Parking behavior analysis method and system based on data mining and storage medium

Abstract

The invention relates to the technical field of data processing, and discloses a parking behavior analysis method and system based on data mining and a storage medium. The method comprises the steps of collecting data through a parking sensor network to construct a parking behavior basic data structure, a local interest point database and a global network configuration database, and performing fusion processing by adopting a local-global feature interaction matrix algorithm to obtain a fusion feature matrix and a dynamic weight vector, and clustering through a fusion density clustering algorithm to obtain a fusion clustering result matrix, and performing pattern recognition based on double-layer space-time pattern mining to obtain a parking behavior space-time fusion pattern library and a demand prediction result. The technical problem that an analysis result and an actual parking use mode lack significant correlation due to the fact that an existing parking behavior analysis method based on data mining cannot effectively fuse local interest point features and global network configuration features is solved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a parking behavior analysis method, system and storage medium based on data mining. Background Art

[0002] Existing parking behavior analysis methods primarily utilize single-dimensional data mining techniques. For example, road network analysis methods based on space syntax evaluate the spatial configuration characteristics of parking areas by calculating metrics such as road integration and selectivity. Parking hotspot identification methods based on the DBSCAN clustering algorithm identify the spatial distribution patterns of parking space usage intensity through density clustering. Parking behavior analysis methods based on social media data mining identify parking preferences and sentiment by mining users' online behavior data. These existing technologies possess considerable analytical capabilities within their respective fields and can reveal some patterns and characteristics of parking behavior from different perspectives.

[0003] Existing methods often use a single analysis dimension. Space syntax methods overly rely on global network configuration features and ignore the influence of local points of interest. Cluster analysis methods lack consideration of temporal dynamic changes. Social media analysis methods are difficult to effectively associate with actual physical parking data. Secondly, existing technologies lack an effective feature fusion mechanism and are unable to balance the interaction between local factors (such as POI density and public transportation accessibility) and global factors (such as road network integration and connectivity), resulting in significant differences between analysis results and actual parking usage patterns. In addition, existing methods lack depth in spatiotemporal pattern mining and find it difficult to simultaneously capture the short-term fluctuation characteristics and long-term trend characteristics of parking behavior. Summary of the Invention

[0004] The present application provides a parking behavior analysis method, system and storage medium based on data mining, which are used to solve the technical problem that the existing parking behavior analysis method based on data mining cannot effectively integrate local point of interest features with global network configuration features, resulting in a lack of significant correlation between the analysis results and actual parking usage patterns.

[0005] In the first aspect, the present application provides a parking behavior analysis method based on data mining, which includes: collecting and processing data from paid parking areas through a parking sensor network to obtain a basic data structure of parking behavior, and constructing a local point of interest database and a global network configuration database; fusing the basic parking behavior data structure, the local point of interest database, and the global network configuration database through a local-global feature interaction matrix algorithm to obtain a fused feature matrix and a dynamic weight vector; clustering the fused feature matrix and the dynamic weight vector through a fusion density clustering algorithm to obtain a fused clustering result matrix; performing pattern recognition processing based on the fused clustering result matrix through double-layer spatiotemporal pattern mining to obtain a parking behavior spatiotemporal fusion pattern library and demand prediction results.

[0006] In a second aspect, the present application provides a parking behavior analysis system based on data mining, the parking behavior analysis system based on data mining comprising:

[0007] The acquisition module is used to collect and process data from paid parking areas through the parking sensor network, obtain the basic data structure of parking behavior, and build a local point of interest database and a global network configuration database;

[0008] a fusion module, configured to perform fusion processing on the parking behavior basic data structure, the local point of interest database, and the global network configuration database by using a local-global feature interaction matrix algorithm to obtain a fusion feature matrix and a dynamic weight vector;

[0009] A clustering module is used to cluster the fusion feature matrix and the dynamic weight vector through a fusion density clustering algorithm to obtain a fusion clustering result matrix;

[0010] The recognition module is used to perform pattern recognition processing through double-layer spatiotemporal pattern mining based on the fusion clustering result matrix to obtain a spatiotemporal fusion pattern library of parking behavior and demand prediction results.

[0011] In a third aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned parking behavior analysis method based on data mining.

[0012] In the technical solution provided by this application, the parking behavior basic data structure, local point of interest database, and global network configuration database constructed by collecting and processing data from paid parking areas through a parking sensor network lay a multi-dimensional data foundation for subsequent feature fusion analysis, solving the problems of single data source and incomplete information dimensions in the existing technology, enabling parking behavior analysis to simultaneously consider multiple influencing factors such as physical parking status, point of interest distribution, and road network configuration. The fused feature matrix and dynamic weight vector obtained by fusion processing through the local-global feature interaction matrix algorithm innovatively solve the core technical problem of the inability to effectively fuse local features and global features in the existing technology. The algorithm quantifies the interactive relationship between local POI density features and global network integration features, and adopts a dynamic weight allocation mechanism to adaptively adjust the influence weights of local and global factors according to the feature distribution of different parking spaces, significantly improving the accuracy and comprehensiveness of feature expression, so that the fused feature matrix can more realistically reflect the comprehensive impact of parking spaces in complex urban environments. The fused density clustering algorithm effectively addresses the technical limitation of traditional clustering methods, which cannot simultaneously identify both locally and globally driven parking patterns, by processing the fused feature matrix and the dynamic weight vector to generate a fused clustering matrix. This algorithm dynamically adjusts the density parameter based on the local-global balance factor, accurately identifying both small-scale fine clusters influenced by POIs and large-scale macro clusters influenced by network configuration within the same parking area, significantly improving clustering accuracy compared to traditional methods. A two-layer spatiotemporal pattern mining algorithm, which performs pattern recognition on the fused clustering matrix to generate a spatiotemporal fusion pattern library and demand forecast for parking behavior, overcomes the technical bottleneck of insufficient spatiotemporal pattern analysis in existing technologies. This method uses a hierarchical processing mechanism to extract POI influence patterns at the local driving layer and network configuration patterns at the global driving layer. A pattern fusion algorithm then organically combines local short-term fluctuations with global long-term trends, constructing a comprehensive prediction model that reflects both micro-behavioral characteristics and macro-developmental laws.

