CN111489064B

CN111489064B - Df-PBS (direct-flow-coupled system) -oriented public bicycle station dynamic planning method and system

Info

Publication number: CN111489064B
Application number: CN202010226676.8A
Authority: CN
Inventors: 陈建国; 李肯立; 刘刚; 彭继武; 李克勤
Original assignee: Hunan University
Current assignee: Hunan University
Priority date: 2020-03-27
Filing date: 2020-03-27
Publication date: 2023-05-12
Anticipated expiration: 2040-03-27
Also published as: CN111489064A

Abstract

The invention discloses a Df-PBS system-oriented public bicycle station dynamic planning method, wherein the adopted bicycle parking point clustering method can fully utilize the position information record of a large-scale historical public bicycle to mine the actual bicycle running renting requirement, running track and bicycle parking point distribution condition; the method for constructing the bicycle station point diagram model and the diagram sequence model can effectively establish and analyze an abstract model of a bicycle station, and can effectively capture the updating and changing conditions of the bicycle station position and the bicycle quantity in the station between different time periods; the public bicycle station prediction method based on the gate control map neural network can fully utilize the deep learning technology to mine the large-scale map structure data, and achieve accurate prediction of bicycle station layout; the bicycle station dynamic planning method can effectively ensure that the recommended bicycle station layout is beneficial to city management, and meanwhile, the income of suppliers and the convenience of users are improved.

Description

Df-PBS (direct-flow-coupled system) -oriented public bicycle station dynamic planning method and system

Technical Field

The invention belongs to the technical field of traffic network data mining, and particularly relates to a public bicycle station dynamic planning method and system for a Df-PBS system.

Background

By virtue of the advantages of zero carbon emissions and convenient driving experience, public bicycles (sharing bicycles) have significant advantages in urban short-distance travel and are widely used as public transportation to solve the problem of the last kilometer. In many cities around the world, there are many public bicycle sharing (SD-PBS) systems based on parking spots/piles to provide public bicycles to citizens. In the SD-PBS system, there are multiple stationary parking posts per bicycle station, which greatly limits the mobility of the bicycle station locations and the proliferation of the number of bicycles in each station, thereby limiting the convenience of users to rent and return public bicycles.

In recent years, a pile-free public bicycle sharing (Df-PBS) system has begun to be popularized in china and many other countries, which provides convenient and very personalized services during bicycle renting and returning. As shown in fig. 1, the Df-PBS system includes a public bike, a non-stub bike parking spot (also referred to as a bike station), a global positioning system (Global Positioning System, GPS for short), a Quick Response (QR) code-based bicycle lock module, and a mobile Application (APP), wherein:

(1) Public bicycle: public bicycles have a unique appearance that can be quickly identified by a user. Each bicycle is equipped with a GPS for recording its location information in real time and a QR code based bicycle lock module.

(2) Pile-free bicycle parking spot: df-PBS suppliers deploy different numbers of public bicycles in areas of various cities where parking of bicycles is allowed (e.g., beside roads, at park entrances, community entrances, or near shopping centers). Each dense bicycle parking spot is referred to as a non-stake bicycle parking spot.

(3) QR code: each bicycle has a unique QR code that the user can scan through the mobile application to unlock the bicycle and pay for the lease.

(4) Mobile application: the mobile application is an important component of the Df-PBS system that provides functions including bicycle GPS positioning, QR code scanning, unlocking, payment, and bicycle tracking.

The Df-PBS system suppliers flexibly deploy public bicycles in a plurality of places in a city according to the people flow statistical information, the positions of bicycle parking points can be dynamically shifted, and the number of the bicycles deployed at each parking point is elastically changed. Thanks to the low bicycle deployment costs and parking spot migration costs, these public bicycles can be deployed and moved anywhere in each city, and the position of the bicycle parking spots and the number of bicycles per parking spot can be dynamically updated according to the respective places and real-time bicycle rental travel requirements. Therefore, almost ubiquitous public bicycles bring very convenient travel services for users. The user can find the nearest public bike by downloading the APP of the Df-PDS system provider and park and return the bike anywhere near the travel destination.

However, the existing Df-PBS system has some non-negligible technical problems: firstly, the layout of the bicycle parking points is unreasonable, which leads to unbalanced bicycle supply, namely, part of the parking points can be borrowed without the bicycle, while other parking points have the phenomena of bicycle redundancy and empty space, and on the other hand, the workload of public bicycle scheduling is increased; secondly, a large number of redundant and low-efficiency bicycle parking points exist in the conventional Df-PBS system, so that urban public space is wasted, problems in urban road management, traffic safety and the like are easily caused, and the operation and maintenance cost of a Df-PBS system provider is increased, so that serious waste of public bicycle resources is caused.

Disclosure of Invention

Aiming at the defects or improvement demands of the prior art, the invention provides a dynamic planning method and a dynamic planning system for public bicycle stations of a Df-PBS (direct-current-direct-current) system, which aim to solve the technical problems that the existing Df-PBS system is unbalanced in bicycle supply and increases bicycle scheduling burden due to unreasonable layout of parking points, and the technical problems that urban public space is wasted, urban road management and traffic safety problems are easily caused, the operation and maintenance cost of a Df-PBS system provider is increased and serious waste of public bicycle resources is caused due to a large number of redundant and inefficient bicycle parking points.

To achieve the above object, according to one aspect of the present invention, there is provided a public bicycle station dynamic planning method for Df-PBS system, comprising the steps of:

(1) Acquiring historical position information data sets of all bicycles from a Df-PBS system, dividing the historical position information data sets into a plurality of space subsets according to administrative regions, further dividing each space subset into a plurality of space-time subsets according to time periods, constructing a position information data subset X corresponding to each space-time subset according to the running tracks of all bicycles in the space-time subset, and processing the position information data subset X by using a density peak clustering algorithm to obtain a dense parking point set as a candidate bicycle site set corresponding to the space-time subset;

(2) Constructing a corresponding bicycle station weighted directed graph according to the candidate bicycle station set corresponding to each space-time subset obtained in the step (1), wherein all candidate bicycle stations in the candidate bicycle station set are used as vertex sets in the weighted directed graph, bicycle running tracks among the bicycle stations are used as directed edges among corresponding vertexes in the vertex sets, low-efficiency bicycle stations are determined based on benefits and utilization rates of the bicycle stations in the weighted directed graph, and the directed edges are deleted from the weighted directed graph to obtain a bicycle station point graph model G corresponding to each space-time subset, bicycle station point graph models corresponding to all space-time subsets in one space subset are combined into a graph sequence model, and a time dimension association is established for all bicycle station point graph models adjacent to each other in the graph sequence model according to running behaviors of all bicycles in adjacent space-time subsets and dynamic updating of the bicycle stations, so that an updated public bicycle station point graph sequence model is obtained;

(3) Inputting the updated graph sequence model obtained in the step (2) into a trained gating graph neural network GGNN model so as to predict and obtain a bicycle station point diagram model of the next time period;

(4) Updating the bicycle station point diagram model of the next time period obtained in the step (3) according to a plurality of legal parking areas which are preset, so as to obtain a final bicycle station point diagram model.

