CN104778508A - Public bicycle renting forecasting method based on multi-source data fusion - Google Patents

Abstract

The invention discloses a public bicycle renting forecasting method based on multi-source data fusion. According to the method, historical data about public bicycle renting/returning records, weather, temperature, holidays, festivals and the like are cleaned and preprocessed, and training datasets are acquired; the datasets are classified with a clustering algorithm, and different renting modes of public bicycles are divided; the classified datasets are used to establish a Bayesian classifier used for forecasting the renting modes according to conditions of holidays, festivals, weather and air temperature of one day in the future; a self-adaptive particle swarm neural network model corresponding to each mode is trained for different modes of datasets respectively; finally, the renting mode of one day is forecasted by the aid of the Bayesian classifier, a corresponding particle swarm neural network model is selected to forecast the renting law of public bicycles. The forecasting accuracy is high, the operation speed is high, reference basis is provided for bicycle renting and returning by a user, the duration time of the unbalanced state of a public bicycle station is shortened, and the users' satisfaction is improved.

Description

A kind of public bicycles based on multisource data fusion leases Forecasting Methodology

Technical field

The invention belongs to city intelligent public transportation system technical field, relate to a kind of public bicycles based on multisource data fusion and lease Forecasting Methodology, particularly the method that quantity predicts is leased to Different periods bicycle in each public bicycles website.

Background technology

Public bicycles system, as a part for urban public transport, has the advantages such as pollution-free, mobility strong, effectively can alleviate urban traffic pressure, reduces CO2 emission, improves urban environment.Due to mobility and the tide of citizens' activities, there will be the problem can borrowed without bicycle, can go back without room in some time period.City ITS (IntelligentTransportation System, be called for short ITS) be apply to whole traffic management system by effectively integrated to the infotech of advanced person, mechanics of communication, sensing technology, control technology and computer technology etc., and set up a kind of on a large scale in, comprehensively to play a role, in real time, multi-transportation and management system accurately and efficiently.May be used for the problems solving the appearance of urban public transport field.

Public bicycles leases the part that quantitative forecast is City ITS, its objective is and predicts that in each public bicycles website different time sections, quantity leased by bicycle exactly, thus launch Operation Measures in advance, effectively solves " renting/return the car difficulty " problem.

At present, less in the research in this field both at home and abroad, and major part is the prediction for automotive field.In these researchs, the main method of proposition comprises: history averaging model, autoregressive model, multivariate regression model, nonparametric Regression Model, Kalman filter model, neural network model, supporting vector machine model etc.But due to the singularity of public bicycles, these methods are not be well suited for for the prediction in this field.History averaging model and regression analysis model all suppose that the magnitude of traffic flow follows strict periodicity, and public bicycles is leased and be there is uncertainty, non-linear, acyclic feature.Kalman filter model is not be well suited on the traffic flow forecasting at short period interval.Supporting vector machine model shows poor when using a large amount of True Data to carry out traffic flow forecasting.And neural network model carries out modeling according to input and output, be a kind of parallel computation model, arithmetic speed is very fast, and has good non-linear mapping capability and the property learnt by oneself, adaptive ability.But traditional neural network model also also exists the shortcoming that speed of convergence slowly, is easily absorbed in local optimum.

Summary of the invention

The object of the invention is the deficiency in order to overcome Individual forecast model, proposing a kind of public bicycles based on multisource data fusion and leasing Forecasting Methodology.The present invention improves classic method and merges, the basis of cluster analysis use Naive Bayes Classification Algorithm carry out leasing model prediction, respective self-adaptation PSO Neural Network model is trained respectively to different patterns of leasing, and use these model predictions difference to lease public bicycles under pattern lease quantity, improve and lease prediction accuracy.

The concrete steps that the present invention solves the technical scheme that its technical matters adopts are as follows:

The present invention specifically comprises the steps:

Step (1) data acquisition and procession:

Read the user stored in public bicycles system to hire a car record, calculate each website and remain public bicycles quantity every time period t ime, set up vectorial N _date=[n ₁, n ₂..., n _j..., n _m], M is time period sum; Obtain from internet weather history situation, temperature and festivals or holidays situation, set up environment attribute vector S={S ₁, S ₂..., S _i..., S _l, S _i=[d, w, t], L are the number that data sample concentrates sample; M, L, i and j are positive integer, wherein j≤M, i≤L;

