CN109993205A - Time Series Forecasting Methods, device, readable storage medium storing program for executing and electronic equipment - Google Patents
Time Series Forecasting Methods, device, readable storage medium storing program for executing and electronic equipment Download PDFInfo
- Publication number
- CN109993205A CN109993205A CN201910152919.5A CN201910152919A CN109993205A CN 109993205 A CN109993205 A CN 109993205A CN 201910152919 A CN201910152919 A CN 201910152919A CN 109993205 A CN109993205 A CN 109993205A
- Authority
- CN
- China
- Prior art keywords
- prediction
- index
- historical
- predicted
- target
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/211—Selection of the most significant subset of features
- G06F18/2113—Selection of the most significant subset of features by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/251—Fusion techniques of input or preprocessed data
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
This disclosure relates to a kind of Time Series Forecasting Methods, device, readable storage medium storing program for executing and electronic equipment.The described method includes: the time series data at current time is separately input into multiple decision models corresponding with index to be predicted, multiple first prediction results of the index to be predicted for subsequent time are obtained;The multiple first prediction result is merged, determines the second prediction result of the index to be predicted for the subsequent time.Thus, corresponding decision model need to be only used when carrying out time prediction, without carrying out Mathematical Fitting to data, simultaneously, multiple first prediction results are merged to obtain the second prediction result, the accuracy that the second prediction result of prediction index can be treated with lift pins, the accuracy of the time series forecasting of prediction index is treated so as to lift pins.
Description
Technical field
This disclosure relates to field of computer technology, and in particular, to a kind of Time Series Forecasting Methods, readable are deposited device
Storage media and electronic equipment.
Background technique
Time series forecasting technology can be based on the data with time correlation, and the development trend of estimated data is to solve reality
Problem.Currently, time series forecasting is widely used in different industries, for example, for the pre- of bank transaction amount situation of change
Survey, for exchange's the change of stock price rule prediction, be directed to application system key index (for example, when CPU, memory, http response
Between etc.) prediction etc. of future trend.
Existing time series forecasting carries out the changing rule that Mathematical Fitting carrys out prediction data generally by data, from
And it is predicted.But with the fast development of computer software technology, data volume, data scale are increasing, time series
The complexity of data is also higher and higher, for example, multi objective time series, existing time series data prediction technique is being applied
When cause predictablity rate not high due to complex data, be not able to satisfy the demand of practical application.
Summary of the invention
Purpose of this disclosure is to provide a kind of Time Series Forecasting Methods, device, readable storage medium storing program for executing and electronic equipment, with
Realize the time series forecasting for being directed to multi objective time series data.
To achieve the goals above, according to the disclosure in a first aspect, provide a kind of Time Series Forecasting Methods, the side
Method includes:
The time series data at current time is separately input into multiple decision models corresponding with index to be predicted, is obtained
For multiple first prediction results of the index to be predicted of subsequent time;
The multiple first prediction result is merged, determines the index to be predicted for being directed to the subsequent time
Second prediction result.
Optionally, multiple decision models corresponding with the index to be predicted construct in the following way:
Obtain historical time sequence data;
According to the historical time sequence data and the index to be predicted, multiple training datasets are constructed;
The decision model is constructed for training dataset described in each.
Optionally, the historical time sequence data includes a variety of pre-set levels in corresponding index of multiple historical junctures
Value, the index to be predicted are one of described a variety of pre-set levels;
It is described to construct multiple training datasets according to the historical time sequence data and the index to be predicted, it wraps
It includes:
Circulation execute from the historical juncture in the multiple historical juncture in addition to a nearest historical juncture choose to
Few historical juncture, and choose a variety of pre-set levels, for selected each historical juncture, according to selected a variety of pre-
If index the historical juncture corresponding index value and the index to be predicted the historical juncture subsequent time institute
Corresponding index value constructs the operation of a training dataset, until the sum of the training dataset constructed reaches
Until preset quantity, the preset quantity is the positive integer greater than 1.
Optionally, the historical time sequence data got is X=[x1, x2, x3..., xt], wherein xi=
[xi,r1, xi,r2..., xi,rn]T, xiIndicate n kind pre-set level in i-th of historical juncture corresponding index value, wherein n is
Positive integer greater than 1;
The circulation execution is selected from the historical juncture in the multiple historical juncture in addition to a nearest historical juncture
At least one historical juncture is taken, and chooses a variety of pre-set levels, for selected each historical juncture, according to selected more
Kind of pre-set level the historical juncture corresponding index value and the index to be predicted the historical juncture lower a period of time
Corresponding index value is carved, the operation of a training dataset is constructed, until the sum of the training dataset constructed
Until reaching preset quantity, comprising:
First historical juncture corresponding to the historical time sequence data is each into the t-1 historical juncture
Moment is successively used as the target histories moment, and is directed to each target histories moment, from the historical time sequence data
A variety of pre-set levels are selected to go through in target histories moment corresponding index value, and by selected index value with described
Next historical juncture corresponding index value group of the index to be predicted described in history time series data at the target histories moment
As target histories moment corresponding column vector described in q-th of training dataset, the initial value of q is 1;
If q is less than the preset quantity, after q=q+1, return corresponding to the historical time sequence data the
Each moment of one historical juncture into the t-1 historical juncture is successively used as the target histories moment, and for each described
The target histories moment selects a variety of pre-set levels respectively right at the target histories moment from the historical time sequence data
The index value answered, and by index to be predicted described in selected index value and the historical time sequence data in the target
It is corresponding that next historical juncture of historical juncture corresponding index value group becomes the target histories moment described in q-th of training dataset
Column vector the step of, until q be equal to the preset quantity until.
