CN103747477A - Network flow analysis and prediction method and device - Google Patents

Abstract

The invention discloses a network flow analysis and prediction method and device. The method comprises the steps of extracting the overall features of flow time sequences of each substation to be measured; clustering according to the extracted overall features; collecting the attribute features of the flow data according to the clustering result; finally performing flow prediction according o the attribute features of the flow data and the flow at the last moment. According to the network flow analysis and prediction method and device, the overall features of the time sequences are extracted, the similarity of the time sequences is reflected by using the similarity of the overall features, the dynamic features of time sequences changed along with time are mastered to obtain a more reasonable result, and meanwhile, the large-sized time sequence is described by less features, thus improving the robustness for judging the similarity result and reducing the complexity in the clustering operation process. Various attribute features related to the flow data are collected according to the clustering result, the flow data are predicted by the flow and the attribute features together, the prediction information amount is large, the prediction precision is correspondingly improved and the reasonable resource configuring is performed on a network.

Description

Network traffic analysis and Forecasting Methodology and device

Technical field

The present invention relates to communication technical field, particularly relate to a kind of network traffic analysis and Forecasting Methodology and device.

Background technology

In communication network optimize, network traffic analysis and prediction are very important links, significant to distributing rationally of Internet resources.Accurately whether volume forecasting, whether the interpretation predicting the outcome and predicting the outcome conforms to actual flow data, all directly affect investment and the construction scale of network, and to the preliminary analysis of flow, be the key of volume forecasting, directly affect the accuracy of volume forecasting.

In prior art, utilize original time series to analyze flow, adopt the similitude between Euclidean distance measuring period sequence, then according to this similitude, carry out cluster; Meanwhile, during predicted flow rate, by historical data on flows, predict unknown flow rate data, adopt traditional Regression Forecast, time series analysis etc.

Existing method is only paid attention to the difference of time series value on corresponding time point; Adopt the similitude between euclidean distance metric time series, thereby cause result to be vulnerable to the impact of value on indivedual time points, lost the robustness of result; Only utilize data on flows, thereby caused the result poor-performing of prediction.

Summary of the invention

Based on above-mentioned situation, the present invention proposes a kind of network traffic analysis and Forecasting Methodology, can improve precision of prediction, network is carried out to rational resource distribution.

To achieve these goals, technical scheme of the present invention is:

A kind of network traffic analysis and Forecasting Methodology, comprise the following steps:

Extract the flow seasonal effect in time series global characteristics of each base station to be measured;

According to extracted global characteristics, carry out cluster;

According to the result of institute's cluster, gather the attributive character of data on flows;

According to the attributive character of described data on flows and the flow in a upper moment, carried out volume forecasting.

For prior art problem, the invention allows for a kind of network traffic analysis and prediction unit, improve existing flow analysis robustness poor, the problem that volume forecasting precision is low, is applicable to practical application.

Specific implementation is: a kind of network traffic analysis and prediction unit, comprising:

Extraction module, for extracting the flow seasonal effect in time series global characteristics of each base station to be measured;

Cluster module, for carrying out cluster according to extracted global characteristics;

Acquisition module, for according to the result of institute's cluster, gathers the attributive character of data on flows;

Prediction module, for according to the attributive character of described data on flows and the flow in a upper moment, carried out volume forecasting.

Compared with prior art, beneficial effect of the present invention is: network traffic analysis of the present invention and Forecasting Methodology and device, first extract the flow seasonal effect in time series global characteristics of each base station to be measured; Then according to extracted global characteristics, carry out cluster; According to the result of institute's cluster, gather the attributive character of data on flows again; Last according to the attributive character of described data on flows and the flow in a upper moment, carry out volume forecasting.Use after technology of the present invention, extraction time sequence global characteristics, by global characteristics similitude, carry out the similitude of reflecting time sequence, catch the time dependent behavioral characteristics of time series, obtain more rational result, by describing large-scale time series by a small amount of feature, improve the robustness of judging analog result simultaneously, reduce the complexity in cluster calculation process; The various attributive character relevant to data on flows according to cluster result collection, according to flow and the common predicted flow rate data of attributive character, containing much information of prediction, has correspondingly improved precision of prediction, and network is carried out to rational resource distribution.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of network traffic analysis and Forecasting Methodology in an embodiment;

Fig. 2 is the structural representation of network traffic analysis and prediction unit in an embodiment.

