CN105894136A - Category inventory prediction method and prediction device - Google Patents

Abstract

The invention provides a category inventory prediction method and a prediction device. On one hand, the existing prediction technology gets rid of the dependence on a condition of stable inventory sequence, and on the other hand, limitation of data size scale insufficiency on learning algorithm can also be avoided, so that an unstable category inventory sequence can be greatly predicted. According to a technical scheme provided by the invention, not only can the operation capability of E-business enterprises be improved, but also reasonable and optimal distribution of a social resource is importantly influenced. The category inventory prediction method comprises: obtaining the category inventory sequence from a data platform, and performing standard processing on the sequence; decomposing the standard sequence into a plurality of intrinsic mode functions; predicting each intrinsic mode function to obtain multiple prediction results; collecting multiple prediction results, thereby obtaining a prediction result of the category inventory.

Description

The Forecasting Methodology of a kind of category quantity in stock and prediction means

Technical field

The present invention relates to computer and software technology field thereof, particularly to the Forecasting Methodology of a kind of category quantity in stock and pre- Survey device.

Background technology

In recent years, along with the modern information technologies with the Internet as representative, particularly mobile payment, social networks, search Developing rapidly of the technology such as engine and cloud computing, the Internet starts progressively to cause the concern of the whole society.Each electricity business's platform is also met Carry out its optimal opportunity to develop.Stock control as the stronger activity of a systematicness, its cash flow with electricity commercial business industry, letter The links such as breath stream and logistics are closely bound up, indivisible, determine the resources allocation ability of electricity business, are related to the life of electricity commercial business industry Arteries and veins.Efficient inventory management system can not only promote the operation ability of electricity commercial business industry, the also optimal sorting to social resources Join and play material impact.The prediction of category quantity in stock is as the key link of stock control, and it is the non-linear of a high complexity System, this is because the quantity in stock of category is affected by many factors, has great randomness and uncertainty.At present, Theoretical research in terms of the Forecast of Nonlinear Time Series such as all about category quantity in stock, it is stable for being built upon forecasting sequence On this basic assumption.But absolutely big number system be not belonging to ballast system.Up to now, for being produced by non-stationary system The forecasting problem of sequence still lack the theoretical research of system.

At the beginning of the seventies, Bock thinks (Box) and Charles Jenkins (Jenkins) proposes a famous time series forecasting mould Type ARMA model (Autoregressive Integrated Moving Average Model), referred to as ARIMA model, the basic ideas of ARIMA model are: the data sequence formed predicting object to elapse in time is considered as one Random sequence.Based on seasonal effect in time series autocorrelation analysis.This sequence of approximate description is carried out with certain mathematical model.This Model just can predict future value from seasonal effect in time series past value and present value after identified.ARIMA model is in economy Both economic phenomenon interdependence in time series had been considered during prediction, it is contemplated that the interference of random fluctuation, for The predictablity rate of economical operation short-term trend is higher, is Application comparison one of method widely in recent years.And along with interconnection netting index According to violent expansion, a lot of forecasting problems can solve by learning algorithm.Such as neural network model, support vector machine (SVM) Algorithm and degree of depth study scheduling algorithm.These algorithms have the ability of preferably Approximation of Arbitrary Nonlinear Function and adaptive study Ability, can by study primary demand, periodic factors, marketing activity, random factor in season etc. affect demand various because of Complicated interaction between element solidifies in a network with the form of weights, thus improves precision of prediction.At present, these algorithms exist A lot of large enterprises answer land used widely.

In prior art, the primary condition that ARIMA model is applied in category quantity in stock forecasting problem is requirement category storehouse Storage sequence meets the condition of stationarity, i.e. individual values and to fluctuate up and down around serial mean, it is impossible to have significantly rise or under Fall trend.Category quantity in stock sequence is occurred rising or falling trend, although ARIMA model has, and original series is carried out difference Tranquilization pretreatment mechanism, but the adaptive processing method of this shortage cannot make the stationarity of sequence be guaranteed, therefore The result that prediction obtains still suffers from deficiency on accuracy.On the other hand, although learning algorithm receives in recent years and much closes Note, but it is higher to the scale requirements of training sample, only when training sample scale is sufficiently large, it was predicted that result just can be more Accurately；For the problem that sequence scale is smaller, this kind of algorithm still cannot overcome this difficulty non-stationary of sequence.

On the one hand in sum, it is the dependence of this condition smoothly that prior art cannot be broken away from quantity in stock sequence, separately Scale to training sample has the requirement that comparison is high the most again.Therefore, non-stationary series or sequence scale are not reached The accuracy predicted the outcome of the sequence required also cannot be guaranteed.This operation ability not simply failing to promote electricity commercial business industry, But also the reasonable optimum allocation to social resources affects.

Summary of the invention

In view of this, the present invention provides Forecasting Methodology and the prediction means of a kind of category quantity in stock, on the one hand can break away from Pre existing survey technology is the dependence of this condition smoothly to quantity in stock sequence, is on the other hand also avoided that data volume scale is not enough Limitation on learning algorithm, it is possible to preferably the category quantity in stock sequence of non-stationary is predicted.The technology of the present invention side Case can not only promote the operation ability of electricity commercial business industry, also distributes social resources rationally and plays material impact.

For achieving the above object, according to an aspect of the invention, it is provided the Forecasting Methodology of a kind of category quantity in stock.

The Forecasting Methodology of the category quantity in stock of the present invention, including: obtain category quantity in stock sequence from data platform, and to this Sequence is standardized processing；With optimizing average decomposition method, the sequence after standardization is decomposed into several intrinsic mode letters Number；It is predicted obtaining multiple predicting the outcome to intrinsic mode function each described；The plurality of predicting the outcome is collected, Thus obtain predicting the outcome of described category quantity in stock.

