constructed vehicle information knowledge graph, according to information such as the license plate number, the attribute information and associated entities (such as vehicle owners, historical parking records, etc.) of the current vehicle are queried through pattern matching. Exemplarily, a vehicle of brand A with a license plate number of "Jing A12345" entering the parking lot is recognized by a license plate recognition camera. Its vehicle model parameters are a length of 4.69 meters, a width of 1.85 meters, and a battery capacity of 75 kWh. Through querying the knowledge graph, it is learned that the vehicle owner is Zhang San and the historical average parking duration is 2 hours;

[0086] Regarding the vehicle information knowledge graph as a heterogeneous information network, define the node types (such as vehicles, vehicle owners, vehicle models, etc.) and the semantic relationship types of the edges (such as "belongs to", "is parked at", etc.) within it. Through network embedding representation learning algorithms (such as metapath2vec, etc.), learn the link topology structure and node attribute information of the heterogeneous information network to obtain the comprehensive feature representation of each vehicle node, reflecting the semantic information of the vehicle in the knowledge graph. Use the graph convolutional network model, with the comprehensive features of the vehicle nodes as the input, and through multi-layer graph convolutional operations, learn the low-dimensional dense vector representation of the vehicle nodes, reflecting the mutual influence between vehicles. When aggregating the features of neighbor nodes in the graph convolutional network, introduce an attention mechanism to dynamically adjust the aggregation weights of neighbor nodes according to the correlation between vehicle nodes, and incorporate the attribute information into the node representation through an attribute feature projection matrix. When training the graph convolutional network model, use node classification (such as vehicle model classification) and relationship prediction (such as parking relationship prediction) as supervision signals, and combine with the cross-entropy loss function for end-to-end training, so that the learned node representation can reflect the comprehensive features of the vehicle. Exemplarily, through metapath2vec, learn the vehicle information knowledge graph to obtain a 128-dimensional comprehensive feature vector for each vehicle node. Then use a 2-layer graph convolutional network model for representation learning of the vehicle nodes, with the output dimension of each layer being 64 dimensions. During training, use vehicle model classification and parking relationship prediction as tasks, the loss function is cross-entropy, and introduce an attention mechanism and attribute features. Finally, obtain a 64-dimensional embedding vector representation for each vehicle node;

[0087] For the current vehicle, calculate its similarity with other vehicles in the parking lot using cosine similarity, select the vehicle with the highest similarity as a reference, calculate the similarity between the current vehicle and the parking space of the reference vehicle as the vehicle-space similarity, and use matrix factorization technology to construct a parking area scoring model. For each parking area, combine the preferences of the current vehicle (such as distance, parking space size, etc.) to calculate the scoring vector for each parking space, sort the parking spaces in descending order according to the scoring vector, and generate a recommended parking area list for the current vehicle. Exemplarily, for the vehicle with license plate number "Jing A12345", through similarity calculation, find the vehicle "Jing B67890" that is most similar to it, which is parked in area A. Calculate the similarity between "Jing A12345" and each parking space in area A, and combine the attribute of this vehicle preferring large parking spaces to score and sort the parking spaces in area A, generating a recommended list of parking spaces, such as [a-001, a-012, a-003];

[0088] Based on the 3D electronic map of the parking lot, model the parking lot roads as an undirected weighted graph, where the nodes represent road intersections, the edges represent road segments, and the weights can be travel time, etc. Use a reinforcement learning algorithm (such as Q-learning) for path planning. The state space includes information such as the current vehicle position, road occupancy, and target parking space, and the action space is the vehicle driving direction; the reward function is set according to parking time and difficulty, etc. Through Q-learning exploration and learning, obtain the optimal driving path from the current vehicle position to the recommended parking space, as well as the parking space selection strategy, guide the vehicle to drive and park according to the planned path, and update the parking space occupancy information in the parking lot. Exemplarily, for the vehicle "Jing A12345", its current position is at the parking lot entrance, and the target parking space is a-012. Through Q-learning exploration on the parking lot road map, the optimal path is obtained as: entrance → South Road → East Road → Second Column in Area A → a-012. Guide the vehicle to drive and park according to this path, and mark a-012 as occupied;

[0089] The data center constructs a charging scheduling model with the goal of maximizing charging efficiency and minimizing vehicle waiting time based on information such as the battery capacity, current power, and expected parking duration of the vehicle, solves the charging scheduling model, obtains the optimal charging time and charging duration for each vehicle, generates a vehicle charging scheduling plan, controls the charging pile to charge the vehicle according to the scheduling plan, and monitors the charging process in real time to update the vehicle power information. Exemplarily, there are 3 charging piles in the parking lot, and currently 5 electric vehicles are parked. The data center solves the charging scheduling model according to the vehicle information and obtains the optimal scheduling plan as: Vehicle A charges first for 1 hour, then Vehicles B and C charge simultaneously for 2 hours, Vehicle D charges last for 30 minutes, and Vehicle E does not need to charge. Control the charging piles to charge the vehicles in sequence according to this plan;

[0090] In this embodiment, a vehicle information collection device and a panoramic camera are used to realize a comprehensive perception of the static attributes and dynamic positions of vehicles entering the parking lot. Heterogeneous information network embedding and graph convolution network technology are introduced to learn the comprehensive feature representation of the vehicle from the vehicle information knowledge graph, which takes into account both the vehicle's own attributes and the correlation between vehicles, making the vehicle feature representation more accurate and comprehensive, and providing a basis for personalized services. Through the parking area scoring model based on vehicle similarity calculation and matrix decomposition, personalized parking space recommendations are realized in combination with vehicle preferences, which improves the parking experience and satisfaction of car owners, reduces the parking space search time and traffic congestion in the parking lot, and based on the reinforcement learning algorithm and the parking lot electronic map, the vehicle's current position is realized. The optimal path planning and parking space selection strategy for the target parking space are implemented, which reduces the blind driving of vehicles and invalid parking space searches in the parking lot, improves the parking efficiency and the traffic capacity of the parking lot. By constructing a charging scheduling model and considering factors such as charging efficiency and vehicle waiting time, the intelligent charging scheduling of electric vehicles is realized, the charging waiting time is reduced, the utilization rate of charging facilities is improved, and the promotion and application of new energy vehicles are promoted. In summary, this embodiment is based on solving practical problems such as difficulty in parking and difficulty in charging new energy vehicles. It realizes the intelligent upgrading and transformation of parking lots from multiple dimensions, and has achieved significant efficiency improvement and experience optimization effects. It is of great significance to alleviate the "difficulty in parking", promote the promotion of new energy vehicles, and build smart cities, and has outstanding technical effects.

