CN115936240A

CN115936240A - Shared bicycle demand forecasting and delivering scheduling method

Info

Publication number: CN115936240A
Application number: CN202211665231.5A
Authority: CN
Inventors: 郭洪飞; 任亚平; 赵淑曼; 何智慧; 童曦; 赵治宇
Original assignee: Jinan University
Current assignee: Jinan University
Priority date: 2022-12-23
Filing date: 2022-12-23
Publication date: 2023-04-07
Anticipated expiration: 2042-12-23
Also published as: CN115936240B

Abstract

The invention relates to a method for forecasting demand of a shared bicycle and dispatching delivery, which comprises the following steps: s1, establishing an XGboost decision tree, and gathering similar and adjacent sites into a cluster; s2, correcting the actual demand of a single site, and bringing the corrected demand into an XGboost decision tree to be trained to obtain an optimized site clustering result; s3, forecasting the vehicle borrowing/returning requirement of each station according to the trained XGboost decision tree; s4, considering the capacity limit and the required arrival distribution of each station, and calculating the initial throwing amount of the single vehicle; s5, dividing the dispatching partitions of the city; and S6, establishing a single-vehicle scheduling model in the region to obtain the optimal selection of the scheduling path. The method starts from demand prediction and scheduling optimization of the shared bicycle, considers various practical situations and practical problems, designs a model and a solving method, provides basis for operation decision of the urban shared bicycle system, is favorable for quickly realizing the demand of borrowing/returning the bicycle for urban residents, and is convenient for life and healthy travel.

Description

Shared bicycle demand prediction and delivery scheduling method

Technical Field

The invention relates to the technical field of shared bicycle operation, in particular to a method for demand forecasting and release scheduling of shared bicycles.

Background

At present, shared bicycles are integrated into daily life and need to be matched with urban management to become a part of public transportation service. The main effects of related management measures issued by governments in part of regions are as follows: (1) Governments in various places limit operators to release shared bicycles, and enterprises can only want to increase the utilization rate of vehicles in order to keep the satisfaction degree of users; (2) The policy stipulates that the sharing bicycle platform cannot collect the deposit of the user, the characteristics of financial products of the sharing bicycle are weakened, and the focus of market competition is shifted to fine management, intelligent travel experience and high-quality service creation; (3) The government withholds the sharing single vehicle which is illegally parked, in order to reduce the loss of the redemption money, the sharing single vehicle platform marks a parking area and a non-parking area, and vehicles which are illegally parked need to be transferred to the parking area in time.

Therefore, the management mode of the existing bicycle enterprises has been changed from the original mode of scale expansion to the fine management. In the current single-vehicle market, only by improving the operation efficiency, improving the user satisfaction and controlling the operation and maintenance cost, the enterprises can be continuously developed. The key points are to solve the problems of unbalanced distribution of station single vehicle resources and illegal parking, scientific and reasonable redistribution of shared single vehicles is needed, and the single vehicles are transported to a demand point from a supply point and a stoppable point from the stoppable point by formulating a reasonable scheme so as to match the requirements of users. In fact, the problem of low utilization rate of the single vehicles in the shared single vehicle system generally exists, and the enterprise scheduling balance still has a large space.

Disclosure of Invention

The invention mainly aims to provide a shared bicycle demand forecasting and releasing scheduling method, which can forecast the distribution situation of user borrowing/returning demands in cities at different times, determine the place, time, quantity and route of the dispatched bicycles by a design scheme, ensure the supply and demand matching of the bicycles and the operation efficiency of a system, and further solve the technical problems provided by the background technology.

In order to achieve the aim, the invention provides a method for forecasting and dispatching shared bicycle demands, which comprises the following steps:

s1, establishing an XGboost decision tree, and aggregating similar and adjacent sites into a cluster;

s2, correcting the real requirement of a single site, and bringing the corrected real requirement into an XGboost decision tree to obtain an optimized site clustering result after training;

s3, forecasting the vehicle borrowing/returning requirement of each station according to the trained XGboost decision tree;

s4, considering the capacity limit and the required arrival distribution of each station, and calculating the initial throwing amount of the single vehicle;

s5, dividing the dispatching partitions of the city;

and S6, establishing a single-vehicle scheduling model in the region to obtain the optimal selection of the scheduling path.

Preferably, the establishing the XGBoost decision tree in step S1, the aggregating similar and adjacent sites into a cluster includes:

respectively counting the net traffic flow number y and the characteristic variable lon in each station h time period in the past d working days _i 、lat _i 、tem _i 、rain _i 、speed _i And collecting N samples in total, wherein the XGboost decision tree model is described as follows:

in the formula (1), the number of subtrees is represented by k,

customer borrowing/returning demand predicted value, lon, representing station i _i Longitude representing station i，lat _i Indicates the latitude, tem, of site i _i Indicating the temperature, rain, of station i _i Indicating the precipitation, speed, of site i _i Representing the wind speed of a station i, the objective function of the XGboost decision tree model is as follows:

according to Taylor's theory of expansion, equation (2) can be converted to:

in the formula (3), T is the total number of leaf nodes, G _lf For samples under lf leaf nodes

Sum of H _lf Is a sample under the lf leaf node->

Summing;

unknown f _k (x)＝ω _q(x) In (d), ω represents a leaf node value of the subtree k, and q (x) represents a structure of the subtree k; equation (3) is considered to relate to the variable ω _lf. The second order function of (1) making the first order derivative zero to obtain the optimal solution omega of the leaf node lf _lf ^* And an optimum target value Obj _lf ^* The following were used:

determining q (x), i.e. the structure of the tree: each time whether a node is divided and which feature variable is used as a division condition is judged from the root node, the Gain is determined, and the Gain is calculated as follows:

and traversing and calculating gains of various splitting modes, selecting a mode with the largest Gain to divide a leaf node into a left node and a right node downwards, and not splitting when all gains are less than or equal to 0.

Preferably, the longitude and latitude of the site of the date to be predicted, the temperature, the precipitation and the wind speed of the site region h period are input into K subtrees obtained by XGboost decision tree model training, and similarity indexes of adjacent sites j and j' are counted

Wherein->

For a Boolean value, a j site and a j' site fall on the same leaf node in subtree k and->

Otherwise->

ω _k Is the weight of subtree k, 0 is less than or equal to M _jj’ ≤1；

With 1-M _jj’ Representing the distance between two stations, with a distance parameter of M _a And (4) obtaining a site clustering result by the hierarchical clustering.

