Intelligent parking dynamic scheduling optimization method based on multi-source data fusion

Abstract

The invention discloses an intelligent parking dynamic scheduling optimization method based on multi-source data fusion, and the method comprises the following steps: S1, collecting multi-source data, and carrying out the preprocessing of the multi-source data; s2, constructing a heterogeneous graph structure, and generating a node feature matrix and a heterogeneous adjacency tensor; s3, extracting a state fusion vector, constructing a joint scoring matrix, and generating a candidate path set; s4, setting target parameters, executing path search and pheromone iteration, and outputting a regional path set; s5, adjusting a node preference weight, executing regional path reordering, and outputting an optimal vehicle parking path; and S6, recording the whole scheduling process, updating the network parameter weight through the sliding window memory pool, and generating an incremental learning sample set. According to the invention, high-precision, low-delay and personalized intelligent parking dynamic scheduling under the driving of multi-source data can be realized, and the parking efficiency and the user experience are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation, and in particular to a smart parking dynamic scheduling optimization method based on multi-source data fusion. BACKGROUND

[0002] Under the background of continuous rise of urban traffic pressure, the problem of parking difficulty has become a key bottleneck restricting the sustainable development of smart city. The traditional parking management system mainly relies on manual guidance, fixed video monitoring, ground magnetic induction or ground induction facilities to guide and schedule parking spaces. Such systems usually have poor real-time performance, data isolation, rough scheduling mechanism and other problems, and are difficult to adapt to the precise parking needs in high-dynamic and high-density urban traffic environment.

[0003] With the continuous development of Internet of Things, big data and artificial intelligence technology, some research attempts to use multi-source data for smart parking management, such as realizing a certain degree of parking prediction and recommendation through vehicle positioning, mobile application, historical parking behavior and other information. However, the existing schemes generally use static models or single optimization targets for scheduling, lack the ability to model multi-type nodes and heterogeneous relationships, and cannot accurately express the complex interaction characteristics between vehicles and parking spaces in the time, space and behavior dimensions, resulting in poor stability of path recommendation and parking space allocation results and low prediction accuracy.

[0004] The common parking path optimization method in the prior art usually uses the shortest path or the minimum waiting time as a single optimization target, ignoring the influence of user individual preferences, regional resource load and multi-objective trade-off, and is easily trapped in local optimum. At the same time, part of the optimization algorithm lacks dynamic regulation mechanism in the search strategy, and fails to update the strategy adaptively combined with the path convergence speed and score priority characteristics in the actual scene, resulting in low search efficiency and non-global robustness of the scheduling result.

[0005] In addition, most of the current path scheduling systems lack a sustainable learning mechanism and feedback update channel, and cannot incrementally update the model parameters according to the evolution process of real-time parking state and user behavior, making it difficult to form a highly adaptive intelligent scheduling capability. The application of graph neural networks in smart transportation is still in its early stages, and most existing models ignore the multi-channel feature fusion and dynamic preference modeling within the heterogeneous graph structure, making it difficult to meet the intelligent parking scheduling optimization needs in large-scale dynamic scenarios.

[0006] Therefore, how to provide a smart parking dynamic scheduling optimization method based on multi-source data fusion is a problem that those skilled in the art need to solve. SUMMARY

[0007] One purpose of the present application is to propose a smart parking dynamic scheduling optimization method based on multi-source data fusion, which fully integrates heterogeneous graph neural network, improved double objective ant colony optimization algorithm and graph ranking learning mechanism, and describes in detail the complete process of realizing high-precision parking path prediction and dynamic scheduling optimization under multi-objective conditions, which has the advantages of fast response speed, high path matching accuracy and high resource utilization efficiency.

[0008] According to the smart parking dynamic scheduling optimization method based on multi-source data fusion, the method comprises the following steps:

[0009] S1, collecting multi-source data and preprocessing;

[0010] S2, based on the preprocessed multi-source data, constructing a heterogeneous graph structure, defining parking space nodes and vehicle nodes, constructing a multi-type edge set connecting vehicle driving trajectories and parking space states, generating a node feature matrix and a heterogeneous adjacency tensor;

[0011] S3, inputting the node feature matrix and the heterogeneous adjacency tensor into the heterogeneous graph neural network, extracting a state fusion vector, and calculating the occupancy probability, the reachability score and the user behavior similarity score of each parking space, constructing a joint score matrix, and generating a candidate path set;

[0012] S4, inputting the candidate path set into the scheduling optimization model, setting the parking distance and the predicted waiting time as the target parameters, performing path search and pheromone iteration, and outputting a regional path set;

[0013] S5, inputting the regional path set into the graph ranking network with a memory unit, adjusting the node preference weight combined with the user historical parking record, performing regional path reordering, and outputting the optimal parking path of the vehicle;

[0014] S6, recording the whole process of vehicle parking scheduling, and updating the parameter weights of the heterogeneous graph neural network and the graph ranking network through the sliding window memory pool, generating an incremental learning sample set and returning to the scheduling optimization model for updating.

[0015] Optionally, the multi-source data includes vehicle driving trajectory, parking space state, user parking preference and historical parking record.