Described user hires a car that record comprises lease duration, leases website, leases knee number, the time of returning the car, website of returning the car, knee number of returning the car;

N _jrepresent residue bicycle quantity in this website of a jth time period t ime;

S _irepresent the environment attribute of i-th day;

D represents attribute festivals or holidays on the same day, and the value of d is

W represents weather conditions, and the value of w is

T represents temperature, and the value of t is

Step (2) K-means cluster:

N _date=[n ₁, n ₂..., n _j..., n _m] as data sample to be sorted concentrate a data tuple; Concentrate random selecting k to put as initial classes center from data sample, use μ respectively ₁, μ ₂..., μ _krepresentation class c ₁, c ₂..., c _kclass center; Wherein, k is positive integer;

K-means clustering algorithm comprises the steps:

2-1. calculates remaining data tuple N _iwith class center μ _kdistance, by remaining data tuple N _iassign in the class the shortest with its distance;

2-2. recalculates all kinds of Xin Lei center μ _k ^*;

2-3. repeats step 2-1 and step 2-2, until each data tuple generic all no longer changes;

Step (3) builds Bayes classifier: after using K-means clustering algorithm to classify to set of data samples, obtain k class, each data tuple that data sample is concentrated will obtain a class label; Bayes Method is based on Bayes' theorem:

$P\left(H\right|X)=\frac{P\left(X\right|H)P\left(H\right)}{P\left(X\right)}$

Calculate P (X), P (H) and P (X|H);

Step (4) self-adaptation PSO Neural Network model training

Use sorted data sample to train affiliated self-adaptation PSO Neural Network model of all categories, the input layer unit number of setting neural network is n, and hidden layer unit number is p, and output layer unit number is q; Then need the weights number N=n × p+p × q adjusted in neural network, by the dimension needing the weights number N problem of representation space adjusted; Use vectorial X _i=[x _i1, x _i2..., x _iN] represent one of all weights of neural network possible value, also represent the position of particle in particle swarm optimization algorithm; PSO Neural Network model training step is as follows:

The position of each particle and speed in 4-1. initialization population population, position and the speed of particle are positioned between [0,1], i.e. stochastic generation G vectorial X _i=[x _i1, x _i2..., x _iN], 1≤i≤G as the position of particle, x _ibetween [0,1]; Stochastic generation G vectorial V again _i=[v _i1, v _i2..., v _iN], G+1≤i≤2 × G as the speed of particle, v _ibetween [0,1]; Setting maximum iteration time iter max, current iteration number of times k=1;

4-2. calculates the fitness value of each particle, stores the current position of each particle for self optimal value P _bi, storing the position that in all particles, fitness value is minimum is global optimum P _bg;

4-3. judges whether k reaches maximum iteration time iter max, if so, forwards step 4-7 to, otherwise forwards 4-4 to;

4-4. upgrades each particle rapidity and position, calculates the fitness value that each particle is new; If the current fitness value of i-th particle is better than P _bi, then P is upgraded _bi; If the optimal location of current particle group is better than P _bg, then P is upgraded _bg; K=k+1 is set;

4-5. upgrades inertia weight ω according to inertia weight computing formula;

If 4-6. global optimum position P _bgin 10 iteration, not change, then forward step 4-7 to, otherwise forward step 4-4 to;

4-7. exports P _bgas the weights of the PSO Neural Network trained;

Step (5) leases model prediction: according to the environment attribute vector S to be predicted of input _prewith the P calculated (H), the value of P (X|H), calculates the probability that this data tuple belongs to of all categories, exports the classification that posterior probability is maximum;

Step (6) public bicycles leases prediction

Lease according to step (5) classification exported in model prediction, select corresponding self-adaptation PSO Neural Network model, the residue public bicycles quantity of a front n time period is inputted this model, export the public bicycles quantity being prediction.

Described in step (2), computing formula is as follows:

N in 2-1 _iwith class center μ _kdistance computing formula as follows:

$D_{N_i,c_k}=\sqrt{\mathrm{\Σ}{\left|\right|N_i-{\mathrm{\μ}}_k\left|\right|}^2}$

Class center calculation formula new in 2-2 is as follows:

${{\mathrm{\μ}}_k}^{*}=\frac{\underset{i=1}{\overset{\left|n_k\right|}{\mathrm{\Σ}}}{N}_{\mathrm{ki}}}{\left|n_k\right|}$

μ _k ^*for class c _kxin Lei center;

| n _k| be class c _kthe number of middle data tuple;

N _kifor class c _kin i-th data tuple.