Embodiment

For making object of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is described in further detail.Should be appreciated that embodiment described herein, only in order to explain the present invention, does not limit protection scope of the present invention.

Network traffic analysis and Forecasting Methodology in an embodiment, as shown in Figure 1, described method comprises:

Step S101: the flow seasonal effect in time series global characteristics that extracts each base station to be measured;

Step S102: carry out cluster according to extracted global characteristics;

Step S103: according to the result of institute's cluster, gather the attributive character of data on flows;

Step S104: according to the attributive character of described data on flows and the flow in a upper moment, carried out volume forecasting.

Known from the above description, this method, according to flow and the common predicted flow rate data of attributive character, improves predicting network flow precision, and network is carried out to rational resource distribution.

As an embodiment, described global characteristics comprises any one or multinomial in tendency feature or seasonal characteristics or kurtosis feature or degree of bias feature or auto-correlation coefficient feature or nonlinear characteristic or spectrum signature.

As an embodiment, described flow time series is by gathering a data on flows for each base station to be measured by sky, and continuous acquisition obtains half a year.

As an embodiment, described tendency feature is weighed by Z statistic, and Z statistic is greater than zero, is ascendant trend; Z statistic is less than zero, is downward trend; The computing formula of Z statistic is: $Z=\left\{\begin{array}{cc}\frac{S-1}{\sqrt{\mathrm{Var}\left(S\right)}}& S>0\\ 0& S=0,\\ \frac{S+1}{\sqrt{\mathrm{Var}\left(S\right)}}& S<0\end{array}\right.$ The statistic that wherein S is Normal Distribution, the variance that Var (S) is S, the computing formula of S is: the computing formula of Var (S) is: Var (S)=T (T-1) (2T+5)/18; Flow time series x _t, t=1,2 ... T, T is flow seasonal effect in time series length, x _jfor flow time series is at the value in j moment, x _kfor flow time series is in the value in k moment, sign function sgn (x _j-x _k) computing formula be: $\mathrm{sgn}(x_j-x_k)=\left\{\begin{array}{cc}+1& (x_j-x_k)>0\\ 0& (x_j-x_k)=0;\\ -1& (x_j-x_k)<0\end{array}\right.$

Described seasonal characteristics is by reflecting average period, and the calculation procedure of average period is: to flow time series x _tcarry out fast fourier transform, i.e. FFT conversion, t=1,2 ... T, T is flow seasonal effect in time series length, obtains: $F\left(k\right)=\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{x}_t{e}^{-j2\mathrm{\&pi;}(k-1)\frac{t-1}{T}},$ The frequency of wherein using is: $f_t=2\mathrm{\&pi;}\frac{t-1}{T};$ Further calculating average frequency is: $f_{\mathrm{tm}}=\frac{\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{f}_t}{T}=\frac{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)}{T^2},$ Calculating average period is: $T_{\mathrm{tm}}=\frac{1}{f_{\mathrm{tm}}}=\frac{T^2}{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)};$

In described kurtosis feature, the computing formula of kurtosis is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length,

for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

In described degree of bias feature, the computing formula of the degree of bias is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length,

for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

Described auto-correlation coefficient feature is weighed by Ljung-Box-Q statistic, and Ljung-Box-Q statistic detects whether flow time series is white-noise process, and the computing formula of Ljung-Box-Q statistic is:

wherein T is flow seasonal effect in time series length, and p is the maximum hysteresis exponent number being considered, and τ is hysteresis issue, r _τfor flow seasonal effect in time series auto-correlation coefficient; r _τcomputing formula be:

wherein x _tfor flow time series, t=1,2 ... T,

for flow seasonal effect in time series average;

Described nonlinear characteristic reflects by BDS test statistics, and BDS test statistics detects whether flow time series is independent same distribution, for flow time series x _t, t=1,2 ... T, moment s, the observed value of w is x _sand x _w, all observed value (x _s, x _w) by being configured to:

