CN102360388A - Time series forecasting method and system based on SVR (Support Vector Regression) - Google Patents
Time series forecasting method and system based on SVR (Support Vector Regression) Download PDFInfo
- Publication number
- CN102360388A CN102360388A CN2011103204825A CN201110320482A CN102360388A CN 102360388 A CN102360388 A CN 102360388A CN 2011103204825 A CN2011103204825 A CN 2011103204825A CN 201110320482 A CN201110320482 A CN 201110320482A CN 102360388 A CN102360388 A CN 102360388A
- Authority
- CN
- China
- Prior art keywords
- svr model
- historical data
- svr
- difference
- time series
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention discloses a time series forecasting method and system based on SVR (Support Vector Regression). The time series forecasting method based on SVR comprises the following steps of: selecting historical data from an existing time series data set and obtaining a plurality of training data sets; determining regular parameters and Gaussian nuclear parameters of SVR modules to be constructed, and constructing an SVR module corresponding to each training data set; selecting T historical data between a moment of t-T+1 to a current moment t; and under the condition that a first difference of a moment to be forecasted and the current moment is less than or equal to the number of SVR modules, selecting the SVR module corresponding to the first difference, and obtaining a forecast value of the moment to be forecasted from the T historic data by directly utilizing the SVR module. Compared with a method for obtaining the forecast value of the moment to be forecasted through multi-step forecasting in the prior art, in the method for obtaining the forecast value through a one-step forecasting disclosed by the invention, the accumulation of forecast errors is reduced, and thus the accuracy of obtaining the forecast value is improved.
Description
Technical field
The application relates to the time series forecasting field, particularly relates to a kind of time series forecasting method and system based on support vector regression.
Background technology
Time series is an orderly observation data sequence, and it has embodied some statistical indicators distribution situation in time of certain phenomenon, like the power load data of somewhere 1 day to No. 15 October.The time series forecasting method then is to utilize acquired historical time arrangement set, analyzes the inherent The statistical properties and the rule of development of the historical data in the set, obtains the development trend of predicted data with video data.
Time series forecasting method commonly used at present is based on single SVR (Support Vector Regression; Support vector regression) time forecasting methods; This method is at first chosen historical data from existing time series data collection; Draw a training dataset; Wherein training dataset comprises a plurality of existing time series data subclass and the predicted value corresponding with each time series data subclass, and arbitrary time series data subclass is the historical data that time series data is concentrated with its corresponding predicted value; Secondly; Confirm the regular parameter and the gaussian kernel parameter of SVR model to be made up,, training dataset is trained according to said regular parameter and gaussian kernel parameter; Make up the SVR model; Choose t-d+1 afterwards constantly to the input data of d the historical data of current t between the moment as the SVR model, prediction t+1 data constantly, wherein d is a time delay; At last; The t+1 that dopes data are constantly joined in d the historical data, and, repeat above-mentioned steps until obtaining by prediction predicted value constantly simultaneously with the data deletion that obtains at first in d the historical data; Wherein, the data that obtain at first are acquisition time data the earliest.Promptly when the data in the prediction t+2 moment, will include t+1 d historical data constantly predicts as the input data of SVR model.That is to say when the data in the prediction t+i moment, will include t+i-1 d historical data constantly and predict that this prediction mode is called the step time forecasting methods based on single SVR model as the input data of SVR model.
Yet; All can there be a predicated error in step time forecasting methods based on single SVR model in the one-step prediction process; This predicated error is in the process that iterates of data prediction; Be that the multi-step prediction process can be accumulated along with the increase of prediction number of times, the accumulation of predicated error directly reduces the degree of accuracy of predicting the data that draw.
Summary of the invention
In view of this; The application embodiment discloses a kind of time series forecasting method and system based on support vector regression; Go on foot time forecasting methods in forecasting process to solve one of existing single SVR model; The increase of prediction number of times and accumulating, predicated error increases with the prediction number of times accumulates, and this accumulation directly reduces the problem of the degree of accuracy of the data that prediction draws.Technical scheme is following:
Based on the one side of the application embodiment, a kind of time series forecasting method based on support vector regression is disclosed, comprising:
Choose historical data from existing time series data is concentrated, draw a plurality of training datasets;
Confirm the regular parameter and the gaussian kernel parameter of SVR model to be made up,, respectively each training dataset is trained, make up self corresponding support vector regression SVR model according to said regular parameter and gaussian kernel parameter;
Choose t-T+1 constantly to T the historical data of current t between the moment, wherein, T is the number of said SVR model;
By prediction constantly with the situation of first difference smaller or equal to the number of said SVR model of current time under, choose the SVR model corresponding with said first difference, directly utilize this SVR model to obtain to be predicted the predicted value in the moment to a said T historical data.