[0013] In the application field of parking behavior analysis, the core contribution of the local-global feature interaction matrix algorithm lies in its ability to automatically identify and quantify the interaction strength between features at different spatial scales. Through matrix multiplication operations and interaction coefficient training, this algorithm enables the parking behavior analysis system to accurately determine whether a parking space is more influenced by the local attractiveness of surrounding commercial facilities or the global convenience of road network accessibility, thereby providing a scientific basis for the formulation of precise parking management strategies. The unique value of the fusion density clustering algorithm in urban parking management applications lies in its ability to simultaneously handle parking clusters at multiple scales. By integrating distance metric functions, this algorithm comprehensively considers parking duration similarity, location proximity, and point of interest influence similarity. The clustering results can identify both small-scale high-frequency use areas formed by commercial activity clusters and large-scale traffic distribution areas formed by transportation hubs, providing a precise spatial division basis for differentiated parking space allocation and pricing strategies. The key role of the dual-layer spatiotemporal pattern mining algorithm in parking demand prediction applications lies in its ability to separate and reconstruct temporal patterns under different driving mechanisms. By performing time series decomposition of POI influence on the local driving layer and time series analysis of network configuration on the global driving layer, the algorithm can accurately predict the changing trends of parking demand under specific spatiotemporal conditions. The local pattern captures short-term demand fluctuations affected by local factors such as commercial activities and public transportation, while the global pattern identifies long-term demand trends affected by global factors such as road network structure and urban development planning. The fusion prediction results of the two models provide data support for the dynamic allocation of parking resources and the optimized operation of the intelligent guidance system. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0015] Figure 1 Schematic diagram of an embodiment of a parking behavior analysis method based on data mining in an embodiment of the present application;

[0016] Figure 2 This is a schematic diagram of an embodiment of a parking behavior analysis system based on data mining in an embodiment of the present application. DETAILED DESCRIPTION

[0017] The embodiments of the present application provide a parking behavior analysis method, system and storage medium based on data mining. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices.

[0018] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of the parking behavior analysis method based on data mining includes:

[0019] Step S101: Data from paid parking areas is collected and processed through a parking sensor network to obtain a basic parking behavior data structure, and a local point of interest database and a global network configuration database are constructed;

[0020] Step S102: performing fusion processing on the parking behavior basic data structure, the local point of interest database, and the global network configuration database using a local-global feature interaction matrix algorithm to obtain a fused feature matrix and a dynamic weight vector;

[0021] Step S103: clustering the fused feature matrix and the dynamic weight vector using a fusion density clustering algorithm to obtain a fusion clustering result matrix;

[0022] Step S104: Based on the fusion clustering result matrix, pattern recognition processing is performed through double-layer spatiotemporal pattern mining to obtain a parking behavior spatiotemporal fusion pattern library and demand prediction results.

[0023] It is understandable that the execution subject of this application can be a parking behavior analysis system based on data mining, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0024] Specifically, parking occupancy detection sensors record the occupancy status of each parking space at a 30-second frequency, generating a raw parking dataset containing vehicle ID, location ID, start time, end time, and parking duration. This raw data is then cleaned using an outlier detection algorithm. Sensor failure data is removed, and missing data is filled using a time series interpolation method, forming a structured basic data structure for parking behavior. Simultaneously, based on the geographic coordinates of the parking spaces, POI information is collected from commercial facilities, public transportation stops, and office buildings within a 500-meter radius. The Euclidean distance from each parking space to the nearest point of interest is calculated to construct a local point of interest database. Spatial syntax analysis is then performed on parking areas based on the road network topology, calculating road connectivity indexes, integration, and selectivity parameters to construct a global network configuration database.

[0025] The parking behavior basic data structure is used to construct feature vectors for parking duration, location coordinates, and occupancy status. This is combined with POI distance data from the local POI database and network integration data from the global network configuration database to generate a multidimensional feature vector for parking behavior containing five eigenvalues. The local-global feature interaction matrix algorithm first extracts local POI density eigenvalues ​​and global network integration eigenvalues ​​from the multidimensional feature vector of parking behavior for feature separation. Local and global feature subvectors are constructed, respectively, to produce dimensionally matched local and global feature matrices. The local and global feature matrices are then transposed and multiplied. The product of each local feature and each global feature is calculated according to the matrix multiplication rule to obtain an initial interaction relationship matrix. The initial interaction relationship matrix is ​​then trained with interaction coefficients based on historical parking data. Regression analysis is used to calculate the impact of each feature combination on parking behavior, resulting in an interaction coefficient vector reflecting the weight of the influence of different feature combinations. The initial interaction relationship matrix is ​​then element-wise multiplied with the interaction coefficient vector, multiplying each element by the corresponding interaction coefficient value to obtain a feature interaction coefficient matrix. The feature interaction coefficient matrix is ​​then input into the sigmoid function for weight normalization. The ratio of local feature weights to global feature weights is dynamically adjusted according to the POI density and network integration of different parking spaces to obtain a dynamic weight vector with a value range of 0 to 1. The multidimensional feature vector of parking behavior is weightedly fused based on the dynamic weight vector. The local eigenvalues ​​multiplied by the local weights and the global eigenvalues ​​multiplied by the global weights are summed to obtain the fused feature matrix.

[0026] A fusion density clustering algorithm is used to calculate parking space density by inputting the fusion feature matrix and dynamic weight vector. The density radius parameters of different parking areas are adjusted according to the local-global balance factor to obtain a density distribution map reflecting the degree of parking space aggregation. Based on the density distribution map, core points of parking spaces are identified. Parking spaces with density values ​​exceeding the set threshold and containing the least number of parking spaces in their neighborhood are marked as core points, resulting in a core point set containing core parking space identifiers and density values. Based on the core point set, parking space similarity is calculated using a fusion distance metric function. The comprehensive similarity between parking spaces is calculated by comprehensively considering the similarity of parking time, location proximity, and POI influence similarity, resulting in a similarity matrix reflecting the strength of parking space association. The similarity matrix is ​​input into the cluster expansion algorithm for parking area division. Cluster expansion is performed from the core point as the starting point to parking spaces with high similarity. Parking spaces with similar parking behavior patterns are merged into the same cluster to obtain a fusion clustering result matrix.

