CN107274009A - A kind of time series data multistep forecasting method and system based on correlation - Google Patents
A kind of time series data multistep forecasting method and system based on correlation Download PDFInfo
- Publication number
- CN107274009A CN107274009A CN201710390314.0A CN201710390314A CN107274009A CN 107274009 A CN107274009 A CN 107274009A CN 201710390314 A CN201710390314 A CN 201710390314A CN 107274009 A CN107274009 A CN 107274009A
- Authority
- CN
- China
- Prior art keywords
- data
- time series
- series data
- correlation
- measured
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Marketing (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Development Economics (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention relates to a kind of time series data multistep forecasting method based on correlation and system, including：According to the correlation between adjacent time point data in time series data to be measured, output length and regression order are set, and time series data to be measured is split as into inputoutput data to set according to output length and regression order；Inputoutput data is combined into the training data of multi output Gaussian process model to collection, training generation forecast model, by time series data input prediction model to be measured, obtains final predicted value, it is more than or equal to prediction step until the total number of final predicted value is no, exports final predicted value.This method utilizes the correlation between adjacent time point data in time series data to be measured, based on multiple-input and multiple-output strategy, multiple future values are predicted simultaneously using multi output Gaussian process model, while progressively being predicted backward with the mode of iteration, with high prediction accuracy.
Description
Technical field
The present invention relates to time series data multi-step prediction field, more particularly to a kind of time series number based on correlation
According to multistep forecasting method and system.
Background technology
Time series data almost covers the related any field of scientific and engineering.Time series data prediction is utilized
Related forecast model, using historical data as input, future value is predicted based on historical perspective data.Time series data is pre-
Survey has obtained critically important application in terms of financial market analysis, economic forecasting, environmental monitoring.How to find and describe the time
The changing rule of sequence data, sets up corresponding forecast model and time series data is predicted so as to reduce the predicated error of model
It is most important.
Time series data multi-step prediction is related to predicting strategy and selects two aspects with model.It is main in terms of predicting strategy
There are 5 kinds of conventional predicting strategies, i.e. iterative strategy (Recursive), direct strategy (Direct), multiple-input and multiple-output strategy
(MIMO), iterative strategy is combined (DirREC) and direct strategy with direct strategy and is combined (DIRMO) with MIMO.In prediction
In performance, the estimated performance of MIMO strategies generally to be preferred over other predicting strategies.For same forecast model, direct strategy
Precision of prediction is higher than the precision of prediction of iterative strategy, because iterative strategy can cause prediction using predicted value as mode input
Error accumulation.In terms of model selection, conventional forecast model has autoregression (AR) model, least square method supporting vector machine at present
(LSSVM) model, neuroid (NN) model and Gaussian process (GP) model etc..AR models are only suitable for linear session sequence
Data modeling and direct strategy and iterative strategy can only be used.The problems such as LSSVM can be good at non-linear solution, high dimension,
LSSVM models can only equally use direct strategy and iterative strategy.For NN models, its flexible structure can use all plans
Omit, therefore be widely used in time series data multi-step prediction.However, NN structure design also it is unified it is theoretic according to
According to, such as network number of plies, neuron number, the selection of transmission function etc., therefore have certain difficulty on the structure of network.GP is 20 generation
The machine learning method that discipline end grows up, this method achieves extremely successful application in classification and recurrence field.At present,
No matter using which kind of model, which kind of strategy, the correlation between time series data different time point data is not all accounted for, i.e.,
Correlation is there is between the data sometime put and the data of its surrounding time point.
In summary, because every kind of model has its own limitation, in addition to neuroid, the above can not be used
5 kinds of predicting strategies.A future value can only be predicted every time using direct strategy or iterative strategy model, and prediction multistep needs multiple
Independently predicted between training pattern, therefore the future value of the different step-lengths of prediction.Though how defeated neuroid can be used
Enter multi output strategy, but only increase the dimension of output, its inner link is not considered.Gaussian process is being used as a ratio
Newer machine learning method, is widely used in recurrence field.Data to be carried out using Gaussian process modeling pre- at present
Survey is all single output, so direct strategy or iterative strategy can only be used, how defeated also nobody is Gaussian process and multi input
Go out strategy and be combined progress time series data prediction.