Computing formula and the variable implication of use in step (3) are as follows:

$P\left(H\right)=\frac{\left|c_{k,T}\right|}{L},$

| c _k,T| for belonging to class c in set of data samples T _ksample number;

L is the total sample number in set of data samples T;

P(X|H)＝P(d _i|c _k)P(w _i|c _k)P(t _i|ck)，

$P\left(d_i\right|c_k)=\frac{\left|d_{i,c_k}\right|}{\left|c_{k,T}\right|},$

$P\left(w_i\right|c_k)=\frac{\left|w_{i,c_k}\right|}{\left|c_{k,T}\right|}$

$P\left(t_i\right|c_k)=\frac{\left|t_{i,c_k}\right|}{\left|c_{k,T}\right|}$

for class c _kthe value of middle attribute d is d _inumber of samples;

for class c _kthe value of middle attribute w is w _inumber of samples;

for class c _kthe value of middle attribute t is t _inumber of samples.

Computing formula and the variable implication of use in step (4) are as follows:

Fitness value fitness (X _g) computing formula be:

$\mathrm{fitness}\left(X_g\right)=\frac{1}{2M}\underset{l=1}{\overset{M}{\mathrm{\Σ}}}\underset{k=1}{\overset{q}{\mathrm{\Σ}}}((\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}(1/(1+\exp (\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_{\mathrm{ij}}-{\mathrm{\θ}}_{\mathrm{hj}})))-{\mathrm{\θ}}_{\mathrm{ok}})-d_{\mathrm{lk}})$

M is that data sample concentrates data tuple number;

X _libe i-th input of l data sample;

θ _hjfor the threshold value of a hidden layer jth unit;

θ _okfor the threshold value of an output layer kth unit;

W _ijbe the connection weights of i-th input layer and a jth hidden layer;

W _jkfor the connection weights of a jth hidden layer and a kth output layer;

D _lkit is a kth output of l data sample;

Speed computing formula is:

${V_i}^{k+1}=w\×{V_i}^k+c_1\×r_1\×({P_i}^k-{X_i}^k)+c_2\×r_2\×({P_g}^k-{X_i}^k)$

V _i ^kfor particle X _iin the speed of kth time iteration;

R ₁, r ₂the random number of value between [0,1];

C ₁, c ₂be one group of random number, generally get c ₁=c ₂=2.5;

ω is inertia weight;

Inertia weight ω _kcomputing formula is:

${\mathrm{\ω}}_k=\mathrm{\ω}\max -\frac{\mathrm{\ω}\max -\mathrm{\ω}\min }{\mathrm{it}\max }\×k$

ωmax＝0.9，ωmin＝0.2；

The position calculation formula of particle is:

X _i ^k+1＝X _i ^k+V _i ^k+1；

X _i ^kthe position of i-th particle during iteration secondary to kth.

The computing formula used in step (6) is as follows:

It is specific as follows that neural network model exports O:

$O=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_{j1}(1/(1+\exp (\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_{\mathrm{li}}-{\mathrm{\θ}}_{\mathrm{hj}})))-{\mathrm{\θ}}_{\mathrm{ok}}.$

Beneficial effect of the present invention is as follows:

The present invention hires a car according to user the data such as record, weather, temperature, festivals or holidays, predicts each website public bicycles flow, and is not only use history loan data.Present invention generally provides a kind of Forecasting Methodology of multisource data fusion, each website public bicycles volume residual of certain day certain time period can be predicted accurately and efficiently.Specifically, following target is achieved:

Can hire a car to record from original user and extract each website residue bicycle quantity of each time period, be convenient to later stage cluster analysis, structure Bayes classifier and training PSO Neural Network.

In conjunction with multiple data sources, static prediction can be carried out to the public bicycles pattern of leasing, be conducive to public bicycles operator and carry out scheduling preliminary work in advance.

Performance prediction can be carried out according to the residue bicycle quantity of n time period before certain day, improve public bicycles and lease quantitative forecast degree of accuracy, for public bicycles operator provides scheduling time and the suggestion of concrete scheduling quantity.Scheduling promptness can be improved, reduce scheduling cost.