Preferably, the step S2 of correcting the real requirement of the single site, and bringing the corrected real requirement into the XGBoost decision tree to obtain the optimized site clustering result after training includes:

let the real demand of site i be x for a given time period h and n sites in the system _i I =1,2, …, n, and the number of records in the system is f _i I =1,2, …, n; defining alpha to represent the willingness of a user to transfer when encountering an unbalanced site, w _ij ＝α/D _ij Represents the transfer rate between the ith and jth stations, where D _ij Represents the distance between site i and site j;

S ₁ set of sites representing over-imbalance occurring during time period h, S ₂ When it is indicatedSet of sites within segment h where no imbalance has occurred, | S ₁ |＝n ₁ ，|S ₂ |＝n ₂ ，n ₁ +n ₁ ＝n，S ₁ ∪S ₂ ＝S；

Belong to S ₂ Station of x _i And f _i The relationship of (a) to (b) is as follows:

belong to S ₁ Station of x _i And f _i The relationship of (a) to (b) is as follows:

in equations (6) and (7), trans (i, j) represents the demand for borrowing/returning vehicles transferred from station i to station j, and pool (i, j) represents the demand for borrowing/returning vehicles if station j is an unbalanced station i at t _i And if not, the bol (i, j) =0, and the station i belongs to S ₁ ；t _i Indicating that site i is in an out-of-balance time zone;

for each of the groups S ₁ The sites of (A) are:

the constraints are as follows:

f _j -x _j ≥0 j∈S ₂ (10)

x _i ≥0, α≥0 i∈S (12)

the objective function is as follows:

in formula (13), S _c Represents the site set, | S, contained by the cluster C in the clustering result _c |＝n _c 。

Preferably, the step S3 of predicting the vehicle borrowing/returning demand of each station according to the trained XGBoost decision tree includes:

recording the number f of each site _i Is replaced by x _i And acquiring demand data after deviation rectification, training the XGboost decision tree model by using the demand data after deviation rectification as a label, and predicting the real demand of the site i in the h time period by inputting characteristic variables in the trained model as follows:

preferably, the step S4 of calculating the initial parking amount of the single vehicle by considering the capacity limit and the demand arrival distribution of each station includes:

estimation of the number of rejected demands in a single time period at a certain site:

given site j and time period h = [ t ] ₀ ^h ,t _s ^h ]Generating random values by the probability distribution of the demand for borrowing and the probability distribution of the demand for returning vehicles in h time period according to the Poisson distribution

And &>

Then the net change per vehicle in h>

Let site capacity be C, at initial time t ₀ ^h The refused vehicle borrowing requirement generated in the h time period under the condition that the number of the station single vehicles is qDesired value of a quantity pick>

Expressed as: />

In formula (15), u ₁ Representing the required quantity of the refused borrowed vehicles;

expected value of the required number of refused vehicles generated in h period

Expressed as:

in formula (16), u ₂ Representing the required quantity of the rejected vehicles;

under the condition that the initial single vehicle putting amount is q, the expected value of the rejected demand quantity in the h period

Comprises the following steps:

calculating q from 0 to C separately

Fetch and hold>

And q corresponding to the minimum value is the optimal initial putting amount of the station j in the h period.

Preferably, the estimation of the number of rejected demands in successive time periods at a station:

if measuring and calculating continuous n time intervals h ₁ ,..h _m ,..h _n Intra site j generationExpected value of the rejected demand quantity, h _m Is h _m-1 The end point inventory of (1) has three cases:

inventory p of single vehicles at h-time station _h Probability of becoming 0 from q:

inventory p of single vehicles at h-time station _h Probability of changing from q to C:

inventory p of single vehicles at h-time station _h Probability of changing from q to q':

under the condition that the initial single vehicle putting amount is q, the expected value of the rejected demand quantity generated by the station in continuous n periods is E (delta) _f )：

Take E (delta) _f ) And q corresponding to the minimum value is the optimal initial putting amount of the station in n continuous time intervals.

Preferably, the step S5 of dividing the scheduling partition of the city includes:

s51, screening out sites with overlarge net borrowing demand absolute values, namely extreme demand sites, sequentially taking the sites as centers, searching all the sites within 1 km near the sites, respectively calculating the distances between the sites and the center sites, collecting a zero demand site set from the nearest site, and terminating searching when the sum of the net borrowing demands of the site set is zero or slightly larger than zero to construct the zero demand site set for the next extreme demand site;

s52, if all the peripheral sites are absorbed and still do not reach the termination condition, the current site central position is taken as a virtual site, the sum of the net borrowing demand of the site in the set is taken as the net borrowing demand of the virtual site, the process is repeated, and new sites are absorbed;

s53, if the capacity of the site set reaches the threshold value and still does not reach the termination condition, releasing the site set, and directly constructing a zero-demand site set for the next extremely-demanding site; if no site which is not added with the site set is within 1 kilometer of a certain site, releasing the site set, adding the site into a 'drop order' site set, and directly constructing a next 'zero-demand' site set;

s54, after the site set of the extremely-demanding sites is completed, finding out peripheral sites on the geographical position in the rest sites, sequentially constructing a 'zero-demand' site set for the sites, and circulating the process until the number of the sites without the added site set is the same as that of the drop site set;

s55, regarding the single station as a zero-demand station set, calculating the geographical center positions of all the zero-demand station sets, replacing one station set with one virtual station, merging the virtual station with the highest longitude with the surrounding virtual stations, and stopping merging to serve as a scheduling partition when the number of the actual stations exceeds 40 and the sum of the inventory of the single vehicles of all the stations in the area is between the upper bound and the lower bound of the optimal single vehicle input amount of the area;

and S56, repeating S51-S54 to finish the dispatching partition division of all sites in the city.

Preferably, the step S6 of establishing the in-region single vehicle scheduling model, and obtaining the optimal selection of the scheduling path includes:

having several shared bicycle stations in a defined area, using set S ₁ = {1,2, … …, N } for dispatch operation by standard delivery truck, S for departure point of delivery truck ₂ And {0 }. The upper limit of the capacity of each truck is Q. In the whole scheduling operation process, the real-time loading e of each operation truck must meet the condition that e is more than or equal to 0 and less than or equal to Q;

objective function 1: minimization of commissioning costs and carbon emissions:

objective function 2: minimization of user dissatisfaction penalty:

constraint conditions are as follows:

L _ij ＝L _ki +q _i ,k∈{0,…,N-1},R _ki ＝1,i∈S ₁ ,j∈{2,…,N}∪{0}(27)

L _ij ≤M·R _ij (28)

L _0j ＝L _j0 ＝ 0 (30)

q _i is an integer (36)>

S represents the dispatching path of dispatching truck, S = S ₁ ∪S ₂ = {0,1,2, …, N }, where 0 denotes the departure point of a dispatch truck and the others denote sites; q. q.s _i Representing the number of bicycles dropped at the site, q _i <0 indicates a fetch of-q at site i _i A single vehicle; r _i,j Is a variable from 0 to 1, if R is in the dispatch route of the truck from station i to station set j _i,j =1, otherwise R _i,j ＝0；g _i Representing the number of the single vehicles of the station i before the dispatching work is carried out; f. of _i The number of the single vehicles of the station i after the dispatching work is finished is represented; f (x, i) represents the number of loss demands expected to be generated when the number of the single vehicles of the station i is x; l is _ij Representing the number of the single vehicles carried by the dispatching truck in the process of driving from the station i to the station j; q represents the capacity of the dispatch truck; d _ij Representing the transport distance between station i to station j; c ^e The fuel consumption of the dispatching truck in kilometers and kilograms is represented; c ^d Indicating an unsatisfied penalty for each demand that generates a rejected demand; c _i Represents the capacity of site i; m represents an infinite positive value; b denotes an auxiliary variable of the cancellation sub-loop.