Preferably, also comprise: by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under, utilize a plurality of SVR models to obtain to a said T historical data by the prediction predicted value in the moment.
Preferably; Saidly utilize a plurality of SVR models to obtain directly to be predicted that predicted value constantly comprises to T historical data: by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under; The T that utilization is chosen historical data distinguished corresponding SVR model, draws the predicted value of T historical data self;
The predicted value of the said T that a draws historical data self is updated to the t-T+1 that chooses constantly to T the historical data of current t between constantly, and with second difference of the number of said first difference and SVR model as the tested moment;
Whether the difference of judging this tested moment and current time is greater than the number of said SVR model; Under the situation of said difference smaller or equal to the number of said SVR model; The SVR model corresponding with said first difference chosen in execution; Directly utilize this SVR model to obtain to a said T historical data by the step of prediction predicted value constantly; Under the situation of said difference greater than the number of said SVR model, carry out and utilize the corresponding respectively SVR model of choosing of T historical data, draw the step of the predicted value of T historical data self.
Preferably, adopt cross validation method, confirm the regular parameter and the gaussian kernel parameter of SVR model to be made up.
Preferably, the time series data subclass number that comprises of different training data set is different.
Based on the application embodiment on the other hand, disclose a kind of time series forecasting system, it is characterized in that, comprising based on support vector regression:
Data set draws module, is used for choosing historical data from existing time series data is concentrated, draws a plurality of training datasets;
The SVR model construction module is used to confirm the regular parameter and the gaussian kernel parameter of SVR model to be made up, and according to said regular parameter and gaussian kernel parameter, respectively each training dataset is trained, and makes up the support vector regression SVR model of self correspondence;
The data decimation module is used to choose t-T+1 constantly to T the historical data of current t between the moment, and wherein, T is the number of said SVR model;
First prediction module; Be used for by prediction constantly with the situation of first difference smaller or equal to the number of said SVR model of current time under; Choose the SVR model corresponding, directly utilize this SVR model to obtain by prediction predicted value constantly to a said T historical data with said first difference.
Preferably, also comprise: second prediction module, be used for by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under, utilize a plurality of SVR models to obtain to a said T historical data by the prediction predicted value in the moment.
Preferably, said second prediction module comprises:
Predicted value draws the unit, be used for by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under, utilize T historical data the choosing SVR model of correspondence respectively, draw the predicted value of T historical data self;
Updating block is used for predicted value with the said T that a draws historical data self and is updated to the t-T+1 that chooses constantly to T the historical data of current t between constantly, and with second difference of the number of said first difference and SVR model as the tested moment;
Whether judging unit, the difference that is used to judge this tested moment and current time be greater than the number of said SVR model;
Trigger is used under the situation of said difference smaller or equal to the number of said SVR model, triggering said first prediction module, under the situation of said difference greater than the number of said SVR model, triggers said second prediction module.
Preferably, said SVR model construction module specifically adopts cross validation method, confirms the regular parameter and the gaussian kernel parameter of SVR model to be made up.
Preferably, to draw the time series data subclass number that different training data set that module draws comprises different for data set.
Use technique scheme; By prediction constantly with the situation of first difference smaller or equal to the number of SVR model of current time under; Choose the SVR model corresponding with difference; To T the historical data of the extremely current constantly t of the t-T+1 that chooses between the moment, directly utilize this SVR model to obtain by the prediction predicted value in the moment.The mode that the application's one-step prediction obtains predicted value obtains the mode that quilt is predicted predicted value constantly with respect to the prior art multi-step prediction, and the accumulation of predicated error reduces, and then obtains the degree of accuracy raising of predicted value.
Description of drawings
In order to be illustrated more clearly in the application embodiment or technical scheme of the prior art; To do to introduce simply to the accompanying drawing of required use in embodiment or the description of the Prior Art below; Obviously, the accompanying drawing in describing below only is some embodiment that put down in writing among the application, for those of ordinary skills; Under the prerequisite of not paying creative work, can also obtain other accompanying drawing according to these accompanying drawings.
Fig. 1 is a kind of process flow diagram of the disclosed time series forecasting method based on support vector regression of the application embodiment;
Fig. 2 is the another kind of process flow diagram of the disclosed time series forecasting method based on support vector regression of the application embodiment;
Fig. 3 is a kind of structural representation of the disclosed time series forecasting system based on support vector regression of the application embodiment;
Fig. 4 is the another kind of structural representation of the disclosed time series forecasting system based on support vector regression of the application embodiment;
Fig. 5 is the disclosed structural representation based on second prediction module in the time series forecasting system of support vector regression of the application embodiment.