Alternatively, with optimization average decomposition method, the sequence after standardization is decomposed into the step of several intrinsic mode functions Suddenly include:

Find out the category quantity in stock sequence after standardizationAll Local Extremum WhereinAnd by these extreme points by intervalIt is divided into:

${\{B_j^x:=\[a_j^x,a_{j+1}^x\]\}}_{j=1}^{m^x}$

Wherein, i is the time point of sequence x, and span is m^x+ 1 is the total number of extreme point of sequence x；The total number of time point for sequence；x_iFor time point i institute in category quantity in stock sequence Corresponding category quantity in stock；

Average Mx of sequence x is obtained by solving following minimum optimization problem:

$\left\{\begin{array}{c}Mx:=\underset{g}{min}\mathrm{\α}\left|\right|D^1(x-g)|{|}_1+{\mathrm{\α}}_1|\left|D^{1}g\right|{|}_1+(1-{\mathrm{\α}}_1)\left|\right|D^{3}g|{|}_1\\ \begin{array}{cc}s.t.& A(x-g)=0\\ & Cg=b,\end{array}\end{array}\right.$

Wherein,Represent segmentation summation operator, its element a_hl=1, whenOtherwise, a_hl =0；Its element definition is Wherein c_hl=1, whenOtherwise c_hl=0；AndIt is respectively 1 rank and 3 jumps divide Matrix；Set for all real number compositions；G is unknown number to be solved；α and α₁Respectively punish parameter；

Sequence of calculation x and the difference of its average Mx, i.e. s₁:=x-Mx, s₁First intrinsic mode function for sequence x；

By intrinsic mode function s₁Separate from sequence x, obtain surplus r₁=x-s₁；Judge r₁Whether it is dull letter Number, if then terminating decomposition method, otherwise, by r₁Replace the step before sequence x repeats to obtain next eigen mode State function, finally gives all eigenfunctions of this sequenceL is eigen mode The number of state function.

Alternatively, described be predicted intrinsic mode function each described obtains multiple step bag predicted the outcome Include:

Use following Holt-Winters exponential smoothing model to described intrinsic mode functionIt is predicted:

$s_{k,\stackrel{\‾}{N}+m}=(u_{\stackrel{\‾}{N}}+{\mathrm{mb}}_{\stackrel{\‾}{N}})s_{k,\stackrel{\‾}{N}-K+m},m=1,2,\mathrm{...}$

Wherein: k represents kth intrinsic mode function, m is the time at intervals number of moment distance present moment to be predicted；

$\left\{\begin{array}{c}{u}_{\stackrel{\‾}{N}}=\mathrm{\η}\frac{s_{k,\stackrel{\‾}{N}}}{F_{\stackrel{\‾}{N}-K}}+(1-\mathrm{\η})(u_{\stackrel{\‾}{N}-1}+b_{\stackrel{\‾}{N}-1})\\ b_{\stackrel{\‾}{N}}=\mathrm{\β}(u_{\stackrel{\‾}{N}}-u_{\stackrel{\‾}{N}-1})+(1-\mathrm{\β})b_{\stackrel{\‾}{N}-1}\\ F_{\stackrel{\‾}{N}}=\mathrm{\γ}\frac{s_{k,\stackrel{\‾}{N}}}{u_{\stackrel{\‾}{N}}}+(1-\mathrm{\γ})F_{\stackrel{\‾}{N}-K}\end{array}\right.$

Wherein:ForThe periodic term in moment, is the time series exponential smoothing average removing seasonal variations impact； ForThe trend term in moment, is the exponential smoothing average of time series variation trend；ForThe item in season in moment, is season The exponential smoothing average of the factor；ForThe actual value in moment；K is length or time cycle in season；η, β, γ represent respectively Smoothing factor, value between (0,1).

Alternatively, described multiple steps collected that predict the outcome are included: collect all intrinsic mode according to equation below Predicting the outcome of function, it may be assumed that

$\hat{x}(\stackrel{\‾}{N}+m)=\underset{k=1}{\overset{L}{\Σ}}{s}_{k,\stackrel{\‾}{N}+m}$

Wherein,Represent that this value is prediction quantity in stock,For current date, m represents the time interval of distance current time.

According to another aspect of the present invention, it is provided that the prediction means of a kind of category quantity in stock.

The prediction means of the category quantity in stock of the present invention, including: acquisition module, for obtaining category stock from data platform Amount sequence, and be standardized this sequence processing；Decomposing module, for optimizing average decomposition method by the sequence after standardization Row are decomposed into several intrinsic mode functions；Prediction module, is predicted obtaining many to intrinsic mode function each described Individual predict the outcome；Summarizing module, for the plurality of predicting the outcome being collected, thus obtains the prediction knot of described category quantity in stock Really.