[0016] Optionally, the preprocessing includes uniform coding, timestamp alignment, missing value filling and outlier removal.

[0017] Optionally, the heterogeneous graph neural network includes a parking space state channel, a trajectory evolution channel and a preference transfer channel, which respectively extract multi-dimensional interaction features between corresponding nodes, fuse channel outputs through a multi-channel attention gate mechanism, train parameters using a graph structure mask contrast loss function, and output a state fusion vector of the parking space node.

[0018] Optionally, the scheduling optimization model is based on an improved dual-objective ant colony optimization algorithm, and adopts an adaptive pheromone attenuation factor and a guidance function enhancement mechanism. The adaptive pheromone attenuation factor dynamically adjusts the volatility coefficient according to the path convergence rate to prevent premature falling into local optimality. The guidance function enhancement mechanism constructs an heuristic function priority queue based on the joint scoring matrix of candidate areas, guiding the ant colony to preferentially expand the search path in high-scoring areas, and combines the distributed parallel search framework to realize multi-path concurrent optimization, thereby improving the scheduling optimization model's ability to find the optimal solution under multi-objective trade-offs.

[0019] Optionally, the S2 specifically includes:

[0020] S21, based on the pre-processed multi-source data, set the node set V = {v1, v2, ..., v m}, where each node v i ∈V represents an entity object, m represents the number of nodes, and the node types include parking space nodes and vehicle nodes. The parking space node represents the physical parking space, and the vehicle node represents the mobile terminal to which parking resources are to be allocated;

[0021] S22, construct edge set E = {e1, e2, ..., e n}, where each edge e j ∈E represents the starting node v s ∈E to target node v t Directed connection, n represents the number of edges, and defines the edge type identifier τ j , used to represent vehicle trajectory connection edges, parking status change edges or user preference transmission edges. There are three types of edges, which are used to describe dynamic paths, static states and behavior correlations respectively;

[0022] S23, construct node feature matrix X, where X i,k ∈X represents the value of the i-th node in the k-th feature dimension, which includes the encoded location coordinates, parking status label, vehicle type number, user parking frequency statistics, and time series feature embedding results;

[0023] S24. Construct a heterogeneous adjacency tensor A. When A i,j,p ∈A and A i,j,p =1, indicating that there is a slave node v under the pth edge type i To node v j The connection relationship, on the contrary A i,j,p =0, there are three edge types in total, numbered as p=1, p=2, and p=3, corresponding to trajectory edges, state edges, and preference edges.

[0024] Optionally, the S3 specifically includes:

[0025] S31, input the node feature matrix X and the heterogeneous adjacency tensor A into the berth state channel, the trajectory evolution channel and the preference transition channel respectively, the channel numbers are p=1, p=2 and p=3, each channel performs a heterogeneous graph convolution operation, and the three-channel outputs are fused by using a gated attention mechanism to generate a state fusion matrix of the parking berth nodes, satisfying the formula:

[0026]

[0027] wherein H f represents the state fusion vector of each node after fusing the three channels, forming a state fusion matrix, A (p) represents the adjacency matrix slice of edge type number p, W (p) represents the feature mapping weight matrix of edge type number p, B (p) represents the bias item matrix of edge type number p, μ p is a learnable channel attention weight coefficient, and LeakyReLU(·) represents a linear unit activation function.

[0028] S32, for the state fusion matrix, a joint score matrix M is constructed by using a score mapping function, which is used to describe the comprehensive scheduling priority of each parking berth node, and the score dimensions include the occupancy probability P o , the reachability score S r and the user behavior similarity score S u , and the calculation formula is:

[0029]

[0030] wherein M is a joint score matrix, each row corresponds to a score vector of a parking berth node, ψ(·) is a score normalization function, σ(·) is a Sigmoid function, η1, η2 and η3 are three types of score weight coefficients respectively, W o , W r and W u are feature mapping matrices of occupancy prediction, reachability modeling and behavior preference similarity respectively, Q represents a user historical behavior embedding vector, Q T represents the transpose vector of Q, ‖·‖ represents the Euclidean norm, and tanh(·) is a hyperbolic tangent activation function.

[0031] S33, according to the joint score matrix M, a sorting and optimization operation is performed on the parking berth nodes, a candidate node subset with a high joint score is extracted, and a candidate path generation is performed with the current position of the vehicle as the starting point to form a candidate path set R c .