Embodiment
For above-mentioned purpose, the feature and advantage that make the application can be more obviously understandable, the application is done further detailed explanation below in conjunction with accompanying drawing and embodiment.
An embodiment
See also Fig. 1, show the process flow diagram of a kind of online service request recognition methods embodiment 1 of the application, can may further comprise the steps:
S101: choose historical data from existing time series data is concentrated, draw a plurality of training datasets.
Wherein, each training dataset comprises a plurality of existing time series data subclass and the predicted value corresponding with each time series data subclass, and arbitrary time series data subclass is the historical data that said time series data is concentrated with its corresponding predicted value.
Suppose that it is x (k) that time series data is concentrated historical data, k=0,1 ..., t-1, t is a current time, and the number of the training dataset that draws is T, and j training dataset can be used S
jExpression, wherein
J training dataset S
jComprise t-d-j+1 time series data subclass u
k, u
k=[x (k-d+1) ... X (k-1) x (k)]
T∈ R
d, each time series data subclass u
kTo a predicted value y should be arranged
k, y
k=x (k+j) ∈ R, R representes set of real numbers, and arbitrary time series data subclass is the historical data that said time series data is concentrated with its corresponding predicted value, and d is a time delay, representes each time series data subclass u
kThe sample space dimension, d=T.
Above-mentioned a plurality of training data is concentrated, because the j value of different training data set is different, so the time series data subclass number that the different training data set comprises is different.
S102: confirm the regular parameter and the gaussian kernel parameter of SVR model to be made up,, respectively each training dataset is trained, make up self corresponding support vector regression SVR model according to said regular parameter and gaussian kernel parameter.
Regular parameter and the gaussian kernel parameter of confirming SVR model to be made up can adopt existing method, confirm regular parameter and gaussian kernel parameter like cross validation method.
Train the process that makes up with its corresponding SVR model identical for arbitrary training dataset, with j training dataset S with existing building process based on SVR model in the step time forecasting methods of single SVR model
j, promptly can utilize the time series data subclass u in it
kWith each time series data subclass u
kCorresponding y
kTrain, obtain j training dataset S
jJ corresponding SVR model, concrete training process is no longer set forth.
S103: choose t-T+1 constantly to T the historical data of current t between the moment.
Wherein, the T that a chooses historical data is respectively x (t-T+1) ..., x (t-1), x (t), it forms a time series data subclass u
t, u
t=[x (t-T+1) ... X (t-1) x (t)]
T
S104: judge by prediction constantly with first difference of current time whether greater than the number of SVR model, by prediction constantly with the situation of first difference smaller or equal to the number of said SVR model of current time under, execution in step S105; By prediction constantly with the situation of first difference greater than the number of said SVR model of current time under, execution in step S106.
Suppose that prediction step is τ, the then tested moment is t+ τ, and first difference of the tested moment with current time then is prediction step τ.The number of training dataset is T, and then the number of SVR model also is T.Therefore, judging that first difference whether during greater than the number of SVR model, promptly can judge through the size of judging τ and T.
S105: choose the SVR model corresponding, directly utilize this SVR model to obtain by prediction predicted value constantly to T historical data with first difference.
Wherein: choose the SVR model corresponding for choosing τ SVR model with first difference.Directly utilize τ SVR model to obtain to T historical data again by prediction predicted value constantly.
S106: utilize a plurality of SVR models to obtain by prediction predicted value constantly to T historical data.
Step S106 is under the situation of first difference greater than the number of said SVR model; Utilize a plurality of SVR models to obtain by prediction predicted value constantly; In this process, need upgrade T the historical data and the tested moment; Whether greater than the number of SVR model, and under different situations, carry out various process in the difference that rejudges the tested moment and current time, specifically see also Fig. 2.Fig. 2 is the another kind of process flow diagram of the disclosed time series forecasting method based on support vector regression of the application embodiment, and it has introduced step S106 in detail on Fig. 1 basis.
Step S106 can comprise the steps:
S1061: utilize the corresponding respectively SVR model of choosing of T historical data, draw the predicted value of T historical data self.
Wherein, T the corresponding respectively SVR model of historical data is: historical data x (t-T+1) is corresponding to training dataset S
1The 1st the SVR model that training draws, historical data x (t-T+2) is corresponding to training dataset S
2The 2nd the SVR model that training draws ..., by that analogy, historical data x (t) is corresponding to training dataset S
TT the SVR model that training draws.Each historical data uses self corresponding SVR model to predict respectively, draws predicted value.