[0027] A two-layer spatiotemporal pattern mining approach was used to cluster the fused clustering matrix according to local and global dominance. Clusters with high local dominance were classified as local driver layers, and clusters with high global dominance were classified as global driver layers, resulting in a hierarchical clustering matrix. Based on the hierarchical clustering matrix, the local driver layer was decomposed into a POI impact time series, extracting short-term fluctuation trends and cyclical changes influenced by commercial facilities and public transportation to obtain a local spatiotemporal pattern set. Based on the hierarchical clustering matrix, the global driver layer was analyzed for network configuration time series, identifying long-term trend changes and macro-flow patterns influenced by road integration and connectivity to obtain a global spatiotemporal pattern set. The local and global spatiotemporal pattern sets were input into a pattern fusion algorithm to calculate their interaction effects. A weighted fusion operation was used to construct a prediction model that comprehensively considers local short-term fluctuations and global long-term trends, resulting in a spatiotemporal fusion pattern library for parking behavior and demand forecasting results.

[0028] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0029] The parking occupancy detection sensor monitors the status of each parking space and records the changes in parking space occupancy status at a 30-second acquisition frequency to obtain an original parking dataset containing vehicle identification, location identification, start time, end time, and parking duration.

[0030] The original parking data set is input into the outlier detection algorithm for data cleaning, sensor failure data is removed, and the time series interpolation method is used to fill the missing data to obtain the basic data structure of structured parking behavior;

[0031] Based on the geographic coordinates of parking spaces, POI information is collected and processed for commercial facilities, public transportation stations, and office buildings within a 500-meter radius. The Euclidean distance from each parking space to the nearest point of interest is calculated to obtain a local point of interest database.

[0032] The parking areas are analyzed using spatial syntax according to the road network topology, and the road connectivity index, integration degree and selectivity parameters are calculated to obtain a global network configuration database that records the network configuration characteristics.

[0033] Specifically, the parking occupancy detection sensor uses magnetic induction technology to detect the presence of metal vehicles above parking spaces. It scans the status every 30 seconds and records a timestamp. When a vehicle enters, the start time and vehicle ID are recorded, and when a vehicle leaves, the end time is recorded. The parking duration is calculated using the time difference, and the location ID is recorded in conjunction with the GPS coordinates of the parking space, forming a raw parking dataset containing five fields. The outlier detection algorithm uses statistical methods to identify anomalous data points. Specifically, it calculates the standard deviation and mean of parking duration. Data that exceeds three times the standard deviation of the normal range is marked as outliers and removed. Time series interpolation is then used to complete missing data. The interpolation algorithm performs linear interpolation based on the parking status change trend at adjacent time points, reorganizing the processed data into a structured basic parking behavior data structure.

[0034] POI information collection and processing uses a GIS (Geographic Information System) to obtain point-of-interest (POI) data within a 500-meter radius of a parking space. This includes attributes such as the latitude and longitude coordinates, type identification, and importance rating of commercial facilities, the location and route number of public transportation stops, and the floor area and number of businesses within office buildings. Euclidean distance calculation uses a two-dimensional coordinate system to calculate the straight-line distance between the parking space coordinates and the coordinates of each POI through square root operations. The closest POI is selected as the primary influencing factor for that parking space, and the POI type, distance value, and influence weight are stored in a local POI database.

[0035] The spatial syntax analysis process first constructs a road network topology for the parking area, treating road intersections as nodes and road segments as edges to form a network graph. The road connectivity index is calculated by counting the number of edges connecting each node. The integration parameter is calculated by analyzing the shortest path length from a node to all other nodes in the network, with smaller values ​​indicating a higher centrality. The selectivity parameter is calculated by counting the number of shortest paths passing through a node, reflecting the strength of that node's intermediary role in the network. These calculation results are then linked according to the road segments where the parking spaces are located and stored in a global network configuration database.

[0036] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0037] The parking behavior basic data structure is used to construct a feature vector of parking duration, location coordinates, and occupancy status. This is combined with POI distance data from the local point of interest database and network integration data from the global network configuration database to obtain a multidimensional feature vector of parking behavior containing five-dimensional feature values.

[0038] Based on the multidimensional feature vector of parking behavior, the interaction relationship is calculated and processed using the local-global feature interaction matrix algorithm. The local POI density feature and the global network integration feature are matrix-producted and multiplied by the interaction coefficient to obtain the feature interaction coefficient matrix that quantifies the mutual influence strength between local and global features.

[0039] The feature interaction coefficient matrix is ​​input into the sigmoid function for weight normalization. The ratio of local feature weights to global feature weights is dynamically adjusted according to the POI density and network integration of different parking spaces to obtain a dynamic weight vector with a value range between 0 and 1.

[0040] The multidimensional feature vectors of parking behaviors are weightedly fused according to the dynamic weight vector, and the local eigenvalues ​​multiplied by the local weights and the global eigenvalues ​​multiplied by the global weights are summed to obtain a fused feature matrix containing the fused eigenvalues.

[0041] Specifically, the feature vector construction process extracts parking duration from the parking behavior basic data structure and performs numerical normalization, converting the raw duration data into a standardized value between 0 and 1. The latitude and longitude coordinates of the parking space are extracted and converted into a position vector in a plane coordinate system. The occupancy status is converted into numerical features through binary encoding. POI distance data from the local point of interest database is converted into distance feature values ​​through logarithmic transformation. Network integration data from the global network configuration database is directly used as global feature values. The feature values ​​of these five dimensions are arranged in sequence to form a multidimensional feature vector for parking behavior.

