Tidal parking group intelligence optimization method based on geohash algorithm

Abstract

The application relates to a tidal parking group intelligence optimization method based on a GeoHash algorithm, which comprises the following steps: step S1: analyzing parking area geographic coordinate source data as sample data to obtain a longitude and latitude range corresponding to the geographic coordinate source data; step S2: according to the longitude and latitude range corresponding to the geographic coordinate source data, using region coding segmentation of the GeoHash algorithm and then through Base32 coding conversion, converting two-dimensional longitude and latitude into one-dimensional strings to obtain coordinate region point space distribution coding in the form of strings; step S3: according to a geographic coordinate point density clustering method, executing a density clustering algorithm based on GeoHash to obtain corresponding coordinate coding clusters, and the union of all the clusters being a coordinate coding result cluster of global density clustering of a feature region; and step S4: according to the obtained coordinate coding result cluster, using a KNN classification recognition algorithm to perform parking group intelligence optimization selection. The application realizes parking group intelligence optimization selection.

Description

Technical Field

[0001] This invention relates to the field of data processing methods, specifically to a tidal parking crowdsourcing optimization method based on the GeoHash algorithm. Background Technology

[0002] In recent years, various government agencies and institutions in cities have taken numerous measures to address the urban parking problem. One of the most visible and significant measures is in urban planning, such as widening roads and providing a large number of parking spaces. However, a significant "tidal phenomenon" still exists. This "tidal phenomenon" refers to the pattern of changes in urban pedestrian and vehicle traffic at different times of the day. During the morning rush hour, a large number of vehicles enter commercial areas and office buildings, while during the evening rush hour, these vehicles surge into residential areas like a tide. This phenomenon is one of the main sources of urban traffic pressure. The above problem is actually caused by the contradiction between the demand for parking spaces and the supply and demand of parking spaces. On the one hand, the "tidal phenomenon" during peak hours leads to huge differences in the demand for parking spaces in different areas; on the other hand, within a certain period of time, the trend of user travel demand tends to be similar, and the demand for parking spaces fluctuates greatly in a short period of time, resulting in a spatial and temporal imbalance between the supply and demand of parking spaces. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a tidal parking crowd optimization method based on the GeoHash algorithm to achieve intelligent crowd optimization selection of parking spaces.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A crowdsourced optimization method for tidal parking based on the GeoHash algorithm includes the following steps: Step S1: Parse the geographic coordinate source data of the parking area as sample data to obtain the latitude and longitude range corresponding to the geographic coordinate source data; Step S2: Based on the latitude and longitude range corresponding to the geographic coordinate source data, the region is segmented using the GeoHash algorithm and then converted using Base32 encoding to transform the two-dimensional latitude and longitude into a one-dimensional string, thus obtaining the spatial distribution code of the coordinate region points in string form; Step S3: Based on the geographic coordinate point density clustering method, execute the GeoHash-based density clustering algorithm to obtain the corresponding coordinate-coded cluster set. The union of all cluster sets is the coordinate-coded result cluster set of the global density clustering of the feature region. Step S4: Based on the obtained coordinate encoding result clusters, use the KNN classification and recognition algorithm to perform intelligent optimization selection of parking spaces.

[0005] Furthermore, the geographic coordinate source data of the parking area includes parking area gate data, vehicle entry and exit data, and vacant parking space ratio data; pedestrian flow data around the characteristic parking area and motor vehicle trajectory distribution data.

[0006] Furthermore, step S2 specifically includes: Step S2.1: GeoHash encoding decomposes the region into multiple sub-blocks based on different precision levels, and uses GeoHash encoding to solve the problem of collecting and retrieving the latitude and longitude of points; Step S2.2: Divide the space into two dimensions, latitude and longitude. If a point falls in the left interval after the division, record it as 0; if a point falls in the right interval after the division, record it as 1. Step S2.3: Recursively follow the partitioning process in step S2.2. As the interval [a, b] gradually shrinks with each iteration, the region will get closer and closer to point P. The higher the precision, the smaller the geographical area represented by the GeoHash string, and the more accurate the location represented by the GeoHash string. Step S2.4: Group 5 bits of binary data into a single visible character based on Base32 information. As the number of visible characters increases, the range [a, b] will gradually expand towards point P.