Preferably, the step (1) specifically comprises:

firstly, dividing a position information data set of a public bicycle into a plurality of space subsets according to administrative regions, and then further dividing each space subset into a plurality of space-time subsets according to time periods;

then, for the running tracks of all bicycles in one space-time subset, extracting 2M position information of all bicycles from M running tracks of each bicycle, thereby constructing a position information data subset X of the bicycles in the space-time subset, wherein M is a natural number;

then, the position information containing the same bicycle ID and the same latitude and longitude is subjected to data deduplication processing, so that a position information data subset X= { X of the bicycle in the current time period is obtained ₁ ，...，x _N The total number N of elements in the data set meets M is less than or equal to N and less than or equal to 2M;

And finally, clustering the position information data subset X by using a density peak clustering algorithm to obtain a candidate bicycle station set corresponding to the space-time subset.

Preferably, the clustering process is specifically:

first, for each data point x _i E X, calculate the data point X _i Local density ρ of (2) _i Equal to:

wherein, the value ranges of i and j are 1, N]And i.noteq.j, dc is the cut-off distance, if d _ij ＜d _c Sigma (d) _ij -d _c ) =1; otherwise, σ (d _ij -d _c )＝0；

Subsequently, x is calculated _i Local density ratio X to position information data subset X _i Every other data point x of height _j Distance d between _ij And obtaining the minimum value from the obtained distances as x _i Delta distance Delta of (2) _i ：

Wherein if x _i The data point with the highest local density is the Delta distance value

A bicycle parking point clustering algorithm decision map is then generated from the local density ρ and Delta distance δ for each data point, with ρ as the x-axis and δ as the y-axis.

Then, ρ in the decision chart of the bicycle parking point clustering algorithm is performed _i ＞θ _ρ And delta _i ＞θ _δ Is designated as a cluster center point while at the same time satisfying

And delta _i ＞θ _δ Is designated as an outlier, the remaining data points are designated as remaining data points, wherein the threshold +. >

Threshold->

Then, for each remaining data point, a neighbor data point with a higher local density value than the remaining data point is found from a plurality of data points close to the remaining data point, then one neighbor data point closest to the remaining data point is selected from the found plurality of neighbor data points, and the cluster where the neighbor data point is located is allocated to the remaining data point. Finally, repeating the above process for the remaining data points, thereby obtaining a series of candidate bicycle parking point clusters c= { C ₁ ，...，c _n As a set of candidate bicycle stations, where c _n Representing the nth candidate bicycle station.

Preferably, the process of constructing the weighted directed graph in step (2) from the set of candidate bicycle stations corresponding to each space-time subset obtained in step (1) is specifically:

first, the candidate bicycle station set c= { C obtained according to step (1) ₁ ，...，c _n Constructing a corresponding vertex set v= { V ₁ ，...，v _n Each bicycle station v _i Containing three attributes, namely the longitude of the bicycle station _i And latitude of

Number of bicycles n owned by the bicycle station _i ；

Then, cluster c is found _i And assigning the longitude and latitude values of the central point to the bicycle stations v _i Longitude psi as the bicycle stations respectively _i And latitude of

Then counting cluster c _i Corresponding to the number of bicycles contained and assigning it to the bicycle station v _i Number of bicycles n owned by the bicycle station _i I.e. n _i ＝|c _i I, thus for each bicycle station v _i Attribute value +.>

All vertexes are formed into a vertex set V;

then, for two vertices v _i And v _j The clusters in their corresponding candidate set of bicycle stations C are C respectively _i And c _j Cluster c _i And c _j All bicycles in (a) are found for one bicycle b _a ∈c _i And another bicycle b _b ∈c _j If there is a driving track b _a →b _b Description is made from vertex c _i To vertex c _j There is a directed path, at v _i And v _j Adds a directed edge e between _ij From v _i Pointing v _j And has e _ij ≠e _ji All of the vertices in vertex set VThe directed edges between vertices form an edge set e= { E _ij }。

Subsequently, each edge E in the edge set E is calculated _ij Length d of (2) _ij I.e. bicycle station v _i And v _j The distance between them, the lengths of all edges in the edge set E constitute a distance set d= { D _ij }. Wherein d is _ij ＝d _ji 。

Subsequently, for each directed edge e _ij Calculating slave bicycle station v _i Departure and arrival at bicycle station v _j And takes the number of the bicycle running tracks as the directed edge e _ij Weight w of (2) _ij And has w _ij ≠w _ji The weights of all directed edges in edge set E constitute a weight set w= { W _ij }；

Finally, a weighted directed graph g= (V, E, D, W) is created from the vertex set V, directed edge set E, distance set D, and weight set W obtained as described above. .

Preferably, the process of determining the inefficient bicycle stations based on the benefits and availability of the bicycle stations in the weighted directed graph in step (2) and deleting them and their associated directed edges from the weighted directed graph to obtain the updated bicycle station point diagram model G specifically comprises the sub-steps of:

(a1) Acquiring each bicycle station v from a weighted directed graph _i Is to (1) the benefits of:

wherein α represents a running cost per unit distance;

(b1) Acquiring each bicycle station v from a weighted directed graph _i And calculates each bicycle station v from the acquired throughput _i Is used for the utilization rate of the (a);

wherein each bicycle station v _i The throughput of (2) is:

wherein w is _j′i Is from any bicycle station v _j′ Departure to bicycle station v _i Directed edge e of (2) _j′i Is a weight of (2).

Each bicycle station v _i The utilization ratio of (2) is equal to:

wherein TP is _G ＝∑w _ij∈W (w _ij ) Representing the throughput of all bike stations in vertex set V.

(c1) Will weight the benefit P in the directed graph _i Below a given threshold value theta _P And its utilization ratio U _i Below a given threshold value theta _U The bicycle stations, and their associated directed edges, are deleted from the weighted directed graph to obtain a high quality bicycle station point diagram model G.

Preferably, in the step (2), the bicycle station point diagram models corresponding to all the space-time subsets in one space subset are combined into one diagram sequence model, and the correlation in the time dimension is built for all the adjacent bicycle station point diagram models in the diagram sequence model according to the running behaviors of all the bicycles in the adjacent space-time subsets and the dynamic update of the bicycle stations, so that the updated diagram sequence model is obtained, and the process comprises the following substeps:

(a2) Constructing a subset sequence xs= { X from all space-time subsets of a space subset ₁ ，...，X _t ，...，X _T Each subset of positional information data X in the sequence of pairs _t Clustering bicycle parking points to construct a bicycle station point diagram model G corresponding to the space-time subset _t Combining bicycle station point diagram models corresponding to all space-time subsets in the subset sequence XS into a graph sequence model (shown in figure 4) GS= { G ₁ ，...，G _t ，...，G _T T.e.1, T }, where t.e.1]T represents the total number of space-time subsets in the spatial subset;

(b2) For each driving station graph model G in the graph sequence model GS constructed in step (a 2) _t Acquiring the driving station diagram model G _t Middle bicycle station v _i Bicycle flow n at _i (t)：

Wherein n0 _i (t) represents a traffic station map model G _t Middle bicycle station v _i An initial number of cycles to be used for the bicycle,

graph model G representing driving station _t During the current time period t to the bicycle station v _i Bicycle number of->

Graph model G representing driving station _t Slave bicycle station v within the current time period t _i Number of bicycles left, N (v) _i ) Graph model G representing driving station _t Neutralizing bicycle station v _i There is a set of all bicycle stations that connect the sides.