{ (x _s, x _w), (x _s+1, x _w+1), (x _s+2, x _w+2) ... (x _s+m-1, x _w+m-1), wherein m embeds interval; The computing formula of BDS statistic is: $W(N,m,r)=\sqrt{N}\frac{C(N,m,r)-C{(N,1,r)}^m}{{\mathrm{\&sigma;}}^{\&prime;}(N,m,r)},$ Wherein r is interval size, and C (N, m, r) is correlation intergal, σ ' (N, m, r) be C (N, m, r)-C (N, 1, r) ^mthe estimation of progressive standard deviation; C (N, m, r) computing formula is: $C(N,m,r)=\frac{2}{N(N-1)}\underset{w<s}{\mathrm{\&Sigma;}}H(r-||x_w^m-x_s^m|\left|\right),$ Wherein $x_w^m,x_s^m$ All m dimensional vectors, $x_w^m=(x_{w-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_w),x_s^m=(x_{s-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_s)H(r-||x_t^m-x_s^m|\left|\right)=\left\{\begin{array}{cc}0& r-\left|\right|x_t^m-x_s^m\left|\right|\&le;0\\ 1& r-\left|\right|x_t^m-x_s^m\left|\right|>0\end{array}\right.;$

Described spectrum signature is the front second order coefficient of the discrete Fourier transform (DFT) of extraction, the extraction of spectrum signature adopts discrete Fourier transform (DFT) coefficient, can extract front n rank coefficient as spectrum signature, because the HFS of a signal is unimportant, therefore most of concentration of energy in frequency domain space is on front several coefficients.

As an embodiment, described tendency characteristic use linear trend method obtains, and adopts linear trend method to isolate seasonal effect in time series trend components, and by the slope term of linear function as this seasonal effect in time series trend feature, settling time sequence x _t, t=1,2 ... T is about the regression model of time t, x _t=alpha+beta _t+ ε _t, wherein α is intercept, and β is slope, and ε is error, and the least-squares estimation of β is:

wherein

t represents flow seasonal effect in time series length;

Described seasonal characteristics utilizes H-P filter method to obtain, by computational minimization time series x _twith Trend value y _tbetween difference estimate trend components: $\underset{t=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;T}{\min }\{\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{(x_t-y_t)}^2+\mathrm{\&lambda;}{[(y_{t+1}-y_t)-(y_t-y_{t-1})]}^2\},$ Wherein, T is flow seasonal effect in time series length, and λ is the penalty factor to trend components fluctuation, can obtain thus periodic component: wherein, L is hysteresis operator, works as C _tthere is obvious peak value, can judge time series x _thave cyclic swing composition, the corresponding cycle of peak value is this seasonal effect in time series Cycle Length;

Described nonlinear characteristic adopts McLeod-Li-check or Bispectral check or RESET check or F checks or Neural Network Based Nonlinear test statistics reflects.

Do not get rid of other method in addition and can obtain above-mentioned global characteristics.

As an embodiment, described cluster comprises Kmeans cluster, and using extracted global characteristics as new characteristic vector, the corresponding new characteristic vector of the flow time series of each base station to be measured, carries out K-means cluster to new characteristic vector.

As an embodiment, described cluster comprises FCM cluster, and using extracted global characteristics as new characteristic vector, the corresponding new characteristic vector of the flow time series of each base station to be measured, carries out FCM cluster to new characteristic vector.

Do not get rid of other clustering method in addition.

In order to understand better this method, below elaborate the application example an of this method:

A, by day gather each prediction base station flow time series { x _t, t=1,2 ... T}, continuous acquisition half a year;

B, extract the flow seasonal effect in time series global characteristics of each base station, comprise tendency feature, seasonal characteristics, kurtosis feature, degree of bias feature, auto-correlation coefficient feature, nonlinear characteristic and spectrum signature;

C, using the global characteristics of each base station extracting as new characteristic vector, now the corresponding new characteristic vector of the flow time series of each base station, carries out cluster to new characteristic vector application K-means clustering method;

D, to each the class base station data attribute suitable according to its feature selecting after cluster, if data on flows presents tendency feature, gather the ARPU value relevant to data on flows, 3G permeability; If data on flows presents periodically, gather the ARPU value relevant to data on flows, 3G permeability, total number of users;

E, set up one and there is the BP neural network structure that three-decker, transfer function are tansig and train;

F, the model that previous step is trained, attributive character and the flow in a upper moment of the data on flows that will predict that input gathers, calculate the flow that will predict, for example, input the attributive character of data on flows and the flow of yesterday of the today gathering, can dope the flow of today.