[0032] Optionally, the S4 specifically comprises:

[0033] S41, the candidate path set R c is input to the scheduling optimization model, for each path r i ∈R c , the cumulative parking distance D i and the predicted waiting time T i of the path are calculated, and a double-objective optimization function is constructed, and the form of the objective function is:

[0034]

[0035] Wherein, F i represents the double-objective function value of the i-th path, D i represents the corresponding cumulative parking distance of the path r i , T i represents the corresponding predicted waiting time of the path r i , D min and D max represent the minimum and maximum values of the parking distance in the candidate path set respectively, T min and T max represent the minimum and maximum values of the waiting time in the candidate path set respectively, γ1 and γ2 are the weighting coefficients of the parking distance and the waiting time, and satisfy γ1+γ2=1;

[0036] S42, based on the double-objective optimization function, an improved double-objective ant colony optimization algorithm is initialized, in each round of iteration, pheromone updating operation is performed to update the pheromone intensity τ i on the path r i , and the pheromone updating formula is:

[0037]

[0038] Wherein, τ i (t) represents the pheromone intensity of the i-th path in the t-th round of iteration, ρ(t) represents an adaptive pheromone decay factor, which is dynamically adjusted according to the path convergence rate, Q is a pheromone release constant, F i,a represents the objective function value of the a-th ant selecting the path r i in the current round, ∈ is a smoothing factor to prevent the denominator from being zero, θ i,a represents the guide function priority weight of the a-th ant on the path r i , which is obtained by normalizing the heuristic value of the corresponding path in the joint score matrix M, and N represents the total number of ants;

[0039] S43, a path updating process based on a distributed parallel search framework is constructed, multiple parallel search threads are used to update the path pheromone and the heuristic function queue simultaneously, and the regional path set R o, as the optimal scheduling path solution set obtained under the condition of multi-objective scheduling optimization model.

[0040] Optionally, the improved double-objective ant colony optimization algorithm constructs pheromone update and path selection mechanism based on double-objective optimization function, sets path distance and predicted waiting time as optimization targets, adjusts search priority through guide function enhancement mechanism, controls evaporation rate through adaptive pheromone decay factor, executes multi-path concurrent iteration under parallel search framework, updates path set according to double-objective function result and outputs optimal scheduling solution.

[0041] Optionally, the S5 specifically includes:

[0042] S51, the region path set R o is input into a graph ranking network with a memory unit, R o contains k to-be-evaluated paths, and structure features corresponding to each node in each path are set as h i , a user historical parking record vector is initialized as H u , a preference weight updating process based on a gated memory mechanism is executed, and the following formula is satisfied:

[0043] z i = tanh(W z ·h i +U z ·m t-1 +V z ·H u +b z );

[0044] wherein z i represents a preference vector of the node i in the current round, m t-1 represents a memory state vector of the last iteration of the graph ranking network, W z is a node feature weight matrix, U z is a memory unit transition matrix, V z is a user preference weight matrix, b z is a bias term, and tanh(·) is a hyperbolic tangent activation function.

[0045] S52, a ranking score function of the graph ranking network is constructed, preference vectors z i of all nodes are ranked and scored, a attention aggregation function and a graph node ranking association matrix G are adopted to perform re-ranking weight extraction, and the ranking score function is expressed as:

[0046]

[0047] wherein S r (i) represents a re-ranking score of the i-th node in the path, and αij denotes the attention weight coefficient between node i and node j, G ij denotes the order correlation weight of node i and node j in the path structure in the graph, and β is the order gain coefficient.

[0048] S53, the order score function S r (i) performing path-level aggregation calculation, scoring normalization and ordering for all paths, selecting the path with the highest score cumulative value as the optimal parking path for the vehicle, and memorizing the state vector m t-1 and the node preference vector z i The sliding update queue is input to support adaptive iterative updating of the memory unit parameters.

[0049] The beneficial effects of the present application are:

[0050] Firstly, the present application collects and pre-processes multi-source data such as vehicle driving trajectory, parking space state, user parking preference and historical behavior, constructs a heterogeneous graph structure containing vehicle nodes and parking space nodes, introduces multi-type edge relationship to express the dynamic interaction features between vehicles and parking spaces, and effectively improves the modeling ability of the system for complex urban parking environment, providing a high-quality data basis for subsequent scheduling.

[0051] Secondly, the heterogeneous graph neural network is used to extract a multi-channel state fusion vector, a joint scoring matrix is constructed by combining the occupancy probability, accessibility score and user behavior similarity score, and an improved double-objective ant colony optimization algorithm based on adaptive pheromone decay and guide function enhancement mechanism is introduced to realize efficient search and parallel optimization of regional paths, significantly improving the optimization efficiency and stability of the scheduling model under multi-objective conditions, and avoiding the problem that the traditional algorithm is easy to fall into local optimum.

[0052] Finally, the graph ordering network with memory unit is introduced to dynamically adjust the node preference weight according to the user's historical parking records, reorder the candidate paths, and further improve the personalization and accuracy of the final path recommendation results. At the same time, the sliding window memory pool mechanism supports continuous optimization and incremental learning of model parameters, enhancing the adaptive ability and long-term scheduling performance of the system in large-scale real-time scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0053] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0054] Figure 1 A flowchart of a smart parking dynamic scheduling optimization method based on multi-source data fusion is proposed for the present application;

[0055] Figure 2 A multi-source data driven heterogeneous graph construction and state fusion process of a smart parking dynamic scheduling optimization method based on multi-source data fusion is proposed for the present application;

[0056] Figure 3 An improved double-target ant colony scheduling optimization path generation flowchart of a smart parking dynamic scheduling optimization method based on multi-source data fusion is proposed for the present application;