Need to prove: this step historical data is when predicting, prediction step is different, and the prediction step of historical data x (t-T+1) is 1; The prediction step of historical data x (t-T+2) is 2; ..., by that analogy, the prediction step of historical data x (t) is T.The predicted value that T historical data draws can be respectively
S1062: the predicted value of the said T that a draws historical data self is updated to the t-T+1 that chooses constantly to T the historical data of current t between constantly, and with second difference of the number of said first difference and SVR model as the tested moment.
Above-mentioned mentioning, first difference are prediction step τ, and the then tested moment is τ-T.
S1063: whether the difference of judging this tested moment and current time greater than the number of said SVR model, under the situation of said difference smaller or equal to the number of said SVR model, and execution in step S105; Under the situation of said difference, return execution in step S1061 greater than the number of said SVR model.
Use technique scheme; By prediction constantly with the situation of first difference smaller or equal to the number of SVR model of current time under; Choose the SVR model corresponding with difference; To T the historical data of the extremely current constantly t of the t-T+1 that chooses between the moment, directly utilize this SVR model to obtain by the prediction predicted value in the moment.The mode that the application's one-step prediction obtains predicted value obtains the mode that quilt is predicted predicted value constantly with respect to the prior art multi-step prediction, and the accumulation of predicated error reduces, and then obtains the degree of accuracy raising of predicted value.
Further, by prediction constantly with the situation of first difference greater than the number of SVR model of current time under, T historical data uses the SVR model of self correspondence to predict respectively, draws T predicted value.Simultaneously T historical data is updated to T predicted value; With second difference of the number of first difference and SVR model as the tested moment; In the difference that rejudges the tested moment and current time whether greater than the number of said SVR model, until the number of difference smaller or equal to the SVR model.Though above-mentioned steps has caused the accumulation of predicated error, this predicated error accumulation can improve the degree of accuracy of predicted value equally less than the accumulation of predicated error in the step time forecasting methods of existing single SVR model.
The validity that provides through experiment comparing result checking the application embodiment below based on the time series forecasting method of support vector regression.
Emulated data produces from the Mackey-Glass sequential system; The Mackey-Glass sequential system is the non-linear differential system of a time delay; Its form is
wherein, δ=17.Produce 1000 historical datas, make time delay d=6, then T=6.Begin to obtain historical data from k=5, first time series data subclass is u
5=[x (0) ... X (4) x (5)]
T, the rest may be inferred, 6 training dataset S that the SVR model is corresponding
jSize be: 994,993,992,991,990 and 989.Get preceding 600 historical data composing training training set S respectively
j
Five step predictions, the prediction of ten steps and the prediction of 100 steps are carried out in experiment respectively, i.e. τ=5, τ=10 and τ=100.Experiment adopts dual mode to make up the SVR model, and a kind of is each time series data subclass u
kCorresponding predicted value y
kIn do not add noise, a kind of is predicted value y
kIn add white noise, signal to noise ratio (S/N ratio) is 3: 1.
The parameter of SVR model is set to: regular parameter C=10, the gaussian kernel parameter is 0.25.This experiment is trained respectively 6 SVR models, then different prediction step is calculated predicted value.In the present embodiment if τ=5, prediction step, promptly first difference is less than the number of SVR model, and therefore, this experiment can directly be adopted the 5th SVR model to predict and draw predicted value.And for τ=10 and τ=100, this experiment is adopted a plurality of SVR models to predict and is drawn predicted value.
This experiment is average to 50 predictions, and its checking feature parameter comprises mean square deviation and proving time, and experimental result sees also table 1.
The contrast of table 1 check feature
Can find out from result's contrast of table 1; The estimated performance of the disclosed time series forecasting method based on support vector regression of the application embodiment is higher than step time forecasting methods of existing single SVR model; Especially under noisy situation, estimated performance improves obviously.Simultaneously, under noisy situation, the proving time of the disclosed time series forecasting method based on support vector regression of the application embodiment is less than the proving time of step time forecasting methods of existing single SVR model.
Embodiment is corresponding with said method; The application embodiment also discloses a kind of time series forecasting system based on support vector regression; Its structural representation sees also Fig. 3, comprising: data set draws module 31, SVR model construction module 32, data decimation module 33 and first prediction module 34.Wherein:
Data set draws module 31, is used for choosing historical data from existing time series data is concentrated, draws a plurality of training datasets.
Wherein, each training dataset comprises a plurality of existing time series data subclass and the predicted value corresponding with each time series data subclass, and arbitrary time series data subclass is the historical data that said time series data is concentrated with its corresponding predicted value.