Alternatively, described decomposing module is additionally operable to:

Find out the category quantity in stock sequence after standardizationAll Local Extremum And by these extreme points by intervalIt is divided into:

${\{B_j^x:=\[a_j^X,a_{j+1}^X\]\}}_{j=1}^{m^X}$

Wherein, i is the time point of sequence x, and span is m^x+ 1 is the total number of extreme point of sequence x；The total number of time point for sequence；x_iFor time point i institute in category quantity in stock sequence Corresponding category quantity in stock；

Average Mx of sequence x is obtained by solving following minimum optimization problem:

$\left\{\begin{array}{c}Mx:=\underset{g}{min}\mathrm{\α}\left|\right|D^1(x-g)|{|}_1+{\mathrm{\α}}_1|\left|D^{1}g\right|{|}_1+(1-{\mathrm{\α}}_1)\left|\right|D^{3}g|{|}_1\\ \begin{array}{cc}s.t.& A(X-g)=0\\ & Cg=b,\end{array}\end{array}\right.$

Wherein,Represent segmentation summation operator, its element a_hl=1, whenOtherwise, a_hl =0；Its element definition isIts Middle c_hl=1, whenOtherwise c_hl=0；AndIt is respectively 1 rank and 3 jumps divide square Battle array；Set for all real number compositions；G is unknown number to be solved；α and α₁Respectively punish parameter；

Sequence of calculation x and the difference of its average Mx, i.e. s₁=x-Mx, s₁First intrinsic mode function for sequence x；

By intrinsic mode function s₁Separate from sequence x, obtain surplus r₁=x-s₁；Judge r₁Whether it is dull letter Number, if then terminating decomposition method, otherwise, by r₁Replace the step before sequence x repeats to obtain next eigen mode State function, finally gives all eigenfunctions of this sequenceL is eigen mode The number of state function.

Alternatively, described prediction module is additionally operable to: use following Holt-Winters exponential smoothing model to described Levy mode functionIt is predicted:

$s_{k\stackrel{\‾}{N}+m}=(u_{\stackrel{\‾}{N}}+{\mathrm{mb}}_{\stackrel{\‾}{N}})s_{k\stackrel{\‾}{N}-K+m},m=1,2,...$

Wherein: k represents kth intrinsic mode function, m is the time at intervals number of moment distance present moment to be predicted；

$\left\{\begin{array}{c}{u}_{\stackrel{\‾}{N}}=\mathrm{\η}\frac{s_{k\stackrel{\‾}{N}}}{F_{\stackrel{\‾}{N}-K}}+(1-\mathrm{\η})(u_{\stackrel{\‾}{N}-1}+b_{\stackrel{\‾}{N}-1})\\ b_{\stackrel{\‾}{N}}=\mathrm{\β}(u_{\stackrel{\‾}{N}}-u_{\stackrel{\‾}{N}-1})+(1-\mathrm{\β})b_{\stackrel{\‾}{N}-1}\\ F_{\stackrel{\‾}{N}}=\mathrm{\γ}\frac{s_{k\stackrel{\‾}{N}}}{u_{\stackrel{\‾}{N}}}+(1-\mathrm{\γ})F_{\stackrel{\‾}{N}-K}\end{array}\right.$

Wherein:ForThe periodic term in moment, is the time series exponential smoothing average removing seasonal variations impact；ForThe trend term in moment, is the exponential smoothing average of time series variation trend；ForThe item in season in moment, is season The exponential smoothing average of the joint factor；ForThe actual value in moment；K is length or time cycle in season；η, β, γ table respectively Show smoothing factor, value between (0,1).

Alternatively, described summarizing module is additionally operable to: collect predicting the outcome of all intrinsic mode functions according to equation below, That is:

$\hat{x}(\stackrel{\‾}{N}+m)=\underset{k=1}{\overset{L}{\Σ}}{s}_{k,\stackrel{\‾}{N}+m}$

Wherein,Represent that this value is prediction quantity in stock,For current date, m represents the time interval of distance current time.

According to technical scheme, owing to have employed optimization average decomposition method, sequence is decomposed, therefore, On the one hand break away from the dependence that pre existing survey technology is this condition smoothly to quantity in stock sequence, be on the other hand also avoided that number According to the not enough limitation on learning algorithm of gauge mould, it is possible to preferably the category quantity in stock sequence of non-stationary is predicted, Technical solution of the present invention can not only promote the operation ability of electricity commercial business industry, the also reasonable optimum allocation to social resources and play weight Affect.

Accompanying drawing explanation

Accompanying drawing is used for being more fully understood that the present invention, does not constitute inappropriate limitation of the present invention.Wherein:

Fig. 1 is the schematic diagram of the Forecasting Methodology of a kind of category quantity in stock according to embodiments of the present invention；

Fig. 2 is the sequence chart of the category quantity in stock training set after the standardization of the embodiment of the present invention；

Fig. 3 is the sequence after the standardization of the embodiment of the present invention, first mode eigenfunction and second mode intrinsic letter The sequence chart of number and corresponding density fonction figure；Wherein, the sequence after Fig. 3 (a) is standardization, first mode intrinsic letter Number and the sequence chart of second mode eigenfunction；Fig. 3 (b) be the sequence after standardization, first mode eigenfunction and The density fonction figure that second mode eigenfunction is corresponding；

Fig. 4 is the 3rd mode eigenfunction, the 4th mode eigenfunction and the 5th mode intrinsic of the embodiment of the present invention The sequence chart of function and corresponding density fonction figure；Wherein, Fig. 4 (a) is the 3rd mode eigenfunction, the 4th mode intrinsic Function and the sequence chart of the 5th mode eigenfunction；Fig. 4 (b) be the 3rd mode eigenfunction, the 4th mode eigenfunction with And the 5th density fonction figure corresponding to mode eigenfunction；

Fig. 5 is the 6th mode eigenfunction, the 7th mode eigenfunction and the sequence chart of surplus of the embodiment of the present invention With corresponding density fonction figure；Fig. 5 (a) is the 6th mode eigenfunction, the 7th mode eigenfunction and the sequence of surplus Row figure；Fig. 5 (b) is the density fonction figure that the 6th mode eigenfunction, the 7th mode eigenfunction and surplus are corresponding；

Fig. 6 is the figure that predicts the outcome of technical scheme according to embodiments of the present invention, Holt-Winters and ARIMA model； Wherein, Fig. 6 (a) is that technology and Holt-Winters predict the outcome the comparison diagram with legitimate reading according to embodiments of the present invention； Fig. 6 (b) is the comparison diagram of ARIMA model and legitimate reading；

Fig. 7 is the schematic diagram of the prediction means of a kind of category quantity in stock according to embodiments of the present invention.