[0007] Furthermore, step S3 specifically includes: Step S3.1: Process each coordinate point p using the GeoHash algorithm to obtain its GeoHash code g, and add the coordinate point p to the coordinate point set PS of the corresponding real block ai according to g, finally forming the real block set B = {a1, a2, …, am}; Step S3.2: Using the adjacency relationship of blocks, process the set of real blocks B to generate a set of connected real block regions Ω = {M1, M2, …, Mn}; Step S3.3: For each coordinate point in a connected real block region Mi in Ω, perform a density clustering algorithm to obtain the corresponding cluster set Ci; Step S3.4: Merge the clusters generated in step S3.3 to obtain the final global density clustering result; where ai = (b, PS), b represents the GeoHash code corresponding to the block, and PS represents the set of points contained in the block. If ai . PS ≠ ∅, then ai is called a real block; otherwise, ai is called an empty block. The length l and width w of the block are determined by the length ε of the Geohash string. Assume that the distance threshold in density clustering is δ, and satisfies δ <min{l ,w}。

[0008] Step S3.5: Divide the geospatial area using GeoHash to form rectangular regions of fixed size, called blocks. Each block has a unique GeoHash code, denoted as bi = (g, PS), where g represents the GeoHash code corresponding to the block, and PS represents the set of points contained in the block. If ai . PS ≠ ∅, then bi is called a real block; otherwise, bi is called an empty block. The length l and width w of the block are determined by the length ε of the GeoHash string. Assume that the distance threshold in density clustering is δ, and satisfies δ <min{l ,w}。

[0009] Furthermore, step S4 specifically includes: Step S4.1: Select vehicle stops in the region with stable density distribution fluctuations and close to the divided region, and select several vehicle stops after calculating and classifying the region codes; Step S4.2: Based on the principle of minimum Euclidean distance, the clustering results are divided into candidate vehicle stops. Assume that the given vehicle stop dataset X={x1, x2,···, xN}, xi (i=1,2,···,N) is a two-dimensional variable, and a test point q =(x, y)∈M((x, y)), where dataset X represents the vehicle stops that vehicles can enter in the encoded region, and (x,y) represents the latitude and longitude coordinates; Step S4.3: Using Euclidean distance as the criterion, find N nearest neighbors of q in set X, i.e., determine N candidate vehicle stopping points, minimizing the absolute difference between data points. Euclidean distance calculation formula: In the formula: d(xi, xj) represents the distance between the test set and the training set, and (xi, xj) are the coordinates of the data point (i, j); Step S4.4: Calculate the similarity between the point to be tested q and the selected K candidate docking points respectively; Step S4.5: Filter out the nearest neighbors of the category and sum their similarities. The following formula represents the total similarity between the point to be tested q and category Ci: In the formula, Tj represents the j-th K-nearest neighbor of the test sample q. The total similarity between the test sample q and different categories is calculated and compared. The test sample q belongs to the category with the largest total similarity, and the final category expression of q is obtained: .

[0010] Furthermore, for the feature regions with missing coordinate data sources, a parking availability approximation model combining a semi-supervised recurrent graph neural network is adopted. By combining real-time parking availability data from other surrounding feature regions, the missing parking availability is approximated from both spatial and temporal perspectives through the introduced propagation convolutional block and GRU module.