(c2) The driving station diagram model G obtained according to the step (b 2) _t Middle bicycle station v _i Bicycle flow n at _i (t) calculating a station point diagram model G in the map sequence model GS _t Driving station diagram model G of next adjacent time period (t+1) _t+1 Middle bicycle station v _i Bicycle flow n at _i (t+1) and thus obtaining a bicycle station v _i Bicycle flow rate variation at the location

(d2) Judging the bicycle station v obtained in the step (c 2) _i Bicycle flow rate variation at the location

If equal to 0, then enter step (e 2), otherwise enter step (g 2);

(e2) Determine whether or not there is equation n0 _i (t+1)＝in _i (t+1)＝out _i (t+1) =0, if yes, go to step (f), otherwise the process ends;

(f2) Bicycle station v _i Station point diagram model G from map sequence model GS _t Driving station diagram model G of next adjacent time period _t+1 Delete in the middle;

(g2) Bicycle station point diagram model G in diagram sequence model GS _t Driving station diagram model G of next adjacent time period _t+1 Middle bicycle station v _i Updated to a value of n _i (t+1)。

Preferably, the forward prediction process of the GGNN model is specifically:

first, each bicycle station point diagram model G _t As input to the GGNN model, a gating recursion unit GRU is used as a gating layer network structure of the GGNN model, wherein a hidden layer matrix h= { H of the GRU ₁ ，...，h _t }∈R ^D×n R represents a value range, n represents the number of neurons in the GGNN model, and D represents the dimension of a hidden layer of each GRU;

then, the t-th time period G is calculated using GRU as forward propagation module _t At the intermediate value r of the reset gate _t And updating the intermediate value z of the gate _t ：

r _t ＝σ(W _r ⊙[h _t-1 ，G _t ]+b _r )

z _t ＝σ(W _z ⊙[h _t-1 ，G _t ]+b _z )

Wherein W is _r And b _r Respectively represent the weight parameter and bias matrix of the reset gate, W _z And b _z Respectively representing the weight parameter and the bias matrix of the update gate, and the operator ≡represents element by elementThe prime multiplication, σ () represents a Sigmoid activation function,

Then according to G _t At the intermediate value r of the reset gate _t Update door z _t Calculating the value of the hidden layer;

wherein the intermediate parameters are:

wherein the method comprises the steps of

A network weight parameter matrix for the GGNN model;

finally, calculating a predicted bicycle station point diagram model of the (t+1) th time period according to the value of the hidden layer:

G′ _t+1 ＝σ(W _o ⊙h _t )

wherein W is _o Another network weight parameter matrix representing the GGNN model.

Preferably, the loss function used in the GGNN model training process is:

the loss functions between different network layers and between gating units are:

δ _y，t ＝(G _t+1 -G′ _t+1 )⊙σ′

wherein delta _y，t Representing the loss function, delta, between the output layer and the gating cell _h，t Representing the loss function, delta, between the hidden layer of the previous GRU and the gating unit _z，t Representing the loss function between the update layer and the gating unit,

representing the loss function, delta, between the hidden layer of the current GRU and the gating unit _r，t Representing a loss function between the reset gate and the gating cell; w (W) _rx Is a weight parameter matrix between the reset gate and the input layer, W _rh Is a matrix of weight parameters between the reset gate and the previous hidden layer, W _zx Is to update the weight parameter matrix between the gate and the input layer, W _zh Is to update the weight parameter matrix between the gate and the hidden layer,>

weight parameter matrix between hidden layers representing front and rear GRUs, W _o Is a weight parameter matrix of the output layer;

iteratively updating the weight parameter matrix W of the GGNN model by using the above equation _r ，W _z ，

W _h And W is _o Until the weight parameter matrixes are finally converged, a stable GGNN model is trained.

Preferably, step (4) specifically comprises the following sub-steps:

(4-1) for each bicycle station v in the bicycle station spot diagram model obtained in the step (3) _i In other words, the legal parking area p nearest to the bicycle station is found from a plurality of legal parking areas defined in advance _j E P, calculate bicycle station v _i And legal parking area p _j Distance d between _ij And determine the distance d _ij Whether or not it is equal to or less than a preset threshold value theta _d If yes, go to step (4-2), otherwise go to step (4-4), where p= { P ₁ ，p ₂ ，...，p _m -representing a set of legal parking areas predefined for the current city, m representing the total number of legal parking areas predefined;

(4-2) judging the legal parking area p _j Number n of receivable bicycles _j Whether or not to be equal to or greater than a bicycle station point diagram model G _t+1 Middle bicycle station v _i Number of bicycles at n _i If yes, ending the process, otherwise, turning to the step (4-3);

(4-3) bicycle station v _i N at _j Parking of a vehicle in a legal parking area p _j Of (c) is the remainder (n _i -n _j ) The vehicle bicycle finding another station v close to the bicycle _i Legal parking area p of (2) _j′ And processed in the same manner as the judgment process in the above steps (4-1) and (4-2) until the bicycle station v _i All bicycles are parked in the legal parking area, and then the process is finished;

(4-4) bicycle station v _i Is set as the legal parking area p _j And returns to step (4-2).

According to another aspect of the present invention, there is provided a public bicycle station dynamic programming system for Df-PBS system, comprising:

a first module, configured to acquire a historical position information data set of all bicycles from the Df-PBS system, divide the historical position information data set into a plurality of spatial subsets according to administrative regions, further divide each spatial subset into a plurality of space-time subsets according to time periods, construct a position information data subset X corresponding to each space-time subset according to driving tracks of all bicycles in the space-time subset, and process the position information data subset X by using a density peak clustering algorithm to obtain a dense set of parking points as a candidate bicycle site set corresponding to the space-time subset;

A second module, configured to construct a corresponding weighted directed graph of bicycle stations according to the set of candidate bicycle stations corresponding to each space-time subset obtained by the first module, wherein all candidate bicycle stations in the set of candidate bicycle stations are used as vertex sets in the weighted directed graph, bicycle travel tracks between bicycle stations are used as directed edges between corresponding vertices in the vertex sets, inefficient bicycle stations are determined based on benefits and utilization rates of bicycle stations in the weighted directed graph, and they and associated directed edges are deleted from the weighted directed graph to obtain a bicycle station point diagram model G corresponding to each space-time subset, bicycle station point diagram models corresponding to all space-time subsets in one space subset are combined into a diagram sequence model, and a time dimension association is established for all adjacent bicycle station point diagram models in the diagram sequence model according to the travel behaviors of all bicycles in adjacent space-time subsets and dynamic updates of bicycle stations, so as to obtain an updated public bicycle station point diagram sequence model;

a third module, configured to input the updated graph sequence model obtained by the second module into a trained gated graph neural network GGNN model, so as to predict and obtain a bicycle station point diagram model of a next time period;

And a fourth module for updating the bicycle station point diagram model of the next time period obtained by the third module according to a plurality of legal parking areas preset so as to obtain a final bicycle station point diagram model.

In general, the above technical solutions conceived by the present invention, compared with the prior art, enable the following beneficial effects to be obtained:

(1) The invention adopts the door control graph neural network technology to learn a large number of historical bicycle station point diagram models and accurately predict the bicycle station deployment of the next time period, so that the technical problem that the bicycle supply is unbalanced and the bicycle scheduling burden is increased due to unreasonable parking point layout of the conventional Df-PBS system can be solved.