Preferably, the obtaining of the optimal selection of the scheduling path through the hybrid multi-objective particle swarm optimization algorithm includes:

s61, generating an initial solution set: calculating a distance matrix between stations in the area, randomly selecting one station as a first station of a route, randomly selecting any one station from 4 stations closest to the station as a next destination according to equal probability, and repeating the process to avoid selecting the selected station until all stations are added into the route; taking the generated route as a solution R _ij Inputting the formulas (22) to (35), and calling the scipy of Python to solve the optimal bicycle loading/unloading number q at each station _i And R is _ij Forming a complete initial solution; forming other initial solutions in the same way, and constructing an initial solution set;

s62, fitness function design and individual optimal selection: and comparing the fitness values of the particles to obtain the dominance relation of the particles, wherein the fitness function is from an objective function 1 and an objective function 2 and is specified as follows:

s63, constructing a non-inferior solution set: defining a pareto dominant relationship for a particle z ₁ And z ₂ E.g. f ₁ (z ₁ )≤f ₁ (z ₂ ) And f is ₂ (z ₁ )<f ₂ (z ₂ ) Or f is ₁ (z ₁ )<f ₁ (z ₂ ) And f is ₂ (z ₁ )≤f ₂ (z ₂ ) Then particle z ₁ Dominant particle z ₂ (ii) a Likewise, if f ₁ (z ₂ )≤f ₁ (z ₁ ) And f is ₂ (z ₂ )<f ₂ (z ₁ ) Or f is ₁ (z ₂ )<f ₁ (z ₁ ) And f is ₂ (z ₂ )≤f ₂ (z ₁ ) Then particle z ₂ Dominant particle z ₁ (ii) a Otherwise, particle z ₁ And z ₂ There is no dominance relationship;

for arbitrary particles z _* If and only if f ₁ (z _* )≤f ₁ (z) and f ₂ (z _* )<f ₂ (z) or f ₁ (z _* )<f ₁ (z) and f ₂ (z _* )≤f ₂ When (z) is zero, particle z _* Called non-dominant particles or pareto particles, all of which are recorded in a non-inferior solution set REP;

s64, selecting global optimal selection: the global best particle gbest is selected from the REP, which saves all the non-dominated solutions found in the search process; at the beginning of the search, all the non-dominant particles constructed in the initialization phase are added to REP; in the searching process, comparing the current non-dominant particles found in each iteration with all solutions in the REP one by one; if a current particle is dominated by a particle in the REP, it is discarded; otherwise, such particles may be appended to REP; if particles in the REP are dominated by the new member, deleting the particles from the REP; when the REP exceeds the maximum capacity of the REP, deleting redundant solutions by using a self-adaptive grid method, and keeping the solution well distributed in the REP; for the selection of gbest, the solution space is divided into a number of equal grids, where the grid with fewer particles has a higher chance of being selected; setting the adjusting parameter as beta, the number of grids as n, the number of particles contained in the grid j as n _j That means, the probability that the solution in the ith lattice is selected as gbest is:

/>

s65, designing a crossover operator: updating particles in the algorithm using crossover operators of the genetic algorithm: updating the particle by pbest first, then by gbest, the particle intersection operator

Wherein X _i (t + 1) and X _i (t) represents the position of the ith particle in the current iteration (t) and the next iteration (t + 1), respectively; pbest (t) is the historical optimum position for the ith particle, gbest (t) is the global optimum position for the population; />

Is a crossover operator;

s66, designing a mutation operator: and (3) updating the particles by using a mutation operator combined with VNS, and adopting three neighborhood structures:

(1) And (3) node insertion: randomly selecting a site and removing it from its original location, and then inserting it into another random location;

(2) Node switching: randomly selecting two sites and exchanging the positions of the two sites;

(3) Randomly selecting two sites, and arranging the sites in reverse order;

the three neighborhood structures are randomly executed, and after mutation operation is carried out, if new particles dominate the original particles, the original particles are replaced; if the original particle is dominant, discarding the new particle; if there is no dominance between the two particles, one of them is randomly selected.

In conclusion, the shared bicycle demand forecasting and releasing scheduling method provided by the invention starts with demand forecasting and scheduling optimization of a shared bicycle, considers various practical situations and practical problems, designs a model and a solving method, provides a basis for operation decision of an urban shared bicycle system, is favorable for quickly realizing the demand of borrowing/returning the bicycle for urban residents, and is convenient for life and healthy trip; the service level of the shared bicycle platform is improved, the brand public praise and the client stickiness of enterprises are improved, the control cost is reduced, and the enterprise competitiveness is improved; simultaneously, the method is beneficial to reducing carbon emission of cities, protecting the environment and building smart cities.

Drawings

Fig. 1 is a schematic diagram of clustering sites by an XGBoost decision tree according to an embodiment of the present invention;

FIG. 2 is a full flow diagram of a predictive model of a shared bicycle demand prediction and delivery scheduling method provided by an embodiment of the present invention;

FIG. 3 is a graph showing the variation of the fitting degree under different parameter max _ depth settings in the prediction model;

FIG. 4 is a diagram showing the fitting error variation under different parameter n _ estimator settings in the prediction model;

FIG. 5 is a graph of the importance of various features in the predictive model using cover as an indicator;

FIG. 6 shows the results of clustering sites for vehicle borrowing demand for the shared bicycle system at 1 month, 24 days 8-00 in 2022;

FIG. 7 is a geographic image of a first clustered site;

FIG. 8 is a net demand forecast distribution over a period of time h, 1 month 25 days 2022;

FIG. 9 is a flowchart of estimating expected rejected demand quantity for n consecutive time periods;

FIG. 10 is a schematic diagram of a scheduling partition;

FIG. 11 is an example of a crossover operator;

FIG. 12 is a schematic diagram of three neighborhood structures;

fig. 13 is a scheduling path obtained with the best benefit according to the scheduling model.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.

The embodiment of the invention provides a method for forecasting and dispatching shared bicycle demands, which comprises the following steps:

s2, correcting the actual demand of a single site, and bringing the corrected demand into an XGboost decision tree to be trained to obtain an optimized site clustering result;

s4, considering the capacity limit and the required arrival distribution of each station, and calculating the initial putting amount of the single vehicle;

s5, dividing the dispatching partitions of the city;

and S6, establishing a single vehicle scheduling model in the region to obtain the optimal selection of the scheduling path.