Optionally, selected a variety of pre-set levels are that there are the pre-set levels of correlativity with the index to be predicted.
Optionally, the decision model is decision tree, and first prediction result includes that the first prediction mean value and first are pre-
Variance is surveyed, second prediction result includes the second prediction mean value and the second prediction variance;
It is described to merge the multiple first prediction result, determine the finger to be predicted for being directed to the subsequent time
The second prediction result of target, comprising:
Successively it regard each first prediction result as the first prediction result of target, it is pre- for each target first
It surveys as a result, performing the following operations:
It is pre- according to the first prediction variance and last time fusion resulting second that include in first prediction result of target
Variance is surveyed, determines the weight of the corresponding decision model of first prediction result of target, wherein in initial fusion, described the
Two prediction variances are ∞;
Include according to the resulting second prediction mean value of last time fusion, the weight and the first prediction result of the target
First prediction mean value, determine this fusion it is resulting second prediction mean value, wherein in initial fusion, it is described second prediction
Mean value is 1;
The the first prediction variance and the weight for including according to first prediction result of target, determine this fusion gained
Second prediction variance.
Optionally, described to be merged according to the include in first prediction result of target first prediction variance and last time
Resulting second prediction variance, determines the weight of the corresponding decision model of first prediction result of target, comprising:
(1) determines the weight according to the following formula:
Wherein, knowFor the weight, vnowFor the first prediction variance that first prediction result of target includes, v (s) is
The resulting second prediction variance of last time fusion;
It is described that mean value, the first prediction result of the weight and the target are predicted according to last time fusion resulting second
Including first prediction mean value, determine this fusion it is resulting second prediction mean value, comprising:
(2) determine the resulting second prediction mean value of this fusion according to the following formula:
P (s+1)=know*pnow+(1-know)*p(s) (2)
Wherein, knowFor the weight, pnowFor the first prediction mean value that first prediction result of target includes, p (s) is
The resulting second prediction mean value of last time fusion, p (s+1) are the resulting second prediction mean value of this fusion;
The first prediction variance and the weight for including according to first prediction result of target, determines this fusion
Resulting second prediction variance, comprising:
(3) determine the resulting second prediction variance of this fusion according to the following formula:
V (s+1)=know*vnow (3)
Wherein, knowFor the weight, vnowFor the first prediction variance that first prediction result of target includes, v (s+1)
For the resulting second prediction variance of this fusion.
According to the second aspect of the disclosure, a kind of time series forecasting device is provided, described device includes:
Module is obtained, it is corresponding with index to be predicted multiple for being separately input into the time series data at current time
Decision model obtains multiple first prediction results of the index to be predicted for subsequent time;
Fusion Module determines the institute for being directed to the subsequent time for merging the multiple first prediction result
State the second prediction result of index to be predicted.
Optionally, described device further includes building module, and the building module is for constructing and the index pair to be predicted
The multiple decision models answered, the building module include:
Acquisition submodule, for obtaining historical time sequence data;
First building submodule, for constructing more according to the historical time sequence data and the index to be predicted
A training dataset;
Second building submodule, for constructing the decision model for each described training dataset.
Optionally, the historical time sequence data includes a variety of pre-set levels in corresponding index of multiple historical junctures
Value, the index to be predicted are one of described a variety of pre-set levels;
Described first, which constructs submodule, includes:
Circulation building submodule, executes from the multiple historical juncture in addition to a nearest historical juncture for recycling
Historical juncture in choose at least one historical juncture, and choose a variety of pre-set levels, for selected each historical juncture,
It is gone through in the historical juncture corresponding index value and the index to be predicted at this according to selected a variety of pre-set levels
Index value corresponding to the subsequent time at history moment constructs the operation of a training dataset, until the instruction constructed
Until the sum of white silk data set reaches preset quantity, the preset quantity is the positive integer greater than 1.
Optionally, the historical time sequence data got is X=[x1, x2, x3..., xt], wherein xi=
[xi,r1, xi,r2..., xi,rn]T, xiIndicate n kind pre-set level in i-th of historical juncture corresponding index value, wherein n is
Positive integer greater than 1;
The circulation building submodule is used for first historical juncture corresponding to the historical time sequence data extremely
Each moment in the t-1 historical juncture is successively used as the target histories moment, and is directed to each target histories moment, from
Select a variety of pre-set levels in target histories moment corresponding index value in the historical time sequence data, and will
Selected index value and index to be predicted described in the historical time sequence data are next the target histories moment
Historical juncture corresponding index value group becomes target histories moment corresponding column vector described in q-th of training dataset, and q's is first
Initial value is 1;If q is less than the preset quantity, after q=q+1, return corresponding to the historical time sequence data the
Each moment of one historical juncture into the t-1 historical juncture is successively used as the target histories moment, and for each described
The target histories moment selects a variety of pre-set levels respectively right at the target histories moment from the historical time sequence data
The index value answered, and by index to be predicted described in selected index value and the historical time sequence data in the target
It is corresponding that next historical juncture of historical juncture corresponding index value group becomes the target histories moment described in q-th of training dataset
Column vector, until q be equal to the preset quantity until.
Optionally, selected a variety of pre-set levels are that there are the pre-set levels of correlativity with the index to be predicted.
Optionally, the decision model is decision tree, and first prediction result includes that the first prediction mean value and first are pre-
Variance is surveyed, second prediction result includes the second prediction mean value and the second prediction variance;
The Fusion Module includes:
Submodule is merged, for successively regarding each first prediction result as the first prediction result of target, for every
A the first prediction result of the target, performs the following operations:
It is pre- according to the first prediction variance and last time fusion resulting second that include in first prediction result of target
Variance is surveyed, determines the weight of the corresponding decision model of first prediction result of target, wherein in initial fusion, described the
Two prediction variances are ∞;
Include according to the resulting second prediction mean value of last time fusion, the weight and the first prediction result of the target
First prediction mean value, determine this fusion it is resulting second prediction mean value, wherein in initial fusion, it is described second prediction
Mean value is 1;
The the first prediction variance and the weight for including according to first prediction result of target, determine this fusion gained
Second prediction variance.