[0057] Figure 4 An optimal path output flowchart of a memory graph ordering network of a smart parking dynamic scheduling optimization method based on multi-source data fusion is proposed for the present application. DETAILED DESCRIPTION

[0058] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0059] REFERENCE Figures 1-4 A smart parking dynamic scheduling optimization method based on multi-source data fusion includes the following steps:

[0060] S1, collect multi-source data and perform preprocessing;

[0061] S2, based on the preprocessed multi-source data, construct a heterogeneous graph structure, define parking space nodes and vehicle nodes, construct a multi-type edge set connecting vehicle driving trajectories and parking space states, generate a node feature matrix and a heterogeneous adjacency tensor;

[0062] S3, input the node feature matrix and the heterogeneous adjacency tensor into the heterogeneous graph neural network, extract the state fusion vector, and calculate the occupancy probability, the reachability score and the user behavior similarity score of each parking space, construct a joint score matrix, and generate a candidate path set;

[0063] S4, input the candidate path set into the scheduling optimization model, set the parking distance and the predicted waiting time as the target parameters, perform path search and pheromone iteration, and output a regional path set;

[0064] S5, input the regional path set into the graph ordering network with a memory unit, adjust the node preference weight combined with the user historical parking record, perform regional path reordering, and output the vehicle parking optimal path;

[0065] S6, record the whole process of vehicle parking scheduling, update the parameter weights of the heterogeneous graph neural network and the graph ordering network through the sliding window memory pool, generate an incremental learning sample set and return to the scheduling optimization model for updating.

[0066] The application provides a complete intelligent parking dynamic scheduling optimization method, and builds a full-process scheduling system covering data collection, graph structure modeling, state fusion, path scheduling, ordering optimization and continuous updating, thereby improving parking efficiency and intelligent level in a complex traffic scene.

[0067] In the embodiment, the multi-source data includes vehicle driving trajectory, parking space state, user parking preference and historical parking record.

[0068] The application provides a comprehensive and real behavior basis for subsequent scheduling modeling by fusing multi-source heterogeneous data such as vehicle trajectory, parking space state, user preference and historical record, and enhances the perception ability and adaptability of the system.

[0069] In the embodiment, the preprocessing includes unified coding, timestamp alignment, missing value filling and outlier removal.

[0070] The application adopts standardized preprocessing procedures such as unified coding, time alignment, missing value filling and outlier removal, improves data quality and consistency, and provides accurate and reliable input for subsequent graph neural network modeling and scheduling optimization.

[0071] In the embodiment, the heterogeneous graph neural network contains a parking space state channel, a trajectory evolution channel and a preference transfer channel, which respectively extract multi-dimensional interaction features between corresponding nodes, fuse channel outputs through a multi-channel attention gate mechanism, train parameters using a graph structure mask contrast loss function, and output a state fusion vector of a parking space node.

[0072] The application independently models the parking space state, trajectory evolution and preference transfer using a multi-channel heterogeneous graph neural network, fuses channel features through a gate mechanism, effectively captures multi-dimensional interaction semantics, and improves the node state expression ability.

[0073] In the embodiment, the scheduling optimization model is based on an improved double-objective ant colony optimization algorithm, adopts an adaptive pheromone decay factor and a guide function enhancement mechanism, the adaptive pheromone decay factor dynamically adjusts the evaporation coefficient according to the path convergence rate to prevent premature convergence into a local optimum, and the guide function enhancement mechanism constructs a heuristic function priority queue based on a candidate region joint scoring matrix to guide the ant colony to preferentially expand the search path in a high-score region, realizes multi-path concurrent optimization in combination with a distributed parallel search framework, and improves the optimization ability of the scheduling optimization model for the optimal solution under multi-objective weighting.

[0074] The scheduling optimization model introduced in the application adopts an adaptive pheromone decay and guide function mechanism, improves the convergence efficiency and global optimal ability of the algorithm for complex path search problems, and avoids falling into a local optimal solution.

[0075] In this embodiment, S2 specifically includes:

[0076] S21, based on the pre-processed multi-source data, a node set V={v1, v2,..., v m} is set, wherein each node v i ∈V represents an entity object, m represents the number of nodes, the node type includes a parking space node and a vehicle node, the parking space node represents a physical parking space, and the vehicle node represents a mobile terminal to be allocated a parking resource;

[0077] S22, an edge set E={e1, e2,..., e n} is constructed, wherein each edge e j ∈E represents a directed connection from a starting node v s ∈E to a target node v t , n represents the number of edges, and an edge type identifier τ j is defined, which is used to represent a vehicle trajectory connection edge, a parking state change edge or a user preference transmission edge, and three types of edges are provided, which are respectively used to describe a dynamic path, a static state and a behavior correlation;

[0078] S23, a node feature matrix X is constructed, wherein X i,k ∈X represents the numerical value of the ith node in the kth feature dimension, and the feature dimension includes an encoded position coordinate, a parking state label, a vehicle type number, a user parking frequency statistical value and a time sequence feature embedding result;

[0079] S24, a heterogeneous adjacency tensor A is constructed, when A i,j,p ∈A and A i,j,p =1, it represents that there is a connection relationship from the node v i to the node v j under the pth edge type, otherwise A i,j,p =0, and the total number of edge types is three, which are numbered as p=1, p=2 and p=3, corresponding to the trajectory edge, the state edge and the preference edge.