The invention first aggregates similar and adjacent sites into clusters through XGboost. In the estimation of real demand, the interaction between stations is not negligible. When one station has no available stakes/cars and a customer borrows one at a nearby station, the full/empty station may generate demand spills to adjacent stations. Thus, a single station in a system can be processed without accounting for overflow effects, resulting in underestimation and overestimation of demand. This means that we cannot even regard the borrowing/returning records of stations that are always functioning properly (no imbalance has occurred) as a real demand, since part of their records may come from neighboring full/empty stations. In addition, the demand for stations is influenced not only by natural factors (such as weather conditions, population, land use, etc.), but also by the distribution of surrounding stations. For example, the need for one station will be reduced due to the establishment of a new station nearby. Therefore, focusing on the demand of an area rather than a single station helps to avoid the influence of inter-station transfer and interaction, and more accurately captures local travel patterns, which requires site clustering.

If a certain time period h (one hour) within a fixed day, the difference in demand between sites can be explained by geographical location and weather conditions. Fig. 1 is a schematic diagram of clustering sites by an XGBoost decision tree according to an embodiment of the present invention, and as shown in fig. 1, the establishing of the XGBoost decision tree in step S1 of the present invention, and the clustering of similar and adjacent sites includes:

in the formula (1), the number of subtrees is shown in a k tableAs shown in the figure, the material of the steel wire,

customer borrowing/returning demand prediction value, lon, representing station i _i Indicating the longitude, lat, of site i _i Indicates the latitude, tem, of site i _i Indicating the temperature, rain, of station i _i Indicating the precipitation, speed, of site i _i Representing the wind speed of a station i, the objective function of the XGboost decision tree model is as follows:

according to Taylor's theory of expansion, equation (2) can be converted to:

Sum of H _lf Is a sample under the lf leaf node->

Summing;

Further, the longitude and latitude of the site of the date to be predicted, the temperature, the precipitation amount and the wind speed of the site region h period are input into K subtrees obtained by XGboost decision tree model training, and similarity indexes of adjacent sites j and j' are counted

Wherein +>

Otherwise->

ω _k Is the weight of subtree k, 0 is less than or equal to M _jj’ ≤1；

It should be noted that the maximum depth max _ depth of the training subtree and the number n _ estimators of the subtrees are parameters that need to be determined carefully, too small affects the model accuracy, and too large is prone to overfitting.

The XGboost decision tree model is trained by taking position and weather information as input and taking the number of taxi borrowing/returning records as a target. The above process is the first stage of the prediction model provided by the present invention, that is, a model structure is proposed, and the sites are clustered according to the trained model structure, however, these values as labels are not really required. In the case of an unbalanced station (empty or full), the user cannot borrow or return the car, and the transaction record of the system does not accord with the initial car borrowing/returning requirement of the user, namely a part of the car borrowing/returning requirement is lost. Therefore, the original requirements of each site user need to be estimated, and the recorded data of the system needs to be rectified so that the result is close to the real user requirements.

The variables set in the deviation correction model are set as follows:

in an embodiment of the present invention, the step S2 of correcting the real requirement of a single site, and bringing the corrected real requirement into the XGBoost decision tree to obtain an optimized site clustering result after training includes:

belong to S ₁ Station of x _i And f _i The relationship of (c) is as follows:

for each of the groups S ₁ The sites of (A) are:

the constraints are as follows:

f _j -x _j ≥0 j∈S ₂ (10)

x _i ≥0, α≥0 i∈S (12)

the objective function is as follows:

the above process is the second stage of the prediction model provided by the present invention, that is, for the estimation of the real demand for borrowing, a non-linear programming model is established by equations (13), (6), (7) and (9) - (12), which is based on an assumption: when a site imbalance is encountered, the user will move to t _i The station which is always running is kept in the period. Although there may be a closer station, at t _i Has not been out of balance for part of the time, but it is not selected because it has been in stock at a dangerous level and the user is concerned that it becomes empty during walking.

The estimation of the returning demand is slightly different from the model of the borrowing demand. When the user encounters a full stop, they must return the bicycle to avoid further payment, which means w =1. Therefore, the model for correcting the borrowing demand is deleted from the formula (9), and the alpha/D is ignored _ij The method can be used for the deviation correction work of the vehicle returning demand.

Fig. 2 is a full flowchart of a prediction model of a method for demand forecasting and delivery scheduling of a shared bicycle according to an embodiment of the present invention. As shown in fig. 3, the predicting of the car borrowing/returning demand of each station according to the trained XGBoost decision tree in step S3 of the present invention includes:

recording the number f of each site _i Is replaced by x _i And acquiring the demand data after deviation rectification, training the XGboost decision tree model by using the demand data after deviation rectification as a label, and predicting the real demand of the site i in the h time period by inputting the characteristic variables in the trained model as follows:

the process is the third stage of the prediction model provided by the invention, namely, the XGboost decision tree model is trained through the requirement data after deviation correction, and then the trained model is used for predicting the real requirement of a single site.

The following are examples of XGboost decision tree model parameter debugging and clustering results:

the data processing and analysis work in this example was performed using Python (Anaconda 3.7) encoding and running, and experiments were calculated on Intel Core i7-7500 CPU@2.70GHz, where the Xgboost model was trained by a tensrflow CPU (1.14.0), setting the learning rate eta =0.2 for the model, a complexity penalty (γ =0.5, λ = 1), and trying different max _ depth and n _ estimator adjustment parameters. The study samples were selected from all weekdays from 1/2021 to 25/1/2022, with 20 weekdays as a window, and were in a rolling prediction mode. For example, in predicting the demand for vehicle borrowing/returning at 25/1/2022, the training data for the model should be taken from 17/12/2021 to 24/1/2022.

FIG. 3 shows the variation of the fitting degree under different parameter max _ depth settings in the prediction model. The greater the maximum depth of the tree max _ depth, the higher the fitness of the model, and the more overfitting is likely to occur. When training the xgboost model with the sample set, different max _ dept are set, respectively, and the degree of fitting is as shown in fig. 3, and the degree of fitting increases slowly after depth 6, so that parameter max _ dept =6 is set.

Fig. 4 shows the fitting error variation under different parameter n _ estimator settings in the prediction model. The larger the number n _ estimators of the subtrees, the smaller the fitting error of the model, but at the same time overfitting tends to occur more easily. When the xgboost model is trained using the sample set, different numbers of learners are set, and the root mean square error (rmse) corresponding to the numbers is as shown in fig. 4, and since the rmse increases slowly after the number of subtrees is 8, the parameter n _ estimator =8 is set.