[0042] The local-global feature interaction matrix algorithm first extracts the local POI density feature and the global network integration feature from the multidimensional feature vector of parking behavior to construct two sub-vectors. The local feature sub-vectors are then transposed and matrix-multiplied with the global feature sub-vector. Each local feature element is multiplied by each global feature element to obtain the initial interaction relationship value. The interaction coefficient is obtained by training with historical parking data. Specifically, regression analysis is used to calculate the impact of different feature combinations on parking behavior. The regression coefficient is stored as the interaction coefficient value in the interaction coefficient vector. Each element of the initial interaction relationship matrix is ​​element-wise multiplied with the corresponding interaction coefficient value to obtain a feature interaction coefficient matrix that quantifies the strength of the mutual influence between local and global features.

[0043] Weight normalization involves inputting the feature interaction coefficient matrix into a sigmoid function for mathematical transformation. The sigmoid function uses exponential operations to map any real-valued value to a probability value between 0 and 1. The specific calculation process involves substituting each matrix element as an independent variable into the sigmoid function to obtain the normalized weight value. Dynamic weight adjustment allocates weights proportionally based on the POI density and network integration of different parking spaces. Parking spaces with high POI density are assigned larger local feature weights, while parking spaces with high network integration are assigned larger global feature weights, forming a dynamically adjusted weight vector.

[0044] The weighted fusion process multiplies the local eigenvalues ​​of the parking behavior multidimensional feature vector by the corresponding local weights, and multiplies the global eigenvalues ​​by the corresponding global weights. The two products are then summed to obtain the fused eigenvalues. After the fused eigenvalues ​​for each parking space are calculated, they are arranged in order of parking space number to form a fused feature matrix. Each row in the matrix represents the fused features of a parking space, and each column represents the fusion results of a feature dimension.

[0045] For example, the parking time for a parking space is 3 hours, which is normalized to an eigenvalue of 0.6. The location coordinates are transformed to positional eigenvalues ​​of 0.8 and 0.3. The occupancy state is encoded as 1, the POI distance is logarithmically transformed to an eigenvalue of 0.4, and the network integration value is 0.7, forming a five-dimensional feature vector. A matrix product operation is performed on the local POI density feature 0.4 and the global network integration feature 0.7 to obtain an initial interaction value of 0.28. This is multiplied by the interaction coefficient 1.5 obtained from historical data training to obtain a feature interaction coefficient of 0.42. The interaction coefficient 0.42 is input into a sigmoid function to obtain a normalized weight of 0.6. Based on the low POI density and high network integration characteristics of this parking space, a local weight of 0.3 and a global weight of 0.7 are assigned. Finally, the local eigenvalue 0.4 is multiplied by the local weight 0.3 to obtain 0.12, and the global eigenvalue 0.7 is multiplied by the global weight 0.7 to obtain 0.49. The sum of these two values ​​yields a fused eigenvalue of 0.61, which is stored in the fused feature matrix.

[0046] In a specific embodiment, the process of performing the interactive relationship calculation based on the parking behavior multi-dimensional feature vector using a local-global feature interaction matrix algorithm may specifically include the following steps:

[0047] Extract local POI density eigenvalues ​​and global network integration eigenvalues ​​from the multidimensional feature vector of parking behavior for feature separation, construct local feature subvectors and global feature subvectors respectively, and obtain dimensionally matched local feature matrices and global feature matrices.

[0048] Perform a transposed matrix product operation on the local feature matrix and the global feature matrix, and calculate the product value of each local feature and each global feature according to the matrix multiplication rule to obtain the initial interaction relationship matrix;

[0049] The interaction coefficients of the initial interaction relationship matrix are trained based on historical parking data. The impact of each feature combination on parking behavior is calculated through regression analysis to obtain the interaction coefficient vector reflecting the influence weight of different feature combinations.

[0050] The initial interaction relationship matrix and the interaction coefficient vector are multiplied element by element, and each element in the matrix is ​​multiplied by the corresponding interaction coefficient value to obtain the feature interaction coefficient matrix that quantifies the mutual influence strength between local and global features.

[0051] Specifically, the feature separation process extracts the local POI density eigenvalue and the global network integration eigenvalue from the multidimensional feature vector of parking behavior according to the preset feature index position, where the POI density eigenvalue reflects the density of points of interest around the parking space, and the network integration eigenvalue reflects the accessibility strength of the parking space in the road network. The local feature sub-vector is constructed by traversing the POI density eigenvalues ​​of all parking spaces and arranging them in the order of parking space numbers. The global feature sub-vector is constructed by extracting the network integration eigenvalues ​​of all parking spaces and arranging them in the same order. The dimension matching process ensures that the local feature sub-vector and the global feature sub-vector have the same dimension length. If the dimension does not match, it is adjusted to the same dimension by zero padding or truncation, and then the local feature sub-vector is expanded to a row vector form to construct a local feature matrix, and the global feature sub-vector is expanded to a column vector form to construct a global feature matrix.

[0052] The transposed matrix product operation first transposes the local feature matrix, converting the original row vectors into column vectors. The transposed local feature matrix is ​​then multiplied with the global feature matrix according to the matrix multiplication rules. The matrix multiplication rules require that the number of columns in the first matrix equal the number of rows in the second matrix. The product operation multiplies each row element of the first matrix by each column element of the second matrix, and then sums the corresponding element values ​​in the result matrix. The product value of each local feature and each global feature is calculated by directly multiplying the local eigenvalue with the corresponding global eigenvalue. All product values ​​are arranged according to the matrix position to form the initial interaction relationship matrix.

[0053] The interaction coefficient training process constructs a training dataset based on historical parking data. This historical data includes local and global eigenvalues ​​for different parking spaces over different time periods, along with corresponding parking behavior results, such as parking duration and occupancy frequency. Regression analysis uses the least squares method to establish a mathematical relationship model for the impact of combinations of local and global features on parking behavior. The strength of the impact of each feature combination on parking behavior is quantified by calculating the regression coefficient. The impact strength calculation process uses the local and global eigenvalues ​​as independent variables and the parking behavior results as the dependent variable. The regression coefficients are obtained through regression analysis as the interaction coefficient values. The interaction coefficient vector is constructed by arranging the regression coefficients corresponding to all feature combinations in matrix order. Each element in the vector corresponds to the interaction coefficient value at the same position in the initial interaction relationship matrix.