[0080] The application refines the heterogeneous graph construction process, clearly defines the types and structures of nodes and edges, models multiple types of relationships by using an adjacency tensor and a feature matrix, and provides a high-explainable and high-expressive graph basis for subsequent graph neural network processing.

[0081] In this embodiment, S3 specifically includes:

[0082] S31, the node feature matrix X and the heterogeneous adjacency tensor A are input into a berth state channel, a trajectory evolution channel and a preference transfer channel, the channel numbers are p=1, p=2 and p=3, a heterogeneous graph convolution operation is performed on each channel, a gated attention mechanism is used to fuse the outputs of the three channels, a state fusion matrix of the parking space node is generated, and the following formula is satisfied:

[0083]

[0084] wherein, H f denotes the state fusion vector of each node after fusing three channels, forming a state fusion matrix, A (p) denotes the adjacency matrix slice of edge type number p, W (p) denotes the feature mapping weight matrix of edge type number p, B (p) denotes the bias item matrix of edge type number p, μ p is a learnable channel attention weight coefficient, and LeakyReLU(·) denotes a linear unit activation function;

[0085] S32, for the state fusion matrix, a joint score matrix M is constructed by using a score mapping function, which is used to describe the comprehensive scheduling priority of each parking space node, and the score dimension includes the occupancy probability P o , the reachability score S r and the user behavior similarity score S u , and the calculation formula is:

[0086]

[0087] wherein, M is a joint score matrix, each row corresponds to a score vector of a parking space node, ψ(·) is a score normalization function, σ(·) is a Sigmoid function, η1, η2 and η3 respectively represent three types of score weight coefficients, W o , W r , W u respectively denote the feature mapping matrices of occupancy prediction, reachability modeling and behavior preference similarity, Q denotes a user historical behavior embedding vector, Q T denotes the transpose vector of Q, ‖·‖ denotes the Euclidean norm, and tanh(·) is a hyperbolic tangent activation function;

[0088] S33, according to the joint score matrix M, the parking space nodes are sorted and optimized, a candidate node subset with a high joint score is extracted, and candidate path generation is performed with the current position of the vehicle as the starting point, forming a candidate path set R c .

[0089] The application constructs a joint score matrix based on a state fusion vector, comprehensively considers the occupancy rate, reachability and behavior preference, guarantees the rationality and individualization of the path candidate set, and effectively improves the accuracy and acceptability of scheduling recommendation.

[0090] In the embodiment, the S4 specifically includes:

[0091] S41, the candidate path set R cInput into the scheduling optimization model, for each path r i ∈ R c , calculate the cumulative parking distance D i and the predicted waiting time T i , and construct a double-objective optimization function, which is in the form of:

[0092]

[0093] Where F i represents the double-objective function value of the ith path, D i represents the corresponding cumulative parking distance of path r i , T i represents the corresponding predicted waiting time of path r i , D min and D max represent the minimum and maximum values of the parking distance in the candidate path set, respectively, T min and T max represent the minimum and maximum values of the waiting time in the candidate path set, respectively, and γ1 and γ2 are the weighting coefficients of the parking distance and the waiting time, satisfying γ1 + γ2 = 1.

[0094] S42, based on the double-objective optimization function, initialize the improved double-objective ant colony optimization algorithm, and in each round of iteration, perform pheromone update operation to update the pheromone intensity τ i on path r i , and the pheromone update formula is:

[0095]

[0096] Where τ i (t) represents the pheromone intensity of the ith path in the tth round of iteration, ρ(t) represents an adaptive pheromone decay factor, which is dynamically adjusted according to the path convergence rate, Q is a constant pheromone release, F i,a represents the objective function value of the ath ant selecting path r i in the current round, ∈ is a smoothing factor to prevent the denominator from being zero, θ i,a represents the guide function priority weight of the ath ant on path r i , which is obtained by normalizing the heuristic value of the corresponding path in the joint scoring matrix M, and N represents the total number of ants.

[0097] S43, construct a path update process based on a distributed parallel search framework, use multiple parallel search threads to update the path pheromone and the heuristic function queue simultaneously, and output the regional path set R o as the optimal scheduling path solution set obtained by the scheduling optimization model under multi-objective conditions when the path convergence stability threshold is reached.

[0098] The application integrates parking distance and waiting time into a unified scheduling target by constructing a path dual-objective optimization function, combines an improved ant colony mechanism with a parallel computing framework, and improves the calculation efficiency and result quality of a scheduling algorithm in a large-scale scenario.

[0099] In the embodiment, the improved dual-objective ant colony optimization algorithm constructs an information pheromone update and path selection mechanism based on a dual-objective optimization function, sets path distance and predicted waiting time as optimization targets, adjusts search priority through a guide function enhancement mechanism, controls evaporation rate through an adaptive information pheromone decay factor, executes multi-path concurrent iteration under a parallel search framework, updates a path set according to a dual-objective function result, and outputs an optimal scheduling solution.