[0091] In an optional embodiment,

[0092] The loss function corresponding to path search based on reinforcement learning is shown in the following formula:

[0093] ;

[0094] in, L ( i ) represents the loss value, Indicates in status s Take action a Rewards received after r and the next state s' The expected value of r Indicates the current time step t Rewards received, c represents the discount factor, Indicates that in the next state s' Next, all actions a' The maximum Q value, Indicates the current status s and actions a The Q value of the network is i Calculated.

[0095] In this embodiment, by continuously interacting and learning with the environment, the intelligent agent can gradually optimize its decision-making strategy and find a more efficient and reasonable driving route. The reinforcement learning algorithm can search for the global optimal solution in the state space by balancing exploration and utilization. Reinforcement learning can plan the global optimal parking path through the accumulation and discount of long-term rewards, avoiding short-sighted behavior. The Q network obtained through training has a strong generalization ability, can adapt to different parking lot environments and dynamic changes, can make reasonable decisions based on the learned strategies, and has good migration ability and robustness. Path planning based on reinforcement learning can realize real-time online decision-making and fast response speed, meeting the needs of real-time scheduling of intelligent parking lots. In summary, this embodiment has a strong generalization ability and real-time performance, can significantly improve the intelligence level of parking lot vehicle navigation, shorten parking time, and improve parking efficiency, and provides new ideas and solutions for path planning problems in complex dynamic environments.

[0096] S3. Obtain the short-term power generation corresponding to the solar panel according to the multivariable power generation prediction model, solve the charging power scheduling curve in combination with the vehicle charging scheduling plan, and dynamically update the multivariable power generation prediction model through incremental learning. If the charging power scheduling curve deviates and the deviation is greater than the preset deviation threshold, update the vehicle charging scheduling plan through a rolling optimization strategy, and charge according to the updated charging scheduling plan and fuzzy logic control algorithm. During the charging process, obtain the battery status parameters in real time and filter out the noise through the Kalman filter algorithm. Combine the adaptive neural fuzzy reasoning algorithm and the solar power generation status parameters to dynamically adjust the charging strategy until charging is completed. Generate a charging report according to the charging process and update the vehicle information knowledge graph.

[0097] The short-term power generation refers to the power output expected to be generated in a shorter period of time (such as hours, several hours or days), which is usually estimated based on meteorological data, historical power generation data and demand forecasts. The charging power dispatch curve represents the charging power distribution of charging facilities to electric vehicles within a certain time range. The incremental learning is a machine learning method that allows the model to gradually learn new data without retraining the entire model. The deviation threshold refers to the allowable error range or tolerance in a control system or optimization model. The rolling optimization strategy refers to continuously updating and optimizing decisions within a certain time period to respond to environmental changes or new information. The fuzzy logic control algorithm is a control method based on fuzzy set theory that can handle uncertainty and ambiguity. The Kalman filter algorithm is a recursive estimation method used to extract system status from measurement data with noise. The adaptive neuro-fuzzy inference algorithm combines the advantages of neural networks and fuzzy logic, and can optimize the decision-making process by learning and adjusting fuzzy rules.

[0098] In an optional embodiment,

[0099] According to the multivariable power generation prediction model, the short-term power generation corresponding to the solar panel is obtained, and the charging power scheduling curve is solved in combination with the vehicle charging scheduling plan, and the multivariable power generation prediction model is dynamically updated through incremental learning. If the charging power scheduling curve deviates and the deviation is greater than the preset deviation threshold, the vehicle charging scheduling plan is updated through a rolling optimization strategy, and charging is performed according to the updated charging scheduling plan and the fuzzy logic control algorithm. During the charging process, the battery state parameters are obtained in real time and the noise is filtered through the Kalman filter algorithm. The charging strategy is dynamically adjusted in combination with the adaptive neural fuzzy reasoning algorithm and the solar power generation state parameters until charging is completed. A charging report is generated according to the charging process and the vehicle information knowledge graph is updated, including:

[0100] According to the preset multivariable power generation prediction model, the short-term power generation corresponding to the solar panel is obtained, and the standard charging power scheduling curve is calculated in combination with the vehicle charging scheduling plan. The incremental samples are added to the multivariable power generation prediction model through the incremental learning algorithm, and the gradient of the loss function of each layer is calculated. The parameters are updated in combination with the stochastic gradient descent algorithm to determine the influence weight of the incremental sample on the original loss function. If the influence weight is greater than the preset influence threshold, it is considered that the incremental sample has information gain, and it is added to the training set and incremental learning is performed. The updated prediction mean and variance are determined in combination with the Bayesian theory to obtain the updated multivariable power generation prediction model. According to the updated multivariable power generation prediction model, the actual charging power scheduling curve is obtained by solving the vehicle charging constraints and the power grid constraints.

[0101] The actual charging power dispatch curve is compared with the standard charging power dispatch curve, and the deviation value is calculated and compared with the preset deviation threshold. If the deviation value is greater than the deviation threshold, a hierarchical collaborative optimization model is set through a rolling optimization strategy, wherein the hierarchical collaborative optimization model includes a day-ahead optimization layer, an intraday correction layer, and a real-time control layer. According to the hierarchical collaborative optimization model, the goal is to maximize the solar energy absorption and minimize the charging cost and the peak-to-valley difference, and the battery physical constraints and the grid technical constraints are used as constraints. The vehicle entry time and departure time and the vehicle battery capacity are initialized according to the vehicle information knowledge graph, and the vehicle charging scheduling plan is updated according to the electricity price information and the photovoltaic output information in combination with the hardware utility function;

[0102] Charging is performed based on the updated vehicle charging scheduling plan, and battery status parameters during the charging process are obtained in real time. Noise is filtered out through the Kalman filtering algorithm, and a fuzzy control rule base is constructed in combination with the fuzzy logic control algorithm. The clear value of the control quantity is determined according to the control rules and the center of gravity defuzzification method, and the optimal charging power is obtained by inferring the clear value. At the same time, according to the uncertainty of battery characteristics and photovoltaic fluctuations, the rules in the fuzzy control rule base are updated through an adaptive neural fuzzy inference algorithm, and the update is repeated until convergence. The updated vehicle charging scheduling plan is dynamically adjusted according to the updated rules until charging is completed. After charging is completed, a charging report is generated according to the charging process and the vehicle information knowledge graph is updated.