[0054] The element-by-element product operation multiplies the initial interaction matrix and the interaction coefficient vector at corresponding positions. Specifically, the calculation process multiplies the element in the i-th row and j-th column of the initial interaction matrix by the interaction coefficient value at the corresponding position in the interaction coefficient vector to obtain the new matrix element value. After the product operation is completed, the feature interaction coefficient matrix is ​​formed. Each element in the matrix reflects the mutual influence strength of the corresponding combination of local features and global features. Larger values ​​indicate stronger mutual influence.

[0055] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0056] The fused feature matrix and dynamic weight vector are input into the fused density clustering algorithm to calculate the parking space density. The density radius parameters of different parking areas are adjusted according to the local-global balance factor to obtain a density distribution map reflecting the degree of parking space aggregation.

[0057] Based on the density distribution map, core point identification is performed on parking spaces. Parking spaces with density values ​​exceeding a set threshold and containing the least number of parking spaces in their neighborhood are marked as core points, thereby obtaining a core point set containing core parking space identifiers and density values.

[0058] The similarity of parking spaces is calculated based on the core point set by integrating the distance measurement function. The comprehensive similarity between parking spaces is calculated by comprehensively considering the similarity of parking time, location proximity and POI influence similarity, and a similarity matrix reflecting the correlation strength of parking spaces is obtained.

[0059] The similarity matrix is ​​input into the cluster extension algorithm to divide the parking area. The clustering is extended from the core point to the parking spaces with high similarity. Parking spaces with similar parking behavior patterns are merged into the same cluster to obtain a fusion clustering result matrix including cluster number, parking space list, local dominance and global dominance.

[0060] Specifically, the fusion density clustering algorithm receives a fusion feature matrix and a dynamic weight vector as input data. First, the density calculation function is used to statistically calculate the number of parking spaces in the neighborhood around each parking space. The density calculation searches for the number of other parking spaces within a set radius with each parking space as the center point. The local-global balance factor dynamically adjusts the density radius parameter based on the ratio of the local feature intensity and the global feature intensity of each parking space in the fusion feature matrix. When the local features of the parking space dominate, the density radius is reduced to capture fine-grained local aggregation patterns. When the global features dominate, the density radius is increased to identify large-scale global aggregation patterns. The density distribution map is generated by spatially mapping the density values ​​of all parking spaces according to the geographic coordinates. The color depth or numerical value of each location point in the map reflects the degree of parking space aggregation at that location.

[0061] Core point identification is based on a threshold determination based on the density values ​​in the density distribution map. The density threshold is calculated based on the average density and standard deviation of the overall parking space distribution. Parking spaces with density values ​​exceeding the threshold are included in the candidate core point set. The minimum number of parking spaces in the neighborhood is checked by counting the number of other parking spaces within the neighborhood radius of the candidate core point. Only candidate points with a minimum number of parking spaces in their neighborhood are officially marked as core points. The core point set contains a unique identifier for each core parking space and its corresponding density value. The identifier is used for subsequent cluster expansion operations, and the density value is used to assess the importance of the core point.

[0062] A fusion distance metric function comprehensively calculates multi-dimensional similarities between parking spaces. Parking time similarity is calculated by normalizing the difference in the average historical parking times of two parking spaces. Location proximity is calculated by calculating the Euclidean distance between the geographic coordinates of two parking spaces and performing a reverse normalization. Point of Interest (POI) impact similarity is calculated by comparing the similarity in type and distance between the major POIs surrounding the two parking spaces. The comprehensive similarity calculation weights the similarity values ​​of the three dimensions according to preset weights. The weights are dynamically adjusted based on the characteristic distribution of different parking areas. A similarity matrix is ​​constructed by calculating the comprehensive similarity values ​​between all parking spaces. The rows and columns of the matrix represent different parking spaces, and the matrix elements represent the strength of association between corresponding pairs of parking spaces. The cluster expansion algorithm begins the expansion process with each core point in the core point set as a cluster seed. The expansion operation queries the similarity matrix to find additional parking spaces whose similarity with the current core point exceeds a set threshold and adds these parking spaces to the same cluster. The expansion process uses a breadth-first search strategy. If a newly added parking space also meets the core point criteria, the cluster continues to expand until no new spaces meet the inclusion criteria. Cluster numbers are assigned in the order in which the core points were discovered, and the parking space list records all parking space identifiers included in each cluster. Local dominance is calculated by taking the mean of the local eigenvalues ​​of the parking spaces within a cluster, while global dominance is calculated by taking the mean of the global eigenvalues ​​of the parking spaces within a cluster. These two dominance values ​​reflect the relative strength of the influence of local and global factors on the cluster.

[0063] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0064] The fusion clustering result matrix is ​​clustered and hierarchically processed according to the local dominance and global dominance. The clusters with high local dominance are divided into the local driving layer, and the clusters with high global dominance are divided into the global driving layer, thus obtaining a hierarchical clustering matrix containing two-layer classification identification.

[0065] Based on the hierarchical clustering matrix, the local driving layer is decomposed into the POI influence time series, and the short-term fluctuation trend and periodic changes affected by commercial facilities and public transportation are extracted to obtain a set of local spatiotemporal patterns reflecting the POI driving laws.

[0066] The global driver layer is analyzed and processed for network configuration time series based on the hierarchical clustering matrix to identify long-term trend changes and macro-flow patterns affected by road integration and connectivity, and to obtain a global spatiotemporal pattern set reflecting the impact of network configuration.

[0067] The local spatiotemporal pattern set and the global spatiotemporal pattern set are input into the pattern fusion algorithm for interactive effect calculation and processing. A prediction model that comprehensively considers local short-term fluctuations and global long-term trends is constructed through weighted fusion operations, and a parking behavior spatiotemporal fusion pattern library and demand forecast results that include spatiotemporal fusion rules and parking demand forecast data are obtained.