(2) According to the invention, the benefits and the bicycle utilization rate of each candidate station are calculated in the process of constructing the public bicycle station point diagram model, and the low-efficiency stations with low benefits and low utilization rate are filtered to obtain the limited number of bicycle stations with high utilization rate and high benefits, so that the technical problems that the conventional Df-PBS system wastes urban public space, increases the operation and maintenance cost of a Df-PBS system provider, causes serious waste of public bicycle resources and easily causes urban road management and traffic safety problems due to a large number of redundant and low-efficiency bicycle parking points can be solved.

(3) The method for constructing the bicycle station point diagram model can effectively establish and analyze an abstract model of bicycle parking points, find out dense bicycle parking points through a density peak clustering method and serve as candidate stations in the diagram model, then fully utilize large-scale historical bicycle running track data to construct the links among stations, and calculate the income and bicycle utilization rate of each bicycle station, so that the bicycle station point diagram model of each time period is finally obtained.

(4) The method for constructing the bicycle station point diagram sequence model can effectively capture the bicycle running behavior, the station position and the dynamic updating condition of the number of bicycles at each station between the bicycle station point diagram models at different time periods, accurately calculate the updating between the bicycle station point diagram models at each adjacent time period, and provide a real and reliable theoretical basis for the dynamic and accurate prediction of the station position of the public bicycle and the number of bicycles required by each station.

Drawings

FIG. 1 is a schematic diagram of a conventional Df-PBS system, wherein FIG. 1 (a) is a public bicycle, FIG. 1 (b) is a public bicycle parking spot, FIG. 1 (c) is a conventional QR code-based bicycle lock, FIG. 1 (d) is a conventional code-scanning unlocking and payment function, and FIG. 1 (e) is a mobile application;

FIG. 2 shows a bicycle parking spot clustering process of the present invention, wherein FIG. 2 (a) is a bicycle parking spot clustering process and FIG. 2 (b) is a bicycle parking spot clustering algorithm decision diagram;

FIG. 3 shows the modeling process of the bicycle station point diagram model of the Df-PBS of the present invention, wherein FIG. 3 (a) is the actual Df-PBS system bicycle position information and travel track, and FIG. 3 (b) is the bicycle station weighted directed graph model constructed in accordance with the present invention;

FIG. 4 illustrates a modeling process of the bicycle station point diagram sequence model of the present invention;

FIG. 5 is a gatekeeper neural network model used in step (3) of the method of the present invention;

FIG. 6 illustrates an example of the relationship between predicted bicycle parking points and legal parking areas of the present invention;

FIG. 7 is a flow chart of a public bicycle station dynamic planning method for Df-PBS system of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.

The invention provides a dynamic planning method of public bicycle stations facing to a Df-PBS system, wherein the adopted bicycle parking point clustering method can fully utilize the position information record of a large-scale historical public bicycle, and excavate the actual bicycle running renting requirement, running track and bicycle parking point distribution condition; the construction method of the bicycle station point diagram model and the diagram sequence model can effectively establish and analyze an abstract model of a bicycle station, can effectively capture the updating and changing conditions of the bicycle station position and the bicycle quantity in the station between different time periods, and provides a real and reliable theoretical basis for the dynamic prediction of the bicycle station layout; the public bicycle station prediction method based on the gate control map neural network can fully utilize the deep learning technology to mine the large-scale map structure data, and achieve accurate prediction of bicycle station layout; the bicycle station dynamic planning method can match the predicted bicycle station layout with the smart city construction and management planning, effectively ensures that the recommended bicycle station layout is beneficial to city management, and improves the benefits of suppliers and the convenience of users.

As shown in fig. 7, the invention provides a method for dynamically planning public bicycle stations facing Df-PBS (direct-current-direct-current) system, which comprises the following steps:

(1) Acquiring historical position information (namely GPS) data sets of all bicycles from a Df-PBS system, dividing the historical position information data sets into a plurality of space subsets according to administrative regions, further dividing each space subset into a plurality of space-time subsets according to time periods (such as daily or weekly), constructing a position information data subset X corresponding to each space-time subset according to the running tracks of all bicycles in the space-time subset, and processing the position information data subset X by using a Density-Peak-based clustering (DPC) algorithm to obtain a dense bicycle parking point set as a candidate bicycle site set corresponding to the space-time subset;

large-scale bicycle GPS datasets and bicycle travel tracks were collected from Df-PBS systems deployed in different cities and administrative areas. The bicycle running tracks are stored according to the running behavior of the user, and each bicycle running track contains information such as user ID, bicycle ID, time stamp (Timestamp) of departure (bicycle unlocking) and arrival (bicycle locking), longitude (Longitude) and Latitude (Latitude) information of departure position and arrival position. Table 1 gives examples of the travel tracks of a bicycle.

TABLE 1

Because of the different maps of each city or administrative area, the layout of public bicycle stations is typically specific to a particular city or administrative area, requiring the construction of a corresponding bicycle station point map model for each city or administrative area, respectively, and the prediction of a corresponding station dynamic planning scheme.

The method specifically comprises the following steps: first, the position information data set of the public bike is divided into a plurality of spatial subsets according to administrative areas, and then each spatial subset is further divided into a plurality of spatial-temporal subsets according to time periods (e.g., daily or weekly). Hereinafter, bicycle parking spot clustering and bicycle station spot map model construction are performed for each space-time subset.

Then, for all bicycle travel tracks in one space-time subset, extracting position information of all bicycles (including departure positions and arrival positions of all bicycles) from each bicycle travel track, thereby constructing a position information data subset X of the bicycles in the space-time subset; assuming that there are M tracks of travel of the bicycle, the position information of 2M bicycles can be extracted, and considering that the same bicycle may be used multiple times in the current time period, the position information of the bicycle is repeatedly extracted, so that the position information of the bicycle containing the same bicycle ID and the same latitude and longitude needs to be subjected to data deduplication processing, and after the deduplication processing is finished, a position information data subset x= { X of the bicycle in the current time period is obtained ₁ ，...，x _N The total number N of elements in the dataset satisfies M.ltoreq.N.ltoreq.2M.

Finally, the subset of location information data X is clustered using a Density-Peak-based clustering (DPC) algorithm, the clustering process being shown in fig. 2 (a).

As can be seen from fig. 2 (a), the position information of each bicycle is represented as one circular data point, and the local density and Delta distance of each data point are calculated by using the formula (1) and the formula (2) respectively to find the bicycle with the cluster center point (shown as a black solid circle in the figure), then find the bicycle with the outlier (shown as a bottom circle in the figure), and finally respectively classify each remaining bicycle (shown as a broken-box hollow circle in the figure) into the nearest cluster, thereby forming a plurality of dense bicycle parking point sets.

The density peak clustering algorithm is based on the following assumptions: the local density around the center of each cluster is high and the distance from the center of any other cluster is greater than the distance from the rest of the points in the cluster.

The clustering process is specifically, first, for each data point x _i E X, calculate the data point X _i Local density ρ of (2) _i Equal to:

wherein, the value ranges of i and j are 1, N]And i.noteq.j, d _c Is a cut-off distance, and the value range is [3, 20 ]Preferably a value of 5; if d _ij ＜d _c Sigma (d) _ij -d _c ) =1; otherwise, σ (d _ij -d _c ) =0. Thus, data point x _i Local density ρ of (2) _i It is represented that the position information data is in the subset X, and X _i Distance between the two than cut-off distance d _c A smaller number of data points.