The rmses for the first 8 subtrees were 5.9573,5.6791,5.4581,5.3188,5.1856,5.1139,5.0279 and 4.9403, respectively. Q of tag data is counted according to a quartile method ₁ And Q ₃ Finding a four-bit distance IQR = Q ₃ -Q ₁ =9. It is considered here that the average gap level from the target value is 8, taking values at will without any prediction. Then the contribution of the 1 st sub-tree is 8-5.9573=3.0427 and the other contributions are in turn 0.2782,0.2210,0.1393,0.1332,0.0717,0.086,0.0876. And the weights of the first eight subtrees are respectively calculated to be 74.95%,6.85%,5.44%,3.43%,3.28%,1.77%,2.12% and 2.16%, and the eight subtrees are used for hierarchical clustering.

The trained XGBoost model can be represented by the following table, wherein f 0-f 4 are 5 characteristics of latitude, longitude, temperature, wind speed and rainfall respectively. The information in the table shows that the total number of nodes of all 8 subtrees is 949, where the eighth subtree has 120 tree nodes (the sum of the root node, the split node, and the leaf node). The first row means: the splitting condition of the root node of the first tree is 'latitude < 40.767029', if the condition is met, the second node (the node with the ID of 0-1) of the first subtree is entered, otherwise, the third node (the node with the ID of 0-2) of the first subtree is entered, and the Gain of the node for carrying out sample division by using the condition is 5425.8652 (the Gain is calculated in a formula of 3.1.2, and 3-5). Since the leaf nodes are no longer split, the split condition parts are all null NaN.

TABLE 1 partial demonstration of the trained xgboost model

	Tree	Node	ID	Feature	Split	Yes	No	Gain
									0	0	0	0-0	f0	40.767029	0-1	0-2	5425.8652
1	0	1	0-1	f0	40.736374	0-3	0-4	14389.0361
									2	0	2	0-2	f1	-73.96813	0-5	0-6	10376.0088
3	0	3	0-3	f0	40.731632	0-7	0-8	3230.9951
									4	0	4	0-4	f1	-73.98779	0-9	0-10	11749.0039
...	...	...	...	...	...	...	...	...
									945	7	116	7-116	Leaf	NaN	NaN	NaN	0.3539
946	7	117	7-117	Leaf	NaN	NaN	NaN	-1.4300
									947	7	118	7-118	Leaf	NaN	NaN	NaN	-0.3808
948	7	119	7-119	Leaf	NaN	NaN	NaN	0.1083
									949	7	120	7-120	Leaf	NaN	NaN	NaN	-0.0112

In the xgboost module, there are 3 indexes for measuring the importance degree of the features. Weight represents the number of times a feature is used to split a node in all trees. cover represents the average number of samples a feature processes (covers) when splitting nodes across all trees. Gain represents the average Gain that a feature brings each time a node is split. Fig. 5 shows how important each feature in the predictive model is based on cover, and as shown in fig. 5, it can be seen that the location information is significant because it includes many factors such as land utilization, population density, building facilities, and the like, the second most important is rainfall factor, then temperature, and finally wind speed.

FIG. 6 shows the results of clustering sites for vehicle borrowing demand for the shared bicycle system at 1 month and 24 days 8-00 in 2022. Clustering sites by using a trained xgboost model, taking 8 am 00-9 am on 24 days of 1 month and 2022 as an example, taking 997 active sites on the same day, taking samples on the same day from a DataSet to form a test set input model, obtaining a similarity matrix, carrying out hierarchical clustering by using a distance threshold parameter Ma =0.05, and obtaining a similarity matrix in 2022-01-24 hours h ₈ The station clustering result under the weather condition of (1), as shown in fig. 6.

The following are examples of correction and prediction of demand:

as the deviation correction of the demand takes the whole system as a planning problem, taking the first cluster site as an example, the deviation correction work is shown to be process analysis. FIG. 7 is a geographic image of a first cluster site, which as shown in FIG. 7 includes 8 sites with IDs 7713.11,7727.08,7692.11,7713.01,7680.03,7650.05,7646.04, and 7634.01. The codes in this section are stations 1-8.

During period [8, 00,9 ] of 2022-01-24, the single-vehicle inventory change for each station is shown in table 2.

Table 2 inventory tracking of sites in the first cluster within h = [8,00, 9 ]

As can be seen from table 2, station 1 is out of service in an unbalanced state (empty station) between 8 ₁ ＝[8:00,8:53]In the same way, t ₂ ＝[8:00,9:00]，t ₅ ＝[8:37,8:48]，t ₆ ＝[8:52,9:00]，t ₇ ＝[8:48,9:00].

Stations

3, 4, 8 remain available for h all the time. However, eight sites x will be described according to section 3.2.2 _i And f _i The relationship between them is as follows:

f ₃ -x ₃ ＝trans(2,3)

f ₄ -x ₄ ＝trans(1,4)

f ₈ -x ₈ ＝trans(6,8)

/>

based on the system history record, (f 1=0, f2=0, f3=2, f4=3, f5=3, f6=4, f7=4, f8= 1) and (D1, 4=241.02, D2,3=225.22, D5,6=247.89, D6,8=231.39, D7,5= 474.86) are given. The solution was obtained using the scipy. Optimize module (α =93.4857, x1=3, x2=2, x3=1, x4=2, x5=3, x6=4, x7=4x8= 1). This indicates that during t1, 1 user has transferred from site 1 to

site

4,2, the idea that the user gave up renting a shared bike at site 1; within t1, 1 user transfers from site 2 to 3 sites, and 1 customer gives up renting at site 2; within t7, no customer arrives at site 7; within t5, no customer arrives at site 5; within t6, no customer arrives at site 6.

Since no station has a full unbalance state in this example, the number of the car return records in the system record can be directly regarded as the real car return demand of each station. If a full station occurs, the same method is used (ignoring α/D) _ij ) And correcting the vehicle returning records of all the stations. In the whole system, 41 sites generate full station imbalance and 67 sites generate empty station imbalance on 24 days in 1 month and 2022, and 86 overflow sites are involved in the imbalance period of the two sites. The number of stations affected by the overflow effect is less than the number of stations with imbalance because two unbalanced stations may have an overflow effect on the same station, and a mutual overflow effect of demand may occur between an empty unbalanced station and a full unbalanced station. When the deviation rectifying optimization model of the borrowing demand is established, the quantity to be solved is 68 (the real demand x of 67 stations) _i And transfer coefficient α), the model solution takes 16 seconds; building (C.E.)When the deviation rectifying optimization model of the vertical returning vehicle requirement is adopted, 41 quantities to be solved (the real requirement x of 41 stations) are provided _i ) The model solution took 7 seconds.

Fig. 8 is the net demand forecast profile over time h of 25 days 1 month 2022. Each 20 days of work were completed in h = [8,00,9]Inner data correction with new data x _i Updating the label of the original DataSet DataSet, and training the XGboost model again to obtain the customer demand prediction result between 1 month and 25 days in 2022, 08. The net borrowing demand is obtained by subtracting the returning demand predicted value from the borrowing demand predicted value at each station, as shown in fig. 8.