[0068] Specifically, the clustering stratification process compares the local dominance and global dominance values ​​of each cluster in the fused clustering result matrix to determine the stratification. When the local dominance value of a cluster is greater than the global dominance value, the cluster is marked as a local driving layer; when the global dominance value is greater than the local dominance value, the cluster is marked as a global driving layer. A two-layer classification identifier is generated by adding a hierarchical identifier to the original cluster number. The local driving layer cluster identifier is "L-cluster number" and the global driving layer cluster identifier is "G-cluster number", forming a hierarchical cluster matrix containing hierarchical information.

[0069] The POI impact time series decomposition process extracts cluster data marked as the local driving layer from the hierarchical clustering matrix, obtains historical parking occupancy time series data for each parking space in these clusters, and then classifies and groups the time series data based on the type of commercial facilities and public transportation stations surrounding the parking spaces. The time series decomposition uses an additive decomposition model to decompose the original time series into three components: trend component, cyclical component, and random component. The trend component is calculated using the moving average method to reflect the direction of long-term change. The cyclical component uses a periodicity detection algorithm to identify repetitive patterns such as daily and weekly cycles. The random component is the residual after subtracting the trend and cyclical components from the original data. Short-term fluctuation trends are extracted by analyzing the high-frequency change features in the cyclical component. Periodic changes are identified by statistically analyzing repetitive patterns at different time scales. These features are classified by POI type and stored in a local spatiotemporal pattern set.

[0070] The network configuration time series analysis process extracts cluster data marked as the global driver layer from the hierarchical clustering matrix, obtains historical parking flow time series data for each parking space within these clusters, and performs correlation analysis based on the integration and connectivity parameters of the roads where the parking spaces are located. Long-term trend changes are identified through linear regression analysis of the time series data. The regression slope reflects the long-term growth or decline in parking demand, and the regression intercept reflects the baseline parking demand level. Macro-flow patterns are identified by analyzing parking flow transfer patterns between different road network nodes. The flow correlation coefficient between parking spaces is calculated to quantify the strength of the mutual influence between nodes. These patterns are then classified according to network configuration characteristics and stored in a global spatiotemporal pattern set.

[0071] The pattern fusion algorithm receives local spatiotemporal pattern sets and global spatiotemporal pattern sets as input data to calculate interaction effects. The interaction effects are quantified by calculating the correlation coefficient between local pattern features and global pattern features. The weighted fusion operation multiplies the local short-term fluctuation characteristics by the dynamically assigned local weights, and multiplies the global long-term trend characteristics by the corresponding global weights. The two weighted results are then summed to obtain the fused prediction value. The weight allocation is dynamically adjusted based on the characteristics of different time periods and spatial regions. During weekdays, local weights are increased to highlight the impact of POIs, while during weekends, global weights are increased to highlight the impact of network flow. The parking behavior spatiotemporal fusion pattern library is constructed by storing fusion pattern features under different spatiotemporal conditions. The demand forecast result is calculated by inputting the current spatiotemporal conditions into the fusion prediction model.

[0072] For example, commercial areas contain local driving clusters centered around large shopping malls and global driving clusters centered around major road intersections. A hierarchical analysis labels parking clusters around shopping malls as the local driving layer due to their high local dominance, and parking clusters at road intersections as the global driving layer due to their high global dominance. Time series decomposition of POI impacts in the local driving layer reveals that shopping mall clusters exhibit distinct shopping-oriented fluctuations during weekday evenings and throughout the weekend, while subway station clusters exhibit a commuter-oriented bimodal pattern during weekday morning and evening rush hours. Time series decomposition of these patterns yields corresponding cyclical and trend components. Network configuration analysis of the global driving layer reveals that parking demand at major road intersections is highly correlated with overall urban traffic flow, with its long-term trend following changes in road network capacity. Macroscopic flow patterns reveal a clear shift in parking demand between intersections. The pattern fusion algorithm performs a weighted fusion of local shopping and commuting cycle patterns with global traffic flow trend patterns, increasing the weight of local patterns during peak shopping periods and the weight of global patterns during traffic congestion periods. The resulting fusion prediction model can simultaneously capture short-term demand fluctuations driven by local commercial activities and long-term demand trends driven by the global transportation network.

[0073] In a specific embodiment, the process of performing the POI impact time series decomposition process on the local driving layer based on the hierarchical clustering matrix may specifically include the following steps:

[0074] The parking clusters with high local dominance values ​​are screened from the hierarchical clustering matrix for data extraction and processing, and the historical occupancy time series of parking spaces and POI distance information are obtained to obtain the local driven parking time series dataset.

[0075] The local driven parking time series dataset is classified according to POI type, and the data groups are divided according to the influence of commercial facilities and public transportation stations to obtain the time series data subsets classified by POI type;

[0076] Based on the subset of time series data, the frequency domain decomposition is performed through Fourier transform to extract the spectral components of the daily cycle, weekly cycle and monthly cycle, and obtain the frequency domain feature matrix containing the periodic characteristics;

[0077] The time domain reconstruction processing is performed according to the frequency domain feature matrix, and the pattern classification is performed according to the POI influence intensity to obtain a set of local spatiotemporal patterns containing the POI driving rules.

[0078] Specifically, the data extraction process scans the local dominance field in the hierarchical clustering matrix to identify parking clusters with values ​​above a set threshold. Then, based on the cluster number, it extracts the corresponding parking space occupancy time series data from the historical parking database. This time series data records the occupancy status changes of each parking space on an hourly basis. Point of Interest (POI) distance information is extracted from the local point of interest database based on parking space identifiers. This information includes the distance from each parking space to the nearest commercial facility or public transportation stop, as well as the POI type identifier. This time series data is then correlated and matched with the POI information based on the parking space number to form a localized driven parking time series dataset.

[0079] The POI type classification process divides the localized driven parking time series dataset into different data groups based on the type identifier in the POI distance information. The commercial facility impact group includes parking space data closest to shopping malls, restaurants, and entertainment venues, while the public transportation station impact group includes parking space data closest to subway stations and bus stops. The classification process compares the distance values ​​of each parking space to different POI types to determine the dominant influencing factor and classifies the parking space into the POI type group with the closest distance, forming time series data subsets categorized by POI type. Each data subset contains time series data for multiple parking spaces affected by the same POI type.