Traditional multistep forecasting method to time series data is all based on some models presented hereinbefore and strategy,
Such as deep neural network regression model of multiple-input and multiple-output structure, using Gaussian process iteration or direct strategy model etc.,
There are certain methods to add the factor related to the time series data when carrying out time series data prediction and predict it jointly not
To be worth, but these models all do not account for before and after time series data time point between correlation, and use Gaussian process side
Method is all single output, and nobody uses multiple-input and multiple-output strategy.
The content of the invention
The deficiency existed for current method, a kind of high-precision multistep forecasting method based on correlation of present invention invention,
Not only can using MIMO strategies, and can with the correlation between analysis and utilization output data, when time series data some when
Between the data value put it is very big when, next time point its value will also tend to it is very big, the data of adjacent time point be not it is independent,
The present invention can export multiple predicted values to improve the degree of accuracy of multi-step prediction simultaneously using the correlation, so as to realize to the time
The accurate prediction of sequence data future value.Correlation between the different step-length correspondence prediction data of the invention by com-parison and analysis, and
Time series data is organized into inputoutput data pair, multi output Gaussian process model prediction is used.This method can be more
Input multi output strategy to be applied in Gaussian process model, while improving time series data using the correlation of each outlet chamber
The degree of accuracy of future value prediction.And this method is applied to small sample, sparse data.
Specifically, the invention discloses a kind of time series data multistep forecasting method based on correlation, wherein wrapping
Include：
Step 1：The time series data to be measured being made up of multiple time point data is read, according to the time series number to be measured
According to the correlation between middle adjacent time point data, output length is set, according to the output length and regression order set in advance,
Time series data to be measured is normalized, and input is split as to the time series data to be measured after normalized
Output data is to set, the prediction step needed for user's input, the output number to control final predicted value, the wherein input
Output data is the output length to the number of set；
Step 2：Using the inputoutput data to gathering the training data as multi output Gaussian process model, training generation
Forecast model, the forecast model is inputted by the time series data to be measured, obtains initial predicted value, and the initial predicted value is entered
Row data renormalization obtains final predicted value, and the number of the wherein initial predicted value and the final predicted value is that the output is long
Degree；
Step 3：Judge whether the total number of the final predicted value is more than or equal to the prediction step, if then algorithm knot
Beam, otherwise returns to the step 2 and continues to predict, untill the number of the final predicted value is more than or equal to the prediction step.
The time series data multistep forecasting method based on correlation, wherein step 1 includes：
Step 11, the time series data to be measured is translated x successively, x time series data of generation and this it is to be measured when
Between sequence data form x+1 time series data together, if the Pearson correlation coefficient between the x+1 time series data
More than 0.5, then judge that there is correlation in the time series data to be measured between adjacent x+1 time point data, wherein x is just whole
Number.
The time series data multistep forecasting method based on correlation, wherein in step 1 the output length be less than etc.
In x+1 positive integer.
The time series data multistep forecasting method based on correlation, also includes using sliding window wherein in the step 1
The time series data to be measured is split as inputoutput data pair by the mode of mouth, and the inputoutput data is combined into this to collection
Inputoutput data is to set.
The time series data multistep forecasting method based on correlation, the wherein step 2 were also included the time to be measured
The time point newest time point data inputs the forecast model as input data in sequence data, wherein the input data
Number is the regression order.