[0103] The incremental samples refer to new data samples gradually introduced in machine learning or data analysis. The loss function gradient refers to the derivative of the loss function with respect to the model parameters, which represents the rate of change of the loss function in the parameter space. The stochastic gradient descent algorithm is an optimization algorithm used to minimize the loss function. The Bayesian theory is a probabilistic inference method that updates the probability of events through prior knowledge and observed data. The hierarchical collaborative optimization model divides complex systems into multiple levels for optimization. Each level makes decisions based on specific goals or constraints to improve overall system efficiency and flexibility. It is often used in energy management and resource scheduling. The day-ahead optimization layer refers to the optimization plan made in advance in power scheduling based on predicted load and renewable energy generation in order to effectively allocate resources and meet future power demand. The intraday correction layer refers to the adjustment of the day-ahead optimization results based on real-time data in power scheduling to cope with actual load fluctuations and the uncertainty of renewable energy output, and to ensure the stability of the power system. The real-time control layer Responsible for dynamically adjusting the operating status of the power system according to real-time monitoring data, ensuring the balance of supply and demand and optimizing resource allocation to cope with emergencies and real-time changes. The maximization of solar energy absorption refers to optimizing the scheduling and operation of photovoltaic power generation systems to maximize the use of solar energy resources, reduce dependence on traditional energy, and improve the utilization rate of renewable energy. The battery physical constraints refer to the technical limitations and performance indicators of the battery, such as charging and discharging rates, energy capacity and cycle life. The grid technical constraints include the physical characteristics and operating limitations of the grid, such as transmission capacity, frequency stability and load balance. The fuzzy control rule base is a set of control rules based on fuzzy logic for dealing with uncertainty and ambiguity. The centroid defuzzification method is a commonly used defuzzification technique in fuzzy logic control, which aims to convert fuzzy outputs into certain values, usually by calculating the centroid of the fuzzy set. The uncertainty of photovoltaic fluctuations refers to the volatility of photovoltaic power generation output due to changes in meteorological conditions (such as clouds, sunshine intensity, etc.).

[0104] According to the pre-set multivariable power generation prediction model, the short-term power generation forecast value of the solar panel is obtained. Combined with the vehicle charging scheduling plan, the standard charging power scheduling curve is calculated as the optimization target. Through the incremental learning algorithm, the newly acquired actual power generation data is added to the prediction model as an incremental sample, the loss function gradient of each layer of the neural network is calculated, and the parameters are updated using the stochastic gradient descent algorithm. By calculating the influence weight of the incremental sample on the original loss function, it is determined whether it has significant information gain. If the influence weight exceeds the preset threshold, the incremental sample is added to the training set for incremental learning. The Bayesian theory is used to update the mean and variance of the prediction model according to the newly added samples to obtain an updated multivariable power generation prediction model. For example, it is assumed that the original power generation prediction model is trained based on the meteorological and power generation data of the past year. When a new day comes, the actual meteorological parameters and power generation data of that day are obtained. These new data are taken as incremental samples, and their influence weight on the model loss function is calculated to be 0.2, which exceeds the preset threshold of 0.1. Therefore, these incremental samples are added to the training set, the model parameters are updated through the incremental learning algorithm, and the prediction mean and variance are adjusted using the Bayesian theory to obtain an updated power generation prediction model;

[0105] According to the updated multivariable power generation prediction model, the solar power generation in the future is re-predicted. Combined with the vehicle charging constraints (such as battery capacity, charging time window, etc.) and the grid operation constraints (such as load curve, electricity price, etc.), a charging power optimization model is constructed. The optimization model is solved to obtain the optimized actual charging power dispatch curve, so that the solar energy absorption is maximized, while taking into account the charging cost and load balance. The actual charging power dispatch curve is compared with the standard curve, and the deviation value is calculated. If the deviation value exceeds the preset deviation threshold, the rolling optimization strategy is started to re-optimize the charging dispatch plan. The rolling optimization adopts a hierarchical collaborative optimization model, including a day-ahead optimization layer, an intraday correction layer, and a real-time control layer. The goal is to maximize the solar energy absorption and minimize the charging cost and peak-to-valley difference. The battery physical constraints and grid technical constraints are used as constraints. The vehicle information knowledge graph is used to initialize the vehicle parameters. According to the information such as electricity price and photovoltaic output, the charging scheduling plan is iteratively updated through the hardware utility function. For example, assuming that the standard charging power scheduling curve of a certain day is [10kW, 15kW, 20kW, 15kW, 10kW], according to the updated power generation prediction model, the actual scheduling curve is optimized to be [12kW, 17kW, 18kW, 13kW, 8kW]. The deviation value of the two curves is 5kW, which exceeds the preset threshold of 3kW. The rolling optimization is started, and a three-layer optimization model is adopted: the day-ahead layer roughly formulates the scheduling plan according to the forecast information of the next day, the intraday layer makes corrections based on the day-ahead plan according to the actual situation of the day, and the real-time layer performs real-time control according to the latest data every 15 minutes. The optimized scheduling curve is [11kW, 16kW, 19kW, 14kW, 9kW], which not only improves the utilization rate of solar energy, but also balances the charging cost and grid load.