[0011] Furthermore, the parking availability approximation model, which incorporates a semi-supervised recurrent graph neural network and real-time parking availability data from other surrounding feature areas, is used to capture the missing parking availability from both spatial and temporal perspectives by introducing a propagation convolutional block and reusing the GRU module. Specifically: (1): Define a model expression: In the formula: P l This represents a set of parking lots with sensor data capabilities. P u p1, p2...pn represents the set of parking lots without sensor data capability, while p1, p2...pn represents the union of all parking lots; (2) Given a parking lot Pi∈P, at time step (t), the parking availability (pa), Let i be the number of vacant parking spaces defined in p; (3) Use This indicates that the observation of parking lot P in time step t is... l Changes. Thus, by utilizing contextual data in parking lot P and partially observed real-time parking availability data l for Pl, the number of pa values ​​i for all parking lots P can be predicted; (4) For parking lots that lack real-time parking availability, parking capacity is estimated in terms of time and space based on parking lots with real-time parking availability information. A parking availability distribution estimate is learned, and finally, the time distribution estimate and the spatial distribution estimate are fused through an entropy-based mechanism to obtain the final parking availability distribution estimate.

[0012] Furthermore, step (4) specifically includes: Spatial approximation: The propagation convolutional block is used to compute parking availability in space, similar to context graph convolution, but with more relaxed handling of connectivity between nodes, thereby expanding the connectivity of the adjacency matrix. For approximate time, the parking capability in time is directly reused from the GRU module, using the hidden state of the previous time period.

[0013] Entropy-based fusion is a better approach than direct averaging of PA distributions. Because the uncertainty of fewer approximations is more critical, an estimated PA distribution with the smaller entropy value is defined, and its entropy expression is: Where xi(j) represents the j-th dimension i of x; Fusion and The expression is derived from the PA distribution: Through the and The fused PA approximate distribution is used for two tasks: first, it is concatenated into the previously learned context graph convolution representation, and then input into soft clustering convolutional blocks to learn latent node representations; second, it is combined with the outputs of context graph convolution (CxtConv block) and soft clustering graph convolution (SCConv block) as latent nodes, as the overall representation of each parking lot at time step t, and input into the GRU module for computation to generate the final PA prediction result.

[0014] Compared with the prior art, the present invention has the following advantages: 1. This invention enables intelligent optimization of parking selection to solve the "tidal phenomenon" of vehicle parking; 2. This invention addresses parking space availability in characteristic areas where IoT operational big data cannot be provided. Based on GCN (Graph Convolutional Neural Network), it proposes a parking availability approximation module that combines a semi-supervised recurrent graph neural network (SHARE). By combining real-time parking availability data from other surrounding characteristic areas, it captures the availability of missing parking data from both spatial and temporal dimensions through the introduction of propagation convolutional blocks and the reuse of GRU modules. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a comparison of the precision representation range after GeoHash transformation partitioning in one embodiment of the present invention; Figure 3 This is the precision obtained before base32 encoding based on GeoHash encoding in one embodiment of the present invention; Figure 4 In one embodiment of the present invention, the GeoHash precision change is based on ( Figure 3 Comparison after base32 encoding. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] Please refer to Figure 1 This invention provides a crowdsourced optimization method for tidal parking based on the GeoHash algorithm, comprising the following steps: Step S1: Parse the geographic coordinate source data of the parking area as sample data to obtain the latitude and longitude range corresponding to the geographic coordinate source data; Step S2: Based on the latitude and longitude range corresponding to the geographic coordinate source data, the region is segmented using the GeoHash algorithm and then converted using Base32 encoding to transform the two-dimensional latitude and longitude into a one-dimensional string, thus obtaining the spatial distribution code of the coordinate region points in string form; In this embodiment, step S2 specifically includes: Step S2.1: GeoHash encoding decomposes the region into multiple sub-blocks based on different precision levels. Sub-blocks within a certain latitude and longitude range will have the same encoding. Under appropriate precision, GeoHash encoding is used to solve the problem of collecting and retrieving the latitude and longitude of points.