[0100] The application constructs an ant colony information pheromone update and path selection mechanism for a dual-objective scheduling scenario, realizes multi-path concurrent optimization through dynamic information pheromone adjustment and guide priority control, and enhances the diversity and stability of path search.

[0101] In the embodiment, the S5 specifically includes:

[0102] S51, inputting a regional path set R o to a graph ranking network with a memory unit, wherein the R o contains k to-be-evaluated paths, a structure feature corresponding to each node in each path is set as h i , a user historical parking record vector is initialized as H u , a preference weight update process based on a gated memory mechanism is executed, and the following formula is satisfied:

[0103] z i =tanh(W z ·h i +U z ·m t-1 +V z ·H u +b z );

[0104] wherein z i represents a preference vector of the node i in the current round, m t-1 represents a memory state vector of a last round of iteration on the graph ranking network, W z is a node feature weight matrix, U z is a memory unit transfer matrix, V z is a user preference weight matrix, b z is a bias term, and tanh(·) is a hyperbolic tangent activation function.

[0105] S52, constructing a ranking score function of the graph ranking network, and inputting the preference vectors zi The ranking score is sorted, the attention aggregation function and the graph node ranking correlation matrix G are used to perform reordering weight extraction, and the ranking score function is expressed as:

[0106]

[0107] Wherein, S r (i) represents the reordering score of the i-th node in the path, alpha ij represents the attention weight coefficient between node i and node j, G ij represents the order correlation weight of node i and node j in the path structure, and beta is the ranking gain coefficient.

[0108] S53, the ranking score function S r (i) performs path-level aggregation calculation, scores normalization and ranking are performed on all paths, the path with the highest score cumulative value is selected as the optimal parking path of the vehicle, and the memory state vector m t-1 and the node preference vector z i The sliding update queue is input to support adaptive iterative update of the memory unit parameters.

[0109] The application designs a graph ranking network with a memory mechanism, dynamically adjusts the path node preference according to the user historical behavior, realizes the personalized ranking of the optimal path of the vehicle, and combines the sliding update mechanism to improve the self-learning ability and adaptability of the system.

[0110] Embodiment 1:

[0111] In order to verify the feasibility of the application in the implementation, the application is applied to a certain large comprehensive commercial area, the vehicle flow amount of the area is large every day, the parking resources are unevenly distributed, the parking space occupancy rate is close to 92% during the peak period under normal circumstances, the average vehicle search time of the user is more than 12 minutes, and there are obvious problems such as low parking scheduling efficiency, frequent path congestion and poor user experience. The traditional parking system only provides parking space recommendation based on distance, and cannot effectively consider the accessibility of parking spaces, dynamic occupancy state and user personalized preference, resulting in unreasonable parking space recommendation, system response lag, uneven traffic load distribution and other problems.

[0112] In this scenario, the intelligent parking dynamic scheduling optimization method based on multi-source data fusion is deployed. First, by deploying vehicle trajectory acquisition terminals, parking space state monitoring devices and user mobile terminal information acquisition interfaces, the system continuously acquires multi-dimensional data such as vehicle position, parking space idle state, user historical behavior and parking preference. After the data is uniformly coded, time aligned, missing value filled and outlier removed on the server side, the heterogeneous graph structure construction process is input, and a graph structure with parking spaces and vehicles as nodes and driving trajectories and state changes as multi-type edges is generated.

[0113] A heterogeneous graph neural network extracts node state fusion vectors, constructs a joint scoring matrix, and comprehensively evaluates each parking space's occupancy probability, accessibility, and user matching, screening the optimal candidate path set in real time. During the scheduling optimization phase, the system inputs candidate paths into a scheduling module based on a dual-objective ant colony optimization algorithm, setting path distance and predicted wait time as target parameters. A guide function priority mechanism guides the search, and path updates are completed within a parallel search framework to generate a regional path set.

[0114] During the final path sorting phase, the graph sorting network incorporates user history records and memory units to adjust path node preferences and generate optimal path recommendations. After each update, the scheduling process feeds back process samples for model training and parameter optimization, forming a closed-loop incremental learning mechanism and enabling adaptive system evolution.

[0115] To validate the performance of our method, we compared changes in core metrics before and after deployment, and conducted a comparative analysis with a traditional distance-priority scheduling strategy. The results showed that after deploying our method, the average time it takes for users to find a car during peak hours decreased from 12.4 minutes to 3.7 minutes, the average parking space turnover rate increased to 3.2 times per hour, and the overall system response latency decreased by 62%. Furthermore, the recommended path hit rate increased significantly, from 58.6% to 91.3%, effectively alleviating congestion in the area and improving parking scheduling efficiency and user satisfaction.

[0116] Table 1 Comparison of performance verification data of the method of the present invention in the smart parking scenario

[0117]

[0118] The data in Table 1 show that the present invention not only has strong robustness in a multi-source data environment, but also effectively integrates personalized preferences with global path scheduling strategies, significantly improving the scheduling accuracy, efficiency, and service experience in complex urban parking scenarios.