[0106] According to the optimized vehicle charging scheduling plan, the vehicle is charged. During the charging process, the battery status parameters, such as voltage, current, temperature, etc., are obtained in real time. The Kalman filter algorithm is used to filter the noise of the collected status parameters to obtain a smooth state estimate. Combined with the fuzzy logic control algorithm, a fuzzy control rule base is constructed. According to factors such as battery status and charging stage, the corresponding charging control strategy is given. The fuzzy control rules are converted into clear control quantity values, such as charging current, charging voltage, etc., through the center of gravity defuzzification method. According to the clear value of the control quantity, the optimal charging power under the current state is inferred. Taking into account the uncertainty of battery characteristics and photovoltaic output, the control rule base is learned and adjusted online using an adaptive neural fuzzy inference system. According to changes in battery status and environmental conditions, the membership function and rule weights in the rule base are adaptively adjusted to enable the control strategy to adapt dynamically. The above process is repeated until the adaptive learning converges to obtain an optimized fuzzy control rule. The optimized rule base is used to fine-tune the charging scheduling plan and control the charging process in real time until the charging is completed. For example, suppose an electric car is charged and the real-time state parameters of the battery are obtained as follows: voltage 3.6V, current 2A, temperature 30℃. The smoothed estimated values ​​obtained by Kalman filtering are: voltage 3.7V, current 1.9A, temperature 29℃. According to the preset fuzzy control rule, if "voltage is low, current is low, and temperature is moderate", then "charging power is increased". Through the center of gravity defuzzification, the control quantity is "charging power increased by 20%", that is, the original charging power of 10kW is adjusted to 12kW. During the charging process, if the battery voltage suddenly increases, the adaptive learning algorithm captures this change and dynamically adjusts the rule base, which may generate a new rule: "voltage is high, current is low, and temperature is moderate", then "charging power is reduced", so that the control strategy can adapt to the new situation and guide the entire charging process according to the adjusted rule base until the charging is completed safely and efficiently.

[0107] After the charging is completed, a complete report of this charging is generated, including indicators such as charging duration, charging power, average power, and efficiency. The charging report data is added to the vehicle information knowledge graph and associated with the vehicle's own attribute information, historical charging records, etc. Through data association and fusion, the knowledge graph is continuously enriched and improved to enable it to more comprehensively and accurately reflect the vehicle's state and charging characteristics. Using the updated knowledge graph, data support and decision-making basis can be provided for subsequent charging scheduling, fault diagnosis, life prediction, etc. Exemplarily, assume that an electric vehicle with a license plate number of "Jing A12345" completes a charging. The charging duration is 2 hours, the charging power is 40 kWh, the average power is 20 kW, and the charging efficiency is 95%. Adding these data to the knowledge graph and associating them with the vehicle's previous information, it can be seen that its historical average charging duration is 2.5 hours and the average power is 50 kWh. Both the charging duration and power have decreased this time. Combining other data such as the vehicle's operating mileage, it can be determined that the battery health status of this vehicle is good and the charging efficiency this time is relatively high.

[0108] In this embodiment, the power generation prediction model is dynamically updated through an incremental learning algorithm to continuously improve the prediction accuracy. Combined with hierarchical collaborative optimization for charging power scheduling, the photovoltaic power generation is maximally utilized, the curtailment of photovoltaic power is reduced, and the photovoltaic power consumption level is improved to achieve the efficient utilization of new energy. Taking minimizing the charging cost as one of the goals, factors such as electricity price and load distribution are comprehensively considered, and the charging load is transferred to the low electricity price period as much as possible to reduce the peak-valley difference, optimize the grid load distribution, and reduce the user's charging cost. In the intelligent charging control link, an adaptive neuro-fuzzy inference system is adopted to dynamically adjust the charging strategy according to the real-time state of the battery and the charging stage, avoid overcharging and over-discharging, reduce battery loss, extend the battery life, and reduce the frequency and cost of battery replacement. Through the continuous update and improvement of the vehicle information knowledge graph, integrating a large amount of vehicle operation and charging data, important data support and decision-making basis are provided for energy management, grid scheduling, charging infrastructure planning, etc., promoting the intelligent development of the energy system. In summary, this embodiment realizes intelligent charging management under multiple goals and has achieved remarkable technical effects in aspects such as the efficient utilization of new energy, grid optimal scheduling, and user experience improvement, providing a beneficial exploration and practice for building a smart energy management system.

[0109] In an alternative embodiment,

[0110] The hierarchical collaborative optimization model updates the vehicle charging scheduling scheme as follows:

[0111] The day-ahead optimization layer obtains the prediction information and combines it with the vehicle information knowledge graph to obtain the power dispatch plan of the current vehicle in the next 24 hours. Based on the dispatch plan, the intraday correction layer obtains the predicted power generation power in the next 4 hours in a rolling manner, and corrects the charging power in combination with the current vehicle data in the parking lot. The field-level charging plan in the next four hours is updated and sent to the charging piles in each parking space. The solar power generation power is tracked on the premise of meeting the charging demand through the rolling time domain optimization strategy.

[0112] The real-time control layer determines the charging power for the current period through fuzzy logic control based on the real-time output of solar power generation, combined with the battery charge state and temperature, and adjusts the vehicle charging scheduling plan based on the generated charging power decision.

[0113] The field-level charging plan refers to a charging scheduling strategy for electric vehicles in a specific charging station. The plan takes into account factors such as the availability of charging piles, charging demand, charging time, and grid load, and aims to optimize the charging process, improve charging efficiency and user experience. The rolling time domain optimization strategy is a dynamic optimization method that continuously updates decisions within a specific time window.

[0114] The day-ahead optimization layer obtains forecast information for the next 24 hours, including solar power generation, grid load, electricity price, etc., extracts the attributes and historical data of all vehicles in the current parking lot from the vehicle information knowledge graph, such as battery capacity, expected stay time, historical charging records, etc., combines the forecast information and vehicle data to build a day-ahead optimization model, and generates a power dispatch plan for the next 24 hours with the goal of maximizing photovoltaic consumption and minimizing charging costs. The dispatch plan includes the total charging power for each time period and the allocated power for each vehicle to meet the vehicle charging needs and grid constraints. For example, assuming that there are 100 electric vehicles in the parking lot, with an average of The battery capacity is 50kWh, and the expected stay time is 8 hours. According to the weather forecast, the photovoltaic power generation in the next 24 hours will fluctuate between 20-80kW, the grid load will vary between 100-500kW, and the time-of-use electricity price will be 0.5 yuan / kWh and 0.8 yuan / kWh. The optimization layer takes these factors into consideration and generates the following scheduling plan: During the peak period of photovoltaic power generation from 10:00 to 14:00, 50 vehicles will be arranged to charge with an average power of 20kW; during the low electricity price period from 23:00 to 6:00, 80 vehicles will be arranged to charge with an average power of 30kW; the rest of the time will be scheduled on demand, and the total power will be controlled within 50kW;