Detailed description of the invention

Below in conjunction with accompanying drawing, the one exemplary embodiment of the present invention is explained, various including the embodiment of the present invention Details is to help understanding, it should they are thought the most exemplary.Therefore, those of ordinary skill in the art it should be noted that Arrive, the embodiments described herein can be made various changes and modifications, without departing from scope and spirit of the present invention.With Sample, for clarity and conciseness, eliminates the description to known function and structure in description below.

Fig. 1 is the schematic diagram of the Forecasting Methodology of a kind of category quantity in stock according to embodiments of the present invention.As it is shown in figure 1, should Method mainly comprises the steps S10 to S13.

Step S10: obtain category quantity in stock sequence from data platform, and be standardized this sequence processing.In this step In Zhou, the technology such as Hadoop/Hive can be used to obtain category quantity in stock sequence from data platform；In order to eliminate the dimension between variable Relation, so that data have comparability, is therefore standardized accessed category quantity in stock sequence processing.

Step S11: the sequence after standardization is decomposed into several intrinsic mode functions with optimizing average decomposition method.? In this step, the sequence after standardization is carried out decomposition and can include following several step:

Find out the category quantity in stock sequence after standardizationAll Local Extremum WhereinAnd by these extreme points by intervalIt is divided into:

${\{B_j^x:=\[a_j^x,a_{j+1}^x\]\}}_{j=1}^{m^x}$

Wherein, i is the time point of sequence x, and span is m^x+ 1 is the total number of extreme point of sequence x；The total number of time point for sequence；x_iFor time point i institute in category quantity in stock sequence Corresponding category quantity in stock；

Average Mx of sequence x is obtained by solving following minimum optimization problem:

$\left\{\begin{array}{c}Mx:=\underset{g}{min}\mathrm{\α}\left|\right|D^1(x-g)|{|}_1+{\mathrm{\α}}_1|\left|D^{1}g\right|{|}_1+(1-{\mathrm{\α}}_1)\left|\right|D^{3}g|{|}_1\\ \begin{array}{cc}s.t.& A(x-g)=0\\ & Cg=b,\end{array}\end{array}\right.$

Wherein,Represent segmentation summation operator, its element a_hl=1, whenOtherwise, a_hl=0；Its element definition is Wherein c_hl=1, whenOtherwise c_hl=0；AndIt is respectively 1 rank and 3 jumps divide Matrix；Set for all real number compositions；G is unknown number to be solved；α and α₁Respectively punish parameter；

Sequence of calculation x and the difference of its average Mx, i.e. s₁:=x-Mx, s₁First intrinsic mode function for sequence x (IMF)；

By intrinsic mode function s₁Separate from quantity in stock sequence x, obtain surplus r₁=x-s₁；If judging r₁For list Letter of transfer number, terminates decomposition method, otherwise, by r₁Replace the step before sequence x repeats to obtain next intrinsic mode Function, finally gives all eigenfunctions of this sequenceL is intrinsic mode The number of function.

Step S12: be predicted obtaining multiple predicting the outcome to intrinsic mode function each described.Real in the present invention Execute and the technical scheme of example uses following Holt-Winters exponential smoothing model to described intrinsic mode functionIt is predicted:

$s_{k,\stackrel{\‾}{N}+m}=(u_{\stackrel{\‾}{N}}+{\mathrm{mb}}_{\stackrel{\‾}{N}})s_{k,\stackrel{\‾}{N}-K+m},m=1,2,\mathrm{...}$

Wherein: k represents kth intrinsic mode function, m is the time at intervals number of moment distance present moment to be predicted；

$\left\{\begin{array}{c}{u}_{\stackrel{\‾}{N}}=\mathrm{\η}\frac{s_{k,\stackrel{\‾}{N}}}{F_{\stackrel{\‾}{N}-K}}+(1-\mathrm{\η})(u_{\stackrel{\‾}{N}-1}+b_{\stackrel{\‾}{N}-1})\\ b_{\stackrel{\‾}{N}}=\mathrm{\β}(u_{\stackrel{\‾}{N}}-u_{\stackrel{\‾}{N}-1})+(1-\mathrm{\β})b_{\stackrel{\‾}{N}-1}\\ F_{\stackrel{\‾}{N}}=\mathrm{\γ}\frac{s_{k,\stackrel{\‾}{N}}}{u_{\stackrel{\‾}{N}}}+(1-\mathrm{\γ})F_{\stackrel{\‾}{N}-K}\end{array}\right.$

Wherein:ForThe periodic term in moment, is the time series exponential smoothing average removing seasonal variations impact；ForThe trend term in moment, is the exponential smoothing average of time series variation trend；ForThe item in season in moment, is season The exponential smoothing average of the joint factor；ForThe actual value in moment；K is length or time cycle in season；η, β, γ table respectively Show smoothing factor, value between (0,1), smoothing factor η, β, γ choose difference, inevitably result in predict the outcome can not By property, thus, utilize computer programming, by exhaustive 3 smoothing factors combination in any in [0,1] interval, respectively according to it Make corresponding prediction, and calculated the quadratic sum of relative error magnitudes by these result of calculations, therefrom choose the relative of minimum Smoothing factor corresponding to error amount quadratic sum is as " optimal smoothing coefficient ".