Wherein, in step B, extract flow seasonal effect in time series global characteristics, extract by the following method:

B1, described tendency feature are weighed by Z statistic, and the computing formula of Z statistic is: $Z=\left\{\begin{array}{cc}\frac{S-1}{\sqrt{\mathrm{Var}\left(S\right)}}& S>0\\ 0& S=0,\\ \frac{S+1}{\sqrt{\mathrm{Var}\left(S\right)}}& S<0\end{array}\right.$ The statistic that wherein S is Normal Distribution, the variance that Var (S) is S, the computing formula of S is:

the computing formula of Var (S) is: Var (S)=T (T-1) (2T+5)/18; Flow time series x _t, t=1,2 ... T, T is flow seasonal effect in time series length, x _jfor flow time series is at the value in j moment, x _kfor flow time series is in the value in k moment, sign function sgn (x _j-x _k) computing formula be: $\mathrm{sgn}(x_j-x_k)=\left\{\begin{array}{cc}+1& (x_j-x_k)>0\\ 0& (x_j-x_k)=0;\\ -1& (x_j-x_k)<0\end{array}\right.$

B2, described seasonal characteristics are by reflecting average period, and the calculation procedure of average period is: to flow time series x _tcarry out fast fourier transform, i.e. FFT conversion, t=1,2 ... T, T is flow seasonal effect in time series length, obtains: $F\left(k\right)=\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{x}_t{e}^{-j2\mathrm{\&pi;}(k-1)\frac{t-1}{T}},$ The frequency of wherein using is: $f_t=2\mathrm{\&pi;}\frac{t-1}{T};$ Further calculating average frequency is: $f_{\mathrm{tm}}=\frac{\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{f}_t}{T}=\frac{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)}{T^2},$ Calculating average period is: $T_{\mathrm{tm}}=\frac{1}{f_{\mathrm{tm}}}=\frac{T^2}{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)};$

In B3, described kurtosis feature, the computing formula of kurtosis is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length,

for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

In B4, described degree of bias feature, the computing formula of the degree of bias is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length,

for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

B5, described auto-correlation coefficient feature are weighed by Ljung-Box-Q statistic, and the computing formula of Ljung-Box-Q statistic is:

wherein T is flow seasonal effect in time series length, and p is the maximum hysteresis exponent number being considered, and τ is hysteresis issue, r _τfor flow seasonal effect in time series auto-correlation coefficient; r _τcomputing formula be:

wherein x _tfor flow time series, t=1,2 ... T,

for flow seasonal effect in time series average;

B6, described nonlinear characteristic reflect by BDS test statistics, for flow time series x _t, t=1,2 ... T, moment s, the observed value of w is x _sand x _w, all observed value (x _s, x _w) by being configured to: { (x _s, x _w), (x _s+1, x _w+1), (x _s+2, x _w+2) ... (x _s+m-1, x _w+m-1), wherein m embeds interval; The computing formula of BDS statistic is: $W(N,m,r)=\sqrt{N}\frac{C(N,m,r)-C{(N,1,r)}^m}{{\mathrm{\&sigma;}}^{\&prime;}(N,m,r)},$ Wherein r is interval size, and C (N, m, r) is correlation intergal, σ ' (N, m, r) be C (N, m, r)-C (N, 1, r) ^mthe estimation of progressive standard deviation; C (N, m, r) computing formula is: $C(N,m,r)=\frac{2}{N(N-1)}\underset{w<s}{\mathrm{\&Sigma;}}H(r-||x_w^m-x_s^m|\left|\right),$ Wherein $x_w^m,x_s^m$ All m dimensional vectors, $x_w^m=(x_{w-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_w),x_s^m=(x_{s-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_s)H(r-||x_t^m-x_s^m|\left|\right)=\left\{\begin{array}{cc}0& r-\left|\right|x_t^m-x_s^m\left|\right|\&le;0\\ 1& r-\left|\right|x_t^m-x_s^m\left|\right|>0\end{array}\right.;$

B7, described spectrum signature are the front second order coefficient of the discrete Fourier transform (DFT) of extraction.

Network traffic analysis and prediction unit in an embodiment, as shown in Figure 2, described device comprises:

Extraction module, for extracting the flow seasonal effect in time series global characteristics of each base station to be measured;

Cluster module, for carrying out cluster according to extracted global characteristics;

Acquisition module, for according to the result of institute's cluster, gathers the attributive character of data on flows;

Prediction module, for according to the attributive character of described data on flows and the flow in a upper moment, carried out volume forecasting.

As shown in Figure 2, this preferred embodiment of installing each module annexation is: extraction module, cluster module, acquisition module and prediction module are linked in sequence successively.