Present invention also offers a kind of time series data multi-step prediction system based on correlation, including：
Setup module：For reading the time series data to be measured being made up of multiple time point data, according to this it is to be measured when
Between correlation in sequence data between adjacent time point data output length is set, according to the output length and set in advance time
Return exponent number, time series data to be measured is normalized, and the time series data to be measured after normalized is torn open
It is divided into inputoutput data to set, the prediction step needed for user's input, the output number to control final predicted value, its
In the inputoutput data be the output length to the number of set；
Prediction module：For using the inputoutput data to gathering the training data as multi output Gaussian process model,
Training generation forecast model, inputs the forecast model by the time series data to be measured, obtains initial predicted value, and preliminary to this
Predicted value carries out data renormalization and obtains final predicted value, and the number of the wherein initial predicted value and the final predicted value is
The output length；
Judge module：For judging whether the total number of the final predicted value is more than or equal to the prediction step, if then
Algorithm terminates, and otherwise calls the prediction module to continue to predict again, until the number of the final predicted value is pre- more than or equal to this
Untill surveying step-length.
The time series data multi-step prediction system based on correlation, wherein setup module includes：
Correlation prediction module, for the time series data to be measured to be translated into x successively, generates x time series number
X+1 time series data is formed according to together with the time series data to be measured, if the skin between the x+1 time series data
You are more than 0.5 at inferior coefficient correlation, then judge there is correlation in the time series data to be measured between adjacent x+1 time point data
Property, wherein x is positive integer.
The time series data multi-step prediction system based on correlation, be provided with the output of this in module length be less than
Positive integer equal to x+1.
The time series data multi-step prediction system based on correlation, is wherein also included in the setup module, using cunning
The time series data to be measured is split as inputoutput data pair by the mode of dynamic window, and by the inputoutput data to set
Be the inputoutput data to set.
The time series data multi-step prediction system based on correlation, the wherein prediction module also include by this it is to be measured when
Between in sequence data the time point newest time point data input the forecast model as input data, the wherein input data
Number is the regression order.
The present invention is trained prediction to the time series data with obvious correlation adjacent time point, by using
Multi output Gaussian process model, can reduce predicated error.The technical characterstic of invention protrudes and is embodied in two sides of function and performance
Face.
Functionally, invention can make accurate prediction to sparse or non-sparse time series data, and user can be with
Regression order, the output relevant parameter such as length and prediction step of prediction are set, and the number that model exports predicted value every time is
Length is exported, is constantly predicted by way of continuous iteration, untill prediction output reaches prediction step.
In performance, this method due to considering correlation between different value when predicting multiple predicted values simultaneously, model
Each output port may be by other output ports to improve the degree of accuracy of prediction.
Brief description of the drawings
Fig. 1 is multi output Gaussian process illustraton of model；
Fig. 2 is that time series data of the present invention is organized into inputoutput data collection schematic diagram：
Fig. 3 translates schematic diagram for the present invention to time series data；
Fig. 4 is the multi-step Predictive Model figure of the invention based on correlation；
Fig. 5 is time series data tendency chart of the present invention；
Fig. 6 is time series data correlation analysis figure of the present invention.
Embodiment
The technical problems to be solved by the invention are, under time series data multi-step prediction scene, how using defeated
Correlation between going out proposes a kind of base while export multiple predicted values, and Gaussian process is combined with multiple-input and multiple-output strategy
In the multistep forecasting method of correlation.Method will simultaneously be exported many using time series data as input according to future time point
Individual predicted value.
The core objective of the present invention is by the phase between adjacent time point corresponding data before and after analysis time sequence data
Guan Xing, it was demonstrated that there is correlation between the corresponding data in consecutive number strong point, and the time series data is organized into suitable defeated
Enter output data in input multi output Gaussian process model, multi output Gaussian process model regards Gaussian process by Gauss as
White noise is obtained by kernel function process of convolution, is shared identical white Gaussian noise by using different output port and is set up output
Dependency relation between port, so as to handle multi output variable problem, therefore to realize time series data pre- by the present invention
Gaussian process is combined with multiple-input and multiple-output strategy in survey, and improves using the correlation between output port prediction essence
Degree.And the step-length of prediction is continuously increased using the mode of iteration, the purpose of multi-step prediction is reached, multi output Gaussian process model is such as
Shown in Fig. 1.