[0018] Step S2.2: The GeoHash encoding method divides the space into two dimensions: latitude and longitude. If a point falls in the left interval after the division, it is recorded as 0; if the point falls in the right interval, it is recorded as 1. For example, the latitude and longitude interval [a, b] of the point P to be encoded is reduced, and the interval [a, b] is continuously divided into left and right parts. If the point belongs to the left half of the interval, it is recorded as binary 0; if it belongs to the right half of the interval, it is recorded as 1.

[0019] Step S2.3: Recursively follow the partitioning process in step 2.2. With each iteration, the interval [a, b] gradually shrinks, and its region gets closer and closer to point P. Higher precision results in a smaller geographical area represented by the GeoHash string, and more accurate location representation. Specific precision ranges are as follows: Figure 2 .

[0020] Step S2.4: As Figure 3 and Figure 4 As shown, GeoHash encoding uses 5 bits of binary data as a group. Based on Base32 information, it can be converted into a single visible character. As the number of visible characters (GeoHash encoding precision) increases, the range [a, b] gradually expands towards point P. The GeoHash algorithm transforms the latitude and longitude grid into small regions, generating the same GeoHash string for points falling within the same region. By converting two-dimensional latitude and longitude data into a one-dimensional string, it simplifies the complexity of geolocation operations.

[0021] Step S3: Based on the geographic coordinate point density clustering method, execute the GeoHash-based density clustering algorithm to obtain the corresponding coordinate-coded cluster set. The union of all cluster sets is the coordinate-coded result cluster set of the global density clustering of the feature region. In this embodiment, step S3 specifically includes: Step S3.1: Process each coordinate point p using the GeoHash algorithm to obtain its GeoHash code g, and add the coordinate point p to the coordinate point set PS of the corresponding real block ai according to g, finally forming the real block set B = {a1, a2, ..., am}; Step S3.2: Using the adjacency relationship of blocks, process the set of real blocks B to generate a set of connected real block regions Ω = {M1, M2, …, Mn}; Step S3.3: For the coordinate points in each connected real block region Mi in Ω, perform density clustering algorithm to obtain the corresponding cluster set Ci; Step S3.4: Merge the clusters generated in step S3.3 to obtain the final global density clustering result; where ai = (b, PS), b represents the GeoHash code corresponding to the block, and PS represents the set of points contained in the block. If ai . PS ≠ ∅, then ai is called a real block; otherwise, ai is called an empty block. The length l and width w of the block are determined by the length ε of the Geohash string. Assume that the distance threshold in density clustering is δ, and satisfies δ <min{l ,w}。

[0022] Step S3.5: Divide the geographic space using GeoHash to form rectangular regions of fixed size, called blocks. Each block has a unique GeoHash code, denoted as bi = (g, PS), where g represents the GeoHash code corresponding to the block, and PS represents the set of points contained in the block. If bi . PS ≠ ∅, then bi is called a real block; otherwise, bi is called an empty block. The length l and width w of the block are determined by the length ε of the GeoHash string. Assume that the distance threshold in density clustering is δ, and satisfies δ <min{l ,w}。

[0023] The advantages of density clustering algorithms are that they embody the divide-and-conquer approach, are easy to implement, significantly improve the time efficiency of density clustering, and ensure the correctness of the results.

[0024] Step S4: Based on the obtained coordinate encoding result clusters, use the KNN classification and recognition algorithm to perform intelligent optimization selection of parking spaces.