[0080] The Fourier transform process converts each time series in the time series data subset from the time domain to the frequency domain for periodic analysis. The transformation process identifies periodic components by calculating the amplitude and phase information of the time series at different frequencies. The daily periodic spectral component is extracted by detecting the frequency components of the twenty-four-hour repeating pattern, the weekly periodic spectral component is extracted by detecting the frequency components of the seven-day repeating pattern, and the monthly periodic spectral component is extracted by detecting the frequency components of the thirty-day repeating pattern. Frequency domain decomposition decomposes the original time series into a combination of sine waves and cosine waves of different frequencies. The amplitude value of each frequency component reflects the intensity of the periodic pattern, and the phase value reflects the time offset of the periodic pattern. The frequency domain feature matrix is ​​constructed by arranging the spectral components of all parking spaces according to frequency and parking space number. The rows of the matrix represent different frequencies, and the columns represent different parking spaces.

[0081] The time-domain reconstruction process converts the main frequency components in the frequency-domain feature matrix back into time-domain signals through an inverse Fourier transform. The reconstruction process selects frequency components with larger amplitude values ​​for superposition calculation to reconstruct the time series. The POI influence strength is determined by calculating the correlation between the reconstructed signal and the POI distance. The closer the parking space is to the POI, the larger the amplitude of the corresponding periodic component in the reconstructed signal, indicating that the POI has a stronger influence. Pattern classification classifies the reconstructed time series patterns according to the influence strength of different POI types. The commercial POI-driven pattern is characterized by high-frequency use on weekends and evenings, while the transportation POI-driven pattern is characterized by bimodal characteristics during the morning and evening rush hours on weekdays. These pattern features are stored in a local spatiotemporal pattern set.

[0082] The above describes the parking behavior analysis method based on data mining in the embodiment of the present application. The following describes the parking behavior analysis system based on data mining in the embodiment of the present application. Figure 2 In the embodiments of the present application, an embodiment of a parking behavior analysis system based on data mining includes:

[0083] The acquisition module 201 is used to collect and process data from paid parking areas through a parking sensor network, obtain a basic data structure of parking behavior, and construct a local point of interest database and a global network configuration database;

[0084] A fusion module 202 is configured to perform a fusion process based on the parking behavior basic data structure, the local point of interest database, and the global network configuration database using a local-global feature interaction matrix algorithm to obtain a fusion feature matrix and a dynamic weight vector;

[0085] The clustering module 203 is used to perform clustering processing on the fusion feature matrix and the dynamic weight vector through a fusion density clustering algorithm to obtain a fusion clustering result matrix;

[0086] The identification module 204 is used to perform pattern recognition processing through double-layer spatiotemporal pattern mining based on the fusion clustering result matrix to obtain a spatiotemporal fusion pattern library of parking behavior and demand prediction results.

[0087] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the parking behavior analysis method based on data mining.

[0088] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. 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 make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A parking behavior analysis method based on data mining, characterized in that: The method comprises: The parking sensor network collects and processes data from paid parking areas to obtain the basic data structure of parking behavior, and constructs a local point of interest database and a global network configuration database; According to the parking behavior basic data structure, the local point of interest database and the global network configuration database, a fusion process is performed using a local-global feature interaction matrix algorithm to obtain a fusion feature matrix and a dynamic weight vector; Clustering the fusion feature matrix and the dynamic weight vector using a fusion density clustering algorithm to obtain a fusion clustering result matrix; Based on the fusion clustering result matrix, pattern recognition processing is performed through double-layer spatiotemporal pattern mining to obtain a parking behavior spatiotemporal fusion pattern library and demand prediction results.

2. The parking behavior analysis method based on data mining according to claim 1 is characterized in that: The method collects and processes data from paid parking areas through a parking sensor network to obtain a basic parking behavior data structure, and constructs a local point of interest database and a global network configuration database, including: The parking occupancy detection sensor monitors the status of each parking space and records the changes in parking space occupancy status at a 30-second acquisition frequency to obtain an original parking dataset containing vehicle identification, location identification, start time, end time, and parking duration. Inputting the original parking data set into the outlier detection algorithm for data cleaning, removing sensor fault data and using time series interpolation method to fill in missing data, thus obtaining a structured parking behavior basic data structure; Based on the geographic coordinates of parking spaces, POI information is collected and processed for commercial facilities, public transportation stations, and office buildings within a 500-meter radius. The Euclidean distance from each parking space to the nearest point of interest is calculated to obtain a local point of interest database. The parking areas are analyzed using spatial syntax according to the road network topology, and the road connectivity index, integration degree and selectivity parameters are calculated to obtain a global network configuration database that records the network configuration characteristics.

3. The parking behavior analysis method based on data mining according to claim 1, characterized in that: The parking behavior basic data structure, the local point of interest database, and the global network configuration database are fused using a local-global feature interaction matrix algorithm to obtain a fused feature matrix and a dynamic weight vector, including: The parking duration, location coordinates, and occupancy status in the parking behavior basic data structure are processed into a feature vector, and combined with the POI distance data in the local point of interest database and the network integration data in the global network configuration database to obtain a multidimensional feature vector of the parking behavior containing five-dimensional feature values; Based on the multidimensional feature vector of parking behavior, the interaction relationship is calculated and processed using a local-global feature interaction matrix algorithm. The local POI density feature and the global network integration feature are matrix-producted and multiplied by the interaction coefficient to obtain a feature interaction coefficient matrix that quantifies the strength of the mutual influence between the local and global features. The feature interaction coefficient matrix is ​​input into the sigmoid function for weight normalization processing, and the ratio of local feature weights to global feature weights is dynamically adjusted according to the POI density and network integration of different parking spaces to obtain a dynamic weight vector with a value range between 0 and 1; The multidimensional feature vector of the parking behavior is weightedly fused according to the dynamic weight vector, and the local feature value multiplied by the local weight and the global feature value multiplied by the global weight are summed to obtain a fused feature matrix containing the fused feature value.