Fig. 1 be refer to it is contemplated that using the real time series data of history, the time series data is inputted mould
In type and set correlation parameter, so as to realize the prediction to the time series data future value.
To allow the features described above and more clearly understandable, the special embodiment below that can illustrate of effect of the present invention, and coordinate
Bright book accompanying drawing is described in detail below.
As shown in Fig. 2 user can be setting regression order p and output length h to pass through constantly to by way of right translation
Time series data is split into h inputoutput data pair, and this h inputoutput data is combined into input and output number to collection
According to input data of the set (training data) together as model.For the correlation between analysis different output port, the present invention
The correlation between the corresponding variate-value of different output port in all data of Direct Analysis is only needed, but is due to time series number
According to particularity, in order to obtain it is different prediction output ports between correlations, the present invention using based on translation mechanism correlation
Property analysis strategy, without time series data first is organized into inputoutput data pair, is then analyzed between output valve again
Correlation, it is specific as follows：The present embodiment considers that output length is 4, but is not limited, but output length should not be too big, with
The correlation between the increase of output length, output port can weaken.It is general that value is carried out according to correlation analysis result.
Correlation between (exporting adjacent 4 step predicted value every time), relatively more adjacent 4 output port predicted values, i.e. analysis time sequence number
According to the correlation between consecutive number strong point, such as time series data [s1,…st,st+1,st+2,st+3…sN], the present invention
Analyze adjacent 4 sampled data st,st+1,st+2,st+3Between correlation, be that this present invention can equal time series data
Move 1~3 time, four time series data series1~series4 are formed together with former time series data, as shown in figure 3,
Then analysis obtains the Pearson correlation coefficient between 4 time series datas, and specially Pearson correlation coefficient is said more than 0.5
It is bright that there is correlation, it is possible to use this method, and the bigger performance boost of correlation is more obvious.Each row correspond to use in figure
(y is by an inputoutput data centering output valve y that multiple-input and multiple-output strategy is organized when predicting the time series data
4 dimension) value, obtain the correlation of 4 output variables in output valve y by comparing the correlation between series1~series4
Property, the prediction effect of correlation more large-sized model is better.Output length h is determined according to the correlation between output port.
Inputoutput data obtained above is combined into training sample input multi output Gaussian process model to collection to instruct
White silk, inputs p newest value of time series data to be measured as prediction, when last p value is that time series data is last p
Between the data value put, if time series data has N number of value, then take last these time point datas of N-p+1 to N as mould
The input of type, i.e. [sN-p+1…sN], h final predicted values can be obtained by obtained forecast model.Judge obtained prediction
Whether the number of data reaches the prediction step that user requires, at the beginning of i, i is constantly increased if being not reaching to by way of iteration
Initial value is 0, increases h every time, and it is the purpose for reaching increase estimation range to set i, because the schematic diagram according to Fig. 2, i is 0
When predict h value nearest in unknown time series, i adds 1 every time can afterwards move prediction " window " step, and window size is solid
Fixed, it is h.I adds h to be to repeat to predict every time.
Increase and training data and training pattern are reorganized after i, reach the purpose of the estimation range of increase model.Until full
Untill the demand of sufficient user, i.e. untill the number of prediction data meets or exceeds prediction step.
The complete prediction process of time series data, overall flow such as Fig. 4 institutes will be directed to an example description below
Show, use water quality parameter (dissolved oxygen) data, such as Fig. 5.