Optionally, the fusion submodule determines the weight for (1) according to the following formula:
Wherein, knowFor the weight, vnowFor the first prediction variance that first prediction result of target includes, v (s) is
The resulting second prediction variance of last time fusion;
The fusion submodule determines the resulting second prediction mean value of this fusion for (2) according to the following formula:
P (s+1)=know*pnow+(1-know)*p(s) (2)
Wherein, knowFor the weight, pnowFor the first prediction mean value that first prediction result of target includes, p (s) is
The resulting second prediction mean value of last time fusion, p (s+1) are the resulting second prediction mean value of this fusion;
The fusion submodule determines the resulting second prediction variance of this fusion for (3) according to the following formula:
V (s+1)=know*vnow (3)
Wherein, knowFor the weight, vnowFor the first prediction variance that first prediction result of target includes, v (s+1)
For the resulting second prediction variance of this fusion.
According to the third aspect of the disclosure, a kind of computer readable storage medium is provided, computer program is stored thereon with,
The step of disclosure first aspect the method is realized when the program is executed by processor.
According to the fourth aspect of the disclosure, a kind of electronic equipment is provided, comprising:
Memory is stored thereon with computer program;
Processor, for executing the computer program in the memory, to realize described in disclosure first aspect
The step of method.
Through the above technical solutions, the time series data at current time is separately input into corresponding with index to be predicted
Multiple decision models obtain multiple first prediction results of the index to be predicted for subsequent time, and pre- by multiple first
It surveys result to be merged, determines the second prediction result of the index to be predicted for subsequent time.In this way, by by current time
Time series data input the corresponding multiple decision models of index to be predicted respectively, can obtain for subsequent time to pre-
Multiple first prediction results for surveying index, then multiple first prediction results are merged, with determine for subsequent time to
Second prediction result of prediction index.Corresponding decision model need to be only used when carrying out time prediction as a result, without to data
Mathematical Fitting is carried out, meanwhile, multiple first prediction results of fusion, can be with lift pins to finger to be predicted to obtain the second prediction result
The accuracy of the second prediction result of target, the accuracy of the time series forecasting of prediction index is treated so as to lift pins.
Other feature and advantage of the disclosure will the following detailed description will be given in the detailed implementation section.
Detailed description of the invention
Attached drawing is and to constitute part of specification for providing further understanding of the disclosure, with following tool
Body embodiment is used to explain the disclosure together, but does not constitute the limitation to the disclosure.In the accompanying drawings:
Fig. 1 is the flow chart of the Time Series Forecasting Methods provided according to an embodiment of the present disclosure;
Fig. 2 is multiple decision models corresponding with index to be predicted in the Time Series Forecasting Methods provided according to the disclosure
A kind of flow chart of example implementations of the building mode of type;
Fig. 3 is multiple decision models corresponding with index to be predicted in the Time Series Forecasting Methods provided according to the disclosure
The flow chart of another example implementations of the building mode of type;
Fig. 4 A is a kind of illustrative scene signal of decision tree in the Time Series Forecasting Methods provided according to the disclosure
Figure;
Fig. 4 B is in the Time Series Forecasting Methods provided according to the disclosure, and the illustrative scene of the another kind of decision tree is shown
It is intended to;
Fig. 4 C is in the Time Series Forecasting Methods provided according to the disclosure, and the illustrative scene of the another kind of decision tree is shown
It is intended to;
Fig. 5 is the block diagram of the time series forecasting device provided according to an embodiment of the present disclosure;
Fig. 6 is the block diagram of a kind of electronic equipment shown according to an exemplary embodiment.
Specific embodiment
It is described in detail below in conjunction with specific embodiment of the attached drawing to the disclosure.It should be understood that this place is retouched
The specific embodiment stated is only used for describing and explaining the disclosure, is not limited to the disclosure.
Before the method that the description disclosure provides, the application scenarios of method of disclosure are briefly described first.When
Between sequence refer to by the numerical value of same statistical indicator by its occur chronological order arrangement made of ordered series of numbers.For time sequence
Column, which carry out analysis, can predict future according to existing historical data, that is, time series forecasting.Time series is pre-
Survey technology can be based on the data with time correlation, and the development trend of estimated data is with solving practical problems.Currently, time series
Prediction is widely used in different industries, for example, becoming for the prediction of bank transaction amount situation of change, for exchange's share price
The prediction of law, the prediction for being directed to application system key index (for example, CPU, memory, http response time etc.) future trend
Deng.When carrying out time series forecasting based on time series data, the data at each moment may include several statistical indicators
In the numerical value at the moment.By taking the time series data for bank transaction amount as an example, synchronization can correspond to multiple statistics and refer to
Trading volume data under mark, such as the trading volume of arm's length dealing, the trading volume that transaction need to be authorized, the trading volume of error-prone transaction
Deng.
Fig. 1 is the flow chart of the Time Series Forecasting Methods provided according to an embodiment of the present disclosure.Such as Fig. 1 institute
Show, this method may comprise steps of.
In a step 11, by the time series data at current time be separately input into it is corresponding with index to be predicted it is multiple certainly
Plan model obtains multiple first prediction results of the index to be predicted for subsequent time.