It can be seen from fig. 8 that the net borrowing demand distribution is shown to be significantly left biased. Because in 8 am to 9 am of a working day, there are individual places such as nearby office buildings where there is a concentrated and large amount of returning demand, and the demand for borrowing may be scattered in various residential areas, with no very obvious demand for borrowing.

And (3) measuring the average deviation level of the predicted values and the recorded values of all the stations in the system by using the average absolute percentage error devi, and calculating to obtain 2.8836 by using the following calculation formula:

in one embodiment of the present invention, the initial placement scenario for a single vehicle is discussed further in view of the capacity limit and demand arrival distribution for each site. Assuming that the arrival of the demand in the unit time obeys the poisson distribution, a set of random values of the car borrowing/returning demand arriving in the unit time can be obtained by using the forecast values of the car borrowing/returning demand of each time interval of the station. Simulating the situation that the borrowing/returning demand reaches the station for many times, tracking the number of the single vehicles and the vehicle piles of the station, and then counting the number delta of the refused borrowing demand occurring in a fixed time period _w And the required number delta of rejected vehicles _r The expected delta of the rejected demand quantity is represented by the sum of the mean values of the two multiple simulation results _f ＝δ _w +δ _r And measuring the user dissatisfaction generated by the site.

Specifically, the step S4 of calculating the initial release amount of the single vehicle, taking into account the capacity limit and the demand arrival distribution of each station, includes:

on one hand, the estimation of the rejected demand quantity in a single time period of a certain site is as follows:

given site j and time period

According to the poisson distribution, random values are generated according to the probability distribution of the borrowing demand and the probability distribution of the returning demand in the h time period>

And &>

Then the net change of the bicycle in h>

Let site capacity be C, at initial time t ₀ ^h The expected value of the number of the rejected vehicle demands generated in the h period is greater than or equal to the expected value of the number of the rejected vehicles in the condition that the number of the single vehicles at the station is q>

Expressed as:

in formula (15), u ₁ Indicating the quantity of the refused vehicle borrowing requirements;

expected value of the required quantity of refused vehicles generated in h period

Expressed as:

Comprises the following steps:

calculating q from 0 to C separately

Taking or combining>

On the other hand, the estimation of the rejected demand quantity in a certain station continuous time period is as follows:

if measuring and calculating continuous n time intervals h ₁ ,..h _m ,..h _n The home site j generates an expected value, h, of the number of denied requests _m Is h _m-1 The end point inventory has three cases:

FIG. 9 is a flow chart of estimating expected values of the rejected demand amounts for n consecutive time periods. hm may have any value from 0 to C and have a corresponding probability, and for each initial vehicle amount, the end of hm period may have a discrete probability distribution of 0 to C inventory (process section 4.1.1). The probability that the inventory of the hm time period end point is each value is calculated by a total probability formula, and the obtained probability distribution is used as the initial condition of the hm +1 time period. The iterative process is shown in figure 9.

Under the condition that the initial single vehicle putting amount is q, the expected value of the rejected demand quantity generated by the station in continuous n time periods is E (delta) _f )：

Take E (delta) _f ) And q corresponding to the minimum value is the optimal initial putting amount of the station in continuous n periods.

The shared bicycle mainly solves the problem of the last kilometer, the riding distance of most people is 500-1000m, and 90% of users ride within two kilometers. The flow and the probability of the single vehicles among cities are within a small range, and the number of the single vehicles in one area can keep a rough balance for a long period of time. Based on this, the invention proposes the concept of urban dispatch zoning. Ideally, the overall supply and demand relationship of a shared bicycle system is balanced, and the bicycle is "borrowed and returned" and the station with the bicycle in short supply in the system is also required to have the station with the bicycle in excess. The partition scheduling idea is to divide a city into a plurality of blocks on an area after the net borrowing demand prediction value of each station is obtained, and the sum of the net borrowing demands of the stations in each block is guaranteed to be zero as much as possible, namely the interior of a station set included in each block can be self-sufficient. Theoretically, 1 dispatching truck is equipped for each person dispatching zone, the truck does not need to carry a single truck when the operation is initial, rebalancing is realized by finishing mutual dispatching among stations in the zones, and the redundant single truck cannot be carried after the operation is finished.

Specifically, in an embodiment of the present invention, the partitioning the city into the scheduling partitions in step S5 includes:

s52, if all the peripheral sites are not absorbed yet to the end condition, taking the central position of the current site as a virtual site, taking the sum of the net borrowing demand of the site in the set as the net borrowing demand of the virtual site, repeating the process, and absorbing the addition of a new site;

Fig. 10 is a schematic diagram of a scheduling partition implemented according to the above method.

In the embodiment of the present invention, the dispatching task of each partition is an independent problem if each partition is equipped with a dispatching truck.

The variables set in the scheduling model are set as follows:

step S6, the establishment of the in-area single vehicle dispatching model and the acquisition of the optimal selection of the dispatching path comprise the following steps:

objective function 1: minimization of commissioning costs and carbon emissions:

objective function 2: minimization of user dissatisfaction penalty:

constraint conditions are as follows:

/>

L _ij ≤M·R _ij (28)

L _0j ＝L _j0 ＝ 0 (30)

constraints (24) to (26) limit the upper limit of the pick-up amount of the truck at the station to the initial inventory of the station, and the upper limit of the drop-out amount cannot exceed the remaining capacity of the station. Constraints (27) to (29) ensure that the number of vehicles carried by the transport vehicle is non-negative and does not exceed the maximum capacity limit of the truck. And the constraint condition (30) constrains the transport vehicles to start in an unloaded mode and return in an unloaded mode, and the total number of the single vehicles in the region is guaranteed to be the same before and after scheduling. Constraints (31) ensure that all sites are represented as being visited at most once to meet demand. Constraints (32) ensure that routes are formed back-and-forth between sites. The constraints (33) ensure that the rebalance vehicle starts from a location and returns to the location. The constraint (34) is a cancellation of the sub-loop constraint. Constraints (35) and (36) define the domain of the variables.