Step 1：The time series data to be measured being made up of multiple time point data is read, according to the time series number to be measured
According to the correlation between middle adjacent time point data, output length is set, according to the output length and regression order set in advance,
Time series data to be measured is normalized, and input is split as to the time series data to be measured after normalized
Output data is to set, the prediction step needed for user's input, the output number to control final predicted value, the wherein input
Output data is the output length to the number of set.Specifically include and correlation is carried out to time series data according to Fig. 2
Analysis, analysis result is shown in Fig. 6, it can be seen that have very strong correlation, Pearson correlation coefficient between adjacent 4 time point datas
There is correlation more than 0.5 explanation.Output length h, h is set to be less than or equal to related adjacent time point data according to correlation
Number, that is, when choosing output length, determine according to the result of correlation analysis.The correlation analysis result such as obtained according to Fig. 6,
All there is correlation between adjacent 4 time point datas, if then h can select integer the series1 and series4 less than or equal to 4
Correlation be less than 0.5 (actual be 0.815), then integer of the h choosings less than or equal to 3, the present embodiment selection h=2.In this implementation
It is less than or equal to 4 in example.Regression order p=4 is set, and prediction step is s=4, length h=2 is exported, wherein prediction step s is small
Regression order is needed not be equal to, prediction step s can need arbitrarily to be set according to user, and wherein the step 1 includes：Step
11, the time series data to be measured is translated x successively, x time series data and the time series data one to be measured is generated
Rise and form x+1 time series data, if the Pearson correlation coefficient between the x+1 time series data is more than 0.5, sentence
There is correlation, wherein x is positive integer, and the output is long in the fixed time series data to be measured between adjacent x+1 time point data
Spend for the positive integer less than or equal to x+1；The step 1 also includes, by the time series data to be measured by the way of sliding window
Inputoutput data pair is split as, and the inputoutput data is combined into the inputoutput data to collection to set.
Step 2：Using the inputoutput data to gathering the training data as multi output Gaussian process model, training generation
Forecast model, the forecast model is inputted by the time series data to be measured, obtains initial predicted value, and the initial predicted value is entered
Row data renormalization obtains final predicted value, and the wherein number of the initial predicted value is the numerical value of the output length.User is defeated
Enter required regression order, to the dimension of control input data, and time point in the time series data to be measured is newest
The time point data inputs the forecast model as input data, and wherein the input data number is the regression order.It is specific right
Time series data is normalized and is organized into training data as shown in Figure 2 and prediction input.Training data is put into
It is trained in multi output Gaussian process model and adjusts model parameter, obtains forecast model.Prediction input is put into the prediction
Initial predicted value (h predicted value of output every time) is obtained in model, data renormalization is carried out to initial predicted value and obtained finally
Predicted value, i.e., circulation obtains h initial predicted value every time, then carries out renormalization to this h initial predicted value, obtains h
Final predicted value.
Step 3：Judge whether the total number of the final predicted value is more than or equal to the prediction step, if then algorithm knot
Beam, otherwise returns to the step 2 and continues to predict, untill the number of the final predicted value is more than or equal to the prediction step, tool
Body is whether the number for judging total predicted value has reached prediction step s, and algorithm terminates if s is reached, if being not reaching to s
Then increase i (i=i+h) return to step 2 continues to predict, untill the number of predicted value meets or exceeds s.
To sum up, this method utilizes the temporal correlation between time series data point, based on multiple-input and multiple-output strategy, uses
Multi output Gaussian process model utilizes the correlation between data point to predict multiple future values simultaneously, while with the mode of iteration progressively
Predict backward.
It is below system embodiment corresponding with above method embodiment, this implementation system can be mutual with above-mentioned embodiment
Coordinate and implement.The above-mentioned relevant technical details mentioned in mode of applying are still effective in this implementation system, in order to reduce repetition, this
In repeat no more.Correspondingly, the relevant technical details mentioned in this implementation system are also applicable in above-mentioned embodiment.