Illustratively, the current share price of trading volume, stock market on the day of time series data can be, for example, bank, application
The data such as the CPU response time of system.The time series data at current time can show as a column vector, the column vector
The corresponding pre-set level of every row, each value of column vector are to correspond to the index value of corresponding pre-set level.For example, when current
The time series data at quarter is [1,2,3]T, then, the corresponding three kinds of pre-set levels of the time series data, and according to from top to bottom
Sequence, the corresponding index value of the first pre-set level is 1, and the corresponding index value of second of pre-set level is 2, the third is default
The corresponding index value of index is 3.By the time series data at current time be separately input into index to be predicted it is corresponding it is multiple certainly
Plan model, for example, if the time series data at current time is [1,2,3]T, there are three decision model (1~moulds of model altogether
Type 3), then by the time series data at current time [1,2,3]TIt is input to model 1 and the time series number by current time
According to [1,2,3]TIt is input to model 2 and the time series data [1,2,3] by current timeTIt is input to model 3.
Illustratively, decision model can be decision tree.The leaf node of each decision tree can correspond to a prediction result,
The time series data at above-mentioned current time is inputted to multiple decision models corresponding with index to be predicted respectively, to obtain needle
To multiple first prediction results of the index to be predicted of subsequent time.
In step 12, multiple first prediction results are merged, determines the index to be predicted for being directed to subsequent time
Second prediction result.
Multiple first prediction results are merged to the second prediction knot to determine the index to be predicted for subsequent time
Fruit.Illustratively, multiple first prediction results obtained through step 11 can be merged one by one, to obtain above-mentioned second prediction
As a result.For example, weight can be distributed for each decision model, after obtaining the first prediction result by decision model, in conjunction with
Each first prediction result and the corresponding weight of each first prediction result (namely obtain determining for each first prediction result
The weight of plan model), multiple first prediction results are merged, to obtain the second prediction result.
Through the above scheme, the time series data at current time is separately input into corresponding with index to be predicted multiple
Decision model obtains multiple first prediction results of the index to be predicted for subsequent time, and multiple first predictions is tied
Fruit is merged, and determines the second prediction result of the index to be predicted for subsequent time.In this way, by by current time when
Between sequence data input the corresponding multiple decision models of index to be predicted respectively, the finger to be predicted for subsequent time can be obtained
Multiple first prediction results of target, then multiple first prediction results are merged, to determine for the to be predicted of subsequent time
Second prediction result of index.Corresponding decision model need to be only used when carrying out time prediction as a result, without carrying out to data
Mathematical Fitting, meanwhile, multiple first prediction results are merged to obtain the second prediction result, can treat prediction index with lift pins
The accuracy of second prediction result treats the accuracy of the time series forecasting of prediction index so as to lift pins.
In order to make those skilled in the art more understand technical solution provided in an embodiment of the present invention, below to above
Corresponding steps are described in detail.
Firstly, the building mode for multiple decision models corresponding with index to be predicted is illustrated.In one kind
In possible embodiment, multiple decision models corresponding with index to be predicted can construct in the following way, such as Fig. 2 institute
Show.
In step 21, historical time sequence data is obtained.
Historical time sequence data may include a variety of pre-set levels in corresponding index value of multiple historical junctures, to be predicted
Index can be one of a variety of pre-set levels.Illustratively, the historical time sequence data got is X=[x1, x2,
x3..., xt], wherein xiIndicate n kind pre-set level in i-th of historical juncture corresponding index value, xi=[xi,r1,
xi,r2..., xi,rn]T, xi,r1~xi,rnN index value corresponding to as above-mentioned i-th of historical juncture pre-set level.Wherein, n
For the positive integer greater than 1, index to be predicted is one of n kind pre-set level.
In step 22, according to historical time sequence data and index to be predicted, multiple training datasets are constructed.
In a kind of possible embodiment, step 22 be may comprise steps of:
Circulation executes and chooses at least one from the historical juncture in multiple historical junctures in addition to a nearest historical juncture
A historical juncture, and a variety of pre-set levels are chosen, for selected each historical juncture, preset according to selected a variety of this
Index is in the historical juncture corresponding index value and index to be predicted corresponding to the subsequent time of the historical juncture
Index value constructs the operation of a training dataset, until the sum of the training dataset constructed reaches preset quantity.Its
In, preset quantity is the positive integer greater than 1.
When choosing at least one history from the historical juncture in multiple historical junctures in addition to a nearest historical juncture
It carves, wherein the selected historical juncture can be in multiple historical junctures that historical time sequence data is included except nearest one
It other all historical junctures except historical juncture, is wrapped alternatively, the selected historical juncture can be historical time sequence data
A part in other historical junctures in the multiple historical junctures contained in addition to a nearest historical juncture.Illustratively, if going through
Contain t historical juncture in history time series data, be followed successively by the 1st historical juncture~t-th of historical juncture, then selected history
Moment can be the 1st historical juncture~the t-1 historical juncture.A variety of pre-set levels are chosen, and for each of selected
Historical juncture exists according to selected a variety of pre-set levels in the historical juncture corresponding index value and index to be predicted
Index value corresponding to the subsequent time of the historical juncture constructs a training dataset.Wherein, index to be predicted can be selected
One of a variety of pre-set levels taken, alternatively, index to be predicted can be in historical time sequence in all pre-set levels except selected
One of other pre-set levels except a variety of pre-set levels taken.Illustratively, if the selected historical juncture is first history
Quarter and second historical juncture, and in pre-set level w1~w5, selected pre-set level is w2 and w3, and index to be predicted is
W5 obtains index w2, w3 respective index value and w5 under first historical juncture and exists then being directed to first historical juncture
Index value corresponding to second historical juncture constitutes column vector 1, meanwhile, for second historical juncture, obtain index w2,
W3 respective index value and w5 index value corresponding to the third historical juncture under second historical juncture, constitute column to
Amount 2, then column vector 1 and column vector 2 collectively form a training dataset.