Step S13: the plurality of predicting the outcome is collected, thus obtain predicting the outcome of described category quantity in stock.At this In step, use and collect predicting the outcome of all intrinsic mode functions according to equation below, it may be assumed that

$\hat{x}(\stackrel{\‾}{N}+m)=\underset{k=1}{\overset{L}{\Σ}}{s}_{k,\stackrel{\‾}{N}+m}$

Wherein,Represent that this value is prediction quantity in stock,For current date, m represents the time interval of distance current time.

Take below certain category quantity in stock sequence in certain warehouse as prediction object, it is made up of the observation of 1200 days. For the effect checking the technology of the present invention to predict, these 1200 observations are divided into two parts: front 1080 observations are training Collection, as in figure 2 it is shown, remaining 120 values are test set.

Due to stationarity is that nearly all traditional time series predicting model notional result treats wanting of forecasting sequence Asking, therefore we first provide a kind of definition of category quantity in stock sequence stationary:

Given category quantity in stock sequence x (t), wherein: t express time sky, x (t) represents that this category was the stock of the t days Amount.To random time τ, if following formula is set up

E(|x²(t)|)<∞,

E (x (t))=m,

C(x(t₁),x(t₂))=C (x (t₁+τ),x(t₂+ τ))=C (t₁-t₂),

Wherein, E () represents expectation, and C () is covariance, and m is a constant, then this category quantity in stock sequence x (t) is called Weakly stationary.

2:(is the most steady in definition)

The another kind of definition of category quantity in stock sequence stationary is: given category quantity in stock sequence x (t), wherein: when t represents Between sky, x (t) represents that this category was at the quantity in stock of the t days.To arbitrary time delay τ and time t_i, (i=1,2 ..., n), as Really [x (t₁),x(t₂),…,x(t_n)] and [x (t₁+τ),x(t₂+τ),…,x(t_n+ τ)] Joint Distribution all keep consistent, then claim This category quantity in stock sequence x (t) is strict stable.

Then have limited second moment E (| x²(t) |) strict steadily quantity in stock sequence be also weakly stationary；Otherwise, then It is false.The definition of both stationarities is the strictest and idealizes.In actual applications, the length of category quantity in stock sequence It is all limited.Therefore, even if check the stationarity of sequence one by one according to the two definition, also can only obtain the judgement knot of approximation Really.To any time sequence, no matter it is derived from natural phenomena still by manual construction, nearly all can not meet on State the condition in two definition.To this end, Yang et al. proposes the technology that a kind of detection sequence is non-stationary, this technology solves sequence The distribution density function of row, according to distribution density function and horizontal degree of approximation, it is possible to intuitively between comparative sequences Non-stationary.Wherein, the calculating process of distribution density function is as follows:

Step A: utilize the delay coordinate structure m-that time delay is n days to tie up phase spaceAnd generate N number of embedding vector Make:

Wherein, subscript T represents transposition computing, andFor spaceIn element number；

Step B: definition phase spaceIn two element s_i,s_j(i, j=1 ..., N, i ≠ j) time natural law be spaced apart:

Wherein, T (s_i)=i represents a s_iThe time point occurred on its track is i-th day；

Step C: for pointFind out and all belong to s_i∈₁Other point of neighborhood, i.e.

Step D: collect s_iWithThe all different time natural law interval value of middle element, and The frequency that accumulative each time natural law interval occursThat is, forIn each time natural law interval, ifThen:

Step F: for given distance radius ∈₁, to s in phase space_iOutside other s_j, (i, j=1 ..., N, J ≠ i), the cumulative time natural law interval frequency

Step G: interval minimum, maximum time natural law interval composition is divided into K+1 minizone, and then obtains specification The density fonction (DDF) changed

Wherein, ξ ∈ 0,1 ..., K} is the normalized time,For normalized time interval,Represent all time interval frequency sums；

Step H: other distance radius ∈ given₂,…,∈_Q, repetition step 43, to step 46, obtains corresponding specification close Degree distribution function

For any one sequence, by superimposed for Q the density fonction obtained, the density that i.e. can obtain this sequence is divided Cloth function, in the technical scheme of the embodiment of the present invention, Q value is 10.

By the average decomposition method that optimizes in embodiment of the present invention technical scheme, the sequence after standardization is decomposed, Its decomposition result shows as shown in Fig. 3, Fig. 4 and Fig. 5.It can be seen that optimize average mark and resolve from Fig. 3, Fig. 4 and Fig. 5 The quantity in stock sequence of non-stationary has been resolved into seven and has approximated stable intrinsic mode function and a trend component by method.Fig. 3, figure Give the sequence after standardization and the density fonction of each intrinsic mode function in 4 and Fig. 5 simultaneously, can from figure Go out, optimize averaging and can effectively carry out non-stationary category quantity in stock sequence being decomposed into the stable sequence of approximation.

Fig. 6 is the figure that predicts the outcome of technical scheme according to embodiments of the present invention, Holt-Winters and ARIMA model； Wherein, Fig. 6 (a) is that technology and Holt-Winters predict the outcome the comparison diagram with legitimate reading according to embodiments of the present invention； Fig. 6 (b) is the comparison diagram of ARIMA model and legitimate reading.From this result it can be seen that the result of the technology of the present invention and true feelings Condition is the most identical.Table 1 weighs the error of above-mentioned three kinds of forecast models and real sequence by several indexs.Table 2 is given respectively The numerical value of three kinds of corresponding the two indexs of prediction.Therefrom it can also be seen that optimize average-Holt-Winters forecast model Result predict the outcome more reasonable than other two kinds of models, its prediction deviation all has substantially reduction than prior art.