Suppose that it is x (k) that time series data is concentrated historical data, k=0,1 ..., t-1, t is a current time, and the number of the training dataset that draws is T, and j training dataset can be used S
jExpression, wherein
J training dataset S
jComprise t-d-j+1 time series data subclass u
k, u
k=[x (k-d+1) ... X (k-1) x (k)]
T∈ R
d, each time series data subclass u
kTo a predicted value y should be arranged
k, y
k=x (k+j) ∈ R, R representes set of real numbers, and arbitrary time series data subclass is the historical data that said time series data is concentrated with its corresponding predicted value, and d is a time delay, representes each time series data subclass u
kThe sample space dimension, d=T.
Above-mentioned a plurality of training data is concentrated, because the j value of different training data set is different, so the time series data subclass number that the different training data set comprises is different.
SVR model construction module 32 is used to confirm the regular parameter and the gaussian kernel parameter of SVR model to be made up, and according to said regular parameter and gaussian kernel parameter, respectively each training dataset is trained, and makes up the support vector regression SVR model of self correspondence.
When SVR model construction module 32 was confirmed regular parameter and the gaussian kernel parameter of SVR model to be made up, SVR model construction module 32 can adopt existing method, confirms regular parameter and gaussian kernel parameter like cross validation method.
Above-mentioned SVR model construction module 32 trains the process that makes up with its corresponding SVR model identical with existing building process based on SVR model in the step time forecasting methods of single SVR model for arbitrary training dataset, with j training dataset S
j, promptly can utilize the time series data subclass u in it
kWith each time series data subclass u
kCorresponding y
kTrain, obtain j training dataset S
jJ corresponding SVR model, concrete training process is no longer set forth.
In the present embodiment, the T that a chooses historical data is respectively x (t-T+1) ..., x (t-1), x (t), it forms a time series data subclass u
t, u
t=[x (t-T+1) ... X (t-1) x (t)]
T
Wherein, suppose that prediction step is τ, the then tested moment is t+ τ, and first difference of the tested moment with current time then is prediction step τ.The number of training dataset is T, and then the number of SVR model also is T.
See also Fig. 4, Fig. 4 is to be the basis with Fig. 3, and the another kind of structural representation of the disclosed time series forecasting system based on support vector regression of the application embodiment also comprises: second prediction module 35.
Second prediction module 35, be used for by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under, utilize a plurality of SVR models to obtain to a said T historical data by the prediction predicted value in the moment.
Utilizing a plurality of SVR models to obtain in the process of quilt prediction predicted value constantly; Second prediction module 35 need be upgraded T the historical data and the tested moment; Whether greater than the number of SVR model, and under different situations, carry out various process in the difference that rejudges the tested moment and current time.Its structural representation sees also Fig. 5, comprising: predicted value draws unit 351, updating block 352, judging unit 353 and trigger 354.
Wherein, predicted value draws unit 351, be used for by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under, utilize T historical data the choosing SVR model of correspondence respectively, draw the predicted value of T historical data self.
T the corresponding respectively SVR model of historical data is: historical data x (t-T+1) is corresponding to training dataset S
1The 1st the SVR model that training draws, historical data x (t-T+2) is corresponding to training dataset S
2The 2nd the SVR model that training draws ..., by that analogy, historical data x (t) is corresponding to training dataset S
TT the SVR model that training draws.Each historical data uses self corresponding SVR model to predict respectively, draws predicted value.
Need to prove: this unit historical data is when predicting, prediction step is different, and the prediction step of historical data x (t-T+1) is 1; The prediction step of historical data x (t-T+2) is 2; ..., by that analogy, the prediction step of historical data x (t) is T.The predicted value that T historical data draws can be respectively
Updating block 352 is used for predicted value with the said T that a draws historical data self and is updated to the t-T+1 that chooses constantly to T the historical data of current t between constantly, and with second difference of the number of said first difference and SVR model as the tested moment.
Whether judging unit 353, the difference that is used to judge this tested moment and current time be greater than the number of said SVR model.Trigger 354 is used under the situation of said difference smaller or equal to the number of said SVR model, triggering said first prediction module, under the situation of said difference greater than the number of said SVR model, triggers said second prediction module.
In the present embodiment; First prediction module 34 by prediction constantly with the situation of first difference smaller or equal to the number of SVR model of current time under; Choose the SVR model corresponding with difference; To T the historical data of the extremely current constantly t of the t-T+1 that chooses between the moment, directly utilize this SVR model to obtain by the prediction predicted value in the moment.Present embodiment obtains the mode that quilt is predicted predicted value constantly with respect to the prior art multi-step prediction, and the accumulation of predicated error reduces, and then obtains the degree of accuracy raising of predicted value.