Circulation executes the operation of above-mentioned one training dataset of building, and every circulation executes the primary operation, then to construct one more
Training dataset can need not then be continued cycling through and execute the operation when the training dataset constructed reaches preset quantity.
In a kind of possible embodiment, as shown in figure 3, circulation is executed removes a nearest history from multiple historical junctures
At least one historical juncture is chosen in historical juncture except moment, and chooses a variety of pre-set levels, for each of selected
Historical juncture exists according to selected a variety of pre-set levels in the historical juncture corresponding index value and index to be predicted
Index value corresponding to the subsequent time of the historical juncture constructs the operation of a training dataset, until the training number constructed
It may comprise steps of until reaching preset quantity according to the sum of collection.
In step 31, by first historical juncture to the t-1 historical juncture corresponding to historical time sequence data
In each moment be successively used as the target histories moment, and each target histories moment is directed to, from historical time sequence data
Select a variety of pre-set levels in target histories moment corresponding index value, and by selected index value and historical time sequence
The corresponding index value group of next historical juncture of index to be predicted at the target histories moment becomes q-th of training data in column data
Concentrate target histories moment corresponding column vector.
Wherein, the initial value of q is 1.Illustratively, if for the historical time sequence data X=[x in above-mentioned example1, x2,
x3..., xt] building training dataset, firstly, by first historical juncture corresponding to historical time sequence data X to t-1
Each historical juncture in a historical juncture is successively used as the target histories moment, selects from historical time sequence data X more
Kind pre-set level is in target histories moment corresponding index value.Assuming that the pre-set level of selection is from xi=[xi,r1,
xi,r2..., xi,rn]TN pre-set level in the m pre-set level (wherein, m≤n) that selects, then, selected index value
It can be [xi,c1, xi,c2..., xi,cm]T, wherein xi,c1~xi,cmFor from xi,r1~xi,rnSeveral pre-set levels pair of middle selection
Several index values answered.M pre-set level is selected from existing pre-set level, utilizes this m pre-set level building training
For data set to construct decision model, m can be less than or equal to the sum of pre-set level, in this way can be with training for promotion data
The applicability of collection.It should be noted that whether xi,c1~xi,cmOr xi,r1~xi,rn, subscript is used only as distinguishing default finger
Target type, no practical reference meaning.Later, index to be predicted in selected index value and historical time sequence data is existed
Corresponding index value group of next historical juncture at target histories moment becomes q-th of training data and concentrates the target histories moment corresponding
Column vector.Illustratively, it is assumed that historical time sequence data X=[x1, x2, x3..., xt] (wherein, xi=[xi,r1,
xi,r2..., xi,rn]T) building q-th of training dataset be Y (q)=[y1, y2, y3..., yt-1], wherein yiIt is this q-th
Training data concentrates i-th of column vector, that is, corresponds to the column vector of the i-th historical juncture.And yi=[xi,c1, xi,c2...,
xi,cm, z]T, [xi,c1, xi,c2..., xi,cm] meaning be given above, z is to be predicted in historical time sequence data
Next historical juncture corresponding index value of the index in the historical juncture, in this example, if index to be predicted is xi,r5It is corresponding
Index, then z is xi+1,r5, that is, i+1 historical juncture xi,r5Index value corresponding to corresponding index.
In the step 32, judge whether q is less than preset quantity.
If determining that q is less than preset quantity through step 32, return step 31 after step 33 is executed;If determining q etc. through step 32
In preset quantity, illustrate that step 22 is completed, then can execute subsequent corresponding steps.
In step 33, q=q+1 is enabled.
Wherein, preset quantity is exactly the sum of training dataset to be built.The sum of training dataset to be built is just
It is the quantity for finally it is expected the training dataset constructed, which, which can be, is artificially arranged.If q is less than preset quantity,
Illustrate the sum of the also insufficient training dataset to be built of the training dataset quantity constructed, therefore can enable q as unit of 1
Above-mentioned steps 31 are returned to from after increasing from increasing, and in q, to establish next training dataset.And when q is equal to preset quantity, explanation
The q training dataset currently established has met the requirement of the sum of training dataset to be built, can not resettle next
A training dataset continues subsequent step thus may determine that step 22 is completed.
It should be noted that when constructing multiple training datasets for index to be predicted, the training data that constructs every time
The quantity for collecting selected a variety of pre-set levels may be the same or different, and the disclosure is to this without limiting.
In a kind of possible embodiment, during constructing multiple training datasets, training data is constructed each time
Selected a variety of pre-set levels can be randomly selected when collection, can guarantee the randomness of selected pre-set level in this way,
It thereby may be ensured that the randomness of training dataset data collection.
In alternatively possible embodiment, during constructing multiple training datasets, building training each time
It is inequality that selected a variety of pre-set levels, which can be randomly selected and constructed each training dataset, when data set
, that is to say, that guarantee again on the basis of randomly choosing pre-set level incomplete each other between each training dataset of building
It is identical.In this way, can on the basis of guaranteeing the randomness of selected pre-set level, guarantee training dataset without repeatability, from
And can guarantee subsequent builds decision model without repeatability, promote the quality of decision model.
In alternatively possible embodiment, during constructing multiple training datasets, training number is constructed each time
Selected a variety of pre-set levels, which can be, when according to collection carries out selection according to preset rules, for example, selecting several adjacent
Pre-set level simultaneously gradually translates, until all pre-set levels were selected.In this manner it is ensured that selected pre-set level is more
Sample guarantees the data area that training dataset is covered, to keep the quality of the decision model constructed in subsequent step higher.