It is fast that concrete implementation result shows as (1) travelling speed, after structure completes Bayes classifier and training self-adaptation PSO Neural Network model, can carry out Static and dynamic prediction within the extremely short time; (2) precision of prediction is high, and use True Data to carry out simulation and forecast experiment, consensus forecast degree of accuracy is more than 96%.

Accompanying drawing explanation

Fig. 1 public bicycles leases Forecasting Methodology figure.

Fig. 2 self-adaptation PSO Neural Network algorithm flow chart.

Concrete embodiment

Below in conjunction with drawings and Examples, the invention will be further described.

As shown in Figure 1, a kind of public bicycles based on multisource data fusion leases Forecasting Methodology, specifically comprises following step:

Step (1) data acquisition and procession:

Read the user stored in public bicycles system (public bicycle system, PBS) to hire a car record, calculate each website and remain public bicycles quantity every time period t ime, set up vectorial N _date=[n ₁, n ₂..., n _j..., n _m], M is time period sum; Obtain from internet weather history situation temperature and festivals or holidays situation, set up environment attribute vector S={S ₁, S ₂..., S _i..., S _l, S _i=[d, w, t], L are the number that data sample concentrates sample; M, L, i and j are positive integer, wherein j≤M, i≤L;

Described user hires a car that record comprises lease duration, leases website, leases knee number, the time of returning the car, website of returning the car, knee number of returning the car;

N _jrepresent residue bicycle quantity in this website of a jth time period t ime;

S _irepresent the environment attribute of i-th day;

D represents attribute festivals or holidays on the same day, and the value of d is

W represents weather conditions, and the value of w is

T represents temperature, and the value of t is

Step (2) K-means cluster:

N _date=[n ₁, n ₂..., n _j..., n _m] as data sample to be sorted concentrate a data tuple; Concentrate random selecting k to put as initial classes center from data sample, use μ respectively ₁, μ ₂..., μ _krepresentation class c ₁, c ₂..., c _kclass center; Wherein, k is positive integer;

K-means clustering algorithm comprises the steps:

2-1. calculates remaining data tuple N _iwith class center μ _kdistance, by remaining data tuple N _iassign in the class the shortest with its distance;

2-2. recalculates all kinds of Xin Lei center μ _k ^*;

2-3. repeats step 2-1 and step 2-2, until each data tuple generic all no longer changes.

Step (3) builds Bayes classifier: after using K-means clustering algorithm to classify to set of data samples, obtain k class, each data tuple that data sample is concentrated will obtain a class label.Bayes Method is based on Bayes' theorem:

$P\left(H\right|X)=\frac{P\left(X\right|H)P\left(H\right)}{P\left(X\right)}$

Calculate P (X), P (H) and P (X|H).

Step (4) as shown in Figure 2, self-adaptation PSO Neural Network model training: use sorted data sample to train affiliated self-adaptation PSO Neural Network model of all categories, the input layer unit number of setting neural network is n, hidden layer unit number is p, and output layer unit number is q.Then need the weights number N=n × p+p × q adjusted in neural network, by the dimension needing the weights number N problem of representation space adjusted.Use vectorial X _i=[x _i1, x _i2..., x _iN] represent one of all weights of neural network possible value, also represent the position of particle in particle swarm optimization algorithm.PSO Neural Network model training step is as follows:

The position of each particle and speed in 4-1. initialization population population, position and the speed of particle are positioned between [0,1], i.e. stochastic generation G vectorial X _i=[x _i1, x _i2..., x _iN], 1≤i≤G as the position of particle, x _ibetween [0,1]; Stochastic generation G vectorial V again _i=[v _i1, v _i2..., v _iN], G+1≤i≤2 × G as the speed of particle, v _ibetween [0,1].Setting maximum iteration time itermax, current iteration number of times k=1;

4-2. calculates the fitness value of each particle, stores the current position of each particle for self optimal value P _bi, storing the position that in all particles, fitness value is minimum is global optimum P _bg;

4-3. judges whether k reaches maximum iteration time itermax, if so, forwards step 4-7 to, otherwise forwards 4-4 to.

4-4. upgrades each particle rapidity and position, calculates the fitness value that each particle is new.If the current fitness value of i-th particle is better than P _bi, then P is upgraded _bi.If the optimal location of current particle group is better than P _bg, then P is upgraded _bg; K=k+1 is set;

4-5. upgrades inertia weight ω according to inertia weight computing formula;

If 4-6. global optimum position P _bgin 10 iteration, not change, then forward step 4-7 to, otherwise forward step 4-4 to;

4-7. exports P _bgas the weights of the PSO Neural Network trained.