Further, predicted value draw unit 351 by prediction constantly with the situation of first difference greater than the number of SVR model of current time under, choose T historical data and use the SVR model of self correspondence to predict respectively, draw T predicted value.Simultaneously updating block 352 is updated to T predicted value with T historical data, with second difference of the number of first difference and SVR model as the tested moment.Judging unit 353 in the difference that rejudges the tested moment and current time whether greater than the number of said SVR model, until the number of difference smaller or equal to the SVR model.Though above-mentioned steps has caused the accumulation of predicated error, this predicated error accumulation can improve the degree of accuracy of predicted value equally less than the accumulation of predicated error in the step time forecasting methods of existing single SVR model.
The practical implementation of the step of said system embodiment sees also the related description among the method embodiment, and this is no longer set forth.
Each embodiment in this instructions all adopts the mode of going forward one by one to describe; Identical similar part is mutually referring to getting final product between each embodiment; Each embodiment stresses all is the difference with other embodiment; Those of ordinary skills promptly can understand and implement under the situation of not paying creative work.
The above only is the application's a embodiment; Should be pointed out that for those skilled in the art, under the prerequisite that does not break away from the application's principle; Can also make some improvement and retouching, these improvement and retouching also should be regarded as the application's protection domain.
At last; Also need to prove; In this article; Relational terms such as first and second grades only is used for an entity or operation are made a distinction with another entity or operation, and not necessarily requires or hint relation or the order that has any this reality between these entities or the operation.And; Term " comprises ", " comprising " or its any other variant are intended to contain comprising of nonexcludability; Thereby make and comprise that process, method, article or the equipment of a series of key elements not only comprise those key elements; But also comprise other key elements of clearly not listing, or also be included as this process, method, article or equipment intrinsic key element.Under the situation that do not having much more more restrictions, the key element that limits by statement " comprising ... ", and be not precluded within process, method, article or the equipment that comprises said key element and also have other identical element.
Claims (10)
1. the time series forecasting method based on support vector regression is characterized in that, comprising:
Choose historical data from existing time series data is concentrated, draw a plurality of training datasets;
Confirm the regular parameter and the gaussian kernel parameter of SVR model to be made up,, respectively each training dataset is trained, make up self corresponding support vector regression SVR model according to said regular parameter and gaussian kernel parameter;
Choose t-T+1 constantly to T the historical data of current t between the moment, wherein, T is the number of said SVR model;
By prediction constantly with the situation of first difference smaller or equal to the number of said SVR model of current time under, choose the SVR model corresponding with said first difference, directly utilize this SVR model to obtain to be predicted the predicted value in the moment to a said T historical data.
2. the time series forecasting method based on support vector regression according to claim 1; It is characterized in that; Also comprise: by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under, utilize a plurality of SVR models to obtain to a said T historical data by the prediction predicted value in the moment.
3. the time series forecasting method based on support vector regression according to claim 2; It is characterized in that; Saidly utilize a plurality of SVR models to obtain directly to be predicted that predicted value constantly comprises to T historical data: by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under; The T that utilization is chosen historical data distinguished corresponding SVR model, draws the predicted value of T historical data self;
The predicted value of the said T that a draws historical data self is updated to the t-T+1 that chooses constantly to T the historical data of current t between constantly, and with second difference of the number of said first difference and SVR model as the tested moment;
Whether the difference of judging this tested moment and current time is greater than the number of said SVR model; Under the situation of said difference smaller or equal to the number of said SVR model; The SVR model corresponding with said first difference chosen in execution; Directly utilize this SVR model to obtain to a said T historical data by the step of prediction predicted value constantly; Under the situation of said difference greater than the number of said SVR model, carry out and utilize the corresponding respectively SVR model of choosing of T historical data, draw the step of the predicted value of T historical data self.
4. according to any described time series forecasting method of claim 1 to 3, it is characterized in that, adopt cross validation method, confirm the regular parameter and the gaussian kernel parameter of SVR model to be made up based on support vector regression.
5. according to any described time series forecasting method of claim 1 to 3, it is characterized in that the time series data subclass number that the different training data set comprises is different based on support vector regression.
6. the time series forecasting system based on support vector regression is characterized in that, comprising:
Data set draws module, is used for choosing historical data from existing time series data is concentrated, draws a plurality of training datasets;
The SVR model construction module is used to confirm the regular parameter and the gaussian kernel parameter of SVR model to be made up, and according to said regular parameter and gaussian kernel parameter, respectively each training dataset is trained, and makes up the support vector regression SVR model of self correspondence;
The data decimation module is used to choose t-T+1 constantly to T the historical data of current t between the moment, and wherein, T is the number of said SVR model;
First prediction module; Be used for by prediction constantly with the situation of first difference smaller or equal to the number of said SVR model of current time under; Choose the SVR model corresponding, directly utilize this SVR model to obtain by prediction predicted value constantly to a said T historical data with said first difference.