In alternatively possible embodiment, during constructing multiple training datasets, training number is constructed each time
It can be that there are the indexs of correlativity with index to be predicted according to selected a variety of pre-set levels when collection.Illustratively, there are phases
The index of pass relationship may, for example, be there are positive correlation, there are negative correlativing relations etc..In this manner it is ensured that selected is default
Index treats prediction index with specific aim, thereby may be ensured that the specific aim of training dataset data collection, is promoted subsequent
Forecasting accuracy.
Through the above scheme, by selecting a variety of pre-set levels from historical time sequence data, and according to selected default
The index value that index value and the subsequent time at each moment of the index at each moment correspond to index to be predicted generates training data
Collection can provide data to constitute the multiple training datasets for being directed to the index to be predicted for the generation of follow-up decision model
Prepare.
After step 22 is all finished, step 23 can be executed.
In step 23, decision model is constructed for each training dataset.
In a kind of possible embodiment, decision model can be decision tree.In this embodiment, for each
A training dataset constructs decision model, that is, constructs decision tree for each training dataset.It illustratively, can will be every
One training dataset as decision tree root node to construct decision tree.It will be carried out briefly for the building of decision tree below
It is bright.
Firstly, using training dataset as the root node of decision tree.
Assuming that a certain training dataset is A, and A=[a1, a2, a3, a4], also, items are as follows in A:
A1=[b1, c1, d1, e2]T;
A2=[b2, c2, d2, e3]T;
A3=[b3, c3, d3, e4]T;
A4=[b4, c4, d4, e5]T。
Wherein, a1~a4 is followed successively by the 1st moment~the 4th moment column vector, and b, c, d are respectively represented from historical time sequence
The three kinds of pre-set levels selected in data, e2~e5 represent in historical time sequence data index to be predicted successively at the 2nd moment
To the 5th moment corresponding index value.It so, can be as shown in Figure 4 A using training dataset A as the root node of decision tree.
Later, the every kind of pre-set level respectively concentrated training data is selecting object, and selected pre-set level is corresponding
Each index value is divided as the column vector that threshold value concentrates training data is divided, and divides concentrate training data each time
Column vector be divided into two parts.It should be noted that not including in the selected various pre-set levels for division here
Treat the selection of prediction index.For the training dataset A in example, b, c, d are to select from historical time sequence data
The three kinds of pre-set levels selected, e are index to be predicted, then the selected pre-set level for division is b, c, d without including
e.For another example if it be the three kinds of pre-set levels selected from historical time sequence data that a training data, which concentrates b, c, d, b for
Prediction index, then the selected pre-set level for division is b, c, d.
Illustratively, it can be divided in the following manner: the index value under selected pre-set level is less than or equal to
Divide threshold value column vector as a part, and using the index value under selected pre-set level be greater than divide threshold value column vector as
Another part.By taking above-mentioned training dataset A as an example, the result after once dividing can be divided into together for a1, a3, together
When a2, a4 be divided into together.Training dataset is divided into two parts column vector, that is, the column that training data is concentrated to
A child node into two child nodes of corresponding present node (being root node under initial situation) is respectively divided in amount, and calculates
Index to be predicted corresponds to the variance of index value in column vector in each child node in corresponding two child nodes of division result, obtains
Two variances, will wherein the greater as the corresponding variance of the division mode.For multiple division modes, by the smallest stroke of variance
Point mode is determined as final division mode.
It, can be respectively by pre-set level b, c, d alternatively object for the root node in Fig. 4 A.For example, pre- in selection
If after index b, the child node for corresponding to the root node can be determined using b1, b2, b3 as division threshold value respectively.Illustratively, if
Using b1 as threshold value is divided, then the column vector that index value corresponding with index b is less than or equal to b1 can be divided to root
In the left child node of node, while the column vector by index value corresponding with index b greater than b1 is divided to the corresponding right side of root node
In child node.If the corresponding column vector of left child node obtained through above-mentioned partition process is a1, a2, right child node is corresponding arrange to
Amount is a3, a4, then can then calculate the variance of the two values of e2 and e3, while calculating the variance of the two values of e4 and e5, and
Will wherein the greater as the corresponding variance of current division mode.It should be noted that corresponding in above-mentioned selected pre-set level
In the case that index value is equal to division threshold value, it can be divided to the left child node of present node, can also be divided to
The right child node of present node, the disclosure is to this without limiting.
, can also be with b2, b3, b4 for above-mentioned example, c1, c2, c3, c4, d1, d2, d3, d4 be respectively divide threshold value into
Row divides.After being respectively to divide threshold value with the above-mentioned various corresponding index values of pre-set level, each variance for will obtaining
It compares, division mode corresponding to minimum variance is determined as to final division mode.If with b1 in above-mentioned example
The corresponding variance of division mode to divide threshold value is minimum, then decision tree can be updated to as shown in Figure 4 B.
Later, for the leaf node of current decision tree, determine whether each leaf node meets preset division eventually
Only condition.Termination condition is divided if meeting, need not be divided again for the leaf node.If being unsatisfactory for dividing and terminating item
Part is then divided continuing with the leaf node, and the mode of division is consistent with the above-mentioned division mode for root node, herein
It is lack of repetition.It should be noted that selected pre-set level can be not when being divided for present node
It is selected, according to above-mentioned example, is if desired divided for the right child node of root node in above-mentioned Fig. 4 B, due to default
Index b had been selected, therefore while dividing to it can select from pre-set level c or pre-set level d.
Illustratively, preset division termination condition can be the whole quilts of pre-set level included in training dataset
It is elected to be the index for division, alternatively, it is zero that preset division termination condition, which can be the corresponding variance of present node,.