First extraction module extracts the flow seasonal effect in time series global characteristics of each base station to be measured; Then cluster module is carried out cluster according to extracted global characteristics; Again by acquisition module according to the result of institute's cluster, gather the attributive character of data on flows; Last prediction module, the flow input neural network structure in the attributive character of described data on flows and a upper moment, was carried out volume forecasting, and the flow analysis of this plant network is more reasonable, the containing much information of prediction, and precision is high, is applicable to applying.

As an embodiment, described global characteristics comprises any one or multinomial in tendency feature or seasonal characteristics or kurtosis feature or degree of bias feature or auto-correlation coefficient feature or nonlinear characteristic or spectrum signature.

As an embodiment, described flow time series is by gathering a data on flows for each base station to be measured by sky, and continuous acquisition obtains half a year.

As an embodiment, described tendency feature is weighed by Z statistic, and Z statistic is greater than zero, is ascendant trend; Z statistic is less than zero, is downward trend; The computing formula of Z statistic is: $Z=\left\{\begin{array}{cc}\frac{S-1}{\sqrt{\mathrm{Var}\left(S\right)}}& S>0\\ 0& S=0,\\ \frac{S+1}{\sqrt{\mathrm{Var}\left(S\right)}}& S<0\end{array}\right.$ The statistic that wherein S is Normal Distribution, the variance that Var (S) is S, the computing formula of S is: the computing formula of Var (S) is: Var (S)=T (T-1) (2T+5)/18; Flow time series x _t, t=1,2 ... T, T is flow seasonal effect in time series length, x _jfor flow time series is at the value in j moment, x _kfor flow time series is in the value in k moment, sign function sgn (x _j-x _k) computing formula be: $\mathrm{sgn}(x_j-x_k)=\left\{\begin{array}{cc}+1& (x_j-x_k)>0\\ 0& (x_j-x_k)=0;\\ -1& (x_j-x_k)<0\end{array}\right.$

Described seasonal characteristics is by reflecting average period, and the calculation procedure of average period is: to flow time series x _tcarry out fast fourier transform, i.e. FFT conversion, t=1,2 ... T, T is flow seasonal effect in time series length, obtains: $F\left(k\right)=\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{x}_t{e}^{-j2\mathrm{\&pi;}(k-1)\frac{t-1}{T}},$ The frequency of wherein using is: $f_t=2\mathrm{\&pi;}\frac{t-1}{T};$ Further calculating average frequency is: $f_{\mathrm{tm}}=\frac{\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{f}_t}{T}=\frac{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)}{T^2},$ Calculating average period is: $T_{\mathrm{tm}}=\frac{1}{f_{\mathrm{tm}}}=\frac{T^2}{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)};$

In described kurtosis feature, the computing formula of kurtosis is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length,

for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

In described degree of bias feature, the computing formula of the degree of bias is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length,

for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

Described auto-correlation coefficient feature is weighed by Ljung-Box-Q statistic, and Ljung-Box-Q statistic detects whether flow time series is white-noise process, and the computing formula of Ljung-Box-Q statistic is:

wherein T is flow seasonal effect in time series length, and p is the maximum hysteresis exponent number being considered, and τ is hysteresis issue, r _τfor flow seasonal effect in time series auto-correlation coefficient; r _τcomputing formula be:

wherein xt is flow time series, t=1, and 2 ... T,

for flow seasonal effect in time series average;

Described nonlinear characteristic reflects by BDS test statistics, and BDS test statistics detects whether flow time series is independent same distribution, for flow time series x _t, t=1,2 ... T, moment s, the observed value of w is x _sand x _w, all observed value (x _s, x _w) by being configured to:

{ (x _s, x _w), (x _s+1, x _w+1), (x _s+2, x _w+2) ... (x _s+m-1, x _w+m-1), wherein m embeds interval; The computing formula of BDS statistic is: $W(N,m,r)=\sqrt{N}\frac{C(N,m,r)-C{(N,1,r)}^m}{{\mathrm{\&sigma;}}^{\&prime;}(N,m,r)},$ Wherein r is interval size, and C (N, m, r) is correlation intergal, σ ' (N, m, r) be C (N, m, r)-C (N, 1, r) ^mthe estimation of progressive standard deviation; C (N, m, r) computing formula is: $C(N,m,r)=\frac{2}{N(N-1)}\underset{w<s}{\mathrm{\&Sigma;}}H(r-||x_w^m-x_s^m|\left|\right),$ Wherein $x_w^m,x_s^m$ All m dimensional vectors, $x_w^m=(x_{w-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_w),x_s^m=(x_{s-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_s)H(r-||x_t^m-x_s^m|\left|\right)=\left\{\begin{array}{cc}0& r-\left|\right|x_t^m-x_s^m\left|\right|\&le;0\\ 1& r-\left|\right|x_t^m-x_s^m\left|\right|>0\end{array}\right.;$

Described spectrum signature is the front second order coefficient of the discrete Fourier transform (DFT) of extraction, the extraction of spectrum signature adopts discrete Fourier transform (DFT) coefficient, can extract front n rank coefficient as spectrum signature, because the HFS of a signal is unimportant, therefore most of concentration of energy in frequency domain space is on front several coefficients.