If x _i The Delta distance of the data point with the highest local density is as follows:

a bicycle parking point clustering algorithm decision map is then generated (as shown in fig. 2 (b)) from the local density ρ and Delta distance δ for each data point, with ρ as the x-axis and δ as the y-axis.

In fig. 2 (b), each data point is plotted at a corresponding position in the graph according to the local density and Delta distance value of each bicycle position data point, so as to form a bicycle parking point clustering algorithm decision graph. It is evident from the figure that these data points exhibit three distributions: the first portion of data points is distributed at the upper right of the graph with a higher ρ (i.e., ρ _i ＞θ _ρ ) And a higher delta (i.e., delta _i ＞θ _δ ) (as shown by the black filled circles); the second portion of the data points is distributed at the top left of the graph with a lower ρ (i.e.

) And a higher delta (i.e., delta _i ＞θ _δ ) (as shown by the circled ground in the figure); the third portion of data is distributed below the figure (as indicated by the open circles in the dashed box).

Then, the clustering algorithm decision diagram is provided with higher rho (namely rho _i ＞θ _ρ ) And a higher delta (i.e., delta _i ＞θ _δ ) Is designated as the cluster center point, while will have a lower p (i.e

) And a higher delta (i.e., delta _i ＞θ _δ ) Is designated as outlier, the remaining data points are designated as remaining data points (wherein the effective threshold is set to +.>

And

) Subsequently, for each remainderFor the data points, a neighbor data point with a local density value higher than that of the remaining data points is searched from a plurality of data points which are close to the neighbor data point, then one neighbor data point which is closest to the remaining data point is selected from the searched plurality of neighbor data points, and the cluster where the neighbor data point is located is distributed to the remaining data point. Finally, repeating the above process for the remaining data points, thereby obtaining a series of candidate bicycle parking point clusters c= { C ₁ ，...，c _n -as a set of candidate bicycle stations, wherein c _n Representing the nth bicycle parking spot cluster (i.e., candidate bicycle stations).

(2) Constructing a corresponding bicycle site weighted directed graph according to the candidate bicycle site set corresponding to each space-time subset obtained in the step (1), wherein all candidate bicycle sites in the candidate bicycle site set are used as vertex sets in the weighted directed graph, bicycle running tracks among the bicycle sites are used as directed edges (shown in fig. 3 (a)) among corresponding vertexes in the vertex sets, inefficient bicycle sites are determined based on benefits and utilization rates of the bicycle sites in the weighted directed graph, and the bicycle sites and the associated directed edges thereof are deleted from the weighted directed graph to obtain bicycle station point graph models G (shown in fig. 3 (b)) corresponding to each space-time subset, bicycle station point models corresponding to all space-time subsets in one space subset are combined into a graph sequence model, time dimension correlations are established for all bicycle station point models adjacent to each other in the graph sequence model according to running behaviors of all bicycles in adjacent space-time subsets and dynamic updates of the bicycle sites, and then updated public bicycle station point sequence models are obtained.

In fig. 3 (a) it can be seen that the bicycle position information in the actual Df-PBS system and the running track of the bicycle, dense bicycle parking spots are clustered into candidate bicycle stations, and the bicycle running track between every two candidate stations is counted and serves as a directed edge between the two stations.

In fig. 3 (b), according to the bicycle station point diagram model of fig. 3 (a), the yield and the utilization rate of each candidate bicycle station (the vertex in the bicycle station point diagram) are calculated, and the low-efficiency station with low station yield and low station utilization rate is deleted from the diagram, thereby obtaining the bicycle station point diagram model with high yield and high utilization rate.

The weighted directed graph constructed in this step corresponding to each space-time subset is denoted g= (V, E, D, W), where the elements in vertex set V are bicycle stations, the elements in edge set E are bicycle travel tracks between bicycle stations, the elements in distance set D are the actual distances between bicycle stations, and the elements in weight set W are the number of bicycle travel tracks between bicycle stations.

In this step, the process of constructing a weighted directed graph from the set of candidate bicycle stations corresponding to each space-time subset obtained in step (1) is specifically:

First, the candidate bicycle station set c= { C obtained according to step (1) ₁ ，...，c _n Constructing a corresponding vertex set v= { V ₁ ，...，v _n Each bicycle station v _i Containing three attributes, namely longitude psii and latitude, of the bicycle station

Number of bicycles n owned by the bicycle station _i . More specifically, the above process is based on the bicycle station v _i Corresponding cluster (i.e., element in candidate set of bicycle stations C) C _i To obtain the three attribute values, i.e. find cluster c _i And assigning the longitude and latitude values of the central point to the bicycle stations v _i Longitude psi as the bicycle stations respectively _i And latitude->

Then counting cluster c _i Corresponding to the number of bicycles contained and assigning it to the bicycle station v _i Number of bicycles n owned by the bicycle station _i I.e. n _i ＝|c _i | a. The invention relates to a method for producing a fibre-reinforced plastic composite. As suchFor each bicycle station v _i Attribute value +.>

And all vertices are grouped into a vertex set V.

Then, a corresponding set of edges is created for vertex set V from the clusters in candidate bicycle site set C, i.e., for two vertices V _i And v _j The clusters in their corresponding candidate set of bicycle stations C are C respectively _i And c _j Cluster c _i And c _j All bicycles in (a) find out, assume for one bicycle b _a ∈c _i And another bicycle b _b ∈c _j If there is a driving track b _a →b _b Description is made from vertex c _i To vertex c _j There is a directed path, at v _i And v _j Adds a directed edge e between _ij From v _i Pointing v _j And has e _ij ≠e _ji Directed edges between all vertices in vertex set V form an edge set e= { E _ij }。

Wherein, specifically, the Haverine method is used for calculating the bicycle station v _i And

distance between: />

Where R is the earth radius, typically set to 6371.0km, and Haverine function H (θ) is defined as:

finally, for each directed edge e _ij Calculating slave bicycle station v _i Departure and arrival at bicycle station v _j And takes the number of the bicycle running tracks as the directed edge e _ij Weight w of (2) _ij And has w _ij ≠w _ji The weights of all directed edges in edge set E constitute a weight set w= { W _ij }。

From the vertex set V, directed edge set E, distance set D, and weight set W obtained in the previous steps, a weighted directed graph model g= (V, E, D, W) of the public bike station is created.

As previously described, in Df-PBS systems, a large number of redundant bicycles are deployed in infrequently used locations due to low deployment costs and vicious competition from the same row. These bicycle parking spots, while having a high number of bicycles, do not have a substantial throughput of bicycles (many of the bicycle parking spots are stationary and are not in use), so that the bicycle parking spots have low utilization and yield, and these candidate bicycle spots need to be detected and removed from the weighted directed graph to maximize the benefits and practicality of the Df-PBS system.

In this step, the inefficient bicycle stations are determined based on the benefits and availability of the bicycle stations in the weighted directed graph, and their associated directed edges are deleted from the weighted directed graph to obtain an updated bicycle station point diagram model G, which is specifically:

(a) Acquiring each bicycle station v from a weighted directed graph _i Is to (1) the benefits of:

wherein α represents a running cost per unit distance;

the benefit of a bicycle station refers to the sum of the running costs of all bicycles from the bicycle station. Since the running cost of a public bike is proportional to the running distance, the benefits of a bike station can be calculated from the distance between two bike stations and the number of runs.