In the manner described above, can construct decision tree for current training dataset, the decision tree constructed it is each
The corresponding variance of a leaf node and a mean value, which, which is that index to be predicted is corresponding in column vector in leaf node, refers to
The variance of scale value, the mean value are that index to be predicted corresponds to the mean value of index value in column vector in leaf node.As shown in Figure 4 C,
There are three leaf nodes, respectively node 41, node 42 and node 43.So the corresponding variance of node 41 be e2 and e3 side
Difference, mean value are the mean value of e2 and e3;The corresponding variance of node 42 is 0, mean value e4;The corresponding variance of node 43 is 0, and mean value is
e5。
It should be noted that the mode of building decision tree belongs to the common knowledge of this field, for those skilled in the art
For be it will be apparent that description above is only used as brief exemplary illustration, the mode for constructing decision tree is not limited to
This, for other building modes, the disclosure is without repeating.
Using aforesaid way, using decision tree as decision model, the advantages of can use decision tree itself, calculation amount it is small and
Accuracy is high.
Below for merging multiple first prediction results in step 12, determine for next moment to pre-
The second prediction result for surveying index is illustrated.
In a kind of possible embodiment, decision model can be decision tree, by the time series data at current time
Multiple decision models corresponding with index to be predicted are separately input into, that is, the time series at current time is separately input into
Multiple decision trees corresponding with index to be predicted.After the time series data at current time is input to decision tree, according to certainly
Corresponding pre-set level is (for example, above for the division under the pre-set level mentioned in the explanation of building decision tree in plan tree
Threshold value) determine corresponding first prediction result of the time series data at current time.First prediction result may include that prediction is equal
Value and prediction variance.
In this embodiment, step 12 may comprise steps of:
Successively it regard each first prediction result as the first prediction result of target, for each the first prediction result of target,
It performs the following operations:
Resulting second prediction side is merged according to the first prediction variance for including in the first prediction result of target and last time
Difference determines the weight of the corresponding decision model of the first prediction result of target;
It is pre- according to last time merges resulting second prediction mean value, weight and the first prediction result of target include first
Mean value is surveyed, determines the resulting second prediction mean value of this fusion;
The the first prediction variance and weight for including according to the first prediction result of target determine that this fusion resulting second is pre-
Survey variance.
Wherein, in initial fusion, the second prediction variance is ∞.And in initial fusion, the second prediction mean value is 1.
Circulation executes aforesaid operations, constantly updates the second prediction variance and the second prediction mean value, it is known that the first all prediction results is equal
It was fused, and obtained the second prediction result.
Illustratively, (1) weight can be determined according to the following formula:
Wherein, knowFor weight, vnowFor the first prediction variance that the first prediction result of target includes, v (s) is last time fusion
Resulting second prediction variance.
Illustratively, (2) the resulting second prediction mean value of this fusion can be determined according to the following formula:
P (s+1)=know*pnow+(1-know)*p(s) (2)
Wherein, knowFor weight, pnowFor the first prediction mean value that the first prediction result of target includes, p (s) is last time fusion
Resulting second prediction mean value, p (s+1) are the resulting second prediction mean value of this fusion.
Illustratively, (3) the resulting second prediction variance of this fusion can be determined according to the following formula:
V (s+1)=know*vnow (3)
Wherein, knowFor weight, vnowFor the first prediction variance that the first prediction result of target includes, v (s+1) melts for this
Close resulting second prediction variance.
Assuming that co-existing in three the first prediction results, respectively mean value p1With variance v1, mean value p2With variance v2, mean value p3With
Variance v3.It is possible to be merged as follows.
Firstly, fusion mean value p1With variance v1:
P (1)=know*p1+(1-know) * p (0)=p1;
V (1)=know*v1=v1;
It is found that this fusion gained second predicts that variance is v1, this second prediction of fusion gained mean value is p1.At this point, also
There are the first prediction results not to be fused, therefore, can continue to merge, such as fusion mean value p2With variance v2:
It is found that this fusion gained second predicts that variance isThis fusion gained second predicts that mean value isAt this point, there is also the first prediction results not to be fused, therefore, it can continue to merge, such as fusion is equal
Value p3With variance v3, the process of fusion is similar to above, herein without description.
Passing through above-mentioned fusion mean value p3With variance v3Afterwards, the first prediction result not being fused has been not present, therefore can
Current fusion results are determined as the second prediction result.
It should be noted that can orderly be merged for multiple first prediction results, can also randomly select not by
First prediction result of fusion is merged one by one, has no influence for the acquisition of the second prediction result, therefore, the disclosure is to this
Without limiting.
By the above-mentioned means, each first prediction result is merged to obtain after obtaining multiple first prediction results
The second prediction result is obtained, can be guaranteed pre- with lift pins to the accuracy of the second prediction result of the index to be predicted of subsequent time
Mass metering.
Fig. 5 is the block diagram of the time series forecasting device provided according to an embodiment of the present disclosure.As shown in figure 5,
The device 50 may include:
Module 51 is obtained, it is corresponding with index to be predicted more for being separately input into the time series data at current time
A decision model obtains multiple first prediction results of the index to be predicted for subsequent time;
Fusion Module 52 is determined for merging the multiple first prediction result for the subsequent time
Second prediction result of the index to be predicted.
Optionally, described device 50 further includes building module, and the building module is for constructing and the index to be predicted
Corresponding multiple decision models, the building module include:
Acquisition submodule, for obtaining historical time sequence data;
First building submodule, for constructing more according to the historical time sequence data and the index to be predicted
A training dataset;
Second building submodule, for constructing the decision model for each described training dataset.