As an embodiment, described tendency characteristic use linear trend method obtains, and adopts linear trend method to isolate seasonal effect in time series trend components, and by the slope term of linear function as this seasonal effect in time series trend feature, settling time sequence x _t, t=1,2 ... T is about the regression model of time t, x _t=alpha+beta _t+ ε _t, wherein α is intercept, and β is slope, and ε is error, and the least-squares estimation of β is:

wherein t represents flow seasonal effect in time series length;

Described seasonal characteristics utilizes H-P filter method to obtain, by computational minimization time series x _twith Trend value y _tbetween difference estimate trend components: $\underset{t=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;T}{\min }\{\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{(x_t-y_t)}^2+\mathrm{\&lambda;}{[(y_{t+1}-y_t)-(y_t-y_{t-1})]}^2\},$ Wherein, T is flow seasonal effect in time series length, and λ is the penalty factor to trend components fluctuation, can obtain thus periodic component:

wherein, L is hysteresis operator, works as C _tthere is obvious peak value, can judge time series x _thave cyclic swing composition, the corresponding cycle of peak value is this seasonal effect in time series Cycle Length;

Described nonlinear characteristic adopts McLeod-Li-check or Bispectral check or RESET check or F checks or Neural Network Based Nonlinear test statistics reflects.

Do not get rid of other method in addition and can obtain above-mentioned global characteristics.

As an embodiment, described cluster comprises Kmeans cluster, and using extracted global characteristics as new characteristic vector, the corresponding new characteristic vector of the flow time series of each base station to be measured, carries out K-means cluster to new characteristic vector.

As an embodiment, described cluster comprises FCM cluster, and using extracted global characteristics as new characteristic vector, the corresponding new characteristic vector of the flow time series of each base station to be measured, carries out FCM cluster to new characteristic vector.

Do not get rid of other clustering method in addition.

The above embodiment has only expressed several execution mode of the present invention, and it describes comparatively concrete and detailed, but can not therefore be 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, can also make some distortion and improvement, these all belong to protection scope of the present invention.Therefore, the protection range of patent of the present invention should be as the criterion with claims.

Claims (10)

1. network traffic analysis and a Forecasting Methodology, is characterized in that, comprises the following steps:

Extract the flow seasonal effect in time series global characteristics of each base station to be measured;

According to extracted global characteristics, carry out cluster;

According to the result of institute's cluster, gather the attributive character of data on flows;

According to the attributive character of described data on flows and the flow in a upper moment, carried out volume forecasting.

2. network traffic analysis according to claim 1 and Forecasting Methodology, it is characterized in that, described global characteristics comprises any one or multinomial in tendency feature or seasonal characteristics or kurtosis feature or degree of bias feature or auto-correlation coefficient feature or nonlinear characteristic or spectrum signature.

3. network traffic analysis according to claim 1 and Forecasting Methodology, is characterized in that, described flow time series is by gathering a data on flows for each base station to be measured by sky, and continuous acquisition obtains half a year.

4. network traffic analysis according to claim 2 and Forecasting Methodology, is characterized in that, described tendency feature is weighed by Z statistic, and the computing formula of Z statistic is: $Z=\left\{\begin{array}{cc}\frac{S-1}{\sqrt{\mathrm{Var}\left(S\right)}}& S>0\\ 0& S=0,\\ \frac{S+1}{\sqrt{\mathrm{Var}\left(S\right)}}& S<0\end{array}\right.$ The statistic that wherein S is Normal Distribution, the variance that Var (S) is S, the computing formula of S is:

the computing formula of Var (S) is: Var (S)=T (T-1) (2T+5)/18; Flow time series x _t, t=1,2 ... T, T is flow seasonal effect in time series length, x _jfor flow time series is at the value in j moment, x _kfor flow time series is in the value in k moment, sign function sgn (x _j-x _k) computing formula be: $\mathrm{sgn}(x_j-x_k)=\left\{\begin{array}{cc}+1& (x_j-x_k)>0\\ 0& (x_j-x_k)=0;\\ -1& (x_j-x_k)<0\end{array}\right.$