1: two performance indications of table and calculation expression thereof

The performance comparison of 2: three kinds of model prediction results of table

Fig. 7 is the schematic diagram of the prediction means of a kind of category quantity in stock according to embodiments of the present invention.As it is shown in fig. 7, this The prediction means 70 of the category quantity in stock of inventive embodiments mainly include acquisition module 71, decomposing module 72, prediction module 73, with And summarizing module 74；Acquisition module 71 is for obtaining category quantity in stock sequence from data platform, and is standardized this sequence Process；Decomposing module 72 for being decomposed into several intrinsic mode letters with the average decomposition method of optimization by the sequence after standardization Number；Prediction module 73 is for being predicted obtaining multiple predicting the outcome to intrinsic mode function each described；Summarizing module 74 for collecting the plurality of predicting the outcome, thus obtains predicting the outcome of described category quantity in stock.

The decomposing module 72 of the prediction means 70 of the category quantity in stock of the embodiment of the present invention can be additionally used in:

Find out the category quantity in stock sequence after standardizationAll Local Extremum WhereinAnd by these extreme points by intervalIt is divided into:

${\{B_j^x:=\&lsqb;a_j^x,a_{j+1}^x\&rsqb;\}}_{j=1}^{m^x}$

Wherein, i is the time point of sequence x, and span is m^x+ 1 is the total number of extreme point of sequence x；The total number of time point for sequence；x_iFor time point i institute in category quantity in stock sequence Corresponding category quantity in stock；

Average Mx of sequence x is obtained by solving following minimum optimization problem:

$\left\{\begin{array}{c}Mx:=\underset{g}{min}\mathrm{\&alpha;}\left|\right|D^1(x-g)|{|}_1+{\mathrm{\&alpha;}}_1|\left|D^{1}g\right|{|}_1+(1-{\mathrm{\&alpha;}}_1)\left|\right|D^{3}g|{|}_1\\ \begin{array}{cc}s.t.& A(x-g)=0\\ & Cg=b,\end{array}\end{array}\right.$

Wherein,Represent segmentation summation operator, its element a_hl=1, whenOtherwise, a_hl =0；Its element definition is Wherein c_hl=1, whenOtherwise c_hl=0；AndIt is respectively 1 rank and 3 jumps divide Matrix；Set for all real number compositions；G is unknown number to be solved；α and α₁Respectively punish parameter；

Sequence of calculation x and the difference of its average Mx, i.e. s₁=x-Mx, s₁First intrinsic mode function (IMF) for sequence x；

By intrinsic mode function s₁Separate from sequence x, obtain surplus r₁=x-s₁；Judge r₁Whether it is dull letter Number, if then terminating decomposition method, otherwise, by r₁Replace the step before sequence x repeats to obtain next eigen mode State function, finally gives all eigenfunctions of this sequenceL is eigen mode The number of state function.

The prediction module 73 of the prediction means 70 of the category quantity in stock of the embodiment of the present invention can be additionally used in:

Use following Holt-Winters exponential smoothing model to described intrinsic mode function It is predicted:

$s_{k,\stackrel{\&OverBar;}{N}+m}=(u_{\stackrel{\&OverBar;}{N}}+{\mathrm{mb}}_{\stackrel{\&OverBar;}{N}})s_{k,\stackrel{\&OverBar;}{N}-K+m},m=1,2,\mathrm{...}$

Wherein: k represents kth intrinsic mode function, m is the time at intervals number of moment distance present moment to be predicted；

$\left\{\begin{array}{c}{u}_{\stackrel{\&OverBar;}{N}}=\mathrm{\&eta;}\frac{s_{k,\stackrel{\&OverBar;}{N}}}{F_{\stackrel{\&OverBar;}{N}-K}}+(1-\mathrm{\&eta;})(u_{\stackrel{\&OverBar;}{N}-1}+b_{\stackrel{\&OverBar;}{N}-1})\\ b_{\stackrel{\&OverBar;}{N}}=\mathrm{\&beta;}(u_{\stackrel{\&OverBar;}{N}}-u_{\stackrel{\&OverBar;}{N}-1})+(1-\mathrm{\&beta;})b_{\stackrel{\&OverBar;}{N}-1}\\ F_{\stackrel{\&OverBar;}{N}}=\mathrm{\&gamma;}\frac{s_{k,\stackrel{\&OverBar;}{N}}}{u_{\stackrel{\&OverBar;}{N}}}+(1-\mathrm{\&gamma;})F_{\stackrel{\&OverBar;}{N}-K}\end{array}\right.$

Wherein:ForThe periodic term in moment, is the time series exponential smoothing average removing seasonal variations impact； ForThe trend term in moment, is the exponential smoothing average of time series variation trend；ForThe item in season in moment, is season The exponential smoothing average of the factor；ForThe actual value in moment；K is length or time cycle in season；η, β, γ represent respectively Smoothing factor, value between (0,1), smoothing factor η, β, γ choose difference, inevitably result in predict the outcome unreliable Property, thus, utilize computer programming, by exhaustive 3 smoothing factors in the combination in any in [0,1] interval, do according to it respectively Go out corresponding prediction, and calculated the quadratic sum of relative error magnitudes by these result of calculations, therefrom choose the relative of minimum and miss Squared difference and corresponding smoothing factor are as " optimal smoothing coefficient ".