[0119] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A smart parking dynamic scheduling optimization method based on multi-source data fusion, characterized by: The steps include: S1. Collect multi-source data and perform preprocessing; S2. Based on the preprocessed multi-source data, a heterogeneous graph structure is constructed, parking space nodes and vehicle nodes are defined, a multi-type edge set is constructed connecting vehicle driving trajectories and parking space status, and a node feature matrix and a heterogeneous adjacency tensor are generated; S3. Input the node feature matrix and heterogeneous adjacency tensor into the heterogeneous graph neural network, extract the state fusion vector, calculate the occupancy probability, accessibility score, and user behavior similarity score of each parking space, construct a joint scoring matrix, and generate a set of candidate paths. S4. Input the candidate path set into the scheduling optimization model, set the parking distance and predicted waiting time as target parameters, perform path search and pheromone iteration, and output the regional path set; S5. Input the regional path set into a graph sorting network with memory units, adjust the node preference weights based on the user's historical parking records, perform regional path re-sorting, and output the optimal parking path for the vehicle; S6. Record the entire process of vehicle parking scheduling, and update the parameter weights of the heterogeneous graph neural network and the graph sorting network through the sliding window memory pool, generate an incremental learning sample set and return it to the scheduling optimization model to complete the update.

2. The method for intelligent parking dynamic scheduling optimization based on multi-source data fusion according to claim 1 is characterized in that: The multi-source data includes vehicle driving trajectory, parking space status, user parking preferences and historical parking records.

3. The method for intelligent parking dynamic scheduling optimization based on multi-source data fusion according to claim 1 is characterized in that: The preprocessing includes unified encoding, timestamp alignment, missing value filling and outlier removal.

4. The method for intelligent parking dynamic scheduling optimization based on multi-source data fusion according to claim 1 is characterized in that: The heterogeneous graph neural network contains a parking space state channel, a trajectory evolution channel, and a preference transfer channel. It extracts multi-dimensional interaction features between corresponding nodes, fuses channel outputs through a multi-channel attention gating mechanism, and uses a graph structure mask contrast loss function for parameter training to output the state fusion vector of the parking space node.

5. The method for intelligent parking dynamic scheduling optimization based on multi-source data fusion according to claim 1 is characterized in that: The scheduling optimization model is based on an improved dual-objective ant colony optimization algorithm and adopts an adaptive pheromone attenuation factor and a guidance function enhancement mechanism. The adaptive pheromone attenuation factor dynamically adjusts the volatility coefficient according to the path convergence rate to prevent premature falling into the local optimum. The guidance function enhancement mechanism constructs a heuristic function priority queue based on the joint scoring matrix of candidate areas, guiding the ant colony to preferentially expand the search path in high-scoring areas. Combined with a distributed parallel search framework, multi-path concurrent optimization is achieved, thereby improving the scheduling optimization model's ability to find the optimal solution under multi-objective trade-offs.

6. The method for intelligent parking dynamic scheduling optimization based on multi-source data fusion according to claim 1 is characterized in that: The S2 specifically includes: S21, based on the pre-processed multi-source data, set the node set V = {v1, v2, ..., v m }, where each node v i ∈V represents an entity object, m represents the number of nodes, and the node types include parking space nodes and vehicle nodes. The parking space node represents the physical parking space, and the vehicle node represents the mobile terminal to which parking resources are to be allocated; S22, construct edge set E = {e1, e2, ..., e n }, where each edge e j ∈E represents the starting node v s ∈E to target node v t Directed connection, n represents the number of edges, and defines the edge type identifier τ j , used to represent vehicle trajectory connection edges, parking status change edges or user preference transmission edges. There are three types of edges, which are used to describe dynamic paths, static states and behavior correlations respectively; S23, construct node feature matrix X, where X i,k ∈X represents the value of the i-th node in the k-th feature dimension, which includes the encoded location coordinates, parking status label, vehicle type number, user parking frequency statistics, and time series feature embedding results; S24. Construct a heterogeneous adjacency tensor A. When A i,j,p ∈A and A i,j,p =1, indicating that there is a slave node v under the pth edge type i To node v j The connection relationship, on the contrary A i,j,p =0, there are three edge types in total, numbered as p=1, p=2, and p=3, corresponding to trajectory edges, state edges, and preference edges.