(b) Acquiring each bicycle station v from a weighted directed graph _i And calculates each bicycle station v from the acquired throughput _i Is used for the utilization rate of the (a);

in particular, throughput of a bike station refers to the number of bikes from and to the bike station. The utilization of a bike station is the ratio of the throughput of that bike station to the throughput of all bike stations in the vertex set V of the weighted directed graph.

Each bicycle station v _i Is defined as:

/>

Each bicycle station v _i The utilization ratio of (2) is equal to:

wherein the method comprises the steps of

Representing the throughput of all bike stations in vertex set V.

(c) Will weight the benefit P in the directed graph _i Below a given threshold value theta _P And its utilization ratio U _i Below a given threshold value theta _U The bicycle stations and their associated directed edges are deleted from the weighted directed graph to obtain a high quality bicycle station point diagram model G;

if bicycle station v _i Revenue P of (2) _i Below a given threshold value theta _P And its utilization ratio U _i Below a given threshold value theta _U It is referred to as an inefficient bicycle station. In the actual operation process of the Df-PBS system, the Df-PBS system in different cities has different layout and density, so the threshold value theta _P And U _i Setting the actual conditions according to the individual bicycle station point diagram models to different values, wherein the threshold value theta _P The range of the values is

Preferably equal to->

Threshold value theta _U The range of the values is

Preferably equal to->

In this step, the bicycle station point diagram models corresponding to all the space-time subsets in a space subset are combined into a bicycle station point diagram sequence model, and according to the running behaviors of all the bicycles in adjacent space-time subsets and the dynamic update of bicycle stations, the correlation in the time dimension is established for all the adjacent bicycle station point diagram models in the diagram sequence model, so as to obtain an updated diagram sequence model, and the process specifically comprises the following steps:

(a) Constructing a subset sequence xs= { X from all space-time subsets of a space subset ₁ ，...，X _t ，...，X _T Each subset of positional information data X in the sequence of pairs _t Clustering bicycle parking points to construct a bicycle station point diagram model G corresponding to the space-time subset _t Combining bicycle station point diagram models corresponding to all space-time subsets in the subset sequence XS into a graph sequence model (shown in figure 4) GS= { G ₁ ，...，G _t ，...，G _T T.e.1, T }, where t.e.1]T represents the total number of space-time subsets in the spatial subset.

In fig. 4, a bicycle station point diagram model of three adjacent time periods is shown, from which it can be seen that the updating of the respective vertices and directed edges in the diagram model between every two adjacent time periods is due to the bicycle running behavior of the respective bicycle stations. And establishing a bicycle station point diagram sequence model by establishing time dimension association on the graph models of adjacent time periods.

Specifically, the bicycle parking point clustering process and the bicycle station point diagram model construction process are identical to the clustering process and the model construction process described above, and are not described in detail herein;

(b) For each bicycle station point diagram model G in the diagram sequence model GS constructed in step (a) _t Obtaining the bicycle station point diagram model G _t Middle bicycle station v _i Bicycle flow n at _i (t)：

Wherein n0 _i (t) represents a bicycle station point diagram model G _t Middle bicycle station v _i An initial number of cycles to be used for the bicycle,

dot pattern model G representing bicycle station _t During the current time period t (which corresponds to the time period in step (1), i.e. the current day or week) to the bicycle station v _i Bicycle number of->

Dot pattern model G representing bicycle station _t Slave bicycle station v within the current time period t _i Number of bicycles left, N (v) _i ) Dot pattern model G representing bicycle station _t Neutralizing bicycle station v _i There is a set of all bicycle stations that connect the sides.

(c) A bicycle station point diagram model G obtained according to step (b) _t Middle bicycle station v _i Bicycle flow n at _i (t) calculating a bicycle station point diagram model G in the diagram sequence model GS _t Bicycle station point diagram model G of next adjacent time period (t+1) _t+1 Middle bicycle station v _i Bicycle flow n at _i (t+1) and thus obtaining a bicycle station v _i Bicycle flow rate variation at the location

The riding behaviour of the bicycle will result in a change in the number of bicycles available per bicycle station and the bicycle flow between bicycle stations. This phenomenon maps to changes in the attribute values (number of bicycles in the bicycle station) and weight values of the directed edges represented as vertices in the graph sequence model.

(d) Judging the bicycle station v obtained in the step (c) _i Bicycle flow rate variation at the location

If equal to 0, then enter step (e), otherwise enter step (g);

(e) Determine whether or not there is equation n0 _i (t+1)＝in _i (t+1)＝out _i (t+1) =0, if yes, go to step (f), otherwise the process ends;

(f) Bicycle station v _i Bicycle station point diagram model G from diagram sequence model GS _t Bicycle station point diagram model G of next adjacent time period _t+1 Delete in the middle;

(g) Bicycle station point diagram model G in diagram sequence model GS _t Next adjacent time periodBicycle station point diagram model G _t+1 Middle bicycle station v _i Updated to a value of n _i (t+1)。

(3) Inputting the updated graph sequence model obtained in the step (2) into a trained gating graph neural network (Gated GraphNeural Network, GGNN for short) model so as to predict and obtain a bicycle station point diagram model of the next time period;

as shown in fig. 5, which shows the GGNN model used in this step.

Public bicycle station map sequence data for each city is trained using a gated map neural network (Gated Graph Neural Network, GGNN) model and public bicycle station layouts for the next time period are predicted. The GGNN-based bicycle station prediction model is shown in fig. 5.

As can be seen from fig. 5, the input data is a graph model { G over all the history periods in the bicycle station point diagram sequence model GS ₁ ，...，G _t (assuming that the current latest time period is t and the corresponding bicycle station point diagram model is G) _t ) Calculating each neural network model in the GGNN model, and finally predicting to obtain a bicycle station point diagram model G of the next time period (t+1) _t+1 。。

The input and output of the GGNN-based bicycle station prediction model are described as follows:

input: a large number of bicycle position information histories are collected from Df-PBS systems of different cities or administrative areas, then the data are divided into a plurality of space-time subsets according to time periods, and bicycle parking point clustering and bicycle station point diagram model construction are carried out on bicycle running data in each space-time subset. According to this step, a bicycle station point diagram sequence (a set of bicycle station point diagram data sets) gs= { G can be obtained ₁ ，...，G _t }. Next, the graph G for each period of time _t The GS serves as an input to the GGNN model.

And (3) outputting: given input G _t The output of the GGNN model is the predicted bicycle station point diagram G of the time slot) _t+1 . Through G _t+1 The predicted position of the bicycle station and each self-position can be obtainedThe number of bicycles required for the driving station.

The prediction and training process of the GGNN-based bicycle station prediction model is described as follows:

(1) Forward prediction process

Given a graph sequence model gs= { G ₁ ，...，G _t ) Dot pattern model G of each bicycle station _t As input to GGNN model, G' _t+1 As the prediction output of the GGNN model, a gating recursion unit (Gated Recurent Unit, abbreviated as GRU) is used as the gating layer network structure of the GGNN model, so that h= { H ₁ ，...，h _t }∈R ^D×n Is a hidden layer matrix of GRUs, wherein R represents a value range, n represents the number of neurons in the GGNN model, and D represents the dimension of a hidden layer of each GRU. Using GRU gating layer as forward propagation module, input data G of the t-th time period _t At reset gate r _t Updating door z _t The intermediate value of (2) is calculated as follows:

wherein W is _r And b _r Weight parameters and bias matrix representing reset gates, W _z And b _z Indicating the weight parameters and bias matrix of the update gate, the operator ∈indicates the element-wise multiplication. σ () represents a Sigmoid activation function,

further calculating values of the hidden layer and the output layer from the reset gate and the update gate:

thus, for each input data G _t The corresponding predicted output G 'can be obtained from the GGNN model' _t+1 As a predictive bicycle station point diagram model for the (t+1) th time period.