Described seasonal characteristics is by reflecting average period, and the calculation procedure of average period is: to flow time series x _tcarry out fast fourier transform, i.e. FFT conversion, t=1,2 ... T, T is flow seasonal effect in time series length, obtains: $F\left(k\right)=\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{x}_t{e}^{-j2\mathrm{\&pi;}(k-1)\frac{t-1}{T}},$ The frequency of wherein using is: $f_t=2\mathrm{\&pi;}\frac{t-1}{T};$ Further calculating average frequency is: $f_{\mathrm{tm}}=\frac{\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{f}_t}{T}=\frac{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)}{T^2},$ Calculating average period is: $T_{\mathrm{tm}}=\frac{1}{f_{\mathrm{tm}}}=\frac{T^2}{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)};$

In described kurtosis feature, the computing formula of kurtosis is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length, for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

In described degree of bias feature, the computing formula of the degree of bias is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length,

for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

Described auto-correlation coefficient feature is weighed by Ljung-Box-Q statistic, and the computing formula of Ljung-Box-Q statistic is:

wherein T is flow seasonal effect in time series length, and p is the maximum hysteresis exponent number being considered, and τ is hysteresis issue, r _τfor flow seasonal effect in time series auto-correlation coefficient; r _τcomputing formula be:

wherein x _tfor flow time series, t=1,2 ... T,

for flow seasonal effect in time series average;

Described nonlinear characteristic reflects by BDS test statistics, for flow time series x _t, t=1,2 ... T, moment s, the observed value of w is x _sand x _w, all observed value (x _s, x _w) by being configured to: { (x _s, x _w), (x _s+1, x _w+1), (x _s+2, x _w+2) ... (x _s+m-1, x _w+m-1), wherein m embeds interval; The computing formula of BDS statistic is: $W(N,m,r)=\sqrt{N}\frac{C(N,m,r)-C{(N,1,r)}^m}{{\mathrm{\&sigma;}}^{\&prime;}(N,m,r)},$ Wherein r is interval size, and C (N, m, r) is correlation intergal, σ ' (N, m, r) be C (N, m, r)-C (N, 1, r) ^mthe estimation of progressive standard deviation; C (N, m, r) computing formula is: $C(N,m,r)=\frac{2}{N(N-1)}\underset{w<s}{\mathrm{\&Sigma;}}H(r-||x_w^m-x_s^m|\left|\right),$ Wherein $x_w^m,x_s^m$ All m dimensional vectors, $x_w^m=(x_{w-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_w),x_s^m=(x_{s-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_s)H(r-||x_t^m-x_s^m|\left|\right)=\left\{\begin{array}{cc}0& r-\left|\right|x_t^m-x_s^m\left|\right|\&le;0\\ 1& r-\left|\right|x_t^m-x_s^m\left|\right|>0\end{array}\right.;$

Described spectrum signature is the front second order coefficient of the discrete Fourier transform (DFT) of extraction.

5. network traffic analysis according to claim 1 and Forecasting Methodology, it is characterized in that, described cluster comprises Kmeans cluster, using extracted global characteristics as new characteristic vector, the corresponding new characteristic vector of flow time series of each base station to be measured, carries out K-means cluster to new characteristic vector.

6. network traffic analysis and a prediction unit, is characterized in that, comprising:

Extraction module, for extracting the flow seasonal effect in time series global characteristics of each base station to be measured;

Cluster module, for carrying out cluster according to extracted global characteristics;

Acquisition module, for according to the result of institute's cluster, gathers the attributive character of data on flows;

Prediction module, for according to the attributive character of described data on flows and the flow in a upper moment, carried out volume forecasting.

7. network traffic analysis according to claim 6 and prediction unit, it is characterized in that, described global characteristics comprises any one or multinomial in tendency feature or seasonal characteristics or kurtosis feature or degree of bias feature or auto-correlation coefficient feature or nonlinear characteristic or spectrum signature.

8. network traffic analysis according to claim 6 and prediction unit, is characterized in that, described flow time series is by gathering a data on flows for each base station to be measured by sky, and continuous acquisition obtains half a year.