The summarizing module 74 of the prediction means 70 of the category quantity in stock of the embodiment of the present invention can be additionally used in:

Predicting the outcome of all intrinsic mode functions is collected according to equation below, it may be assumed that

$\hat{x}(\stackrel{\&OverBar;}{N}+m)=\underset{k=1}{\overset{L}{\&Sigma;}}{s}_{k,\stackrel{\&OverBar;}{N}+m}$

Wherein,Represent that this value is prediction quantity in stock,For current date, m represents the time interval of distance current time.

Technical scheme according to embodiments of the present invention, has been carried out sequence point owing to have employed the average decomposition method of optimization Solve, therefore, on the one hand broken away from the dependence that pre existing survey technology is this condition smoothly to quantity in stock sequence, the most also It is avoided that the not enough limitation on learning algorithm of data volume scale, it is possible to preferably the category quantity in stock sequence of non-stationary is entered Row prediction.Embodiment of the present invention technical scheme can not only promote the operation ability of electricity commercial business industry, also reasonable to social resources Material impact is played in optimum allocation.

Above-mentioned detailed description of the invention, is not intended that limiting the scope of the invention.Those skilled in the art should be bright White, depend on that design requires and other factors, various amendment, combination, sub-portfolio and replacement can occur.Any Amendment, equivalent and the improvement etc. made within the spirit and principles in the present invention, should be included in scope Within.

Claims (8)

1. the Forecasting Methodology of a category quantity in stock, it is characterised in that including:

Obtain category quantity in stock sequence from data platform, and be standardized this sequence processing；

With optimizing average decomposition method, the sequence after standardization is decomposed into several intrinsic mode functions；

It is predicted obtaining multiple predicting the outcome to intrinsic mode function each described；

The plurality of predicting the outcome is collected, thus obtains predicting the outcome of described category quantity in stock.

Method the most according to claim 1, it is characterised in that the sequence after standardization is divided with optimizing average decomposition method Solve the step for several intrinsic mode functions to include:

Find out the category quantity in stock sequence after standardizationAll Local Extremum And by these extreme points by intervalIt is divided into:

${\{B_j^x:=\&lsqb;a_j^x,a_{j+1}^x\&rsqb;\}}_{j=1}^{m^x}$

Wherein, i is the time point of sequence x, and span is m^x+ 1 is the total number of extreme point of sequence x；The total number of time point for sequence；x_iFor time point i institute in category quantity in stock sequence Corresponding category quantity in stock；

Average Mx of sequence x is obtained by solving following minimum optimization problem:

$\left\{\begin{array}{c}Mx:=\underset{g}{min}\mathrm{\&alpha;}\left|\right|D^1(x-g)|{|}_1+{\mathrm{\&alpha;}}_1|\left|D^{1}g\right|{|}_1+(1-{\mathrm{\&alpha;}}_1)\left|\right|D^{3}g|{|}_1\\ \begin{array}{cc}s.t.& A(x-g)=0\\ & Cg=b,\end{array}\end{array}\right.$

Wherein,Represent segmentation summation operator, its element a_hl=1, whenOtherwise, a_hl=0；Its element definition isWherein c_hl=1, whenOtherwise c_hl=0；AndIt is respectively 1 rank and 3 jump sub matrixs；Set for all real number compositions；G is unknown number to be solved；α and α₁Respectively punish parameter；

Sequence of calculation x and the difference of its average Mx, i.e. s₁:=x-Mx, s₁First intrinsic mode function for sequence x；

By intrinsic mode function s₁Separate from sequence x, obtain surplus r₁=x-s₁；Judge r₁Whether it is monotonic function, if It is to terminate decomposition method, otherwise, by r₁Replace the step before sequence x repeats to obtain next intrinsic mode letter Number, finally gives all eigenfunctions of this sequence(k=1,2 ..., L)；L is intrinsic mode letter The number of number.

Method the most according to claim 1, it is characterised in that described intrinsic mode function each described is predicted Include obtaining multiple step predicted the outcome:

Use following Holt-Winters exponential smoothing model to described intrinsic mode function It is predicted:

$s_{k,\stackrel{\&OverBar;}{N}+m}=(u_{\stackrel{\&OverBar;}{N}}+{\mathrm{mb}}_{\stackrel{\&OverBar;}{N}})s_{k,\stackrel{\&OverBar;}{N}-K+m},m=1,2,...$

Wherein: k represents kth intrinsic mode function, m is the time at intervals number of moment distance present moment to be predicted；

$\left\{\begin{array}{c}{u}_{\stackrel{\&OverBar;}{N}}=\mathrm{\&eta;}\frac{s_{k,\stackrel{\&OverBar;}{N}}}{F_{\stackrel{\&OverBar;}{N}-K}}+(1-\mathrm{\&eta;})(u_{\stackrel{\&OverBar;}{N}-1}+b_{\stackrel{\&OverBar;}{N}-1})\\ b_{\stackrel{\&OverBar;}{N}}=\mathrm{\&beta;}(u_{\stackrel{\&OverBar;}{N}}-u_{\stackrel{\&OverBar;}{N}-1})+(1-\mathrm{\&beta;})b_{\stackrel{\&OverBar;}{N}-1}\\ F_{\stackrel{\&OverBar;}{N}}=\mathrm{\&gamma;}\frac{s_{k,\stackrel{\&OverBar;}{N}}}{u_{\stackrel{\&OverBar;}{N}}}+(1-\mathrm{\&gamma;})F_{\stackrel{\&OverBar;}{N}-K}\end{array}\right.$

Wherein:ForThe periodic term in moment, is the time series exponential smoothing average removing seasonal variations impact；For The trend term in moment, is the exponential smoothing average of time series variation trend；ForThe item in season in moment, is seasonal factor Exponential smoothing average；ForThe actual value in moment；K is length or time cycle in season；η, β, γ represent smooth respectively Coefficient, value between (0,1).