7. the time series forecasting system based on support vector regression according to claim 6; It is characterized in that; Also comprise: second prediction module; Be used for by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under, utilize a plurality of SVR models to obtain to a said T historical data by the prediction predicted value in the moment.
8. the time series forecasting system based on support vector regression according to claim 7 is characterized in that, said second prediction module comprises:
Predicted value draws the unit, be used for by prediction constantly with the situation of first difference greater than the number of said SVR model of current time under, utilize T historical data the choosing SVR model of correspondence respectively, draw the predicted value of T historical data self;
Updating block is used for predicted value with the said T that a draws historical data self and is updated to the t-T+1 that chooses constantly to T the historical data of current t between constantly, and with second difference of the number of said first difference and SVR model as the tested moment;
Whether judging unit, the difference that is used to judge this tested moment and current time be greater than the number of said SVR model;
Trigger is used under the situation of said difference smaller or equal to the number of said SVR model, triggering said first prediction module, under the situation of said difference greater than the number of said SVR model, triggers said second prediction module.
9. according to any described time series forecasting system of claim 6 to 8 based on support vector regression; It is characterized in that; Said SVR model construction module specifically adopts cross validation method, confirms the regular parameter and the gaussian kernel parameter of SVR model to be made up.
10. according to any described time series forecasting system of claim 6 to 8, it is characterized in that it is different that data set draws the time series data subclass number that different training data set that module draws comprises based on support vector regression.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2011103204825A CN102360388A (en) | 2011-10-20 | 2011-10-20 | Time series forecasting method and system based on SVR (Support Vector Regression) |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2011103204825A CN102360388A (en) | 2011-10-20 | 2011-10-20 | Time series forecasting method and system based on SVR (Support Vector Regression) |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN102360388A true CN102360388A (en) | 2012-02-22 |
Family
ID=45585717
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2011103204825A Pending CN102360388A (en) | 2011-10-20 | 2011-10-20 | Time series forecasting method and system based on SVR (Support Vector Regression) |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102360388A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103544544A (en) * | 2013-10-29 | 2014-01-29 | 广东工业大学 | Energy consumption forecasting method and device |
| CN103723292A (en) * | 2013-10-23 | 2014-04-16 | 南京工业大学 | Control method for improving performance of automatic weighing and bagging system |
| CN104657749A (en) * | 2015-03-05 | 2015-05-27 | 苏州大学 | Method and device for classifying time series |
| CN105962906A (en) * | 2016-06-14 | 2016-09-28 | 广州视源电子科技股份有限公司 | Body temperature measuring method and device |
| KR101856170B1 (en) * | 2017-09-20 | 2018-05-10 | 주식회사 모비젠 | Apparatus for predicting error generation time of system based on time-series data and method thereof |
| CN109684302A (en) * | 2018-12-04 | 2019-04-26 | 平安科技（深圳）有限公司 | Data predication method, device, equipment and computer readable storage medium |
| CN110390342A (en) * | 2018-04-16 | 2019-10-29 | 北京京东尚科信息技术有限公司 | Time Series Forecasting Methods and device |
| CN110837937A (en) * | 2019-11-13 | 2020-02-25 | 联想(北京)有限公司 | Prediction method, prediction device, calculation device, and medium |
| CN110991774A (en) * | 2019-12-31 | 2020-04-10 | 新奥数能科技有限公司 | Electric quantity load prediction method and device |
| CN111461397A (en) * | 2020-02-26 | 2020-07-28 | 山东浪潮通软信息科技有限公司 | Budget prediction method, equipment and medium based on improved support vector regression |
| CN111813822A (en) * | 2020-05-25 | 2020-10-23 | 国网河南省电力公司 | Method and system for determining polynomial fitting node voltage effective value based on WANS |
| CN111835536A (en) * | 2019-04-16 | 2020-10-27 | 杭州海康威视数字技术股份有限公司 | Flow prediction method and device |
| CN112711615A (en) * | 2019-10-24 | 2021-04-27 | 富士通株式会社 | Information processing apparatus, information processing method, and computer-readable storage medium |
-
2011
- 2011-10-20 CN CN2011103204825A patent/CN102360388A/en active Pending