[0025] In this embodiment, step S4 specifically includes: Select vehicle stops with stable density distribution and close proximity to the divided area. Select several vehicle stops after the region is coded and the clustering results are included as candidate vehicle stops based on the principle of minimizing Euclidean distance. Assume a given vehicle stop dataset X={x1, x2, ..., xN}, xi (i=1,2, ...,N) is a two-dimensional variable, and a test point q =(x, y)∈M((x, y)), where dataset X represents the vehicle stops accessible to vehicles in the coded area. (x, y) represent latitude and longitude coordinates. Using Euclidean distance as the criterion, find N nearest neighbors of q in set X, i.e., determine N candidate vehicle stops, minimizing the absolute difference between data points. The Euclidean distance calculation formula is: In the formula: d(xi, xj) represents the distance between the test set and the training set, and (xi, xj) are the coordinates of the data point (i, j). The similarity between the test point q and the selected K candidate docking points is calculated respectively. Then, the category nearest neighbors are selected and their similarities are summed. The following formula represents the total similarity between the test point q and category Ci: In the formula: Tj represents the j-th K-nearest neighbor of the test sample q. The total similarity of the test sample q with different categories is calculated and compared. The test sample q belongs to the category with the highest total similarity, thus yielding the final category expression for q. The advantage of the KNN classification algorithm is that it can quickly and accurately identify parking space location data with obvious weak features, and combine the connectivity between parking spaces and the supply and demand of parking spaces to complete the best analysis of parking space location data.

[0026] Preferably, in this embodiment, for the feature region of missing coordinate data source, a parking availability approximation module combining a semi-supervised recurrent graph neural network (SHARE) proposed by GCN (Graph Convolutional Neural Network) is used, combined with real-time parking availability data of other surrounding feature regions, to capture the missing parking availability from both spatial and temporal aspects by introducing a propagation convolution block and reusing the GRU module.

[0027] Define a model expression: In the formula: P l This represents a set of parking lots with sensor data capabilities. P up1, p2...pn represents the set of parking lots without sensor data capability, while p1, p2...pn represents the union of all parking lots; Given a parking lot Pi∈P, at time step (t), parking availability (pa), Let i be the number of vacant parking spaces defined in p.

[0028] use Let represent the change in Pl of parking lot P observed over time step t. This allows for the prediction of the number of pa values ​​i for all parking lots P by utilizing contextual data from parking lot P and partially observed real-time parking availability data l for Pl.

[0029] For parking lots lacking real-time parking availability, parking capacity is estimated from both temporal and spatial perspectives, relying on parking lots with real-time parking availability information. Instead of directly obtaining specific parking capacity, it learns a parking availability distribution estimate. Finally, through an entropy-based mechanism, the temporal and spatial distribution estimates are fused to obtain the final parking availability distribution estimate.

[0030] Preferably, spatially approximate, propagating convolutional blocks are used to compute parking availability in space, similar to context graph convolution, but with more relaxed handling of connectivity between nodes, thereby expanding the connectivity of the adjacency matrix.

[0031] For approximate time, the parking capability in time is directly reused from the GRU module, using the hidden state of the previous time period.

[0032] Entropy-based fusion is a better approach than direct averaging of PA distributions. Because the uncertainty of fewer approximations is more critical, an estimated PA distribution with the smaller entropy value is defined, and its entropy expression is: Where xi(j) represents the j-th dimension i of x; Fusion and The expression obtained after applying the PA distribution is: Through the and The fused PA approximate distribution is used for two tasks: first, it is concatenated into the previously learned context graph convolution representation, and then input into soft clustering convolutional blocks to learn latent node representations; second, it is combined with the outputs of context graph convolution (CxtConv block) and soft clustering graph convolution (SCConv block) as latent nodes, as the overall representation of each parking lot at time step t, and input into the GRU module for computation to generate the final PA prediction result.

[0033] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be included in the scope of the present invention.