Method the most according to claim 1, it is characterised in that described multiple steps collected that predict the outcome are included:

Predicting the outcome of all intrinsic mode functions is collected according to equation below, it may be assumed that

$\hat{x}(\stackrel{\&OverBar;}{N}+m)=\underset{k=1}{\overset{L}{\&Sigma;}}{s}_{k,\stackrel{\&OverBar;}{N}+m}$

Wherein,Represent that this value is prediction quantity in stock,For current date, m represents distance current time interval.

5. the prediction means of a category quantity in stock, it is characterised in that including:

Acquisition module, for obtaining category quantity in stock sequence from data platform, and is standardized this sequence processing；

Decomposing module, for being decomposed into several intrinsic mode functions with the average decomposition method of optimization by the sequence after standardization；

Prediction module, is predicted obtaining multiple predicting the outcome to intrinsic mode function each described；

Summarizing module, for the plurality of predicting the outcome being collected, thus obtains predicting the outcome of described category quantity in stock.

Device the most according to claim 5, it is characterised in that described decomposing module is additionally operable to:

Find out the category quantity in stock sequence after standardizationAll Local Extremum And by these extreme points by intervalIt is divided into:

${\{B_j^x:=\&lsqb;a_j^x,a_{j+1}^x\&rsqb;\}}_{j=1}^{m^x}$

Wherein, i is the time point of sequence x, and span is m^x+ 1 is the total number of extreme point of sequence x；The total number of time point for sequence；x_iFor time point i institute in category quantity in stock sequence Corresponding category quantity in stock；

Average Mx of sequence x is obtained by solving following minimum optimization problem:

$\left\{\begin{array}{c}Mx:=\underset{g}{min}\mathrm{\&alpha;}\left|\right|D^1(x-g)|{|}_1+{\mathrm{\&alpha;}}_1|\left|D^{1}g\right|{|}_1+(1-{\mathrm{\&alpha;}}_1)\left|\right|D^{3}g|{|}_1\\ \begin{array}{cc}s.t.& A(x-g)=0\\ & Cg=b,\end{array}\end{array}\right.$

Wherein,Represent segmentation summation operator, its element a_hl=1, whenOtherwise, a_hl=0；Its element definition isWherein c_hl=1, whenOtherwise c_hl=0；AndIt is respectively 1 rank and 3 jump sub matrixs；Set for all real number compositions；G is unknown number to be solved；α and α₁Respectively punish parameter；

Sequence of calculation x and the difference of its average Mx, i.e. s₁=x-Mx, s₁First intrinsic mode function for sequence x；

By intrinsic mode function s₁Separate from sequence x, obtain surplus r₁=x-s₁；Judge r₁Whether it is monotonic function, if It is to terminate decomposition method, otherwise, by r₁Replace the step before sequence x repeats to obtain next intrinsic mode letter Number, finally gives all eigenfunctions of this sequence(k=1,2 ..., L)；L is intrinsic mode letter The number of number.

Device the most according to claim 5, it is characterised in that described prediction module is additionally operable to:

Use following Holt-Winters exponential smoothing model to described intrinsic mode function It is predicted:

$s_{k,\stackrel{\&OverBar;}{N}+m}=(u_{\stackrel{\&OverBar;}{N}}+{\mathrm{mb}}_{\stackrel{\&OverBar;}{N}})s_{k,\stackrel{\&OverBar;}{N}-K+m},m=1,2,...$

Wherein: k represents kth intrinsic mode function, m is the time at intervals number of moment distance present moment to be predicted；

$\left\{\begin{array}{c}{u}_{\stackrel{\&OverBar;}{N}}=\mathrm{\&eta;}\frac{s_{k,\stackrel{\&OverBar;}{N}}}{F_{\stackrel{\&OverBar;}{N}-K}}+(1-\mathrm{\&eta;})(u_{\stackrel{\&OverBar;}{N}-1}+b_{\stackrel{\&OverBar;}{N}-1})\\ b_{\stackrel{\&OverBar;}{N}}=\mathrm{\&beta;}(u_{\stackrel{\&OverBar;}{N}}-u_{\stackrel{\&OverBar;}{N}-1})+(1-\mathrm{\&beta;})b_{\stackrel{\&OverBar;}{N}-1}\\ F_{\stackrel{\&OverBar;}{N}}=\mathrm{\&gamma;}\frac{s_{k,\stackrel{\&OverBar;}{N}}}{u_{\stackrel{\&OverBar;}{N}}}+(1-\mathrm{\&gamma;})F_{\stackrel{\&OverBar;}{N}-K}\end{array}\right.$

Wherein:ForThe periodic term in moment, is the time series exponential smoothing average removing seasonal variations impact；For The trend term in moment, is the exponential smoothing average of time series variation trend；ForThe item in season in moment, is seasonal factor Exponential smoothing average；ForThe actual value in moment；K is length or time cycle in season；η, β, γ represent smooth respectively Coefficient, value between (0,1).

Device the most according to claim 5, it is characterised in that described summarizing module is additionally operable to:

Predicting the outcome of all intrinsic mode functions is collected according to equation below, it may be assumed that

$\hat{x}(\stackrel{\&OverBar;}{N}+m)=\underset{k=1}{\overset{L}{\&Sigma;}}{s}_{k,\stackrel{\&OverBar;}{N}+m}$

Wherein,Represent that this value is prediction quantity in stock,For current date, m represents that distance is when the time interval of time.