Present invention also offers a kind of time series data multi-step prediction system based on correlation, including：
Setup module：For reading the time series data to be measured being made up of multiple time point data, according to this it is to be measured when
Between correlation in sequence data between adjacent time point data output length is set, according to the output length and set in advance time
Return exponent number, time series data to be measured is normalized, and the time series data to be measured after normalized is torn open
It is divided into inputoutput data to set, the prediction step needed for user's input, the output number to control final predicted value, its
In the inputoutput data be the output length to the number of set；
Prediction module：For using the inputoutput data to gathering the training data as multi output Gaussian process model,
Training generation forecast model, inputs the forecast model by the time series data to be measured, obtains initial predicted value, and preliminary to this
Predicted value carries out data renormalization and obtains final predicted value, and the number of the wherein initial predicted value and the final predicted value is
The output length；
Judge module：For judging whether the total number of the final predicted value is more than or equal to the prediction step, if then
Algorithm terminates, and otherwise calls the prediction module to continue to predict again, until the number of the final predicted value is pre- more than or equal to this
Untill surveying step-length.
The time series data multi-step prediction system based on correlation, wherein setup module includes：
Correlation prediction module, for the time series data to be measured to be translated into x successively, generates x time series number
X+1 time series data is formed according to together with the time series data to be measured, if the skin between the x+1 time series data
You are more than 0.5 at inferior coefficient correlation, then judge there is correlation in the time series data to be measured between adjacent x+1 time point data
Property, wherein x is positive integer.
The time series data multi-step prediction system based on correlation, be provided with the output of this in module length be less than
Positive integer equal to x+1.
The time series data multi-step prediction system based on correlation, is wherein also included in the setup module, using cunning
The time series data to be measured is split as inputoutput data pair by the mode of dynamic window, and by the inputoutput data to set
Be the inputoutput data to set.
The time series data multi-step prediction system based on correlation, the wherein prediction module also include by this it is to be measured when
Between in sequence data the time point newest time point data input the forecast model as input data, the wherein input data
Number is the regression order.
Although the present invention is disclosed with above-described embodiment, specific embodiment only to explain the present invention, is not used to limit
The present invention, any those skilled in the art of the present technique without departing from the spirit and scope of the invention, can make the change and complete of some
It is kind, therefore the scope of the present invention is defined by claims.
Claims (10)
1. a kind of time series data multistep forecasting method based on correlation, it is characterised in that including：
Step 1：The time series data to be measured being made up of multiple time point data is read, according in the time series data to be measured
Correlation between adjacent time point data sets output length, according to the output length and regression order set in advance, treats
Survey time series data to be normalized, and input and output are split as to the time series data to be measured after normalized
Data are to set, the prediction step needed for user's input, the output number to control final predicted value, the wherein input and output
Data are the output length to the number of set；
Step 2：Using the inputoutput data to gathering the training data as multi output Gaussian process model, training generation prediction
Model, the forecast model is inputted by the time series data to be measured, obtains initial predicted value, and enter line number to the initial predicted value
Final predicted value is obtained according to renormalization, the number of the wherein initial predicted value and the final predicted value is the output length；
Step 3：Judge whether the total number of the final predicted value is more than or equal to the prediction step, it is no if then algorithm terminates
Then return to the step 2 to continue to predict, untill the number of the final predicted value is more than or equal to the prediction step.
2. the time series data multistep forecasting method as claimed in claim 1 based on correlation, it is characterised in that the step
1 includes：
Step 11, the time series data to be measured is translated x successively, generates x time series data and the time sequence to be measured
Column data forms x+1 time series data together, if the Pearson correlation coefficient between the x+1 time series data is more than
0.5, then judge that there is correlation in the time series data to be measured between adjacent x+1 time point data, wherein x is positive integer.
3. the time series data multistep forecasting method as claimed in claim 2 based on correlation, it is characterised in that step 1
In the output length be positive integer less than or equal to x+1.
4. the time series data multistep forecasting method as claimed in claim 1 based on correlation, it is characterised in that the step
Also include the time series data to be measured is split as into inputoutput data pair by the way of sliding window in 1, and this is defeated
Enter output data and the inputoutput data is combined into collection to set.
5. the time series data multistep forecasting method as claimed in claim 1 based on correlation, it is characterised in that the step
2 also include the newest time point data of time point in the time series data to be measured inputting the prediction mould as input data
Type, wherein the input data number are the regression order.