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103723292A (en) * | 2013-10-23 | 2014-04-16 | 南京工业大学 | Control method for improving performance of automatic weighing and bagging system |
| CN103544544A (en) * | 2013-10-29 | 2014-01-29 | 广东工业大学 | Energy consumption forecasting method and device |
| CN103544544B (en) * | 2013-10-29 | 2016-08-17 | 广东工业大学 | A kind of energy resource consumption Forecasting Methodology and device |
| CN104657749A (en) * | 2015-03-05 | 2015-05-27 | 苏州大学 | Method and device for classifying time series |
| CN105962906A (en) * | 2016-06-14 | 2016-09-28 | 广州视源电子科技股份有限公司 | Body temperature measuring method and device |
| KR101856170B1 (en) * | 2017-09-20 | 2018-05-10 | 주식회사 모비젠 | Apparatus for predicting error generation time of system based on time-series data and method thereof |
| CN110390342A (en) * | 2018-04-16 | 2019-10-29 | 北京京东尚科信息技术有限公司 | Time Series Forecasting Methods and device |
| CN109684302A (en) * | 2018-12-04 | 2019-04-26 | 平安科技（深圳）有限公司 | Data predication method, device, equipment and computer readable storage medium |
| CN109684302B (en) * | 2018-12-04 | 2023-08-15 | 平安科技（深圳）有限公司 | Data prediction method, device, equipment and computer readable storage medium |
| CN111835536A (en) * | 2019-04-16 | 2020-10-27 | 杭州海康威视数字技术股份有限公司 | Flow prediction method and device |
| CN111835536B (en) * | 2019-04-16 | 2022-04-05 | 杭州海康威视数字技术股份有限公司 | Flow prediction method and device |
| CN112711615A (en) * | 2019-10-24 | 2021-04-27 | 富士通株式会社 | Information processing apparatus, information processing method, and computer-readable storage medium |
| CN110837937A (en) * | 2019-11-13 | 2020-02-25 | 联想(北京)有限公司 | Prediction method, prediction device, calculation device, and medium |
| CN110991774A (en) * | 2019-12-31 | 2020-04-10 | 新奥数能科技有限公司 | Electric quantity load prediction method and device |
| CN111461397A (en) * | 2020-02-26 | 2020-07-28 | 山东浪潮通软信息科技有限公司 | Budget prediction method, equipment and medium based on improved support vector regression |
| CN111813822A (en) * | 2020-05-25 | 2020-10-23 | 国网河南省电力公司 | Method and system for determining polynomial fitting node voltage effective value based on WANS |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102360388A (en) | Time series forecasting method and system based on SVR (Support Vector Regression) | |
| CN103250031B (en) | Routing system, routing method and route selection program | |
| Xiang et al. | An active learning method combining deep neural network and weighted sampling for structural reliability analysis | |
| CN104200087B (en) | For the parameter optimization of machine learning and the method and system of feature tuning | |
| US9727533B2 (en) | Detecting anomalies in a time series | |
| Xu et al. | Pipe break prediction based on evolutionary data-driven methods with brief recorded data | |
| CN103942375B (en) | High-speed press sliding block dimension robust design method based on interval | |
| Zeng et al. | Improved dynamic programming for parallel reservoir system operation optimization | |
| Park et al. | A data warehouse-based decision support system for sewer infrastructure management | |
| JP6784780B2 (en) | How to build a probabilistic model for large-scale renewable energy data | |
| CN102779223A (en) | Method and device for forecasting short-term power load | |
| CN105068874A (en) | Resource on-demand dynamic allocation method combining with Docker technology | |
| CN104992244A (en) | Airport freight traffic prediction analysis method based on SARIMA and RBF neural network integration combination model | |
| CN103685347B (en) | Method and device for allocating network resources | |
| CN102346745A (en) | Method and device for predicting user behavior number for words | |
| CN104915251A (en) | Task scheduling method and device | |
| US20130332243A1 (en) | Predictive analytics based ranking of projects | |
| Huang et al. | Exact project reliability for a multi-state project network subject to time and budget constraints | |
| Alfieri et al. | The time buffer approximated Buffer Allocation Problem: A row–column generation approach | |
| CN101866143A (en) | Road traffic service level prediction method based on space-time characteristic aggregation | |
| CN112181659B (en) | Cloud simulation memory resource prediction model construction method and memory resource prediction method | |
| Micheli et al. | Multiple-surrogate models for probabilistic performance assessment of wind-excited tall buildings under uncertainties | |
| CN104699614A (en) | Software defect component predicting method | |
| CN102033936A (en) | Method for comparing similarity of time sequences | |
| CN101699545B (en) | Method and device for automatically generalizing map spatial points |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C12 | Rejection of a patent application after its publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20120222 |