In one embodiment of the present invention, obtaining an optimal selection of a scheduling path by a hybrid multi-objective particle swarm optimization algorithm comprises:

s61, generating an initial solution set: calculating a distance matrix between stations in the area, randomly selecting one station as a first station of a route, randomly selecting any one station from 4 stations closest to the station as a next destination according to equal probability, and repeating the process to avoid selecting the selected station until all stations are added into the route; taking the generated route as a solution R _ij Inputting the formulas (22) - (35), and calling the scipy of Python to solve to obtain the optimal bicycle loading/unloading number q at each station _i And R is _ij Forming a complete initial solution; forming other initial solutions in the same way, and constructing an initial solution set;

s63, constructing a non-inferior solution set: defining pareto dominance relationshipsFor the particle z ₁ And z ₂ E.g. f ₁ (z ₁ )≤f ₁ (z ₂ ) And f is ₂ (z ₁ )<f ₂ (z ₂ ) Or f is ₁ (z ₁ )<f ₁ (z ₂ ) And f is ₂ (z ₁ )≤f ₂ (z ₂ ) Then particle z ₁ Dominant particle z ₂ (ii) a Also, if f ₁ (z ₂ )≤f ₁ (z ₁ ) And f is ₂ (z ₂ )<f ₂ (z ₁ ) Or f is ₁ (z ₂ )<f ₁ (z ₁ ) And f is a ₂ (z ₂ )≤f ₂ (z ₁ ) Then particle z ₂ Dominant particle z ₁ (ii) a Otherwise, particle z ₁ And z ₂ No dominance relationship;

for arbitrary particles z _* If and only if ₁ (z _* )≤f ₁ (z) and f ₂ (z _* )<f ₂ (z) or f ₁ (z _* )<f ₁ (z) and f ₂ (z _* )≤f ₂ When (z) is zero, particle z _* Called non-dominant particles or pareto particles, all of which are recorded in a non-inferior solution set REP;

s64, selecting global optimal selection: the global best particle gbest is selected from the REP, which saves all the non-dominated solutions found in the search process; at the beginning of the search, adding all the non-dominant particles constructed in the initialization phase to the REP; in the searching process, comparing the current non-dominant particles found in each iteration with all solutions in the REP one by one; if a current particle is dominated by a particle in the REP, it is discarded; otherwise, such particles may be appended to REP; if particles in the REP are dominated by the new member, deleting the particles from the REP; when the REP exceeds the maximum capacity of the REP, deleting redundant solutions by using a self-adaptive grid method, and keeping the solution well distributed in the REP; for the selection of gbest, the solution space is divided into a number of equal grids, where the grid with fewer particles has a higher chance of being selected; setting the adjusting parameter as beta, the number of grids as n, the number of particles contained in the grid j as n _j That means, the probability that the solution in the ith grid is selected as gbest is:

Wherein X _i (t + 1) and X _i (t) represents the position of the ith particle in the current iteration (t) and the next iteration (t + 1), respectively; pbest (t) is the historical optimum position of the ith particle, gbest (t) is the global optimum position of the population; />

Is a crossover operator;

FIG. 11 is an example of a crossover operator. As shown in FIG. 11 (a), X is first selected _i (t) and its pbest _i As a parent P ₁ And P ₂ Then, updating the vehicle route by exchanging chromosome segments randomly selected by the chromosome segments, replacing repeated stations in the route with lost stations by using a nearby principle, and then adjusting the number of the bicycles loaded/unloaded at each station; after the crossover operation, two offspring particles O are generated ₁ And O ₂ As shown in fig. 11 (b); comparing them by the fitness value calculated by equation (37), selecting the one that is not dominated to perform the subsequent operation; if they have no dominance relationship, randomly selecting one of them; in FIG. 11, the descendant O ₂ Is selected as a new parent P ₁ ', gbest is another parent P ₂ '; process P in the same way ₁ ' and P ₂ ' Generation of progeny particles O ₁ ' and O ₂ ', as shown in FIG. 11 (c); similarly, non-dominant individuals are considered as renewed particles X _i (t+1)。

S66, designing a mutation operator: updating the particle by using a mutation operator combined with VNS, and adopting three neighborhood structures, fig. 12 is a schematic diagram of the three neighborhood structures, as shown in fig. 12:

(2) Node switching: randomly selecting two sites and exchanging the positions of the sites;

(3) Randomly selecting two sites, and arranging the sequence between the two sites in a reverse order;

the three neighborhood structures are randomly executed, and after mutation operation is carried out, if a new particle dominates the original particle, the original particle is replaced; if the original particles are dominant, discarding the new particles; if there is no dominance between the two particles, one of them is randomly selected.

Fig. 13 is a scheduling path obtained with the best benefit according to the scheduling model. Because the number of the stations of each block is still large after partitioning, for convenience of display, the invention takes four 'zero-demand' station sets as an example and carries out scheduling solution respectively. The optimal result after the algorithm is run for 10 times is taken as the solving result of each example, and the optimal benefit path of each example is shown in fig. 13.

The number of sites for the four examples is 15, 20, 25 and 30, and the corresponding solution run lengths are 18.37 seconds, 27.7 seconds, 42.38 seconds and 59.38 seconds, respectively.

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims

1. A method for forecasting demand and dispatching release of shared bicycle comprises the following steps:

s1, establishing an XGboost decision tree, and gathering similar and adjacent sites into a cluster;

s5, dividing the dispatching partitions of the city;

2. The method of claim 1, wherein the method comprises: establishing the XGboost decision tree in the step S1, and gathering similar and adjacent sites into a cluster comprises the following steps:

in the formula (1), the number of subtrees is represented by k,

according to Taylor's theory of expansion, equation (2) can be converted to:

Sum of H _lf Is a sample under the lf leaf node->

Summing;

unknown f _k (x)＝ω _q(x) In (e), ω represents a leaf node value of the subtree k, q (x) represents a structure of the subtree k; equation (3) is considered to relate to the variable ω _lf. The second order function of (2) to make the first order derivative zero to obtain the optimal solution omega of the leaf node lf _lf ^* And an optimum target value Obj _lf ^* The following:

3. The method of claim 2, wherein the demand forecasting and delivery scheduling method comprises: longitude and latitude of a station of a date to be predicted, and temperature of a station area in a period hInputting the precipitation and the wind speed into K subtrees obtained by XGboost decision tree model training, and counting similarity indexes of adjacent sites j and j

Wherein

Otherwise

ω _k Is the weight of subtree k, M is more than or equal to 0 _jj’ ≤1；

4. The method of claim 1, wherein the method comprises: in the step S2, the correcting the real demand of a single site, and bringing the corrected real demand into the XGBoost decision tree to be trained to obtain an optimized site clustering result includes:

let the real demand of site i be x for a given time period h and n sites in the system _i I =1,2, …, n, the number of records in the system is f _i I =1,2, …, n; defining alpha to represent the willingness of a user to transfer when encountering an unbalanced site, w _ij ＝α/D _ij Represents the transfer rate between the ith and jth stations, where D _ij Represents the distance between site i and site j;

S ₁ set of sites representing over-imbalance occurring during time period h, S ₂ Represents the set of sites that never experience imbalance, | S, within a period h ₁ |＝n ₁ ，|S ₂ |＝n ₂ ，n ₁ +n ₁ ＝n，S ₁ ∪S ₂ ＝S；

Belong to S ₂ Station of x _i And f _i The relationship of (c) is as follows:

for each of the groups S ₁ The sites of (2) have:

the constraints are as follows:

f _j -x _j ≥0 j∈S ₂ (10)

x _i ≥0,α≥0 i∈S (12)

the objective function is as follows:

5. The method of claim 4, wherein the demand forecasting and delivery scheduling method comprises: predicting the vehicle borrowing/returning requirement of each station according to the trained XGboost decision tree in the step S3 comprises the following steps:

6. the method of claim 1, wherein the demand forecasting and delivery scheduling method comprises: in step S4, the calculating of the initial delivery amount of the single vehicle, taking into account the capacity limit and the demand arrival distribution of each station, includes:

estimation of rejected demand quantity in a single time period of a certain site:

given site j and time period h = [ t ] ₀ ^h ,t _s ^h ]According to the Poisson distribution, random values are generated by the probability distribution of the demand for borrowing and the probability distribution of the demand for returning vehicles in the h time period respectively

And &>

Then the net change of the bicycle in h>

Let site capacity be C, at initial time t ₀ ^h Expected value of the quantity of refused borrowing demands generated in h time period under the condition that the quantity of the station single vehicles is q

Expressed as:

expected value of the required number of refused vehicles generated in h period

Expressed as:

Comprises the following steps:

calculating q from 0 to C separately

Fetch and hold>

7. The method of claim 6, wherein the demand forecasting and delivery scheduling method comprises:

estimation of the number of rejected demands in a certain station in consecutive time periods:

if measuring and calculating continuous n time intervals h ₁ ,..h _m ,..h _n The inner site j generates the expected value of the rejected demand quantity, h _m Is h _m-1 The end point inventory has three cases:

inventory p of vehicles at h-time station _h Probability of q becoming 0:

inventory p of vehicles at h-time station _h Probability of changing from q to C:

inventory p of vehicles at h-time station _h Probability of q becoming q':

Taking E (delta) _f ) And q corresponding to the minimum value is the optimal initial putting amount of the station in n continuous time intervals.

8. The method of claim 1, wherein the demand forecasting and delivery scheduling method comprises: the step S5 of dividing the scheduling partition of the city includes:

s53, if the capacity of the site set reaches the threshold value and still does not reach the termination condition, releasing the site set and directly constructing a zero-demand site set for the next extremely-demanded site; if no site which is not added with the site set is within 1 kilometer of a certain site, releasing the site set, adding the site into a 'drop order' site set, and directly constructing a next 'zero-demand' site set;

s55, regarding the drop-off station as a zero-demand station set, calculating the geographical center positions of all the zero-demand station sets, replacing one station set with one virtual station, merging with surrounding virtual stations by taking the virtual station with the highest longitude as the start, and stopping merging when the number of the actual stations exceeds 40 and the sum of the inventory of all the stations in the area is between the upper bound and the lower bound of the optimal single vehicle throwing amount in the area to serve as a scheduling partition;

9. The method of claim 1, wherein the method comprises: step S6, the establishment of the in-area single vehicle dispatching model and the acquisition of the optimal selection of the dispatching path comprise the following steps:

having several shared bicycle stations in a defined area, using set S ₁ = {1,2, … …, N } for dispatch operation by standard delivery truck, S for departure point of delivery truck ₂ And {0 }. The upper capacity limit for each truck is Q. In the whole scheduling operation process, the real-time loading e of each operation truck must meet the condition that e is more than or equal to 0 and less than or equal to Q;

objective function 1: minimization of commissioning costs and carbon emissions:

the objective function 2: minimization of user dissatisfaction penalty:

constraint conditions are as follows:

/>

L _ij ≤M·R _ij (28)

L _0j ＝L _j0 ＝0 (30)

q _i is an integer (36)

S represents the dispatching path of dispatching truck, S = S ₁ ∪S ₂ = {0,1,2, …, N }, where 0 represents the departure point of the pickup truck and the others represent stations; q. q.s _i Representing the number of cars dropped at the site, q _i <0 indicates a fetch of-q at site i _i A single vehicle; r _i,j Is a variable of 0 to 1, if R is in the dispatch route of the truck from station i to station set j _i,j =1, otherwise R _i,j ＝0；g _i Representing the number of the single vehicles of the station i before the dispatching work is carried out; f. of _i Representing the number of the single vehicles of the station i after the dispatching work is finished; f (x, i) represents the number of loss demands expected to be generated when the number of the single vehicles of the station i is x; l is _ij Representing the number of the single vehicles carried by the dispatching truck in the process of driving from the station i to the station j; q represents the capacity of the dispatch truck; d _ij Representing the transport distance between station i to station j; c ^e The fuel consumption of the dispatching truck in kilometers and kilograms is represented; c ^d Indicating an unsatisfied penalty for each demand that generates a rejected demand; c _i Represents the capacity of site i; m represents an infinite positive value; b denotes the auxiliary variable of the cancellation sub-loop.

10. The method of shared-bicycle demand forecasting and delivery scheduling of claim 9, wherein: the optimal selection of the scheduling path is obtained through a hybrid multi-objective particle swarm optimization algorithm, and the method comprises the following steps:

s63, constructing a non-inferior solution set: defining a pareto dominant relationship for a particle z ₁ And z ₂ E.g. f ₁ (z ₁ )≤f ₁ (z ₂ ) And f is ₂ (z ₁ )<f ₂ (z ₂ ) Or f is ₁ (z ₁ )<f ₁ (z ₂ ) And f is ₂ (z ₁ )≤f ₂ (z ₂ ) Then particle z ₁ Dominant particle z ₂ (ii) a Also, if f ₁ (z ₂ )≤f ₁ (z ₁ ) And f is a ₂ (z ₂ )<f ₂ (z ₁ ) Or f is ₁ (z ₂ )<f ₁ (z ₁ ) And f is ₂ (z ₂ )≤f ₂ (z ₁ ) Then particle z ₂ Dominant particle z ₁ (ii) a Otherwise, particle z ₁ And z ₂ No dominance relationship;

s64, selecting global optimal selection: the global best particle gbest is selected from the REP, which saves all the non-dominated solutions found in the search process; at the beginning of the search, all the non-dominant particles constructed in the initialization phase are added to REP; in the searching process, the current non-dominant particles found in each iteration are compared with all solutions in the REP one by one; if a current particle is dominated by a particle in the REP, it is discarded; otherwise, such particles may be appended to REP; if particles in the REP are dominated by the new member, deleting the particles from the REP; when the REP exceeds the maximum capacity of the REP, deleting redundant solutions by using a self-adaptive grid method, and keeping the solution well distributed in the REP; for the selection of gbest, the solution space is divided into multipleAn equal grid, where a grid with fewer particles has a higher chance of being selected; let the adjustment parameter be beta, the number of grids be n, the number of particles included in the grid j is n _j That means, the probability that the solution in the ith lattice is selected as gbest is:

Is a crossover operator;

(3) Randomly selecting two sites, and arranging the sites in reverse order;

the three neighborhood structures are randomly executed, and after mutation operation is carried out, if new particles dominate the original particles, the original particles are replaced; if the original particles are dominant, discarding the new particles; if there is no dominance between the two particles, one of them is randomly selected.