Claims

1. A crowdsourced intelligent optimization method for tidal parking based on the GeoHash algorithm, characterized in that, Includes the following steps: Step S1: Parse the geographic coordinate source data of the parking area as sample data to obtain the latitude and longitude range corresponding to the geographic coordinate source data; Step S2: Based on the latitude and longitude range corresponding to the geographic coordinate source data, the region is segmented using the GeoHash algorithm and then converted using Base32 encoding to transform the two-dimensional latitude and longitude into a one-dimensional string, thus obtaining the spatial distribution code of the coordinate region points in string form; Step S3: Based on the geographic coordinate point density clustering method, execute the GeoHash-based density clustering algorithm to obtain the corresponding coordinate-coded cluster set. The union of all cluster sets is the coordinate-coded result cluster set of the global density clustering of the feature region. Step S4: Based on the obtained coordinate encoding result clusters, use the KNN classification and recognition algorithm to perform intelligent optimization selection of parking spaces; Step S4 specifically involves: Step S4.1: Select vehicle stops in the region with stable density distribution fluctuations and close to the divided region, and select several vehicle stops after calculating and classifying the region codes; Step S4.2: Based on the principle of minimum Euclidean distance, the clustering results are divided into candidate vehicle stop points. Assume a given vehicle stop point dataset X={x1, x2, ..., x...} N }, x i Let be a two-dimensional variable, i=1,2,···,N, and a test point q =(x, y)∈M((x, y)), where dataset X represents the vehicle parking points that vehicles can enter in the encoded region, and (x, y) represents the latitude and longitude coordinates; Step S4.3: Using Euclidean distance as the criterion, find N nearest neighbors of q in set X, i.e., determine N candidate vehicle stopping points, minimizing the absolute difference between data points. Euclidean distance calculation formula: In the formula: d(x) i x j (x) represents the distance between the test set and the training set. i x j ) represents the coordinates of the data point (i, j); Step S4.4: Calculate the similarity between the test point q and the selected K candidate docking points respectively; Step S4.5: Filter out the nearest neighbors of the category and sum their similarities. The following formula represents the similarity between the test point q and the category C. i Total similarity: In the formula, T j' Let represent the j'-th K nearest neighbor of test point q. Calculate and compare the total similarity between test point q and different categories. Test point q belongs to the category with the highest total similarity, thus deriving the final category expression for q: 。 2. The tidal parking crowdsourcing optimization method based on the GeoHash algorithm according to claim 1, characterized in that, The geographic coordinate source data for the parking area includes parking area gate data, vehicle entry and exit data, vacant parking space ratio data, pedestrian flow data around the characteristic parking area, and motor vehicle trajectory distribution data.

3. The tidal parking crowdsourcing optimization method based on the GeoHash algorithm according to claim 1, characterized in that, Step S2 specifically involves: Step S2.1: GeoHash encoding decomposes the region into multiple sub-blocks based on different precision levels, and uses GeoHash encoding to solve the problem of collecting and retrieving the latitude and longitude of points; Step S2.2: Divide the space into two dimensions, latitude and longitude. If a point falls in the left interval after the division, record it as 0; if a point falls in the right interval after the division, record it as 1. Step S2.3: Recursively follow the partitioning process in step S2.

2. As the interval [a, b] gradually shrinks with each iteration, the region will get closer and closer to point P. The higher the precision, the smaller the geographical area represented by the GeoHash string, and the more accurate the location represented by the GeoHash string. Step S2.4: Group 5 bits of binary data into a single visible character based on Base32 information. As the number of visible characters increases, the range [a, b] will gradually expand towards point P.

4. The tidal parking crowdsourcing optimization method based on the GeoHash algorithm according to claim 1, characterized in that, Step S3 specifically involves: Step S3.1: Process each coordinate point p using the GeoHash algorithm to obtain its GeoHash code g, and add the coordinate point p to the coordinate point set PS of the corresponding real block ai according to g, finally forming the real block set B = {a1, a2, ..., am}; Step S3.2: Using the adjacency relationship of blocks, process the set of real blocks B to generate the set of connected real block regions Ω = {M1, M2, …, Mn}; Step S3.3: For each coordinate point in a connected real block region Mi in Ω, perform a density clustering algorithm to obtain the corresponding cluster set Ci; Step S3.4: Merge the clusters generated in step S3.3 to obtain the final global density clustering result; where ai = (b, PS), b represents the GeoHash code corresponding to the block, and PS represents the set of points contained in the block; the length l and width w of the block are determined by the length ε of the Geohash string. Assume that the distance threshold in density clustering is δ, and satisfies δ < min{l, w}; Step S3.5: Use GeoHash to divide the geographic space into rectangular areas of fixed size, which are called blocks. Each block has a unique GeoHash code, denoted as bi = (g, PS), where g represents the GeoHash code corresponding to the block and PS represents the set of points contained in the block. If bi .PS ≠ ∅, then bi is called a real block; otherwise, bi is called an empty block.