[0115] Based on the day-ahead optimization, the intraday correction layer obtains more accurate photovoltaic power generation forecast values ​​for the next 4 hours on a rolling basis, obtains the current actual vehicle data in the parking lot, including information on newly added vehicles and vehicles leaving early, and corrects and optimizes the charging plan for the next 4 hours in combination with the updated forecast values ​​and real-time vehicle data, generates a revised field-level charging plan, and sends it to each charging pile to guide it to track photovoltaic power generation while meeting vehicle demand. The above optimization process is executed on a rolling basis every hour to realize dynamic updating of the charging plan. By way of example, based on the day-ahead plan, the intraday correction layer obtains short-term photovoltaic forecast values ​​of 40kW, 60kW, 50kW, and 30kW for the next 4 hours. At the same time, 5 new vehicles are added to the parking lot and 3 vehicles leave early. Based on this, the charging plan for 10:00-14:00 was revised: the new vehicles were added to the charging, and the average power was adjusted to 25kW; the power was reduced to 50kW during the 12:00-13:00 period to adapt to the downward trend of photovoltaic power generation, and the other periods remained unchanged. The revised plan was sent to each charging pile to guide them to control the charging power to about 60kW during the 11:00-12:00 period, and to about 50kW during the 12:00-13:00 period. The original plan was maintained for the rest of the period. The correction and optimization process was repeated every hour to respond to forecast changes and vehicle updates in a timely manner.

[0116] During the charging process, the real-time control layer obtains the actual output data of solar power generation in real time, collects the real-time status of each vehicle battery through on-board sensors or charging piles, such as the state of charge (SoC), temperature, etc., inputs the power generation data and battery data into the fuzzy logic controller, and makes inference decisions according to the preset fuzzy rule base. The fuzzy rule base contains a series of If-Then rules based on parameters such as power generation, SoC, temperature, etc., which are used to determine the current suitable charging power. The fuzzy controller outputs a clear power value as the actual charging power of each vehicle in the current period. According to the decision result of the fuzzy controller, the original vehicle charging scheduling plan is fine-tuned to ensure real-time optimization. The above real-time control process is repeated every 5-15 minutes to realize dynamic tracking and adjustment of charging power. For example, at a certain moment, the real-time output of solar power generation is 55kW, the SoC of a vehicle battery is 60%, and the temperature is 30°C. According to the fuzzy rule "IF high power generation AND medium SoC AND moderate temperature THEN The fuzzy controller gives the current charging power of the car as 25kW. If the original scheduling plan of the car is 20kW, it will be adjusted to 25kW according to the real-time decision. The same is true for other vehicles. According to their respective battery status and shared power generation, the fuzzy controller gives their respective real-time charging power, dynamically optimizes the charging plan of the entire parking lot, and repeats it every 10 minutes to always keep the real-time matching of charging power and photovoltaic output.

[0117] In this embodiment, a preliminary dispatching plan is generated based on 24-hour forecast information through the day-ahead optimization layer, the charging plan is optimized on a rolling basis through the intraday correction layer, and the charging power is dynamically adjusted according to the actual power generation output through the real-time control layer. Multi-level collaborative optimization maximizes the use of photovoltaic power for electric vehicle charging, improves the photovoltaic consumption level, and realizes the efficient use of new energy. Time-of-use electricity price factors are comprehensively considered during dispatching optimization, and the charging load is transferred to the low-price period as much as possible, reducing the charging cost of users. By optimizing the charging power distribution, the grid load fluctuation is also reduced, playing a role in peak shaving and valley filling. Through intelligent scheduling and dynamic optimization, photovoltaic power is maximized while meeting the charging demand, reducing the dependence on grid power and improving the charging reliability. Through the construction of vehicle information knowledge graph, data integration and sharing are realized. Through the layered collaborative optimization model, the integrated scheduling of energy flow, information flow, and control flow is realized, which promotes the coordinated optimization of the energy system in the parking lot and improves the comprehensive energy management level.

[0118] Figure 2 FIG. 1 is a schematic diagram of a smart parking charging pile positioning system combined with solar energy according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0119] The first unit is used to obtain the three-dimensional spatial information of the parking lot through pre-deployed sensors, combine the image data obtained by the panoramic camera, draw a three-dimensional electronic map of the parking lot through a synchronous positioning and map construction algorithm, mark the coordinates of the parking spaces, and sense the changes in occupancy status through magnetic sensors pre-installed on the parking spaces. According to the environmental information collected by the environmental sensors, combined with historical environmental parameters and historical power generation data of solar panels, a multivariate power generation prediction model is constructed in combination with a long short-term memory network. According to the traffic video data, parked vehicles are identified through target detection and tracking algorithms, parking space occupancy data is obtained and uploaded to the edge computing server, a structured vehicle information knowledge graph is constructed based on vehicle attribute information and vehicle preference information, the association rules between vehicle information are determined through a graph neural network algorithm, and a charging demand prediction model is constructed to generate a vehicle prediction report;

[0120] The second unit is used to obtain the unique identification number of the vehicle entering the parking lot through the information collection device, obtain the real-time location coordinates and battery capacity, and obtain the vehicle attribute information corresponding to the current vehicle in combination with the vehicle information knowledge graph. According to the vehicle attribute information and the remaining resource information of the parking lot, the fitness score is calculated and the optimal parking area of ​​the current vehicle is predicted through the collaborative filtering recommendation algorithm based on graph network embedding, and a parking area recommendation list is generated. Through the multi-objective reinforcement learning algorithm, the optimal parking path and parking space selection strategy are solved with the parking area fitness and parking time as the reward function, the vehicle is parked and the remaining resource information of the parking lot is updated. The data center solves the vehicle charging scheduling plan based on the vehicle battery information with the goal of maximizing the charging efficiency and minimizing the waiting time;

[0121] The third unit is used to obtain the short-term power generation corresponding to the solar panel according to the multivariable power generation prediction model, solve the charging power scheduling curve in combination with the vehicle charging scheduling plan, and dynamically update the multivariable power generation prediction model through incremental learning. If the charging power scheduling curve deviates and the deviation is greater than a preset deviation threshold, the vehicle charging scheduling plan is updated through a rolling optimization strategy, and charging is performed according to the updated charging scheduling plan and fuzzy logic control algorithm. During the charging process, battery status parameters are obtained in real time and noise is filtered through the Kalman filtering algorithm. The charging strategy is dynamically adjusted in combination with the adaptive neural fuzzy inference algorithm and the solar power generation status parameters until charging is completed, and a charging report is generated according to the charging process and the vehicle information knowledge graph is updated.