Step (5) leases model prediction: according to the environment attribute vector S to be predicted of input _prewith the P calculated (H), the value of P (X|H), calculates the probability that this data tuple belongs to of all categories, exports the classification that posterior probability is maximum.

Step (6) public bicycles leases prediction: lease according to step (5) classification exported in model prediction, select corresponding self-adaptation PSO Neural Network model, the residue public bicycles quantity of a front n time period is inputted this model, exports the public bicycles quantity being prediction:

Described in step (2), computing formula is as follows:

N in 2-1 _iwith class center μ _kdistance computing formula as follows:

$D_{N_i,c_k}=\sqrt{\mathrm{\Σ}{\left|\right|N_i-{\mathrm{\μ}}_k\left|\right|}^2}$

Class center calculation formula new in 2-2 is as follows:

${{\mathrm{\μ}}_k}^{*}=\frac{\underset{i=1}{\overset{\left|n_k\right|}{\mathrm{\Σ}}}{N}_{\mathrm{ki}}}{\left|n_k\right|}$

μ _k ^*for class c _kxin Lei center;

| n _k| be class c _kthe number of middle data tuple;

N _kifor class c _kin i-th data tuple.

Computing formula and the variable implication of use in step (3) are as follows:

$P\left(H\right)=\frac{\left|c_{k,T}\right|}{L},$

| c _k,T| for belonging to class c in set of data samples T _ksample number;

L is the total sample number in set of data samples T.

P(X|H)＝P(d _i|c _k)P(w _i|c _k)P(t _i|c _k)，

$P\left(d_i\right|c_k)=\frac{\left|d_{i,c_k}\right|}{\left|c_{k,T}\right|},$

$P\left(w_i\right|c_k)=\frac{\left|w_{i,c_k}\right|}{\left|c_{k,T}\right|}$

$P\left(t_i\right|c_k)=\frac{\left|t_{i,c_k}\right|}{\left|c_{k,T}\right|}$

for class c _kthe value of middle attribute d is d _inumber of samples;

for class c _kthe value of middle attribute w is w _inumber of samples;

for class c _kthe value of middle attribute t is t _inumber of samples;

Computing formula and the variable implication of use in step (4) are as follows:

Fitness value fitness (X _g) computing formula be:

$\mathrm{fitness}\left(X_g\right)=\frac{1}{2M}\underset{l=1}{\overset{M}{\mathrm{\Σ}}}\underset{k=1}{\overset{q}{\mathrm{\Σ}}}((\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}(1/(1+\exp (\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_{\mathrm{ij}}-{\mathrm{\θ}}_{\mathrm{hj}})))-{\mathrm{\θ}}_{\mathrm{ok}})-d_{\mathrm{lk}})$

M is that data sample concentrates data tuple number;

X _libe i-th input of l data sample;

θ _hjfor the threshold value of a hidden layer jth unit;

θ _okfor the threshold value of an output layer kth unit;

W _ijbe the connection weights of i-th input layer and a jth hidden layer;

W _jkfor the connection weights of a jth hidden layer and a kth output layer;

D _lkit is a kth output of l data sample.

Speed computing formula is:

${V_i}^{k+1}=w\×{V_i}^k+c_1\×r_1\×({P_i}^k-{X_i}^k)+c_2\×r_2\×({P_g}^k-{X_i}^k)$

V _i ^kfor particle X _iin the speed of kth time iteration;

R ₁, r ₂the random number of value between [0,1];

C ₁, c ₂be one group of random number, generally get c ₁=c ₂=2.5;

ω is inertia weight;

Inertia weight ω _kcomputing formula is:

${\mathrm{\ω}}_k=\mathrm{\ω}\max -\frac{\mathrm{\ω}\max -\mathrm{\ω}\min }{\mathrm{it}\max }\×k$

ωmax＝0.9，ωmin＝0.2；

The position calculation formula of particle is:

X _i ^k+1＝X _i ^k+V _i ^k+1；

X _i ^kthe position of i-th particle during iteration secondary to kth;

The computing formula used in step (6) is as follows:

It is specific as follows that neural network model exports O:

$O=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_{j1}(1/(1+\exp (\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_{\mathrm{li}}-{\mathrm{\θ}}_{\mathrm{hj}})))-{\mathrm{\θ}}_{\mathrm{ok}}.$

Claims (5)

1. the public bicycles based on multisource data fusion leases a Forecasting Methodology, it is characterized in that comprising the steps:

Step (1) data acquisition and procession:

Read the user stored in public bicycles system to hire a car record, calculate each website and remain public bicycles quantity every time period t ime, set up vectorial N _date=[n ₁, n ₂..., n _j..., n _m]; Obtain from internet weather history situation, temperature and festivals or holidays situation, set up environment attribute vector S={S ₁, S ₂..., S _i..., S _l, S _i=[d, w, t], L are the number that data sample concentrates sample; M, L, i and j are positive integer, wherein j≤M, i≤L;

Described user hires a car that record comprises lease duration, leases website, leases knee number, the time of returning the car, website of returning the car, knee number of returning the car;

N _jrepresent residue bicycle quantity in this website of a jth time period t ime;

S _irepresent the environment attribute of i-th day;

D represents attribute festivals or holidays on the same day, and the value of d is

W represents weather conditions, and the value of w is

T represents temperature, and the value of t is

Step (2) K-means cluster:

N _date=[n ₁, n ₂..., n _j..., n _m] as data sample to be sorted concentrate a data tuple; Concentrate random selecting k to put as initial classes center from data sample, use μ respectively ₁, μ ₂..., μ _krepresentation class c ₁, c ₂..., c _kclass center; Wherein, k is positive integer;

K-means clustering algorithm comprises the steps:

2-1. calculates remaining data tuple N _iwith class center μ _kdistance, by remaining data tuple N _iassign in the class the shortest with its distance;

2-2. recalculates all kinds of Xin Lei center μ _k ^*;

2-3. repeats step 2-1 and step 2-2, until each data tuple generic all no longer changes;

Step (3) builds Bayes classifier: after using k-means clustering algorithm to classify to set of data samples, obtain k class, each data tuple that data sample is concentrated will obtain a class label; Bayes Method is based on Bayes' theorem:

$P\left(H\right|X)=\frac{P\left(X\right|H)P\left(H\right)}{P\left(X\right)}$

Calculate P (X), P (H) and P (X|H);

Step (4) self-adaptation PSO Neural Network model training

Use sorted data sample to train affiliated self-adaptation PSO Neural Network model of all categories, the input layer unit number of setting neural network is n, and hidden layer unit number is p, and output layer unit number is q; Then need the weights number N=n × p+p × q adjusted in neural network, by the dimension needing the weights number N problem of representation space adjusted; Use vectorial X _i=[x _i1, x _i2..., x _iN] represent one of all weights of neural network possible value, also represent the position of particle in particle swarm optimization algorithm; PSO Neural Network model training step is as follows:

The position of each particle and speed in 4-1. initialization population population, position and the speed of particle are positioned between [0,1], i.e. stochastic generation G vectorial X _i=[x _i1, x _i2..., x _iN], 1≤i≤G as the position of particle, x _ibetween [0,1]; Stochastic generation G vectorial V again _i=[v _i1, v _i2..., v _iN], G+1≤i≤2 × G as the speed of particle, v _ibetween [0,1]; Setting maximum iteration time itermax, current iteration number of times k=1;

4-2. calculates the fitness value of each particle, stores the current position of each particle for self optimal value P _bi, storing the position that in all particles, fitness value is minimum is global optimum P _bg;

4-3. judges whether k reaches maximum iteration time itermax, if so, forwards step 4-7 to, otherwise forwards 4-4 to;

4-4. upgrades each particle rapidity and position, calculates the fitness value that each particle is new; If the current fitness value of i-th particle is better than P _bi, then P is upgraded _bi; If the optimal location of current particle group is better than P _bg, then P is upgraded _bg; K=k+1 is set;

4-5. upgrades inertia weight ω according to inertia weight computing formula;

If 4-6. global optimum position P _bgin 10 iteration, not change, then forward step 4-7 to, otherwise forward step 4-4 to;

4-7. exports P _bgas the weights of the PSO Neural Network trained;

Step (5) leases model prediction: according to the environment attribute vector S to be predicted of input _prewith the P calculated (H), the value of P (X|H), calculates the probability that this data tuple belongs to of all categories, exports the classification that posterior probability is maximum;

Step (6) public bicycles leases prediction

Lease according to step (5) classification exported in model prediction, select corresponding self-adaptation PSO Neural Network model, the residue public bicycles quantity of a front n time period is inputted this model, export the public bicycles quantity being prediction.

2. a kind of public bicycles based on multisource data fusion leases Forecasting Methodology as claimed in claim 1, it is characterized in that described in step (2), computing formula is as follows:

N in 2-1 _iwith class center μ _kdistance computing formula as follows:

$D_{N_i,c_k}=\sqrt{\mathrm{\Σ}{\left|\right|N_i-{\mathrm{\μ}}_k\left|\right|}^2}$

Class center calculation formula new in 2-2 is as follows:

${{\mathrm{\μ}}_k}^{*}=\frac{\underset{i=1}{\overset{\left|n_k\right|}{\mathrm{\Σ}}}{N}_{\mathrm{ki}}}{\left|n_k\right|}$

μ _k ^*for class c _kxin Lei center;

| n _k| be class c _kthe number of middle data tuple;

N _kifor class c _kin i-th data tuple.

3. a kind of public bicycles based on multisource data fusion leases Forecasting Methodology as claimed in claim 1, it is characterized in that the computing formula that uses in step (3) and variable implication as follows:

$P\left(H\right)=\frac{\left|c_{k,T}\right|}{L},$

| c _k,T| for belonging to class c in set of data samples T _ksample number;

L is the total sample number in set of data samples T;

P(X|H)＝P(d _i|c _k)P(w _i|c _k)P(t _i|c _k)，

$P\left(d_i\right|c_k)=\frac{\left|d_{i,c_k}\right|}{\left|c_{k,T}\right|},$

$P(w_i,c_k)=\frac{\left|w_{i,c_k}\right|}{\left|c_{k,T}\right|}$

$P\left(t_i\right|c_k)=\frac{\left|t_{i,c_k}\right|}{\left|c_{k,T}\right|}$

for class c _kthe value of middle attribute d is d _inumber of samples;

for class c _kthe value of middle attribute w is w _inumber of samples;

for class c _kthe value of middle attribute t is t _inumber of samples.

4. a kind of public bicycles based on multisource data fusion leases Forecasting Methodology as claimed in claim 1, it is characterized in that the computing formula that uses in step (4) and variable implication as follows:

Fitness value fitness (X _g) computing formula be:

$\mathrm{fitness}\left(X_g\right)=\frac{1}{2M}\underset{l=1}{\overset{M}{\mathrm{\Σ}}}\underset{k=1}{\overset{q}{\mathrm{\Σ}}}((\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}(1/(1+\exp (\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_{\mathrm{li}}-{\mathrm{\θ}}_{\mathrm{hj}})))-{\mathrm{\θ}}_{\mathrm{ok}})-d_{\mathrm{lk}})$

M is that data sample concentrates data tuple number;

X _libe i-th input of l data sample;

θ _hjfor the threshold value of a hidden layer jth unit;

θ _okfor the threshold value of an output layer kth unit;

W _ijbe the connection weights of i-th input layer and a jth hidden layer;

W _jkfor the connection weights of a jth hidden layer and a kth output layer;

D _lkit is a kth output of l data sample;

Speed computing formula is:

V _i ^k+1＝w×V _i ^k+c ₁×r ₁×(P _i ^k-X _i ^k)+c ₂×r ₂×(P _g ^k-X _i ^k)

V _i ^kfor particle X _iin the speed of kth time iteration;

R ₁, r ₂the random number of value between [0,1];

C ₁, c ₂be one group of random number, generally get c ₁=c ₂=2.5;

ω is inertia weight;

Inertia weight ω _kcomputing formula is:

${\mathrm{\ω}}_k=\mathrm{\ω}\max -\frac{\mathrm{\ω}\max -\mathrm{\ω}\min }{\mathrm{it}\max }\×k$

ωmax＝0.9，ωmin＝0.2；

The position calculation formula of particle is:

X _i ^k+1＝X _i ^k+V _i ^k+1；

X _i ^kthe position of i-th particle during iteration secondary to kth.

5. a kind of public bicycles based on multisource data fusion leases Forecasting Methodology as claimed in claim 1, it is characterized in that the computing formula used in step (6) is as follows:

It is specific as follows that neural network model exports O:

$O=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{w}_{j1}(1/(1+\exp (\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{x}_{\mathrm{li}}-{\mathrm{\θ}}_{\mathrm{hj}})))-{\mathrm{\θ}}_{\mathrm{ok}}.$