5. The tidal parking crowdsourcing optimization method based on the GeoHash algorithm according to claim 1, characterized in that, For feature regions with missing coordinate data sources, a parking availability approximation model combining a semi-supervised recurrent graph neural network is adopted. This model combines real-time parking availability data from other surrounding feature regions to capture the missing parking availability from both spatial and temporal perspectives through the introduction of propagation convolutional blocks and GRU modules.

6. The tidal parking crowdsourcing optimization method based on the GeoHash algorithm according to claim 5, characterized in that, The proposed parking availability approximation model combines a semi-supervised recurrent graph neural network with real-time parking availability data from other surrounding feature areas. It captures the missing parking availability from both spatial and temporal perspectives using an introduced propagation convolutional block and GRU module. Specifically: (1): Define a model expression: P=P l ∪P u ={p1,p2...p n' } In the formula: P l P represents a set of parking lots with sensor data capabilities. u This represents a set of parking lots without sensor data capabilities, where p1, p2...p n' It is represented as the union of all parking lots; (2) Given a parking lot P i' ∈P, at time step t, parking availability pa, It is defined as the number of vacant parking spaces in the i'th parking lot at time step t; (3) Use This indicates that the observation of parking lot P in time step t is... l Changes; thus by utilizing contextual data in parking lot P and the changes in P l The number of observed real-time parking availability data l is used to predict the number of parking availability pa values ​​for all parking lots P. (4) For parking lots that lack real-time parking availability, parking capacity is estimated in terms of time and space based on parking lots with real-time parking availability information. A parking availability distribution estimate is learned, and finally, the time distribution estimate and the spatial distribution estimate are fused through an entropy-based mechanism to obtain the final parking availability distribution estimate.

7. The tidal parking crowdsourcing optimization method based on the GeoHash algorithm according to claim 6, characterized in that, For parking lots lacking real-time parking availability, parking capacity is estimated temporally and spatially based on parking lots with real-time parking availability information. A parking availability distribution estimate is learned, and finally, an entropy-based mechanism is used to fuse the temporal and spatial distribution estimates to obtain the final parking availability distribution estimate. Specifically: Spatial approximation, propagating convolutional blocks are used to compute parking availability in space; The time approximation, the time-based parking capability approximation, directly reuses the GRU module, utilizing the hidden state of the previous time period; The entropy mechanism is fused to obtain an approximate PA distribution, which is used for two tasks. First, it is concatenated into the previously learned context graph convolution representation and then input into soft clustering to learn potential node representations. Second, it is combined with the outputs of the context graph convolution and the soft clustering graph convolution as the overall representation of each parking lot at time step t, and then input into the GRU module to generate the final PA prediction result.

8. The tidal parking crowdsourcing optimization method based on the GeoHash algorithm according to claim 7, characterized in that, For the smaller entropy value, define an estimated PA distribution, whose entropy expression is: in m represents the h-th dimension of x; Fusion and The expression is derived from the PA distribution: Fusion through entropy mechanism and The obtained PA approximate distribution is used for two tasks: first, it is concatenated into the previously learned context graph convolution representation and then input into soft clustering to learn latent node representations; second, it is combined with the outputs of the context graph convolution and soft clustering graph convolution as latent nodes, serving as the overall representation of each parking lot at time step t, and then input into the GRU module for computation to generate the final PA prediction result.

Images