(2) Reverse propagation training process

In the forward prediction process, the bicycle station point diagram model G can be obtained according to each time period t _t Bicycle station point diagram model G 'of next time period (t+1) is predicted' _t+1 . Benefiting from a large-scale historical bicycle station point diagram sequence model gs= { G ₁ ，...，G _t ) The GGNN model is continuously trained to stabilize and converge by comparing the predicted map with the actual map for each time period. Let G' _t+1 A predicted self-station point diagram model for the (t+1) th time period, G _t+1 Is an actual bicycle station point diagram model for that time period. Using the mean square error (Mean Square Error, MSE for short) as a loss function of the GGNN model, defined as:

the loss functions between different network layers and between gating units are calculated as follows:

wherein W is _o Is the weight parameter matrix of the output layer, W _rx Is a weight parameter matrix between the reset gate and the input layer, W _rh Is a matrix of weight parameters between the reset gate and the previous hidden layer, W _zx Is to update the weight parameter matrix between the gate and the input layer, W _zh Is to update the weight parameter matrix between the gate and the hidden layer,

weight parameter matrix representing the weight between hidden layer and input layer, < ->

Weight parameter matrix delta between hidden layers representing front and back GRUs _y，t Representing the loss function, delta, between the output layer and the gating cell _h，t Before representationLoss function delta between hidden layer of GRU and door control unit _z，t Representing the loss function between the update layer and the gating unit, < >>

Representing the loss function, delta, between the hidden layer of the current GRU and the gating unit _r，t Representing the loss function between the reset gate and the gating cell.

W _r Is W _rx And W is _rh Is connected to the matrix, i.e. W _r ＝W _rx +W _rh ，W _z Is W _zx And W is _zh Is connected to the matrix, i.e. W _z ＝W _zx +W _zh ，

Is->

And->

Is a matrix of interconnections, i.e.)>

W _h Is W _hx And W is _hh Is connected to the matrix, i.e. W _h ＝W _hx +W _hh 。

Iteratively updating a network weight parameter matrix of the GGNN model using the above equation, including W _r ，W _z ，

W _h And W is _o Until the network weight parameter matrixes are finally converged, a stable GGNN model is trained. Finally, a bicycle station point map model for the next time period is predicted using the trained GGNN model.

(4) Updating the bicycle station point map model of the next time period obtained in the step (3) according to a plurality of legal parking areas preset by government or municipal administration to obtain a final bicycle station point map model (as shown in fig. 6).

Specifically, the plurality of legal parking areas predefined in this step are areas defined by government or municipal administration for the current city, which allow the white-running vehicles to park.

The step (4) specifically comprises the following substeps:

(4-1) for each bicycle station v in the bicycle station spot diagram model obtained in the step (3) _i In other words, the legal parking area p nearest to the bicycle station is found from a plurality of legal parking areas defined in advance _j E P, calculate bicycle station v _i And legal parking area p _j Distance d between _ij (the calculation process is specifically using the formula (3) described above), and the distance d is determined _ij Whether or not it is equal to or less than a preset threshold value theta _d If so, consider bicycle station v _i Located in legal parking area p _j In, then go to step (4-2), otherwise describe bicycle station v _i In an illegal parking position, then go to step (4-4), where p= { P ₁ ，p ₂ ，...，p _m -representing a set of legal parking areas predefined for the current city, m representing the total number of legal parking areas predefined;

threshold value θ in this step _d The value of (2) is in the range of 0 to 20 meters, preferably 5 to 10 meters.

(4-2) judging the legal parking area p _j Number n of receivable bicycles _j Whether or not to be equal to or greater than a bicycle station point diagram model G _t+1 Middle bicycle station v _i Number of bicycles at n _i If yes, the legal parking area p is indicated _j Having sufficient space to accommodate bicycle station v _i At bicycle, the process ends, otherwise the bicycle station v is explained _i Located in legal parking area p _j But legal parking area p _j But is not enoughIs used for parking bicycle station v _i At the bicycle, i.e. bicycle station v _i Where the number of bicycles exceeds the legal parking area p _j The available space is provided, and then the step (4-3) is carried out;

(4-3) bicycle station v _i N at _j Parking of a vehicle in a legal parking area p _j Of (c) is the remainder (n _i -n _j ) The vehicle bicycle finding another station v close to the bicycle _i Legal parking area p of (2) _j′ And processed in the same manner as in steps (4-1) and (4-2) above until the bicycle station v _i All bicycles are parked in the legal parking area, and then the process is finished;

After matching the predicted bicycle stations with the legal parking areas of the current city, all bicycle stations in illegal parking locations are modified to allocate bicycles in the bicycle stations beyond the parking space to the nearest legal parking areas. Finally, the obtained bicycle site layout scheme accords with the current urban management plan, and simultaneously meets the targets of maximizing bicycle site income and maximizing bicycle site utilization rate of Df-DPS suppliers. The Df-DPS personnel will schedule and deploy the public bicycles according to the layout scheme.

In fig. 6, three relationships between a bicycle station and a legal parking area of a current city can be seen in a predicted public bicycle station spot diagram model, in a scenario (a), the bicycle station is located in the legal parking area and the legal parking area has enough space to accommodate the bicycles of the bicycle station, in a scenario (b), the bicycle station is located in the legal parking area but the legal parking area does not have enough space for parking the bicycles of the bicycle station, then it is necessary to park a portion of the bicycles of the bicycle station into the current legal parking area, search for another legal parking area close to the bicycle station for the remaining bicycles, in a scenario (c), the position of the bicycle station is set to be the nearest legal parking area, and it is further determined whether the legal parking area has enough space for parking the bicycles of the current bicycle station.

It will be readily appreciated by those skilled in the art that the foregoing description is merely a preferred embodiment of the invention and is not intended to limit the invention, but any modifications, equivalents, improvements or alternatives falling within the spirit and principles of the invention are intended to be included within the scope of the invention.

Claims

1. A Df-PBS system-oriented public bicycle station dynamic planning method is characterized by comprising the following steps:

2. The method for dynamically planning a public bike station for Df-PBS system according to claim 1, wherein the step (1) specifically includes:

then, the position information containing the same bicycle ID and the same latitude and longitude is subjected to data deduplication processing, so that a position information data subset X= { X of the bicycle in the current time period is obtained ₁ ,…,x _N The total number N of elements in the data set meets M is less than or equal to N and less than or equal to 2M;

3. The method for dynamically planning public bicycle stations facing Df-PBS system according to claim 2, wherein the clustering process is specifically:

wherein, the value ranges of i and j are 1, N]And i.noteq.j, d _c Is the cut-off distance, if d _ij <d _c Sigma (d) _ij -d _c ) =1; otherwise, σ (d _ij -d _c )＝0；

Then, generating a bicycle parking point clustering algorithm decision graph according to the local density rho and Delta distance Delta of each data point, wherein rho is taken as an x-axis, and Delta is taken as a y-axis;

then, ρ in the decision chart of the bicycle parking point clustering algorithm is performed _i >θ _ρ And delta _i >θ _δ Is designated as a cluster center point while at the same time satisfying

And delta _i >θ _δ Is designated as an outlier, the remaining data points are designated as remaining data points, wherein the threshold +.>

Threshold->

Then, for each remaining data point, searching a neighbor data point with a local density value higher than that of the remaining data point from a plurality of data points close to the remaining data point, selecting one neighbor data point closest to the remaining data point from the plurality of found neighbor data points, and distributing a cluster where the neighbor data point is located to the remaining data point; finally, repeating the above process for the remaining data points, thereby obtaining a series of candidate bicycle parking point clusters c= { C ₁ ，...，c _n As a set of candidate bicycle stations, where c _n Representing the nth candidate bicycle station.

