CN106650763A - Calculating method of index selection, weight optimization and channel planning of electric power payment channel analysis - Google Patents

Abstract

A calculating method of index selection, weight optimization and channel planning of electric power payment channel analysis is disclosed. The method comprises the following steps of step1, through a SG186 system or a questionnaire, acquiring data of basic attribute information and payment habit attribute information of a user who pays; step2, using a characteristic weight optimization method to optimize each weight in an individual user figure, acquiring an optimal individual user figure, and through a clustering algorithm, establishing a group user payment behavior figure; step3, using a K-nearest neighbor classification algorithm to establish an index evaluation system; step4, using a genetic annealing algorithm to calculate a weight value of each attribute index; and step5, determining whether the value is an optimal value and determining an optimal payment channel. In the invention, based on user payment large data, a lean management level of marketing electricity charge recovery work is increased and finally a payment service channel through which the user can pay conveniently and which is satisfied by the user is realized.

Description

A kind of index of electric power payment channel analysis is selected, right-value optimization and channel planning Computational methods

Technical field

The present invention relates to communication of power system field, the index selection of more particularly to a kind of electric power payment channel analysis, The computational methods that right-value optimization is planned with channel.

Background technology

With payment channel building variation, the development of way to pay dues diversification, the payment of business office of original electric power mechanism " single to pay dues " pattern is broken, but in real life, the payment of electric power mechanism remains head during users' electricity payment Mode is selected, it is overstaffed during indivedual business office's payment peaks.Between subscriber payment custom, payment demand and payment channel building Contradiction is highlighted, and tariff recovery hidden danger, electric service hidden danger, is gradually manifested the problems such as human resource distribution is unreasonable.

Chinese Government proposes in the U.S.《Big data research and development plan》Also reply " 12 states housekeeping in 2012 Business informatization project planning ", using big data as constructing emphases, gross investment is estimated hundreds of hundred million.Issue within 2013《In State's electric power big data development white paper》, it is proposed that electric power big data is the necessarily mistake of power industry technologies innovation in energy revolution Journey, rather than simple technology category.Electric power big data is not only technological progress, is even more related to whole power system and is counting greatly It is that intelligent electric Force system of future generation exists according to the major transformation of the aspect such as idea of development, management system and technology path under the epoch The form of value rises under the big data epoch.

Under big data environment, magnanimity isomeric data batch is integrated, flow cytometer showed and low time delay ETL integration technologies are systems Key technology, how the refining initial data of precise and high efficiency is one of core competence of KPI indexs and big data system.Big number Analysis payment channel appraisement system, first the producers and consumers' angle from data and service is needed to provide ginseng according to analytical model The various roles analyzed with big data, and information and data are classified from whole analysis and data life-cycle processes, With reference to practical business scene, data, services are formed.

Therefore, excavated by subscriber payment data, it is found that those seem the unrelated transaction data transaction back of the body in database Afterwards certain hiding contact, according to analysis result in-depth study analysis can be carried out to the payment of Electricity customers, found to electricity The valuable channel of power company, the payment channel being most welcomed by customers is found out with this, is had far-reaching significance.

The content of the invention

To solve above-mentioned technical problem, the present invention is in the survey data and electric company end for including userspersonal information On the basis of data, individual consumer's portrait is set up, and set up by the analysis of user's typical behaviour, clustering algorithm and model, can be with Client is understood to payment and the requirement of related service, so as to while cost or reduces cost is kept, improve customer satisfaction.

It is an object of the invention to provide one kind is in electric power payment channel analysis, can formulate different for different value customers Power supply mode provides the computational methods that marketing decision-making supports and make index selection, right-value optimization and the channel of anticipation to plan.

For achieving the above object, the technical scheme of present invention offer is：

A kind of index of electric power payment channel analysis is selected, the computational methods that right-value optimization and channel are planned, including following Step：

Step one obtains payment user's base attribute information and payment custom attribute letter by SG186 systems or questionnaire The data of breath；

Step 2 is optimized using feature weight optimization method to each weight in individual consumer's portrait, obtains optimum individual User draws a portrait, and by clustering algorithm group of subscribers paying behaviors portrait is set up；

Step 3 sets up indicator evaluation system using K arest neighbors sorting algorithms；

Step 4 adopts genetic annealing algorithms, calculates the weighted value of each ATTRIBUTE INDEX；

Step 5 determines whether optimal value, it is determined that optimum payment channel.

Further, in step one, the base attribute information of user includes name, age, sex, home address；Pay Taking custom attribute information includes every time the information of average payment amount of money and pay charge way.

Further, in step 2, the property and payment custom of individual consumer is represented in the form of a label, as individuality User draws a portrait, and feature weight optimization method includes K nearest neighbor algorithms and simulated annealing, is randomly provided the initial of each attribute Weight；Set test set is divided based on sample data set, k nearest neighbor classified calculating is carried out, per data in traversal test set, from instruction Practice to concentrate and extract closest front K datas, compare with the actual value of test data, and statistical error；Carry out application condition, The weight of each attribute is adjusted, if error is less than threshold value, the weight of each attribute is obtained, otherwise, the weight of attribute is adjusted, is entered Row iteration is tested, and obtains the weight of each attribute.

Further, in step 2, the principle of the clustering algorithm is that attribute data to be clustered is randomly placed into one In the environment of individual two-dimensional grid, each attribute data objects has a random initial position, and each ant can be in grid Upper movement, and swarm similarity of the existing object in local environment is measured, swarm similarity is changed by probability transfer function Into the probability of mobile object, object is picked up or put down with this probability；Ant colony joint action causes to belong to same category of attribute Data object can be built up together in same area of space；

So that similar factor of evaluation is gathered for a class, polymerization result is paid the fees channel evaluation index as power system, electricity Force system payment channel factor of evaluation includes coverage rate, the utilization rate of channel, cost, operating efficiency, the subscriber payment row of channel For portrait, convenience, CSAT, channel development trend.

Further, in step 3, the K arest neighbors sorting algorithm is comprised the following steps：

For the test sample in a test set, according to Feature Words test sample vector is formed；

The Sample Similarity of the test sample and each sample in training set is calculated, computing formula is：

Wherein, d_iFor the characteristic vector of test sample, d_jFor the center vector of jth class；M is characterized the dimension of vector；W_kFor The kth dimension of vector；The determination of k values first adopts an initial value, then adjusts K values according to the result of experiment test K；

According to Sample Similarity, concentrate in training sample and select the k sample most like with test sample；

In the individual k neighbours of test sample, the weight per class is calculated successively, computing formula is as follows：

Wherein, x is the characteristic vector of test sample；Sim(x,d_i) it is calculating formula of similarity；B is threshold value, is awaited excellent Change and select；y(d_i,C_j) value be 1 or 0, if d_iBelong to C_j, then functional value is 1, is otherwise 0；

Relatively the weight of class, sample is assigned in that maximum classification of weight.

Further, in step 4, genetic Annealing computational methods are：

Step 4 a setting models each parameter variation ranges, randomly chooses an initial model in the range of this, and Calculate corresponding target function value；

Step 4 b carries out disturbance to "current" model and produces a new model, calculates corresponding target function value, obtains

Δ E=E (m)-E (m₀)；

If step 4 c Δ E ＜ 0, new model is received by m；If Δ E ＞ 0, new model m press probability P=exp (- Δ E/ T) received, T is external influence factor, when model is received, put m₀=m；

Step 4 d under external influence factor T, the disturbance of the certain number of times of repetition and reception process, the i.e. b of repeat step four and Step 4 c；

Step 4 e slowly reduces external influence factor T；

The b of step 4 f repeat step four and four e, till the condition of convergence meets.

Further, in step 5, the judgment formula of optimal value be | Δ E |=| E (m)-E (m₀)|≤0

Wherein, Δ E represents channel optimal solution；E (m) represents the channel value for calculating, E (m₀) represent initial channel value.

Using above-mentioned technical proposal, the present invention has the advantages that：

First, the weight that the present invention is drawn a portrait using the method optimizing user of feature weight optimization, feature weight optimization was both made For the pretreatment stage of data mining, and it combines with specific data mining algorithm by this, so as to construct succinct, essence Really, stable data mining computational methods.

Second, in the present invention, the client of like attribute is gathered for a class by clustering algorithm, and the client in inhomogeneity Attribute it is then different, and set up the model of each class client respectively.The model can operate with subsequent software to the pre- of Future Data Survey and the resolution of the analysis to user preference and payment channel building.

3rd, the present invention cracks that outlet's layout is unreasonable, business window sets on the basis of subscriber payment big data Put that dumb, human resources configuration is uneven, tariff recovery has risk, expense control agreement signs a slow difficult problem, improve marketing The lean managerial skills of tariff recovery work, it is final to realize allowing user conveniently to pay the fees, allow customer satisfaction system paying service channel.

Description of the drawings

Fig. 1 is the computational methods that the index of electric power payment channel analysis of the present invention is selected, right-value optimization and channel are planned Flow chart；

Fig. 2 is the flow chart of KNN algorithms of the present invention；

Fig. 3 is the flow chart of the weight optimization based on KNN and simulated annealing；

Fig. 4 is the flow chart of clustering algorithm；

Fig. 5 is genetic annealing algorithms flow chart.

Specific embodiment

It is below in conjunction with the accompanying drawings and embodiment, right in order that the objects, technical solutions and advantages of the present invention become more apparent The present invention is further elaborated.It should be appreciated that structure chart described herein and specific embodiment are only to explain this Invention, is not intended to limit the present invention.

Embodiment 1

Fig. 1 is the computational methods that the index of electric power payment channel analysis of the present invention is selected, right-value optimization and channel are planned Flow chart, as shown in figure 1, the computational methods that a kind of index of electric power payment channel analysis is selected, right-value optimization and channel are planned, Comprise the following steps：

Step one obtains payment user's base attribute information and payment custom attribute letter by SG186 systems or questionnaire The data of breath；

Step 2 is optimized using feature weight optimization method to each weight in individual consumer's portrait, obtains optimum individual User draws a portrait, and by clustering algorithm group of subscribers paying behaviors portrait is set up；

Step 3 sets up indicator evaluation system using K arest neighbors sorting algorithms,

Step 4 adopts genetic annealing algorithms, calculates the weighted value of each ATTRIBUTE INDEX；

Step 5 determines whether optimal value, it is determined that optimum payment channel.

Embodiment 2

Obtain the data of payment user's base attribute information and payment custom attribute information

Payment customer group's investigation purpose is the data for objectively collecting payment client, is that follow-up work is prepared. Target of investication and study is mainly the electricity charge in units of family and pays client, and each family is represented with power grid user numbering.Investigation mode is Based on survey and electric company's offer data research combine.

Survey mainly acquires name, age, sex, home address, the payment habits information of user, and combines confession The subscriber payment information that electric company provides, sets up individual consumer's portrait.It is specific as follows：

Name：Replaced by Customs Assigned Number

Age：According to family's average age and everyone weight analysis of paying the fees of family, family is paid in the electricity charge and is divided into three kinds, point It is not

Sex：Replace (man with numbering：0, female：1)

Home address：On the basis of survey and data being provided, the address of object is divided into

Payment custom：Payment custom include user arrearage frequency, whether can pay the fees in time, every time averagely payment amount of money with And the information such as pay charge way, specially：

Remarks：Name, sex, home address are individual privacy, need encryption.

On the basis of the data that the survey comprising information above and electric company provide, individual consumer's picture is set up Picture, and set up by the analysis of user's typical behaviour, clustering algorithm and model, it will be appreciated that client wants to payment and related service Ask, so as to while cost or reduces cost is kept, improve customer satisfaction.And can be according to investigation content analysis and mistake The payment information data analysis of 3 years is gone, following 1 year payment information data are predicted.

Obtain optimum individual user portrait

Payment customer group typical behaviour analysis is on the basis of investigation and electric company provide data, to finding Synthesis is analyzed with data, and the data model foundation for customer group is prepared.To paying the fees, customer group's typical behaviour enters Row analysis, it is necessary first to optimization is adjusted to each weight in individual consumer's portrait with feature weight optimization method, is adjusted Optimum individual user portrait afterwards, then optimum individual user portrait is clustered and modeled, obtain group of subscribers portrait and number According to model.

Payment customer group typical behaviour analysis Main Basiss are the payment data that survey and electric company provide, point Analysis content includes：

User's portrait delineates targeted customer, contact user's demand and design also known as user role (Persona) as one kind The effective tool in direction, user's portrait is widely used in each field.Often with most during practical operation The attribute of user, behavior and expectation are tied for plain and closeness to life language.As the virtual representations of actual user, User draw a portrait formed user role be not an off outside product and market it is constructed out, the user role of formation needs Want the main audient and target group of representative energy representative products.User's portrait will be set up on real data, when having When multiple users draw a portrait, the priority for considering user's portrait, and user's portrait is needed to be in constantly amendment.

The core work of user's portrait " is labelled " for user.User tag is had been defined in questionnaire, this The characteristics of a little labels have succinct, simple, facilitate label is extracted and cluster analysis.User tag is illustrated with example, it is as follows：

This illustrates in the form of a label the property and payment custom of payment client, and each little lattice content is in form For a user tag.

Set up the electricity charge pay client user portrait can be divided into three levels：First level is the investigation point of group of subscribers Analysis；Second level is that data analysis is embodied individual description；Third level is the development and application after abstract data modeling.

Determining the way of label weight has multiple：Such as expert's setting method, by artificial setting, have the advantages that easy to adjust； Algorithm optimization method, based on investigation sample and the data sample for providing, must have enough sample training collections, according to simulated target not Together, the weight for obtaining is different.The electricity charge are paid the weight of customer users portrait and are obtained by KNN algorithms and simulated annealing.

K arest neighbors (k-Nearest Neighbor, KNN) sorting algorithm, according to some sample instances and other examples it Between similitude classified.KNN algorithms can be not only used for classification, can be also used for returning, and is one and compares into theory Ripe method, is also one of simplest machine learning algorithm.The thinking of the method is：If a sample is in feature space K most like (i.e. closest in feature space) sample in great majority belong to some classification, then the sample falls within This classification.In KNN algorithms, selected neighbours are the objects correctly classified.The method on class decision-making is determined only according to The classification belonging to sample to be divided is determined according to the classification of one or several closest samples.Although KNN methods are from principle Limit theorem is also relied on, but in classification decision-making, it is only relevant with minimal amount of adjacent sample.Because KNN methods are mainly by week It is with the neighbouring sample of limit, rather than by differentiating the method for class field determining generic, therefore for the intersection of class field Or overlap for more sample set to be divided, KNN methods are more suitable for compared with additive method.Specifically by finding out sample K nearest-neighbors, by the mean value of the attribute of these neighbours the sample is assigned to, so as to obtain the attribute of the sample.KNN algorithm streams Journey figure is as shown in Figure 2.

According to traditional vector space model, sample is characterized the weighted feature vector in space, i.e. D=D by formalization (T1, W1；T2,W2；…；Tn,Wn).For a test sample, the similarity that it concentrates each sample with training sample is calculated, The most like samples of K are found out, the classification according to Weighted distance and belonging to judging test sample.The test sample is calculated with training The similarity of each sample, computing formula is concentrated to be:

In formula:d_iFor the characteristic vector of test sample, d_jFor the center vector of jth class；M is characterized the dimension of vector；W_kFor The kth dimension of vector.The determination of k values typically first adopts an initial value, then adjusts K values according to the result of experiment test K, typically Initial value is set to tens and arrives hundreds of.

Simulated annealing is based on a kind of random optimizing algorithm of Monte-Carlo iterative strategies, its starting point It is the similitude between the annealing process and general combinatorial optimization problem based on solid matter in physics.Simulated annealing is from certain One higher initial temperature is set out, and with the continuous decline of external influence factor parameter, join probability kick characteristic is random in solution space The globally optimal solution of object function is found, i.e., probability can jump out and finally tend to global optimum in locally optimal solution.Simulation Annealing algorithm first makees starting with an arbitrfary point in search space, and each step first selects one " neighbours ", then calculates again from now There is position to reach the probability of " neighbours ".

The algorithm model of simulated annealing is as follows：

That annealing algorithm (SA) access is modeled in the iteration is solution j, and is modeled in (k+1) secondary iteration and moves back What fiery algorithm (SA) accessed is the probability for solving j.It is made up of two independent probability distribution, is produced from solution i in kth time iteration The probability g of solution_ij(T), wherein g_ij(T) it is required to meet normalizing condition:

Solve received probability λ_ij(T) external influence factor when, here T is kth time iteration, for i The situation of ≠ j, the expression formula of transition probability is as follows

Because λ_ij(T) it is not always equivalent to, therefore new explanation has not received possibility, algorithm rests on the probability of solution i and is

Because Ω is a countable set, therefore the random process representated by the stochastic variable of simulated annealing generation is one Markov chain, one walks transition probability two formulas by more than and defines, and one step transition probability of note is：

Then k steps transition probability is

Wherein I be unit matrix, T_tRepresent external influence factor value during the t time iteration.The implication of its matrix element is

P_ij(m, m+k)=P_r{X_m+k=j | X_m=i }

State i, probability of the m+k time iteration in state j are in by m iteration.

Iteration tuning index weights parameter, so as to provide index weights reference for channel planning.As shown in figure 3, setting at random Put the initial weight of each attribute；Set test set is divided based on sample data set, k nearest neighbor classified calculating, traversal test is carried out Concentrate per data, closest front K datas are extracted from training set, compare with the actual value of test data, and count Error；Application condition is carried out, the weight of each attribute is adjusted, if error is less than threshold value, the weight of each attribute is obtained, it is no Then, the weight of attribute is adjusted, test is iterated, the weight of each attribute is obtained.

Cluster analysis

The target of cluster is that important distinguishing cluster is distinguished and extracted in potential data set, up to the present studies people Member develops five kinds of basic clustering methods, partition clustering, hierarchical clustering, density clustering, based on Grid Clustering and Cluster based on model.Wherein, implement and compare because the thought of algorithm itself is simple based on the k-means algorithms for dividing Easily, it is widely used.But, k-means algorithms are sensitive to exceptional value, and need to determine k values in advance.Therefore adopt The improved customer behavior analysis algorithm based on swarm intelligence, the algorithm is with the naive model of ant colony cooperation ant nest classification as base Plinth, analyzes a kind of Self-organization clustering algorithm of customer action, and the method can be such that data easily visualize, it highlight induce one it is emerging The feature of interest.The number of cluster centre is automatically generated from data.

Assume there was only a kind of object, all of object is all randomly dispersed in above two-dimentional lattice, each lattice point only includes one Object, ant is placed at random on two-dimentional lattice, and every time along random one lattice of direction movement, every time after movement, if phase The lattice point answered has if object, and the ant without burden determines that giving following probability picks up an object：

Wherein, λ ants feel to obtain object number, and γ around it₁＞ 0.When only a small amount of object is in ant week When enclosing, i.e. λ ＜ ＜ γ₁, then P_pClose to 1；Therefore, object has larger probability to be lifted, on the other hand, if ant is perceived To many object λ ＞ ＞ γ₁, then P_pClose to 0, the probability that object is lifted is just smaller.

The ant often only born puts down the probability of born object and is given by equation below：

Wherein, should ensure that given corresponding lattice point is empty, γ₂＞ 0, if ant finds around substantial amounts of object, i.e. λ ＞ ＞ γ₂, then P_dClose to 1, the probability for putting down object is very big.If λ is ＜ ＜ γ₂, then P_dClose to 0, the probability for putting down almost does not have Have.

Basic model (BM) based on ant colony clustering algorithm, Lumer and Faieta is generalized to BM with actual element to gather Class data vector, proposes famous LF algorithms.Introduce a similar density function in LF algorithms to weigh two data objects Between similarity degree.

In LF algorithms, data vector is laid randomly on two-dimentional lattice, and in observation n is referred to as_NPathIndividually During point peripheral region, ant randomly moves about in lattice, and moving area is exactly a square fieldI.e. ant is worked as N around the i of front position_N×n_NIndividual place, it is assumed that ant is located at position i in time t, finds data vector O_i, in ant field Interior data vector O_i" local " density f (O_i) computing formula it is as follows：

In formula, α ＞ 0 define data vector O_iAnd O_jDistinctiveness ratio scope.Constant α determines that two objects when should Or should not put together, it is a coefficient for adjusting averag density between data vector, if α is too little, the little of many can be formed Group, the object for belonging to same group different groups are gathered, if α is too big, are likely to result in obscuring between each group, not belonging to Gather together in same group of object.So the number of clusters of the α on being formed has directly impact, as shown in Figure 4.

Using measuring similarity f (O_i), pick up and abandon probability and be defined as follows：

By clustering algorithm, the group of subscribers portrait of payment client can be set up.The group of subscribers portrait is in individual consumer Portrait is set up by cluster.Group of subscribers portrait can describe the age distribution of whole payment client, payment preference and pay Take the label informations such as mode.Conversely, if it is known that the label information of certain individual specimen, should have a class sample cluster therewith in colony There is close property.Therefore, individual consumer's portrait can be described with group of subscribers portrait, it is also possible to which individual consumer draws a portrait to infer Analysis group of subscribers portrait.

By user's portrait, the analysis of payment customer group typical behaviour and subsequent software analysis, the electricity charge are obtained and pay group The information such as the payment preference of body, are that channel planning and deployment of human resources provide support.

Embodiment 3

Indicator evaluation system is set up using K arest neighbors sorting algorithms

The main thought of KNN sorting algorithms is：First calculate between sample to be sorted and the training sample of known class away from From or similarity, find distance or similarity K neighbours nearest with sample data to be sorted；Further according to belonging to these neighbours Classification is judging the classification of sample data to be sorted.If K neighbours of sample data to be sorted belong to a classification, then Sample to be sorted falls within this classification.Otherwise, each candidate categories is scored, determines according to certain rule and treat point The classification of class sample data.

For a test sample, the similarity that it concentrates each sample with training sample is calculated, find out K individual most like Sample, the classification according to Weighted distance and belonging to judging test sample.Specific algorithm step is as follows:

(1) for a test sample, according to Feature Words test sample vector is formed.

(2) Sample Similarity of the test sample and each sample in training set is calculated, computing formula is:

In formula:d_iFor the characteristic vector of test sample, d_jFor the center vector of jth class；M is characterized the dimension of vector；W_kFor The determination of the kth dimension .k values of vector typically first adopts an initial value, then adjusts K values according to the result of experiment test K, typically Initial value is set to tens and arrives hundreds of.

(3) according to Sample Similarity, concentrate in training sample and select the k sample most like with test sample.

(4) in k neighbour of test sample, the weight per class is calculated successively, computing formula is as follows:

In formula：X is the characteristic vector of test sample；

Sim(x,d_i) it is calculating formula of similarity；

B is threshold value, awaits optimum choice；

y(d_i,C_j) value be 1 or 0, if d_iBelong to C_j, then functional value is 1, is otherwise 0.

(5) compare the weight of class, sample is assigned in that maximum classification of weight.

KNN methods are based on analogical learning, are a kind of non-parametric sorting techniques, in the pattern-recognition based on statistics very Effectively, for unknown and Non-Gaussian Distribution can obtain higher classification accuracy, there is robustness, clear concept. But in sample classification, KNN methods there is also deficiency, and such as KNN algorithms are slack sorting algorithms, and at that time air switch pin is big；Calculate During similarity, feature vector dimension is high, does not account for the incidence relation between Feature Words；When sample distance is calculated, each right-safeguarding value phase Together so that the distance between characteristic vector calculates not accurate enough, affects nicety of grading.

Using genetic annealing algorithms, the weighted value of each ATTRIBUTE INDEX is calculated

Simulated annealing (Simulated Annealing) comes from statistical physics, according to statistics thermodynamics, in object The state of each molecule obeys Gibbs distributions, i.e.,：

In formula：E(r_i) be i-th molecule energy function；

Ri is i-th molecule state in which；

K is Boltzmann constant；

T represents external influence factor；

ρ(r_i) be i-th molecule probability density, k=1 is made for convenience's sake.

Simulated annealing is comprised the following steps that：

1) each parameter variation range of setting models, randomly chooses an initial model m in the range of this₀, and count Calculate corresponding target function value E (m₀)；

2) disturbance is carried out to "current" model and produces a new model m, calculate corresponding target function value E (m), obtain Δ E =E (m)-E (m₀)

If 3) Δ E ＜ 0, new model is received by m；If Δ E ＞ 0, new model m are carried out by probability P=exp (- Δ E/T) Receive, T is external influence factor.When model is received, m is put₀=m；

4) under external influence factor T, the disturbance of the certain number of times of repetition and reception process, i.e. repeat step 2), 3)；

5) it is slow to reduce external influence factor T；

6) repeat step 2), 5), until the condition of convergence meet till.

Determine whether optimal value, the judgment formula of optimal value is | Δ E |=| E (m)-E (m0) |≤0, and Δ E represents channel Optimal solution；E (m) represents the channel value for calculating, E (m₀) represent initial channel value.

If it is judged that being no, then returning from this carries out cluster analysis；If it is judged that being yes, obtain the attribute and refer to Target weight optimal value, so that it is determined that optimum payment channel, as shown in Figure 5.

Embodiment described above only expresses embodiments of the present invention, and its description is more concrete and detailed, but can not Therefore it is interpreted as the restriction to the scope of the claims of the present invention.It should be pointed out that for the person of ordinary skill of the art, Without departing from the inventive concept of the premise, some deformations and improvement can also be made, these belong to the protection model of the present invention Enclose.Therefore, the protection domain of patent of the present invention should be defined by claims.

Claims (7)

1. the computational methods that a kind of index of electric power payment channel analysis is selected, right-value optimization and channel are planned, it is characterised in that Comprise the following steps：

Step one obtains payment user's base attribute information and custom attribute information of paying the fees by SG186 systems or questionnaire Data；

Step 2 is optimized using feature weight optimization method to each weight in individual consumer's portrait, obtains optimum individual user Portrait, by clustering algorithm group of subscribers paying behaviors portrait is set up；

Step 3 sets up indicator evaluation system using K arest neighbors sorting algorithms；

Step 4 adopts genetic annealing algorithms, calculates the weighted value of each ATTRIBUTE INDEX；

Step 5 determines whether optimal value, it is determined that optimum payment channel.

2. computational methods according to claim 1, it is characterised in that in step one, the base attribute packet of user Include name, age, sex, home address；Payment custom attribute information includes every time average payment amount of money and pay charge way Information.

3. computational methods according to claim 1, it is characterised in that in step 2, represent in the form of a label individual The property of user and payment custom, used as individual consumer's portrait, feature weight optimization method includes that K nearest neighbor algorithms and simulation are moved back Fiery algorithm, is randomly provided the initial weight of each attribute；Set test set is divided based on sample data set, k nearest neighbor classification is carried out Calculate, per data in traversal test set, closest front K datas, the reality with test data are extracted from training set Value compares, and statistical error；Application condition is carried out, the weight of each attribute is adjusted, if error is less than threshold value, each category is obtained Property weight, otherwise, adjust attribute weight, be iterated test, obtain the weight of each attribute.

4. computational methods according to claim 1, it is characterised in that in step 2, the principle of the clustering algorithm is Attribute data to be clustered is randomly placed in the environment of a two-dimensional grid, each attribute data objects has one at random just Beginning position, each ant can move on grid, and measure swarm similarity of the existing object in local environment, by general Swarm similarity is converted into rate transfer function the probability of mobile object, and object is picked up or put down with this probability；Ant colony is combined Action causes to belong to same category of attribute data objects can be built up together in same area of space；

So that similar factor of evaluation is gathered for a class, polymerization result is paid the fees channel evaluation index as power system, power train System payment channel factor of evaluation includes coverage rate, the utilization rate of channel, cost, operating efficiency, the subscriber payment behavior picture of channel Picture, convenience, CSAT, channel development trend.

5. computational methods according to claim 3, it is characterised in that in step 3, the K arest neighbors sorting algorithm bag Include following steps：

For the test sample in a test set, according to Feature Words test sample vector is formed；

The Sample Similarity of the test sample and each sample in training set is calculated, computing formula is：

$Sim(d_i,d_j)=\frac{\underset{k=1}{\overset{M}{\Σ}}{W}_{ik}\×W_{jk}}{\sqrt{\underset{k=1}{\overset{M}{\Σ}}{W}_{ik}^2}\sqrt{\underset{k=1}{\overset{M}{\Σ}}{W}_{jk}^2}}$

Wherein, d_iFor the characteristic vector of test sample, d_jFor the center vector of jth class；M is characterized the dimension of vector；W_kFor vector Kth dimension；The determination of k values first adopts an initial value, then adjusts K values according to the result of experiment test K；

According to Sample Similarity, concentrate in training sample and select the k sample most like with test sample；

In the individual k neighbours of test sample, the weight per class is calculated successively, computing formula is as follows：

Wherein, x is the characteristic vector of test sample；Sim(x,d_i) it is calculating formula of similarity；B is threshold value, awaits optimization choosing Select；y(d_i,C_j) value be 1 or 0, if d_iBelong to C_j, then functional value is 1, is otherwise 0；

Relatively the weight of class, sample is assigned in that maximum classification of weight.

6. computational methods according to claim 1, it is characterised in that in step 4, genetic Annealing computational methods are：

Step 4 a setting models each parameter variation ranges, randomly chooses an initial model in the range of this, and calculates Corresponding target function value；

Step 4 b carries out disturbance to "current" model and produces a new model, calculates corresponding target function value, obtains

Δ E=E (m)-E (m₀)；

If step 4 c Δ E ＜ 0, new model is received by m；If Δ E ＞ 0, new model m are entered by probability P=exp (- Δ E/T) Row receives, and T is external influence factor, when model is received, puts m₀=m；

Step 4 d repeats disturbance and reception process, the i.e. b of repeat step four and the step of certain number of times under external influence factor T Four c；

Step 4 e slowly reduces external influence factor T；

The b of step 4 f repeat step four and four e, till the condition of convergence meets.

7. computational methods according to claim 1, it is characterised in that in step 5, the judgment formula of optimal value is | Δ E |=| E (m)-E (m₀)|≤0

Wherein, Δ E represents channel optimal solution；E (m) represents this channel value for calculating, E (m₀) represent initial channel value.