[0122] According to a third aspect of the embodiments of the present invention,

[0123] An electronic device is provided, comprising:

[0124] processor;

[0125] a memory for storing processor-executable instructions;

[0126] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0127] A fourth aspect of the embodiments of the present invention is:

[0128] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0129] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart parking charging pile positioning method combined with solar energy, characterized in that: include: The three-dimensional spatial information of the parking lot is obtained through pre-deployed sensors, and the image data obtained by the panoramic camera is combined to draw a three-dimensional electronic map of the parking lot through a synchronous positioning and map construction algorithm. The parking space coordinates are marked and the occupancy status changes are sensed through magnetic sensors pre-installed on the parking spaces. According to the environmental information collected by the environmental sensors, combined with historical environmental parameters and historical power generation data of solar panels, a multivariate power generation prediction model is built in combination with a long short-term memory network. According to the traffic video data, parked vehicles are identified through target detection and tracking algorithms, and parking space occupancy data is obtained and uploaded to the edge computing server. A structured vehicle information knowledge graph is built based on vehicle attribute information and vehicle preference information. The association between vehicle information is determined through a graph neural network algorithm and a charging demand prediction model is built to generate a vehicle prediction report. The unique identification number of the vehicle entering the parking lot is obtained through the information collection device, and the real-time location coordinates and battery capacity are obtained. The vehicle attribute information corresponding to the current vehicle is obtained in combination with the vehicle information knowledge graph. According to the vehicle attribute information and the remaining resource information of the parking lot, the fitness score is calculated and the optimal parking area of ​​the current vehicle is predicted through the collaborative filtering recommendation algorithm based on graph network embedding, and a parking area recommendation list is generated. The optimal parking path and parking space selection strategy are solved by a multi-objective reinforcement learning algorithm with the parking area fitness and parking time as the reward function, the vehicle is parked and the remaining resource information of the parking lot is updated. The data center solves the vehicle charging scheduling plan based on the vehicle battery information with the goal of maximizing charging efficiency and minimizing waiting time; The short-term power generation corresponding to the solar panel is obtained according to the multivariable power generation prediction model, the charging power scheduling curve is solved in combination with the vehicle charging scheduling plan, and the multivariable power generation prediction model is dynamically updated through incremental learning. If the charging power scheduling curve deviates and the deviation is greater than a preset deviation threshold, the vehicle charging scheduling plan is updated through a rolling optimization strategy, and charging is performed according to the updated charging scheduling plan and the fuzzy logic control algorithm. During the charging process, the battery status parameters are obtained in real time and the noise is filtered out through the Kalman filtering algorithm. The charging strategy is dynamically adjusted in combination with the adaptive neural fuzzy inference algorithm and the solar power generation status parameters until charging is completed. A charging report is generated according to the charging process and the vehicle information knowledge graph is updated.

2. The method according to claim 1, characterized in that The three-dimensional spatial information of the parking lot is obtained through pre-deployed sensors, combined with the image data obtained by the panoramic camera, and a three-dimensional electronic map of the parking lot is drawn through a synchronous positioning and map construction algorithm. The parking space coordinates are marked and the occupancy status changes are sensed through magnetic sensors pre-installed on the parking spaces. According to the environmental information collected by the environmental sensors, combined with historical environmental parameters and historical power generation data of solar panels, a multivariate power generation prediction model is built in combination with a long short-term memory network. According to the traffic video data, parked vehicles are identified through target detection and tracking algorithms, parking space occupancy data is obtained and uploaded to the edge computing server, and a structured vehicle information knowledge graph is built based on vehicle attribute information and vehicle preference information. The association between vehicle information is determined through a graph neural network algorithm and a charging demand prediction model is built. The vehicle prediction report generated includes: The parking lot space is fully scanned by a laser radar sensor pre-deployed in the parking lot to obtain the three-dimensional spatial information of the parking lot in the form of a point cloud. Combined with a pre-set panoramic camera, a panoramic color image corresponding to the parking lot is obtained. The three-dimensional spatial information of the parking lot and the panoramic color image are synchronously aligned in time and space. The point cloud data is pre-processed based on a synchronous positioning and map building algorithm, feature points are selected and the initial pose estimation corresponding to the sensor is obtained by tracking the feature points between different frames, a constraint graph between the sensor pose and the feature points is constructed, the pose is optimized by a graph optimization library, and the laser point cloud is inserted into a pre-set octree map to obtain the three-dimensional electronic map of the parking lot. The coordinates of the parking spaces are marked in the three-dimensional electronic map of the parking lot, and the magnetic field strength data in the state of having a car and the state of not having a car are collected. The judgment threshold is solved by a linear discriminant analysis method, and the three-axis magnetic field strength data of the current parking space is obtained in combination with a magnetic sensor pre-installed on the parking space. The change in the occupancy state of the current parking space is judged according to the three-axis magnetic field strength data and the judgment threshold. The environmental parameters at the current moment are collected by a pre-set multi-parameter sensor, wherein the environmental parameters include temperature, humidity, light and wind speed. The historical power generation data of the solar panels in the parking lot are obtained, and a nonlinear mapping relationship between the current environmental parameters and the historical power generation data is constructed by combining a long short-term memory network with the collected environmental parameter time series data as input, and the multivariate power generation prediction model is constructed based on the nonlinear mapping relationship; According to the traffic video data acquired by the panoramic camera, a multi-scale feature map is generated through the feature extraction network in the target detection model, and the target bounding box and category probability are predicted at each position of the multi-scale feature map in combination with the anchor frame mechanism, and the detection result is output. According to the detection result, a deep simple sorting multi-target tracking algorithm is used, and the vehicle bounding box in the detection result is used as an observation. The position and speed of each vehicle target in the next frame are predicted through a Kalman filter, and the vehicle bounding box detected in the current frame is matched with the tracking result of the previous frame by solving the minimum cost matrix to generate a tracking trajectory, and the parked vehicle and parking time are determined based on the spatial relationship between the tracking trajectory and the parking space coordinates, and the parking space occupancy data is obtained by combining and uploading to the edge computing server; For each vehicle entering the current parking lot, the edge node obtains the vehicle attribute information and vehicle preference information corresponding to the current vehicle, wherein the vehicle attribute information includes vehicle category, model, brand, fast charging rate and driver information, and the vehicle preference information includes historical parking location, historical parking duration and charging preference. If the current vehicle enters the current parking lot for the first time, the vehicle preference information is marked as empty. The vehicle attribute information and the vehicle preference information are imported into the database in graph form to generate a structured vehicle information knowledge graph. The charging demand prediction model is constructed through a graph neural network algorithm. The vehicle information knowledge graph is divided into subgraphs centered on a single vehicle and containing multi-hop neighbors. For each subgraph, the node features are updated and the neighbor node features are aggregated through a graph convolutional neural network. The vehicle node feature representation is repeatedly aggregated. The association rules between vehicle information and the influence between vehicle information and charging demand are determined based on the vehicle node feature representation. The hyperparameters of the charging demand prediction model are updated and predicted in combination with the back propagation algorithm to generate the predicted charging time corresponding to the current vehicle. The vehicle prediction report is obtained in combination with the tracking trajectory and vehicle attribute information.