4. The parking behavior analysis method based on data mining according to claim 3 is characterized in that: The interaction relationship calculation process is performed based on the multidimensional feature vector of parking behavior using a local-global feature interaction matrix algorithm. The local POI density feature and the global network integration feature are matrix-producted and multiplied by the interaction coefficient to obtain a feature interaction coefficient matrix that quantifies the mutual influence strength between the local and global features, including: Extracting local POI density eigenvalues ​​and global network integration eigenvalues ​​from the multidimensional feature vector of parking behavior to perform feature separation processing, constructing local feature subvectors and global feature subvectors respectively, and obtaining dimensionally matched local feature matrices and global feature matrices; Performing a transposed matrix product operation on the local feature matrix and the global feature matrix, calculating the product value of each local feature and each global feature according to the matrix multiplication rule, and obtaining an initial interaction relationship matrix; Performing interaction coefficient training on the initial interaction relationship matrix based on historical parking data, calculating the impact of each feature combination on parking behavior through regression analysis, and obtaining an interaction coefficient vector reflecting the impact weight of different feature combinations; The initial interaction relationship matrix is ​​multiplied element by element with the interaction coefficient vector, and each element in the matrix is ​​multiplied by the corresponding interaction coefficient value to obtain a feature interaction coefficient matrix that quantifies the strength of the mutual influence between local and global features.

5. The parking behavior analysis method based on data mining according to claim 1, characterized in that: The fusion feature matrix and the dynamic weight vector are clustered by a fusion density clustering algorithm to obtain a fusion clustering result matrix, including: Inputting the fused feature matrix and the dynamic weight vector into a fused density clustering algorithm to calculate parking space density, adjusting density radius parameters of different parking areas according to a local-global balance factor, and obtaining a density distribution map reflecting the degree of parking space aggregation; Performing core point identification processing on the parking spaces based on the density distribution map, marking parking spaces with density values ​​exceeding a set threshold and having the least number of parking spaces in their neighborhood as core points, and obtaining a core point set including core parking space identifiers and density values; Performing parking space similarity calculation based on the core point set by integrating a distance metric function, calculating the comprehensive similarity between parking spaces by comprehensively considering parking time similarity, location proximity, and POI influence similarity, and obtaining a similarity matrix reflecting the association strength of parking spaces; The similarity matrix is ​​input into the cluster extension algorithm to perform parking area division processing. Cluster extension is performed from the core point as the starting point to the parking spaces with high similarity. Parking spaces with similar parking behavior patterns are merged into the same cluster to obtain a fusion clustering result matrix including cluster number, parking space list, local dominance and global dominance.

6. The parking behavior analysis method based on data mining according to claim 1, characterized in that: The method of performing pattern recognition processing based on the fusion clustering result matrix through double-layer spatiotemporal pattern mining to obtain a parking behavior spatiotemporal fusion pattern library and demand prediction results includes: The fusion clustering result matrix is ​​clustered and hierarchically processed according to the local dominance and the global dominance, the clusters with high local dominance are divided into the local driving layer, and the clusters with high global dominance are divided into the global driving layer, to obtain a hierarchical clustering matrix containing a double-layer classification identifier; Based on the hierarchical clustering matrix, the local driving layer is decomposed into POI influence time series, the short-term fluctuation trend and periodic changes affected by commercial facilities and public transportation are extracted, and a local spatiotemporal pattern set reflecting the POI driving law is obtained; Performing a network configuration time series analysis on the global driver layer based on the hierarchical clustering matrix to identify long-term trend changes and macro-flow patterns affected by road integration and connectivity, and obtaining a global spatiotemporal pattern set reflecting the impact of network configuration; The local spatiotemporal pattern set and the global spatiotemporal pattern set are input into the pattern fusion algorithm for interactive effect calculation and processing, and a prediction model that comprehensively considers local short-term fluctuations and global long-term trends is constructed through weighted fusion operations to obtain a parking behavior spatiotemporal fusion pattern library and demand prediction results that include spatiotemporal fusion laws and parking demand prediction data.

7. The parking behavior analysis method based on data mining according to claim 6, characterized in that: The POI influence time series decomposition process is performed on the local driving layer based on the hierarchical clustering matrix to extract the short-term fluctuation trend and periodic changes affected by commercial facilities and public transportation, and obtain a local spatiotemporal pattern set reflecting the POI driving law, including: Screening parking clusters with high local dominance values ​​from the hierarchical clustering matrix for data extraction and processing, obtaining parking space historical occupancy time series and POI distance information, and obtaining a local driven parking time series dataset; Classifying the local drive parking time series data set according to POI type, dividing the data group according to the influence of commercial facilities and public transportation stations, and obtaining a time series data subset classified by POI type; Performing frequency domain decomposition processing based on the time series data subset through Fourier transform, extracting the spectral components of the daily cycle, weekly cycle and monthly cycle, and obtaining a frequency domain feature matrix containing periodic features; A time domain reconstruction process is performed according to the frequency domain feature matrix, and pattern classification is performed according to the POI influence intensity to obtain a local spatiotemporal pattern set containing POI driving rules.

8. A parking behavior analysis system based on data mining, characterized in that: For implementing the parking behavior analysis method based on data mining according to any one of claims 1 to 7, the parking behavior analysis system based on data mining comprises: The acquisition module is used to collect and process data from paid parking areas through the parking sensor network, obtain the basic data structure of parking behavior, and build a local point of interest database and a global network configuration database; a fusion module, configured to perform fusion processing on the parking behavior basic data structure, the local point of interest database, and the global network configuration database by using a local-global feature interaction matrix algorithm to obtain a fusion feature matrix and a dynamic weight vector; A clustering module is used to cluster the fusion feature matrix and the dynamic weight vector through a fusion density clustering algorithm to obtain a fusion clustering result matrix; The recognition module is used to perform pattern recognition processing through double-layer spatiotemporal pattern mining based on the fusion clustering result matrix to obtain a spatiotemporal fusion pattern library of parking behavior and demand prediction results.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the parking behavior analysis method based on data mining according to any one of claims 1 to 7.

Images