9. network traffic analysis according to claim 7 and prediction unit, is characterized in that, described tendency feature is weighed by Z statistic, and the computing formula of Z statistic is: $Z=\left\{\begin{array}{cc}\frac{S-1}{\sqrt{\mathrm{Var}\left(S\right)}}& S>0\\ 0& S=0,\\ \frac{S+1}{\sqrt{\mathrm{Var}\left(S\right)}}& S<0\end{array}\right.$ The statistic that wherein S is Normal Distribution, the variance that Var (S) is S, the computing formula of S is:

the computing formula of Var (S) is: Var (S)=T (T-1) (2T+5)/18; Flow time series x _t, t=1,2 ... T, T is flow seasonal effect in time series length, x _jfor flow time series is at the value in j moment, x _kfor flow time series is in the value in k moment, sign function sgn (x _j-x _k) computing formula be: $\mathrm{sgn}(x_j-x_k)=\left\{\begin{array}{cc}+1& (x_j-x_k)>0\\ 0& (x_j-x_k)=0;\\ -1& (x_j-x_k)<0\end{array}\right.$

Described seasonal characteristics is by reflecting average period, and the calculation procedure of average period is: to flow time series x _tcarry out fast fourier transform, i.e. FFT conversion, t=1,2 ... T, T is flow seasonal effect in time series length, obtains: $F\left(k\right)=\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{x}_t{e}^{-j2\mathrm{\&pi;}(k-1)\frac{t-1}{T}},$ The frequency of wherein using is: $f_t=2\mathrm{\&pi;}\frac{t-1}{T};$ Further calculating average frequency is: $f_{\mathrm{tm}}=\frac{\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{f}_t}{T}=\frac{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)}{T^2},$ Calculating average period is: $T_{\mathrm{tm}}=\frac{1}{f_{\mathrm{tm}}}=\frac{T^2}{2\mathrm{\&pi;}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(t-1)};$

In described kurtosis feature, the computing formula of kurtosis is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length,

for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

In described degree of bias feature, the computing formula of the degree of bias is:

wherein x _tfor flow time series, t=1,2 ... T, T is flow seasonal effect in time series length,

for flow seasonal effect in time series average, σ is flow seasonal effect in time series sample standard deviation;

Described auto-correlation coefficient feature is weighed by Ljung-Box-Q statistic, and the computing formula of Ljung-Box-Q statistic is: wherein T is flow seasonal effect in time series length, and p is the maximum hysteresis exponent number being considered, and τ is hysteresis issue, r _τfor flow seasonal effect in time series auto-correlation coefficient; r _τcomputing formula be:

wherein x _tfor flow time series, t=1,2 ... T,

for flow seasonal effect in time series average;

Described nonlinear characteristic reflects by BDS test statistics, for flow time series x _t, t=1,2 ... T, moment s, the observed value of w is x _sand x _w, all observed value (x _s, x _w) by being configured to: { (x _s, x _w), (x _s+1, x _w+1), (x _s+2, x _w+2) ... (x _s+m-1, x _w+m-1), wherein m embeds interval; The computing formula of BDS statistic is: $W(N,m,r)=\sqrt{N}\frac{C(N,m,r)-C{(N,1,r)}^m}{{\mathrm{\&sigma;}}^{\&prime;}(N,m,r)},$ Wherein r is interval size, and C (N, m, r) is correlation intergal, σ ' (N, m, r) be C (N, m, r)-C (N, 1, r) ^mthe estimation of progressive standard deviation; C (N, m, r) computing formula is: $C(N,m,r)=\frac{2}{N(N-1)}\underset{w<s}{\mathrm{\&Sigma;}}H(r-||x_w^m-x_s^m|\left|\right),$ Wherein $x_w^m,x_s^m$ All m dimensional vectors, $x_w^m=(x_{w-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_w),x_s^m=(x_{s-m+1},\&CenterDot;\&CenterDot;\&CenterDot;,x_s)H(r-||x_t^m-x_s^m|\left|\right)=\left\{\begin{array}{cc}0& r-\left|\right|x_t^m-x_s^m\left|\right|\&le;0\\ 1& r-\left|\right|x_t^m-x_s^m\left|\right|>0\end{array}\right.;$

Described spectrum signature is the front second order coefficient of the discrete Fourier transform (DFT) of extraction.

10. network traffic analysis according to claim 6 and prediction unit, it is characterized in that, described cluster comprises Kmeans cluster, using extracted global characteristics as new characteristic vector, the corresponding new characteristic vector of flow time series of each base station to be measured, carries out K-means cluster to new characteristic vector.

Images