3. The method according to claim 2, characterized in that The following formula is used to update node features and aggregate neighbor node features through graph convolutional neural network: ； in, Indicates l + 1 Tier i The feature representation of a node, σ represents the activation function, N ( i ) represents a node i The set of neighbor nodes of a represents the attention vector, T represents transpose, W Indicates l The weight matrix of the layer, Indicates l Tier i The feature representation of a node, Indicates l Tier j The feature representation of a node.

4. The method according to claim 1, characterized in that The unique identification number of the vehicle entering the parking lot is obtained through the information collection device, and the real-time location coordinates and battery capacity are obtained. The vehicle attribute information corresponding to the current vehicle is obtained in combination with the vehicle information knowledge graph. According to the vehicle attribute information and the remaining resource information of the parking lot, the fitness score is calculated and the optimal parking area of ​​the current vehicle is predicted through the collaborative filtering recommendation algorithm based on graph network embedding, and a parking area recommendation list is generated. The optimal parking path and parking space selection strategy are solved by a multi-objective reinforcement learning algorithm with the parking area fitness and parking time as the reward function, the vehicle is parked and the remaining resource information of the parking lot is updated. The data center solves the vehicle charging scheduling scheme based on the vehicle battery information with the goal of maximizing charging efficiency and minimizing waiting time, including: The unique identification number corresponding to the vehicle entering the parking lot is obtained through a pre-set information collection device, and the vehicle model parameters and battery capacity are obtained in combination with the vehicle model library in the network. At the same time, the real-time position coordinates of the current vehicle are obtained through the panoramic camera. Based on the vehicle information knowledge graph, the vehicle attribute information corresponding to the current vehicle is queried, and the vehicle attribute information and associated entities corresponding to the current vehicle are obtained according to pattern matching; Based on the vehicle attribute information and parking lot remaining resource information, combined with the network embedding representation learning algorithm, the vehicle information knowledge graph is used as a heterogeneous information network, a heterogeneous node set and a semantic relationship edge set are defined, and the link topology and node attribute information of the heterogeneous information network are learned to obtain the comprehensive feature representation corresponding to the current vehicle. Based on the comprehensive feature representation, the low-dimensional dense vector of each vehicle node is learned through the graph convolution network model, and the latent vector matrix is ​​generated and the node embedding matrix is ​​obtained by stacking multiple layers of graph convolution. The aggregation weights of the neighbor nodes corresponding to the current vehicle node are dynamically adjusted in combination with the attention mechanism, and the neighbor nodes are aggregated in combination with the attribute feature projection matrix. According to the aggregation results, irrelevant vehicle-vehicle edges are judged and shielded. Through the multi-task learning algorithm, as well as node classification and relationship prediction as the goal, the graph convolution network model is trained in combination with the cross entropy loss function to obtain a trained graph convolution network; For the current vehicle, the vehicle similarity is calculated by cosine similarity, and the vehicle with the highest similarity to the current vehicle in the parking lot is selected as the comparison vehicle according to the vehicle similarity, and the vehicle-parking space similarity between the parking spaces corresponding to the current vehicle and the comparison vehicle is calculated. The regional scoring model is constructed by combining the matrix decomposition technology. For each area, the scoring vector of each parking space is calculated in combination with the current vehicle preference. According to the scoring vector, the parking spaces are sorted in descending order and the parking area recommendation list is generated; According to the pre-acquired three-dimensional electronic map of the parking lot, the current parking lot road is modeled to obtain an undirected weighted graph. According to the reinforcement learning algorithm, the state space is constructed in combination with the vehicle position, road occupancy and target parking space conditions. The vehicle driving is used as an action and the action space is constructed. According to the parking time and parking difficulty, a reward function is set and a path search is performed to obtain the optimal parking path and parking space selection strategy. The vehicle is parked and the remaining resource information of the parking lot is updated. According to the vehicle battery information, the data center solves the vehicle charging scheduling plan with the goal of maximizing the charging efficiency and minimizing the waiting time.

5. The method according to claim 4, characterized in that The loss function corresponding to path search based on reinforcement learning is shown in the following formula: ； in, L ( θ ) represents the loss value, Indicates in status s Take action a Rewards received after r and the next state s' The expected value of r Indicates the current time step t Rewards received, γ represents the discount factor, Indicates that in the next state s' Next, all actions a' The maximum Q value, Indicates the current status s and actions a The Q value of the network is θ Calculated.