6. a kind of time series data multi-step prediction system based on correlation, it is characterised in that including：
Setup module：For reading the time series data to be measured being made up of multiple time point data, according to the time sequence to be measured
Correlation in column data between adjacent time point data sets output length, according to the output length and recurrence rank set in advance
Number, time series data to be measured is normalized, and the time series data to be measured after normalized is split as
Inputoutput data wherein should to set, the prediction step needed for user's input, the output number to control final predicted value
Inputoutput data is the output length to the number of set；
Prediction module：For the inputoutput data, to gathering the training data as multi output Gaussian process model, to be trained
Forecast model is generated, the time series data to be measured is inputted into the forecast model, initial predicted value is obtained, and to the tentative prediction
Value carries out data renormalization and obtains final predicted value, and the number of the wherein initial predicted value and the final predicted value is that this is defeated
Go out length；
Judge module：For judging whether the total number of the final predicted value is more than or equal to the prediction step, if then algorithm
Terminate, otherwise call the prediction module to continue to predict again, until the number of the final predicted value is walked more than or equal to the prediction
Untill length.
7. the time series data multi-step prediction system as claimed in claim 1 based on correlation, it is characterised in that the setting
Module includes：
Correlation prediction module, for the time series data to be measured to be translated into x successively, x time series data of generation and
The time series data to be measured forms x+1 time series data together, if the Pearson came between the x+1 time series data
Coefficient correlation is more than 0.5, then judges there is correlation in the time series data to be measured between adjacent x+1 time point data, its
Middle x is positive integer.
8. the time series data multi-step prediction system as claimed in claim 7 based on correlation, it is characterised in that mould is set
The output of this in block length is the positive integer less than or equal to x+1.
9. the time series data multi-step prediction system as claimed in claim 1 based on correlation, it is characterised in that the setting
Also include in module, the time series data to be measured is split as inputoutput data pair by the way of sliding window, and will
The inputoutput data is combined into the inputoutput data to collection to set.
10. the time series data multi-step prediction system as claimed in claim 1 based on correlation, it is characterised in that this is pre-
Surveying module also includes the newest time point data of time point in the time series data to be measured being somebody's turn to do as input data input
Forecast model, wherein the input data number are the regression order.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710390314.0A CN107274009A (en) | 2017-05-27 | 2017-05-27 | A kind of time series data multistep forecasting method and system based on correlation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710390314.0A CN107274009A (en) | 2017-05-27 | 2017-05-27 | A kind of time series data multistep forecasting method and system based on correlation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107274009A true CN107274009A (en) | 2017-10-20 |
Family
ID=60065641
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710390314.0A Pending CN107274009A (en) | 2017-05-27 | 2017-05-27 | A kind of time series data multistep forecasting method and system based on correlation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107274009A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109298351A (en) * | 2018-09-30 | 2019-02-01 | 清华大学深圳研究生院 | A kind of new energy on-vehicle battery remaining life estimation method based on model learning |
| CN109408545A (en) * | 2018-09-30 | 2019-03-01 | 中科驭数（北京）科技有限公司 | Target polymerization operation processing method and device |
| CN110555537A (en) * | 2018-05-30 | 2019-12-10 | 微软技术许可有限责任公司 | Multi-factor multi-time point correlated prediction |
| CN111046338A (en) * | 2018-10-11 | 2020-04-21 | 国际商业机器公司 | Multi-step predictive advance using complex valued vector autoregression |