7. The method for intelligent parking dynamic scheduling optimization based on multi-source data fusion according to claim 1 is characterized in that: The S3 specifically includes: S31. Input the node feature matrix X and the heterogeneous adjacency tensor A into the parking space state channel, trajectory evolution channel, and preference transfer channel respectively. The channels are numbered p=1, p=2, and p=3. Each channel performs a heterogeneous graph convolution operation once. The gated attention mechanism is used to fuse the three-channel outputs to generate the state fusion matrix of the parking space node, which satisfies the formula: Among them, H f Represents the state fusion vector of each node after fusing three channels, forming a state fusion matrix, A (p) represents the adjacency matrix slice with edge type number p, W (p) Represents the feature map weight matrix of edge type number p, B (p) Represents the bias matrix of edge type number p, μ p is the learnable channel attention weight coefficient, LeakyReLU(·) represents the linear unit activation function; S32. For the state fusion matrix, a scoring mapping function is used to construct a joint scoring matrix M, which is used to describe the comprehensive scheduling priority of each parking space node. The scoring dimension includes the occupancy probability P o , accessibility score S r Similarity score S with user behavior u , the calculation formula is: Where M is the joint rating matrix, each row corresponds to the rating vector of a parking space node, ψ(·) is the rating normalization function, σ(·) is the Sigmoid function, η1, η2, and η3 represent the three types of rating weight coefficients respectively, and W o 、W r 、W u They represent the feature mapping matrices of occupancy prediction, accessibility modeling, and behavior preference similarity, Q represents the user historical behavior embedding vector, and Q T represents the transposed vector of Q, ‖·‖ represents the Euclidean norm, and tanh(·) is the hyperbolic tangent activation function; S33: Perform a sorting and optimization operation on the parking space nodes according to the joint score matrix M, extract a subset of candidate nodes with the highest joint score, and generate candidate paths starting from the current position of the vehicle to form a candidate path set R. c .

8. The method for intelligent parking dynamic scheduling optimization based on multi-source data fusion according to claim 1 is characterized in that: The S4 specifically includes: S41, set the candidate path R c Input to the scheduling optimization model, for each path r i ∈R c , calculate the cumulative stopping distance D of the path i and the predicted waiting time T i , and construct a dual-objective optimization function, the objective function form is: Among them, F i Denotes the value of the dual objective function of the i-th path, D i Represents the path r i The corresponding cumulative stopping distance, T i Represents the path r i The corresponding predicted waiting time, D min 、D max Respectively represent the minimum and maximum values ​​of the stopping distance in the candidate path set, T min 、T max They represent the minimum and maximum waiting time in the candidate path set, respectively. γ1 and γ2 are the trade-off coefficients between the stopping distance and the waiting time, satisfying γ1+γ2=1. S42, based on the dual-objective optimization function, initialize the improved dual-objective ant colony optimization algorithm, perform the pheromone update operation in each round of iteration, and update the path r i Pheromone intensity τ on i , the pheromone update formula is: Among them, τ i (t) represents the pheromone intensity of the i-th path in the t-th iteration, ρ(t) represents the adaptive pheromone attenuation factor, which is dynamically adjusted according to the path convergence rate, Q is the pheromone release constant, and F i,a Indicates that the ath ant chooses path r in the current round i The objective function value, ∈ is the smoothing factor to prevent the denominator from being zero, θ i,a Indicates that the ath ant is on path r i The priority weight of the guidance function on is obtained by normalizing the heuristic value of the corresponding path in the joint scoring matrix M, where N represents the total number of ants; S43. Construct a path update process based on a distributed parallel search framework, use multiple parallel search threads to update the path pheromone and heuristic function queues simultaneously, and output the regional path set R after reaching the path convergence stability threshold. o , as the optimal scheduling path solution set obtained by the scheduling optimization model under multi-objective conditions.

9. The method for intelligent parking dynamic scheduling optimization based on multi-source data fusion according to claim 8 is characterized in that: The improved dual-objective ant colony optimization algorithm constructs a pheromone update and path selection mechanism based on a dual-objective optimization function, sets path distance and predicted waiting time as optimization objectives, adjusts search priority through a guidance function enhancement mechanism, controls the volatilization rate through an adaptive pheromone attenuation factor, performs multi-path concurrent iterations under a parallel search framework, updates the path set according to the dual-objective function results, and outputs the optimal scheduling solution.

10. The method for intelligent parking dynamic scheduling optimization based on multi-source data fusion according to claim 1 is characterized in that: The S5 specifically includes: S51, regional path set R o Input to the graph sorting network with memory unit, R o There are k paths to be evaluated, and the structural features corresponding to each node in each path are represented as h i , initialize the user's historical parking record vector to H u , execute the preference weight update process based on the gated memory mechanism, satisfying the following formula: z i =tanh(W z ·h i +U z m t-1 +V z ·H u +b z ); Among them, z i represents the preference vector of node i in the current round that integrates the user's historical preferences, m t-1 Represents the memory state vector of the last iteration of the graph sorting network, W z is the node feature weight matrix, U z is the memory unit transfer matrix, V z is the user preference weight matrix, b z is the bias term, tanh(·) is the hyperbolic tangent activation function; S52, construct the ranking score function of the graph ranking network, and set the preference vector z of all nodes i To perform ranking and scoring, the attention aggregation function and the graph node ranking association matrix G are used to perform re-ranking weight extraction. The ranking score function is expressed as: Among them, S r (i) represents the re-ranking score of the i-th node in the path, α ij represents the attention weight coefficient between node i and node j, G ij It represents the order correlation weight between node i and node j in the path structure of the graph, and β is the ranking gain coefficient; S53, sorting score function S r (i) Perform path-level aggregation calculation, normalize and sort the scores of all paths, select the path with the highest cumulative score as the optimal parking path for the vehicle, and store the memory state vector m in the current round. t-1 and the node preference vector z i Input sliding update queue to support adaptive iterative update of memory unit parameters.

Images