6. The method according to claim 1, characterized in that According to the multivariable power generation prediction model, the short-term power generation corresponding to the solar panel is obtained, and the charging power scheduling curve is solved in combination with the vehicle charging scheduling plan, and the multivariable power generation prediction model is dynamically updated through incremental learning. If the charging power scheduling curve deviates and the deviation is greater than the preset deviation threshold, the vehicle charging scheduling plan is updated through a rolling optimization strategy, and charging is performed according to the updated charging scheduling plan and the fuzzy logic control algorithm. During the charging process, the battery state parameters are obtained in real time and the noise is filtered through the Kalman filter algorithm. The charging strategy is dynamically adjusted in combination with the adaptive neural fuzzy reasoning algorithm and the solar power generation state parameters until charging is completed. A charging report is generated according to the charging process and the vehicle information knowledge graph is updated, including: According to the preset multivariable power generation prediction model, the short-term power generation corresponding to the solar panel is obtained, and the standard charging power scheduling curve is calculated in combination with the vehicle charging scheduling plan. The incremental samples are added to the multivariable power generation prediction model through the incremental learning algorithm, and the gradient of the loss function of each layer is calculated. The parameters are updated in combination with the stochastic gradient descent algorithm to determine the influence weight of the incremental sample on the original loss function. If the influence weight is greater than the preset influence threshold, it is considered that the incremental sample has information gain, and it is added to the training set and incremental learning is performed. The updated prediction mean and variance are determined in combination with the Bayesian theory to obtain the updated multivariable power generation prediction model. According to the updated multivariable power generation prediction model, the actual charging power scheduling curve is obtained by solving the vehicle charging constraints and the power grid constraints. The actual charging power dispatch curve is compared with the standard charging power dispatch curve, and the deviation value is calculated and compared with the preset deviation threshold. If the deviation value is greater than the deviation threshold, a hierarchical collaborative optimization model is set through a rolling optimization strategy, wherein the hierarchical collaborative optimization model includes a day-ahead optimization layer, an intraday correction layer, and a real-time control layer. According to the hierarchical collaborative optimization model, the goal is to maximize the solar energy absorption and minimize the charging cost and the peak-to-valley difference, and the battery physical constraints and the grid technical constraints are used as constraints. The vehicle entry time and departure time and the vehicle battery capacity are initialized according to the vehicle information knowledge graph, and the vehicle charging scheduling plan is updated according to the electricity price information and the photovoltaic output information in combination with the hardware utility function; Charging is performed based on the updated vehicle charging scheduling plan, and battery status parameters during the charging process are obtained in real time. Noise is filtered out through the Kalman filtering algorithm, and a fuzzy control rule base is constructed in combination with the fuzzy logic control algorithm. The clear value of the control quantity is determined according to the control rules and the center of gravity defuzzification method, and the optimal charging power is obtained by inferring the clear value. At the same time, according to the uncertainty of battery characteristics and photovoltaic fluctuations, the rules in the fuzzy control rule base are updated through an adaptive neural fuzzy inference algorithm, and the update is repeated until convergence. The updated vehicle charging scheduling plan is dynamically adjusted according to the updated rules until charging is completed. After charging is completed, a charging report is generated according to the charging process and the vehicle information knowledge graph is updated.

7. The method according to claim 6, characterized in that The hierarchical collaborative optimization model updates the vehicle charging scheduling scheme including: The day-ahead optimization layer obtains the prediction information and combines it with the vehicle information knowledge graph to obtain the power dispatch plan of the current vehicle in the next 24 hours. Based on the dispatch plan, the intraday correction layer obtains the predicted power generation power in the next 4 hours in a rolling manner, and corrects the charging power in combination with the current vehicle data in the parking lot. The field-level charging plan in the next four hours is updated and sent to the charging piles in each parking space. The solar power generation power is tracked on the premise of meeting the charging demand through the rolling time domain optimization strategy. The real-time control layer determines the charging power for the current period through fuzzy logic control based on the real-time output of solar power generation, combined with the battery charge state and temperature, and adjusts the vehicle charging scheduling plan based on the generated charging power decision.

8. An intelligent parking charging pile positioning system combined with solar energy, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the three-dimensional spatial information of the parking lot through pre-deployed sensors, combine the image data obtained by the panoramic camera, draw a three-dimensional electronic map of the parking lot through a synchronous positioning and map construction algorithm, mark the coordinates of the parking spaces, and sense the changes in occupancy status through magnetic sensors pre-installed on the parking spaces. According to the environmental information collected by the environmental sensors, combined with historical environmental parameters and historical power generation data of solar panels, a multivariate power generation prediction model is constructed in combination with a long short-term memory network. According to the traffic video data, parked vehicles are identified through target detection and tracking algorithms, parking space occupancy data is obtained and uploaded to the edge computing server, a structured vehicle information knowledge graph is constructed based on vehicle attribute information and vehicle preference information, the association rules between vehicle information are determined through a graph neural network algorithm, and a charging demand prediction model is constructed to generate a vehicle prediction report; The second unit is used to obtain the unique identification number of the vehicle entering the parking lot through the information collection device, obtain the real-time location coordinates and battery capacity, and obtain the vehicle attribute information corresponding to the current vehicle in combination with the vehicle information knowledge graph. According to the vehicle attribute information and the remaining resource information of the parking lot, the fitness score is calculated and the optimal parking area of ​​the current vehicle is predicted through the collaborative filtering recommendation algorithm based on graph network embedding, and a parking area recommendation list is generated. Through the multi-objective reinforcement learning algorithm, the optimal parking path and parking space selection strategy are solved with the parking area fitness and parking time as the reward function, the vehicle is parked and the remaining resource information of the parking lot is updated. The data center solves the vehicle charging scheduling plan based on the vehicle battery information with the goal of maximizing the charging efficiency and minimizing the waiting time; The third unit is used to obtain the short-term power generation corresponding to the solar panel according to the multivariable power generation prediction model, solve the charging power scheduling curve in combination with the vehicle charging scheduling plan, and dynamically update the multivariable power generation prediction model through incremental learning. If the charging power scheduling curve deviates and the deviation is greater than a preset deviation threshold, the vehicle charging scheduling plan is updated through a rolling optimization strategy, and charging is performed according to the updated charging scheduling plan and fuzzy logic control algorithm. During the charging process, battery status parameters are obtained in real time and noise is filtered through the Kalman filtering algorithm. The charging strategy is dynamically adjusted in combination with the adaptive neural fuzzy inference algorithm and the solar power generation status parameters until charging is completed, and a charging report is generated according to the charging process and the vehicle information knowledge graph is updated.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images