| CN112070324A (en) * | 2020-11-11 | 2020-12-11 | 江西鹭鹭行科技有限公司 | Site passenger flow volume prediction method and device, readable storage medium and computer equipment |
| CN113297439A (en) * | 2021-05-31 | 2021-08-24 | 中电福富信息科技有限公司 | Time series data feature extraction and classification method |
-
2017
- 2017-05-27 CN CN201710390314.0A patent/CN107274009A/en active Pending
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110555537A (en) * | 2018-05-30 | 2019-12-10 | 微软技术许可有限责任公司 | Multi-factor multi-time point correlated prediction |
| CN109298351A (en) * | 2018-09-30 | 2019-02-01 | 清华大学深圳研究生院 | A kind of new energy on-vehicle battery remaining life estimation method based on model learning |
| CN109408545A (en) * | 2018-09-30 | 2019-03-01 | 中科驭数（北京）科技有限公司 | Target polymerization operation processing method and device |
| CN109408545B (en) * | 2018-09-30 | 2019-10-25 | 中科驭数（北京）科技有限公司 | Target polymerization operation processing method and device |
| CN111046338A (en) * | 2018-10-11 | 2020-04-21 | 国际商业机器公司 | Multi-step predictive advance using complex valued vector autoregression |
| US11308414B2 (en) | 2018-10-11 | 2022-04-19 | International Business Machines Corporation | Multi-step ahead forecasting using complex-valued vector autoregregression |
| CN111046338B (en) * | 2018-10-11 | 2024-04-19 | 国际商业机器公司 | Multi-step advance prediction using complex-valued vector autoregressions |
| CN112070324A (en) * | 2020-11-11 | 2020-12-11 | 江西鹭鹭行科技有限公司 | Site passenger flow volume prediction method and device, readable storage medium and computer equipment |
| CN113297439A (en) * | 2021-05-31 | 2021-08-24 | 中电福富信息科技有限公司 | Time series data feature extraction and classification method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107274009A (en) | A kind of time series data multistep forecasting method and system based on correlation | |
| CN109165664B (en) | Attribute-missing data set completion and prediction method based on generation of countermeasure network | |
| CN110135635B (en) | Regional power saturated load prediction method and system | |
| CN109241291A (en) | Knowledge mapping optimal path inquiry system and method based on deeply study | |
| CN111079989B (en) | DWT-PCA-LSTM-based water supply amount prediction device for water supply company | |
| CN111126562B (en) | Target algorithm fitting method based on neural network, terminal and application | |
| CN108876044B (en) | Online content popularity prediction method based on knowledge-enhanced neural network | |
| CN109325613B (en) | Method for online predicting static voltage stability margin by using parallel sequential extreme learning machine | |
| CN115577872B (en) | Structured data prediction optimization method based on multi-energy agent deep reinforcement learning | |
| CN113988358A (en) | Carbon emission index prediction and treatment method based on transfer reinforcement learning | |
| CN113935489A (en) | Variational quantum model TFQ-VQA based on quantum neural network and two-stage optimization method thereof | |
| CN115759321A (en) | Intelligent reactive power compensation method taking minimum line loss as optimization target | |
| CN116107279A (en) | Flow industrial energy consumption multi-objective optimization method based on attention depth neural network | |
| CN110717264A (en) | Improved strength pareto evolutionary algorithm for multi-objective optimization design of product appearance | |
| CN115906303A (en) | Planar microwave filter design method and device based on machine learning | |
| CN112308298A (en) | Multi-scenario performance index prediction method and system for semiconductor production line | |
| Nguyen et al. | InfoCNF: An efficient conditional continuous normalizing flow with adaptive solvers | |
| JPH06337852A (en) | Time series prediction method by neural network | |
| US20230342626A1 (en) | Model processing method and related apparatus | |
| CN111861002A (en) | Building cold and hot load prediction method based on data-driven Gaussian learning technology | |
| Jie et al. | Regularized flexible activation function combination for deep neural networks | |
| CN115081609A (en) | Acceleration method in intelligent decision, terminal equipment and storage medium | |
| CN114638421A (en) | Method for predicting requirement of generator set spare parts | |
| Teslyuk et al. | Synthesis of Neurocontroller for Intellectualization Tasks of Process Control Systems | |
| CN113111308A (en) | Symbolic regression method and system based on data-driven genetic programming algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171020 |