4. A method for dynamic planning of public bike stations for Df-PBS system according to claim 3, wherein the process of constructing weighted directed graph according to the candidate bike station set corresponding to each space-time subset obtained in step (1) in step (2) is specifically:

first, the candidate bicycle station set c= { C obtained according to step (1) ₁ ,...,c _n Constructing a corresponding vertex set v= { V ₁ ,...,v _n Each bicycle station v _i Containing three attributes, namely the longitude of the bicycle station _i And latitude of

Number of bicycles n owned by the bicycle station _i ；

Then, cluster c is found _i Center of (2)Point and respectively assign the longitude and latitude values of the central point to bicycle stations v _i Longitude psi as the bicycle stations respectively _i And latitude of

All vertexes are formed into a vertex set V;

then, for two vertices v _i And v _j The clusters in their corresponding candidate set of bicycle stations C are C respectively _i And c _j Cluster c _i And c _j All bicycles in (a) are found for one bicycle b _a ∈c _i And another bicycle b _b ∈c _j If there is a driving track b _a →b _b Description is made from vertex c _i To vertex c _j There is a directed path, at v _i And v _j Adds a directed edge e between _ij From v _i Pointing v _j And has e _ij ≠e _ji Directed edges between all vertices in vertex set V form an edge set e= { E _ij }；

Subsequently, each edge E in the edge set E is calculated _ij Length d of (2) _ij I.e. bicycle station v _i And v _j The distance between them, the lengths of all edges in the edge set E constitute a distance set d= { D _ij -a }; wherein d is _ij ＝d _ji ；

Finally, a weighted directed graph g= (V, E, D, W) is created from the vertex set V, directed edge set E, distance set D, and weight set W obtained as described above.

5. The method of dynamic planning of public bike stations for Df-PBS system according to claim 4, wherein the step (2) of determining inefficient bike stations based on the profits and availability of bike stations in the weighted directed graph and deleting them and their associated directed edges from the weighted directed graph to obtain the updated bike station point graph model G specifically includes the following sub-steps:

wherein α represents a running cost per unit distance;

wherein each bicycle station v _i The throughput of (2) is:

wherein w is _j′i Is from any bicycle station v _j′ Departure to bicycle station v _i Directed edge e of (2) _j′i Weights of (2);

each bicycle station v _i The utilization ratio of (2) is equal to:

wherein the method comprises the steps of

Representing throughput of all bicycle stations in vertex set V;

6. The method for dynamically planning public bike stations for Df-PBS system according to claim 5, wherein in step (2), the bike station point diagram models corresponding to all the space-time subsets in one space subset are combined into one diagram sequence model, and the correlation in time dimension is established for all the bike station point diagram models adjacent to each other in the diagram sequence model according to the running behaviors of all the bicycles in the adjacent space-time subsets and the dynamic update of the bike stations, so that the updated diagram sequence model is obtained by the following sub-steps:

(a2) Constructing a subset sequence xs= { X from all space-time subsets of a space subset ₁ ,…,X _t ,…,X _T Each subset of positional information data X in the sequence of pairs _t Clustering bicycle parking points to construct a bicycle station point diagram model G corresponding to the space-time subset _t Combining bicycle station point diagram models corresponding to all space-time subsets in the subset sequence XS into a graph sequence model GS= { G ₁ ,…,G _t ,…,G _T T.e.1, T }, where t.e.1]T represents the total number of space-time subsets in the spatial subset;

Graph model G representing driving station _t Slave bicycle station v within the current time period t _i Number of bicycles left, N (v) _i ) Graph model G representing driving station _t Neutralizing bicycle station v _i A set of all bicycle stations with connecting sides;

If equal to 0, then enter step (e 2), otherwise enter step (g 2);

(e2) Determine whether or not there is equation n0 _i (t+1)＝in _i (t+1)＝out _i (t+1) =0, if yes, go to step (f 2), otherwise the process ends;

7. The method for dynamically planning a public bike station for Df-PBS system according to claim 6, wherein the forward prediction process of the GGNN model is specifically:

first, each bicycle station point diagram model G _t As input to the GGNN model, a gating recursion unit GRU is used as a gating layer network structure of the GGNN model, wherein a hidden layer matrix h= { H of the GRU ₁ ,…,h _t }∈R ^D×n R represents a value range, n represents the number of neurons in the GGNN model, and D represents the dimension of a hidden layer of each GRU;

r _t ＝σ(W _r ⊙[h _t-1 ，G _t ]+b _r )

z _t ＝σ(W _z ⊙[h _t-1 ，G _t ]+b _z )

Wherein W is _r And b _r Respectively represent the weight parameter and bias matrix of the reset gate, W _z And b _z Respectively representing the weight parameter and the bias matrix of the update gate, and the operator ≡represents element-by-element multiplicationσ () represents a Sigmoid activation function,

/>

wherein the intermediate parameters are:

wherein the method comprises the steps of

A network weight parameter matrix for the GGNN model;

G′ _t+1 ＝σ(W _o ⊙h _t )

8. The method for dynamic planning of public bike station for Df-PBS system according to claim 7,

the loss function used in the GGNN model training process is:

δ _y,t ＝(G _t+1 -G′ _t+1 )⊙σ′

wherein delta _y,t Representing the loss function, delta, between the output layer and the gating cell _h,t Representing the loss function, delta, between the hidden layer of the previous GRU and the gating unit _z,t Representing the loss function between the update layer and the gating unit,

representing the loss function, delta, between the hidden layer of the current GRU and the gating unit _r,t Representing a loss function between the reset gate and the gating cell; w (W) _rx Is a weight parameter matrix between the reset gate and the input layer, W _rh Is a matrix of weight parameters between the reset gate and the previous hidden layer, W _zh Is to update the weight parameter matrix between the gate and the hidden layer,>

9. The method for dynamic planning of public bike stations for Df-PBS system according to claim 8, wherein the step (4) specifically includes the following sub-steps:

(4-1) for each bicycle station, v, in the bicycle station point diagram model obtained in the step (3) _i In other words, the legal parking area p nearest to the bicycle station is found from a plurality of legal parking areas defined in advance _j E P, calculate bicycle station, v _i And legal parking area p _j Distance d between _ij And determine the distance d _ij Whether or not it is equal to or less than a preset threshold value theta _d If yes, go to step (4-2), otherwise go to step (4-4), where p= { P ₁ ，p ₂ ，…，p _m -representing a set of legal parking areas predefined for the current city, m representing the total number of legal parking areas predefined;

(4-4) bicycle station v _i Is set as the legal parking area p _j Is of the order of (2)And (4) setting and returning to the step (4-2).

10. A public bicycle station dynamic programming system for Df-PBS systems, comprising: