CN103617459A - Commodity demand information prediction method under multiple influence factors - Google Patents

Abstract

The invention discloses a commodity demand information prediction method under multiple influence factors. The method comprises the steps that historical data are collected at first; the data are processed to form a training sample set, wherein the processing process comprises processing the data in a smooth mode, removing singular values, processing various influence factors in a fuzzy mode, processing the data in a normalization mode, and preventing data calculation overflow; then the formed training sample set is placed into a support vector machine to conduct study, and parameters in the prediction algorithm are adjusted to be optimal values; external information of nodes with prediction demands is input to a fuzzy processing system, and information to be predicted is obtained. According to the commodity demand information prediction method under the multiple influence factors, various factors influencing commodity information are considered comprehensively, and the influence of each factor on the commodity information is processed in a quantification mode, the problem that external information is ignored in a traditional prediction algorithm is solved better, overall optimization of the prediction algorithm can be achieved, and therefore people can grasp and know the commodity information more accurately and conveniently.

Description

Demand for commodity information forecasting method under a kind of many influence factors

Technical field

The present invention relates to demand for commodity information forecasting method under a kind of many influence factors, take historical merchandise news and extraneous data as basis, and data are processed, thereby realize, the information of commodity is predicted.Belong to information prediction technical field.

Background technology

At present, along with the trend of economic globalization, information becomes the key factor that enterprise obtains competitive power in society gradually, that is to say, who is the acquired information of morning more, and who can obtain information more accurately, and who has just occupied absolute advantage in competition of the same trade.And the Forecasting Methodology of applied merchandise news on present market but seldom considers that extraneous factor is for the impact of merchandise news, even if some method has been considered this impact, be also difficult to hold accurately influence factor, realize long-term prediction comparatively accurately.In this case, designing and a kind ofly can realize total optimization, consider the Forecasting Methodology of external influence factor, predict merchandise news, is vital.

Summary of the invention

Goal of the invention: the problem existing in the method for existing merchandise news prediction, the invention provides demand for commodity information forecasting method under a kind of many influence factors.First, historical data need to be collected, guarantee reliability and the adequacy of data.Secondly, data are processed, formed the sample set of training, processing procedure comprises carries out smoothing processing to data, rejects singular value, and various influence factors are carried out Fuzzy processing and data are normalized, and prevents data calculation overflow.Then, the training sample set of formation is put into support vector machine and learn, the parameter adjustment in prediction algorithm is arrived to optimal value.And then, the external information that has the node of forecast demand is input in Fuzzy Processing system and is gone.Finally, obtain the information that will predict.By this process, carry out merchandise news prediction, can hold comparatively accurately the dynamic of merchandise news.

Technical scheme: demand for commodity information forecasting method under a kind of many influence factors, comprising:

1. influence factor is judged

The influence factor of commodity mainly comprises: weather, temperature, season, festivals or holidays, individual preference, special event etc.These influence factors can be divided into two classes: objective history data and demand environment data.

(1) objective history data

Objective history data mainly refer to the sales data in like product past.According to correlation analysis, the previous day of prediction day and the correlativity on the same day of last week in product demand prediction are compared.

(2) demand environment data

With consumption demand environmental correclation because have weather, temperature, season, date type (working day, festivals or holidays) and special event etc.

2. the Fuzzy processing of influence factor

Environmental factor is converted into fuzzy quantity through subordinate function.For linear input, the number of subordinate function can be got less, and for nonlinear output relation, needs to establish some subordinate functions more.

3. the structure of support vector machine

By a kind of like this algorithm that minimizes structure risk of SVM, predict.The basic thought of this Forecasting Methodology may be summarized to be: first by nonlinear transformation, input vector x is mapped to a more space Z for higher-dimension, then in this new space, ask for optimum linearity classifying face, this nonlinear transformation is to realize by defining suitable inner product function, and this inner product function is constructed by kernel function.

Prediction steps:

(1) objective history data are carried out to smoothing processing and normalized, demand environment data are carried out to fuzzy quantization processing, then form sample set;

(2) with training sample, set up the objective function suc as formula (3-11);

(3) utilize SVM training algorithm to solve (3-11), obtain separating α _iwith

, i=1,2...n;

(4), by the Lagrange multiplier substitution formula (3-12) obtaining, recycling sample is predicted tomorrow requirement amount.

Beneficial effect: compared with prior art, demand for commodity information forecasting method under many influence factors provided by the invention, on the basis of the historical information of comprehensive commodity and the external information of history, processes historical information, form training sample set, in support vector machine, learn.The external information of the time period of simultaneously needs being predicted is processed by Fuzzy Processing system, the sample set of formation is input in support vector machine, and then prediction of output information.This Forecasting Methodology can be better in conjunction with affecting the various factors of merchandise news in actual life, can meet to greatest extent the accuracy predicting the outcome simultaneously, thereby for related personnel provides more accurate and merchandise news timely, for the many industries that comprise logistic industry, there is great meaning.

Accompanying drawing explanation

Fig. 1 is the method flow diagram of the embodiment of the present invention;

Fig. 2 is the schematic diagram of the embodiment of the present invention;

Fig. 3 is the comparative graph of SVM prediction sales volume and true sales volume.

Embodiment

Below in conjunction with specific embodiment, further illustrate the present invention, should understand these embodiment is only not used in and limits the scope of the invention for the present invention is described, after having read the present invention, those skilled in the art all fall within the application's claims limited range to the modification of the various equivalent form of values of the present invention.

As shown in Figure 1-2, demand for commodity information forecasting method under many influence factors, comprises early-stage preparations and two stages of prediction, and wherein the early-stage preparations stage comprises influence factor judgement, the Fuzzy processing of influence factor and the structure of support vector machine; Detailed process is as follows:

Early-stage preparations:

1. influence factor is judged

The variation of merchandise news has certain seasonality, and consumer's selection has diversity and substitutability, makes certain class particular commodity information predict the factor that has a lot of impacts.These factors mainly comprise: weather, temperature, season, festivals or holidays, individual preference, special event etc.These influence factors can be divided into two classes: objective history data and demand environment data.

(1) objective history data

Objective history data mainly refer to the sales data in like product past.Its reflects the level of consumption of these type of commodity in a certain period.According to correlation analysis, in product demand prediction, the previous day of prediction day and the correlativity on the same day of last week are relatively good.

(2) demand environment data

With consumption demand environmental correclation because have weather, temperature, season, date type (working day, festivals or holidays) and special event etc.

2. the Fuzzy processing of influence factor

Except objective history data, the demand environment data such as weather condition, temperature conditions, date type all exert an influence to demand.If consider the impact of these environmental factors in demand forecast, need these environmental datas to process, these environmental factors are converted into fuzzy quantity through subordinate function here.The design of subordinate function can affect the robustness of system and the precision of prediction, and general for linear input, the number of subordinate function can be got less, and for nonlinear output relation, needs to establish some subordinate functions more.(note: in this official documents and correspondence we to take temperature and week be example, summary influence factor Fuzzy processing process)

The subordinate function of temperature is adopted to trapezoidal profile:

(1) subordinate function for low temperature condition adopts type trapezoidal profile less than normal:

$u_{t1}=\left\{\begin{array}{cc}0& t>10\\ \frac{10-t}{10-0}& 0\&le;t\&le;10\\ 1& t<0\end{array}\right.$

(2) subordinate function of centering temperature adopts osculant trapezoidal profile:

$u_{t2}=\left\{\begin{array}{cc}& \\ 0& t>25\\ \frac{t-5}{15-5}& 5\&le;t\&le;15\\ \frac{25-t}{25-15}& 15\&le;t\&le;25\end{array}\right.$

(3) membership function of high temperature is adopted to type trapezoidal profile bigger than normal:

$u_{t3}=\left\{\begin{array}{cc}0& t<20\\ \frac{t-20}{40-20}& 20\&le;t\&le;40\\ 1& t>40\end{array}\right.$

To week, the subordinate function of type adopts half rectangular distribution:

(1) to workaday subordinate function, be:

(2) to the membership function of two-day weekend, be:

By the maximum temperature T of of the same type day _h, minimum temperature T _lbring into respectively three of temperature subordinate functions, obtain the degree of membership { T for low temperature, middle temperature, three states of high temperature _h1, T _h2, T _h3, { T _l1, T _l2, T _l3.According to the maximum principle of degree of membership, get T _h=max{T _h1, T _h2, T _h3and T _l=max{T _l1, T _l2, T _l3, known maximum temperature T _h, minimum temperature T _laffiliated fuzzy set.

According to the fuzzy set under said temperature, consider the factor of weather, adopt temperature-weather quantization parameter method.As following table 1:

Table 1 temperature-weather condition quantization parameter

3. the structure of support vector machine

Traditional machine learning method all, using empirical risk minimization as the target of making great efforts, although can reach very high accuracy on sample set, differs greatly when unknown data is predicted, so-called Generalization Ability is poor.Therefore statistics introduced the concept of extensive error bound, and real exactly risk should be portrayed by two parts content, and the one, empiric risk, has represented the error of sorter on given sample; The 2nd, put trade wind danger, represented the error of sorter on unknown data sample.Put trade wind danger relevant with two amounts, the one, sample size, obviously given sample size is larger, and the structure of machine learning is more possible correct, now puts trade wind dangerous less; The 2nd, the VC dimension of classification function, obviously VC dimension is larger, and Generalization Ability is poorer.Put trade wind and can become by inches large.

The formula of extensive error bound is: R (C)≤R _emp(C)+φ (n/h)

R in formula (C) is exactly real risk, R _emp(C) be exactly empiric risk, φ (n/h) puts trade wind danger.The target of statistical learning has become and has sought empiric risk with that put trade wind danger and minimum from empirical risk minimization, and SVM is a kind of like this algorithm that minimizes structure risk just.

The basic thought of this Forecasting Methodology may be summarized to be: first by nonlinear transformation, input vector x is mapped to a more space Z for higher-dimension, then in this new space, ask for optimum linearity classifying face, this nonlinear transformation is to realize by defining suitable inner product function, and this inner product function is constructed by kernel function.

A given data set

d wherein _iexpectation value, x _ibe the data vector of input, n is the number of training sample.SVM adopts following formula as regression function:

y＝f(x)＝wφ(x)+b （3-1）

φ in formula (x) is the Nonlinear Mapping from input vector collection to high-dimensional feature space, and w and b are coefficient, by minimum risk function R (C), is estimated:

$\mathrm{Minimize}:R\left(C\right)=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{L}_{\mathrm{\&epsiv;}}(d_i,y_i)---(3-2)$

$L_{\mathrm{\&epsiv;}}(d,y)=\left\{\begin{array}{cc}0& |d-y|<\mathrm{\&epsiv;}\\ |d-y|-\mathrm{\&epsiv;}& |d-y|\&GreaterEqual;\mathrm{\&epsiv;}\end{array}\right.---(3-3)$

In formula

be that C is free constant to the measuring of how much intervals between support vector, be used for balance model substantially to measure weight parameter with training error item.ε is given error parameter, error term L _ε(d, y) is the insensitive loss function of ε, and this loss function determines that in feature space one centered by lineoid y=f (x), and thick is the thin plate region of 2 ε.When sample data falls into this region, mean that the difference between predicted value and actual value is less than ε, risk of loss equals 0; In the time of outside sample data falls into this region, it is carried out to linearity punishment.

In order to seek coefficient w and b, introduce slack variable ζ and ζ ^*:

$\begin{array}{c}\mathrm{Minimize}:R(w,\mathrm{\&zeta;},{\mathrm{\&zeta;}}^{*})=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&zeta;}}_i+{\mathrm{\&zeta;}}_i^{*})\\ \mathrm{Subjected\; to}:\left\{\begin{array}{c}{d}_i-\mathrm{w\&phi;}\left(x_i\right)-b\&le;\mathrm{\&epsiv;}+{\mathrm{\&zeta;}}_i\\ \mathrm{w\&phi;}\left(x_i\right)+b-d_i\&le;\mathrm{\&epsiv;}+{\mathrm{\&zeta;}}_i^{*}\\ {\mathrm{\&zeta;}}_i,{\mathrm{\&zeta;}}_i^{*}\&GreaterEqual;0\end{array}\right.\end{array}---(3-4)$

According to Lagrange (Lagrange) principle of duality, introduce dual variable and meet non-negative condition, i.e. α _i,

η _i,

constructed fuction is as follows:

$\begin{array}{c}L=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&zeta;}}_i+{\mathrm{\&zeta;}}_i^{*})-\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i(\mathrm{\&epsiv;}+{\mathrm{\&zeta;}}_i-d_i+\mathrm{w\&phi;}\left(x_i\right)+b)\\ -\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^{*}(\mathrm{\&epsiv;}+{\mathrm{\&zeta;}}_i^{*}+d_i-\mathrm{w\&phi;}\left(x_i\right)-b)-\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&eta;}}_i{\mathrm{\&zeta;}}_i+{\mathrm{\&eta;}}_i^{*}{\mathrm{\&zeta;}}_i^{*})\end{array}---(3-5)$

In optimizing problem, Lagrange function has a saddle point, and according to saddle point condition, L function is respectively with respect to original variable (w, b, ζ _i, ζ ^*) partial derivative be 0, that is:

${\&PartialD;}_{w}L=w-\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})\mathrm{\&phi;}\left(x_i\right)=0---(3-6)$

${\&PartialD;}_{b}L=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&alpha;}}_i)=0---(3-7)$

${\&PartialD;}_{{\mathrm{\&zeta;}}_i}L=C-{\mathrm{\&alpha;}}_i-{\mathrm{\&eta;}}_i=0---(3-8)$

${\&PartialD;}_{{\mathrm{\&zeta;}}_i^{*}}L=C-{\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&eta;}}_i^{*}=0---(3-9)$

By formula (3-6), (3-7), (3-8), (3-9) substitution (3-5), utilize Wolfe antithesis skill, obtain following formula:

$\begin{array}{c}\max \mathrm{imize}:R({\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*})=-\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})({\mathrm{\&alpha;}}_j-{\mathrm{\&alpha;}}_j^{*})<\mathrm{\&phi;}\left(x_i\right),\mathrm{\&phi;}\left(x_j\right)>\\ -\mathrm{\&epsiv;}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i+{\mathrm{\&alpha;}}_i^{*})+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{d}_i({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})\end{array}---(3-10)$

$\mathrm{Subjected\; to}:\left\{\begin{array}{c}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})=0\\ {\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*}\&Element;[0,C]\end{array}\right.$

For nonlinear regression problem, can problem be transformed in a new space by the new proper vector of structure, this need to be converted into the linear problem in another higher dimensional space by nonlinear transformation by nonlinear problem.In this transformation space, only need the inner product operation after definition conversion, need not know the form of the nonlinear transformation of employing, this algorithm has overcome the dimension disaster problem that may cause, and has reduced the complexity of calculating.In Statistical Learning Theory, according to Hilbert-Schmidt principle, as long as a kind of computing meets Mercer condition, just can be used as inner product and use.Therefore SVM method is by introducing suitable inner product kernel function K (x _i, x _j) realize the linear regression after nonlinear transformation.

Kernel function K (x _i, x _j) selection determined the structure of feature space, its value equals two vector x _iand x _jat feature space φ (x _i) and φ (x _j) inner product, i.e. K (x _i, x _j)=φ (x _i) * φ (x _j).Choosing of kernel function has a great impact precision of prediction, and main kernel function has polynomial kernel function K (x, x _i)=(xx _i+ 1) ^q, exponential type radial basis function K (x, x _i)=exp (| x-x _i| ²/ σ ²) and sigmoid function K (x, x _i)=tanh (v (xx _i)+c).Wherein, q is the exponent number of polynomial kernel, σ ²be the width parameter of radial basis core, c is constant.Adopt comparatively accurate radial basis kernel function herein.Therefore, objective function (3-10) becomes:

$\begin{array}{c}\max \mathrm{imize}:R({\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*})=-\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})({\mathrm{\&alpha;}}_j-{\mathrm{\&alpha;}}_j^{*})K(x_i,x_j)\\ -\mathrm{\&epsiv;}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i+{\mathrm{\&alpha;}}_i^{*})+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{d}_i({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})\end{array}---(3-11)$

$\mathrm{Subjected\; to}:\left\{\begin{array}{c}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})=0\\ {\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*}\&Element;[0,C]\end{array}\right.$

Corresponding regression function formula (2-1) is:

$y=f(x,{\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*})=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})K(x,x_i)+b---(3-12)$

By the character of SVM regression function,

and only has minority α _i,

non-vanishing, vector corresponding to these parameters is referred to as support vector, and regression function is completely by its decision.

Prediction steps:

(1) objective history data are carried out to smoothing processing and normalized, demand environment data are carried out to fuzzy quantization processing, then form sample set;

(2) with training sample, set up the objective function suc as formula (3-11);

(3) utilize SVM training algorithm to solve (3-11), obtain separating α _iwith

i=1,2...n;

(4), by the Lagrange multiplier substitution formula (3-12) obtaining, recycling sample is predicted tomorrow requirement amount.

Example demonstration

The input of agricultural product dynamic need forecast model comprises following variable:

T _{max d}: the maximum temperature on same day prediction day

T _{min d}: the minimum temperature on same day prediction day

T _{max (d-1)}: the maximum temperature of the previous day prediction day

T _{min (d-1)}: the minimum temperature of the previous day prediction day

W _d: the weather condition (fine day, rainy day etc.) of prediction day

W _d-1: the weather condition of the previous day prediction day

Wea _d: the week type (weekend or working day) of prediction day

Wea _d-1: the week type of the previous day prediction day

S _d-1: the sales volume of the previous day prediction day

S _d-2: predict the sales volume of the day before yesterday a few days ago

S _d-7: the sales volume of previous this day in week of prediction day

SVM data center based on radial basis kernel function (SBF) is support vector, and network weight is these parameters are all produced automatically by SVM, and SVM algorithm can be summed up as a constrained quadratic programming problem, and its arbitrary solution is globally optimal solution.In the prediction of agricultural product dynamic need, input vector x=[x ₁, x ₂... x _n] ^t=[T _{max d}, T _{min d}... S _d-7] ^t, be output as

the number that wherein n is support vector.

1. collect historical data (table 2) and predicted data (table 3), wherein above 400 groups of data as historical data, after 30 groups of data as test data.

Table 2

Table 3

2. pair historical data and predicted data are carried out respectively Fuzzy Processing (table 2.1) and are carried out level and smooth normalized (table 2.2) again, then form sample set (table 4);

Table 2.1

0.4	0.4	0.2	0.25	1	1	1514	1525	1624
									0.9	0.95	0.4	0.4	0	1	1645	1514	1352
0.2	0.25	0.9	0.95	0	0	1423	1645	1327
									0	0.05	0.2	0.25	0	0	1286	1423	1431
0.2	0.25	0	0.05	0	0	1355	1286	1352
									0	0.05	0.2	0.25	0	0	1287	1355	1622
0	0	0	0.05	1	0	1466	1287	1553
									0.3	0.2	0	0	1	1	1529	1466	1645
0.9	0.9	0.3	0.2	0	1	1563	1529	1423
									0.4	0.45	0.9	0.9	0	0	1326	1563	1286
0	0.05	0.4	0.45	0	0	1245	1326	1355
									0.9	0.9	0	0.05	0	0	1465	1245	1287
0.9	0.9	0.9	0.9	0	0	1333	1465	1466
									0.2	0.2	0.9	0.9	1	0	1487	1333	1529
…	…	…	…	…	…	…	…	…
									0.1	0	0.9	0.9	0	1	1382	1385	1395
0.3	0.2	0.1	0	0	0	1332	1382	1333
									0.3	0.2	0.3	0.2	0	0	1359	1332	1358
0.3	0.2	0.3	0.2	0	0	1289	1359	1295
									0.3	0.2	0.3	0.2	0	0	1296	1289	1395
1	0.9	0.3	0.2	1	0	1329	1296	1385
									1	0.9	1	0.9	1	1	1346	1329	1382

Table 2.2

1	1
		2	0.439394
3	0.093434
		4	0.267677
5	0.09596
		6	0.54798
7	0.707071
		8	0.792929
9	0.194444
		10	-0.0101
11	0.545455
		12	0.212121
13	0.60101
		14	0.191919
…	…

394	0.209596
		395	0.277778
396	0.10101
		397	0.118687
398	0.20202
		399	0.244949
400	0.184343

Table 4

0.4	0.4	0.2	0.25	1	1	0.669192	0.69697	0.94697
									0.9	0.95	0.4	0.4	0	1	1	0.669192	0.260101
0.2	0.25	0.9	0.95	0	0	0.439394	1	0.19697
									0	0.05	0.2	0.25	0	0	0.093434	0.439394	0.459596
0.2	0.25	0	0.05	0	0	0.267677	0.093434	0.260101
									0	0.05	0.2	0.25	0	0	0.09596	0.267677	0.941919
0	0	0	0.05	1	0	0.54798	0.09596	0.767677
									0.3	0.2	0	0	1	1	0.707071	0.54798	1
0.9	0.9	0.3	0.2	0	1	0.792929	0.707071	0.439394
									0.4	0.45	0.9	0.9	0	0	0.194444	0.792929	0.093434
0	0.05	0.4	0.45	0	0	-0.0101	0.194444	0.267677
									0.9	0.9	0	0.05	0	0	0.545455	-0.0101	0.09596
0.9	0.9	0.9	0.9	0	0	0.212121	0.545455	0.54798
									0.2	0.2	0.9	0.9	1	0	0.60101	0.212121	0.707071
…	…	…	…	…	…	…	…	…
									0.1	0	0.9	0.9	0	1	0.335859	0.343434	0.368687
0.3	0.2	0.1	0	0	0	0.209596	0.335859	0.212121
									0.3	0.2	0.3	0.2	0	0	0.277778	0.209596	0.275253
0.3	0.2	0.3	0.2	0	0	0.10101	0.277778	0.116162
									0.3	0.2	0.3	0.2	0	0	0.118687	0.10101	0.368687
1	0.9	0.3	0.2	1	0	0.20202	0.118687	0.343434
									1	0.9	1	0.9	1	1	0.244949	0.20202	0.335859

3. pair predicted data is carried out Fuzzy Processing (table 5.1) and is carried out level and smooth normalized (table 5.2) again, then forms sample set (table 6);

Table 5.1

0.9	0.9	1	0.9	0	1	1322	1346	1332
									0.9	0.9	0.9	0.9	0	0	1266	1322	1359
0.5	0.4	0.9	0.9	0	0	1222	1266	1289
									0.1	0	0.5	0.4	0	0	1243	1222	1296
0.1	0	0.1	0	0	0	1322	1243	1329
									0.3	0.2	0.1	0	1	0	1265	1322	1346
0.3	0.2	0.3	0.2	1	1	1385	1265	1322
									0.5	0.4	0.3	0.2	0	1	1333	1385	1266

0.9	0.9	0.5	0.4	0	0	1243	1333	1222
									0.9	0.9	0.9	0.9	0	0	1193	1243	1243
0.9	0.9	0.9	0.9	0	0	1125	1193	1322
									0.3	0.2	0.9	0.9	0	0	1086	1125	1265
0.3	0.2	0.3	0.2	1	0	1196	1086	1385
									0.1	0	0.3	0.2	1	1	1245	1196	1333
0.3	0.2	0.1	0	0	1	1353	1245	1243
									0.3	0.2	0.3	0.2	0	0	1239	1353	1193
0.3	0.2	0.3	0.2	0	0	1193	1239	1125
									0.5	0.4	0.3	0.2	0	0	1127	1193	1086
1	0.9	0.5	0.4	0	0	1083	1127	1196
									1	0.9	1	0.9	1	0	1022	1083	1245
0.3	0.2	1	0.9	1	1	1158	1022	1353
									1	0.9	0.3	0.2	0	1	1222	1158	1239
0.9	0.9	1	0.9	0	0	1085	1222	1193
									1	0.9	0.9	0.9	0	0	1023	1085	1127
1	0.9	1	0.9	0	0	957	1023	1083
									0.3	0.2	1	0.9	0	0	938	957	1022
0.1	0	0.3	0.2	1	0	1028	938	1158
									0.1	0	0.1	0	1	1	1184	1028	1222
0.3	0.2	0.1	0	0	1	1122	1184	1085
									0.3	0.2	0.3	0.2	0	0	1138	1122	1023

Table 5.2

1	0.042929
		2	-0.06818
3	-0.01515
		4	0.184343
5	0.040404
		6	0.343434
7	0.212121
		8	-0.01515
9	-0.14141
		10	-0.31313
11	-0.41162
		12	-0.13384
13	-0.0101
		14	0.262626
15	-0.02525
		16	-0.14141
17	-0.30808
		18	-0.41919
19	-0.57323

20	-0.2298
		21	-0.06818
22	-0.41414
		23	-0.57071
24	-0.73737
		25	-0.78535
26	-0.55808
		27	-0.16414
28	-0.32071
		29	-0.2803
30	-0.4899

Table 6

0.9	0.9	1	0.9	0	1	0.184343	0.244949	0.209596
									0.9	0.9	0.9	0.9	0	0	0.042929	0.184343	0.277778
0.5	0.4	0.9	0.9	0	0	-0.06818	0.042929	0.10101
									0.1	0	0.5	0.4	0	0	-0.01515	-0.06818	0.118687
0.1	0	0.1	0	0	0	0.184343	-0.01515	0.20202
									0.3	0.2	0.1	0	1	0	0.040404	0.184343	0.244949
0.3	0.2	0.3	0.2	1	1	0.343434	0.040404	0.184343
									0.5	0.4	0.3	0.2	0	1	0.212121	0.343434	0.042929
0.9	0.9	0.5	0.4	0	0	-0.01515	0.212121	-0.06818
									0.9	0.9	0.9	0.9	0	0	-0.14141	-0.01515	-0.01515
0.9	0.9	0.9	0.9	0	0	-0.31313	-0.14141	0.184343
									0.3	0.2	0.9	0.9	0	0	-0.41162	-0.31313	0.040404
0.3	0.2	0.3	0.2	1	0	-0.13384	-0.41162	0.343434
									0.1	0	0.3	0.2	1	1	-0.0101	-0.13384	0.212121
0.3	0.2	0.1	0	0	1	0.262626	-0.0101	-0.01515
									0.3	0.2	0.3	0.2	0	0	-0.02525	0.262626	-0.14141
0.3	0.2	0.3	0.2	0	0	-0.14141	-0.02525	-0.31313
									0.5	0.4	0.3	0.2	0	0	-0.30808	-0.14141	-0.41162
1	0.9	0.5	0.4	0	0	-0.41919	-0.30808	-0.13384
									1	0.9	1	0.9	1	0	-0.57323	-0.41919	-0.0101
0.3	0.2	1	0.9	1	1	-0.2298	-0.57323	0.262626
									1	0.9	0.3	0.2	0	1	-0.06818	-0.2298	-0.02525
0.9	0.9	1	0.9	0	0	-0.41414	-0.06818	-0.14141
									1	0.9	0.9	0.9	0	0	-0.57071	-0.41414	-0.30808
1	0.9	1	0.9	0	0	-0.73737	-0.57071	-0.41919
									0.3	0.2	1	0.9	0	0	-0.78535	-0.73737	-0.57323
0.1	0	0.3	0.2	1	0	-0.55808	-0.78535	-0.2298
									0.1	0	0.1	0	1	1	-0.16414	-0.55808	-0.06818
0.3	0.2	0.1	0	0	1	-0.32071	-0.16414	-0.41414
									0.3	0.2	0.3	0.2	0	0	-0.2803	-0.32071	-0.57071

4. with training sample, set up objective function, in this example with kernel function form K (x, x _i)=exp (γ * | x-x _i| ²) replace original radial basis kernel function K (x, x _i)=exp (| x-x _i| ²/ σ ²), wherein γ and 1/ σ ²played effect of equal value.Three parameters can drafting voluntarily in algorithm are ε, C, γ.

In test process, find, the value of parameter ε is very little on precision of prediction impact, therefore fixing ε=0.01.In actual measurement, find, the value of parameters C, γ is larger on precision of prediction impact.Here according to average percent error e and prediction accuracy A _cfor performance index, take the get parms optimal value of C, γ of the method for cross validation.The result is as following table:

The affect situation of table 7 parameter on precision of prediction

By upper table 7, can be obtained, when parameter γ is fixed on 1, change the value of parameters C, will obtain the forecast model of different accuracy.When C increases to 0.1 by 0.01, prediction accuracy increases, and C is in 0.1 process that continues to increase 100, and the accuracy of prediction is in continuous minimizing, and when C is greater than after 100, prediction accuracy decline rate will increase.When preset parameter C is 0.1, makes γ constantly reduce from 0.25 to both sides or constantly increase the equal variation of prediction accuracy.Therefore this time the parameters optimal value of test is ε=0.01, C=0.1, γ=0.25.

5. utilize SVM training algorithm to solve objective function, obtain separating α _iwith

i=1,2...n;

6. by the Lagrange multiplier substitution obtaining, recycling sample is predicted tomorrow requirement amount, obtains table 8.

Table 8

1	1225.065	16	1224.134
				2	1183.957	17	1169.204
3	1170.928	18	1103.858
				4	1255.546	19	1061.722
5	1325.055	20	1117.751
				6	1373.847	21	1231.923
7	1376.96	22	1163.791
				8	1276.974	23	1040.04
9	1161.522	24	989.3822
				10	1103.24	25	963.1091
11	1095.081	26	988.0927
				12	1101.476	27	1195.036
13	1329.1	28	1253.879
				14	1335.225	29	1121.063
15	1268.517	30	1106.855

As shown in Figure 2, both values are very approaching for the curve of the true sales volume that test obtains under the method and prediction sales volume, and as can be seen here, the prediction sales volume result that method provided by the invention is carried out has higher accuracy.

Claims (5)

1. a demand for commodity information forecasting method under influence factor more than, is characterized in that:

First, need to collect commodity historical data and external information, commodity historical data is carried out to data smoothing processing, data Fuzzy Processing and data normalization and process, information is carried out data Fuzzy Processing and data normalization processing to external world simultaneously, forms training sample set; Then, the training sample set of formation is put into support vector machine and learn, the parameter adjustment in prediction algorithm is arrived to optimal value; Obtain the information that will predict.

2. demand for commodity information forecasting method under many influence factors as claimed in claim 1, is characterized in that:

Commodity historical data mainly refers to the sales data in like product past, according to correlation analysis, the previous day of prediction day and the correlativity on the same day of last week in product demand prediction is compared;

Historical external information mainly refers to and the factor of consumption demand environmental correclation, comprises weather, temperature, season, date type and special event.

3. demand for commodity information forecasting method under many influence factors as claimed in claim 1, is characterized in that: the Fuzzy processing of external information refers to, environmental factor is converted into fuzzy quantity through subordinate function; For linear input, the number of subordinate function can be got less, and for nonlinear output relation, needs to establish some subordinate functions more.

4. demand for commodity information forecasting method under many influence factors as claimed in claim 1, is characterized in that:

In the structure of support vector machine,

The formula of extensive error bound is: R (C)≤R _emp(C)+φ (n/h)

R in formula (C) is exactly real risk, R _emp(C) be exactly empiric risk, φ (n/h) puts trade wind danger;

First by nonlinear transformation, input vector x is mapped to a more space Z for higher-dimension, then in this new space, ask for optimum linearity classifying face, this nonlinear transformation is to realize by defining suitable inner product function, and this inner product function is constructed by kernel function;

A given data set d wherein _iexpectation value, x _ibe the data vector of input, n is the number of training sample; SVM adopts following formula as regression function:

y＝f(x)＝wφ(x)+b （3-1）

φ in formula (x) is the Nonlinear Mapping from input vector collection to high-dimensional feature space, and w and b are coefficient, by minimum risk function R (C), is estimated:

$\mathrm{Minimize}:R\left(C\right)=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{L}_{\mathrm{\&epsiv;}}(d_i,y_i)---(3-2)$

$L_{\mathrm{\&epsiv;}}(d,y)=\left\{\begin{array}{cc}0& |d-y|<\mathrm{\&epsiv;}\\ |d-y|-\mathrm{\&epsiv;}& |d-y|\&GreaterEqual;\mathrm{\&epsiv;}\end{array}\right.---(3-3)$

In formula

be that C is free constant to the measuring of how much intervals between support vector, be used for balance model substantially to measure

weight parameter with training error item; ε is given error parameter, error term L _ε(d, y) is the insensitive loss function of ε, and this loss function determines that in feature space one centered by lineoid y=f (x), and thick is the thin plate region of 2 ε.When sample data falls into this region, mean that the difference between predicted value and actual value is less than ε, risk of loss equals 0; In the time of outside sample data falls into this region, it is carried out to linearity punishment;

In order to seek coefficient w and b, introduce slack variable ζ and ζ ^*:

$\begin{array}{c}\mathrm{Minimize}:R(w,\mathrm{\&zeta;},{\mathrm{\&zeta;}}^{*})=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&zeta;}}_i+{\mathrm{\&zeta;}}_i^{*})\\ \mathrm{Subjected\; to}:\left\{\begin{array}{c}{d}_i-\mathrm{w\&phi;}\left(x_i\right)-b\&le;\mathrm{\&epsiv;}+{\mathrm{\&zeta;}}_i\\ \mathrm{w\&phi;}\left(x_i\right)+b-d_i\&le;\mathrm{\&epsiv;}+{\mathrm{\&zeta;}}_i^{*}\\ {\mathrm{\&zeta;}}_i,{\mathrm{\&zeta;}}_i^{*}\&GreaterEqual;0\end{array}\right.\end{array}---(3-4)$

According to Lagrange duality principle, introduce dual variable and meet non-negative condition, i.e. α _i,

η _i,

constructed fuction is as follows:

$\begin{array}{c}L=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&zeta;}}_i+{\mathrm{\&zeta;}}_i^{*})-\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i(\mathrm{\&epsiv;}+{\mathrm{\&zeta;}}_i-d_i+\mathrm{w\&phi;}\left(x_i\right)+b)\\ -\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^{*}(\mathrm{\&epsiv;}+{\mathrm{\&zeta;}}_i^{*}+d_i-\mathrm{w\&phi;}\left(x_i\right)-b)-\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&eta;}}_i{\mathrm{\&zeta;}}_i+{\mathrm{\&eta;}}_i^{*}{\mathrm{\&zeta;}}_i^{*})\end{array}---(3-5)$

In optimizing problem, Lagrange function has a saddle point, and according to saddle point condition, L function is respectively with respect to original variable (w, b, ζ _i, ζ ^*) partial derivative be 0, that is:

${\&PartialD;}_{w}L=w-\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})\mathrm{\&phi;}\left(x_i\right)=0---(3-6)$

${\&PartialD;}_{b}L=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&alpha;}}_i)=0---(3-7)$

${\&PartialD;}_{{\mathrm{\&zeta;}}_i}L=C-{\mathrm{\&alpha;}}_i-{\mathrm{\&eta;}}_i=0---(3-8)$

${\&PartialD;}_{{\mathrm{\&zeta;}}_i^{*}}L=C-{\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&eta;}}_i^{*}=0---(3-9)$

By formula (3-6), (3-7), (3-8), (3-9) substitution (3-5), utilize Wolfe antithesis skill, obtain following formula:

$\begin{array}{c}\max \mathrm{imize}:R({\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*})=-\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})({\mathrm{\&alpha;}}_j-{\mathrm{\&alpha;}}_j^{*})<\mathrm{\&phi;}\left(x_i\right),\mathrm{\&phi;}\left(x_j\right)>\\ -\mathrm{\&epsiv;}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i+{\mathrm{\&alpha;}}_i^{*})+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{d}_i({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})\end{array}---(3-10)$

$\mathrm{Subjected\; to}:\left\{\begin{array}{c}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})=0\\ {\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*}\&Element;[0,C]\end{array}\right.$

For nonlinear regression problem, can problem be transformed in a new space by the new proper vector of structure, this need to be converted into the linear problem in another higher dimensional space by nonlinear transformation by nonlinear problem; In this transformation space, only need the inner product operation after definition conversion; According to Hilbert-Schmidt principle, as long as a kind of computing meets Mercer condition, just can be used as inner product and use; Therefore SVM method is by introducing suitable inner product kernel function K (x _i, x _j) realize the linear regression after nonlinear transformation;

Kernel function K (x _i, x _j) selection determined the structure of feature space, its value equals two vector x _iand x _jat feature space φ (x _i) and φ (x _j) inner product, i.e. K (x _i, x _j)=φ (x _i) * φ (x _j); Choosing of kernel function has a great impact precision of prediction, and main kernel function has polynomial kernel function K (x, x _i)=(xx _i+ 1) ^q, exponential type radial basis function K (x, x _i)=exp (| x-x _i| ²/ σ ²) and sigmoid function K (x, x _i)=tanh (v (xx _i)+c); Wherein, q is the exponent number of polynomial kernel, σ ²it is the width parameter of radial basis core; Adopt comparatively accurate radial basis kernel function herein; Therefore, objective function (3-10) becomes:

$\begin{array}{c}\max \mathrm{imize}:R({\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*})=-\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})({\mathrm{\&alpha;}}_j-{\mathrm{\&alpha;}}_j^{*})K(x_i,x_j)\\ -\mathrm{\&epsiv;}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i+{\mathrm{\&alpha;}}_i^{*})+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{d}_i({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})\end{array}---(3-11)$

$\mathrm{Subjected\; to}:\left\{\begin{array}{c}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})=0\\ {\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*}\&Element;[0,C]\end{array}\right.$

Corresponding regression function formula (2-1) is:

$y=f(x,{\mathrm{\&alpha;}}_i,{\mathrm{\&alpha;}}_i^{*})=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i-{\mathrm{\&alpha;}}_i^{*})K(x,x_i)+b---(3-12)$

By the character of SVM regression function,

and only has minority α _i,

non-vanishing, vector corresponding to these parameters is referred to as support vector, and regression function is completely by its decision.

5. demand for commodity information forecasting method under many influence factors as claimed in claim 4, is characterized in that: with training sample, set up the objective function suc as formula (3-11); Utilize SVM training algorithm to solve (3-11), obtain separating α _iwith

, i=1,2...n; By in the Lagrange multiplier substitution formula (3-12) obtaining, recycling sample is predicted tomorrow requirement amount.

Images