Optionally,
The historical time sequence data includes a variety of pre-set levels in corresponding index value of multiple historical junctures, it is described to
Prediction index is one of described a variety of pre-set levels;
Described first, which constructs submodule, includes:
Circulation building submodule, executes from the multiple historical juncture in addition to a nearest historical juncture for recycling
Historical juncture in choose at least one historical juncture, and choose a variety of pre-set levels, for selected each historical juncture,
It is gone through in the historical juncture corresponding index value and the index to be predicted at this according to selected a variety of pre-set levels
Index value corresponding to the subsequent time at history moment constructs the operation of a training dataset, until the instruction constructed
Until the sum of white silk data set reaches preset quantity, the preset quantity is the positive integer greater than 1.
Optionally, the historical time sequence data got is X=[x1, x2, x3..., xt], wherein xi=
[xi,r1, xi,r2..., xi,rn]T, xiIndicate n kind pre-set level in i-th of historical juncture corresponding index value, wherein n is
Positive integer greater than 1;
The circulation building submodule is used for first historical juncture corresponding to the historical time sequence data extremely
Each moment in the t-1 historical juncture is successively used as the target histories moment, and is directed to each target histories moment, from
Select a variety of pre-set levels in target histories moment corresponding index value in the historical time sequence data, and will
Selected index value and index to be predicted described in the historical time sequence data are next the target histories moment
Historical juncture corresponding index value group becomes target histories moment corresponding column vector described in q-th of training dataset, and q's is first
Initial value is 1;If q is less than the preset quantity, after q=q+1, return corresponding to the historical time sequence data the
Each moment of one historical juncture into the t-1 historical juncture is successively used as the target histories moment, and for each described
The target histories moment selects a variety of pre-set levels respectively right at the target histories moment from the historical time sequence data
The index value answered, and by index to be predicted described in selected index value and the historical time sequence data in the target
It is corresponding that next historical juncture of historical juncture corresponding index value group becomes the target histories moment described in q-th of training dataset
Column vector, until q be equal to the preset quantity until.
Optionally, selected a variety of pre-set levels are that there are the pre-set levels of correlativity with the index to be predicted.
Optionally, the decision model is decision tree, and first prediction result includes that the first prediction mean value and first are pre-
Variance is surveyed, second prediction result includes the second prediction mean value and the second prediction variance;
The Fusion Module includes:
Submodule is merged, for successively regarding each first prediction result as the first prediction result of target, for every
A the first prediction result of the target, performs the following operations:
It is pre- according to the first prediction variance and last time fusion resulting second that include in first prediction result of target
Variance is surveyed, determines the weight of the corresponding decision model of first prediction result of target, wherein in initial fusion, described the
Two prediction variances are ∞;
Include according to the resulting second prediction mean value of last time fusion, the weight and the first prediction result of the target
First prediction mean value, determine this fusion it is resulting second prediction mean value, wherein in initial fusion, it is described second prediction
Mean value is 1;
The the first prediction variance and the weight for including according to first prediction result of target, determine this fusion gained
Second prediction variance.
Optionally, the fusion submodule determines the weight for (1) according to the following formula:
Wherein, knowFor the weight, vnowFor the first prediction variance that first prediction result of target includes, v (s) is
The resulting second prediction variance of last time fusion;
The fusion submodule determines the resulting second prediction mean value of this fusion for (2) according to the following formula:
P (s+1)=know*pnow+(1-know)*p(s) (2)
Wherein, knowFor the weight, pnowFor the first prediction mean value that first prediction result of target includes, p (s) is
The resulting second prediction mean value of last time fusion, p (s+1) are the resulting second prediction mean value of this fusion;
The fusion submodule determines the resulting second prediction variance of this fusion for (3) according to the following formula:
V (s+1)=know*vnow (3)
Wherein, knowFor the weight, vnowFor the first prediction variance that first prediction result of target includes, v (s+1)
For the resulting second prediction variance of this fusion.
About the device in above-described embodiment, wherein modules execute the concrete mode of operation in related this method
Embodiment in be described in detail, no detailed explanation will be given here.
Fig. 6 is the block diagram of a kind of electronic equipment shown according to an exemplary embodiment.For example, electronic equipment 1900 can be with
It is provided as a server.Referring to Fig. 6, electronic equipment 1900 includes processor 1922, and quantity can be one or more, with
And memory 1932, for storing the computer program that can be executed by processor 1922.The computer stored in memory 1932
Program may include it is one or more each correspond to one group of instruction module.In addition, processor 1922 can be by
It is configured to execute the computer program, to execute above-mentioned Time Series Forecasting Methods.
In addition, electronic equipment 1900 can also include power supply module 1926 and communication component 1950, the power supply module 1926
It can be configured as the power management for executing electronic equipment 1900, which can be configured as realization electronic equipment
1900 communication, for example, wired or wireless communication.In addition, the electronic equipment 1900 can also include that input/output (I/O) connects
Mouth 1958.Electronic equipment 1900 can be operated based on the operating system for being stored in memory 1932, such as Windows
ServerTM, Mac OS XTM, UnixTM, LinuxTM etc..
In a further exemplary embodiment, a kind of computer readable storage medium including program instruction is additionally provided, it should
The step of above-mentioned Time Series Forecasting Methods are realized when program instruction is executed by processor.For example, the computer-readable storage
Medium can be the above-mentioned memory 1932 including program instruction, and above procedure instruction can be by the processor of electronic equipment 1900
1922 execute to complete above-mentioned Time Series Forecasting Methods.
The preferred embodiment of the disclosure is described in detail in conjunction with attached drawing above, still, the disclosure is not limited to above-mentioned reality
The detail in mode is applied, in the range of the technology design of the disclosure, a variety of letters can be carried out to the technical solution of the disclosure
Monotropic type, these simple variants belong to the protection scope of the disclosure.
It is further to note that specific technical features described in the above specific embodiments, in not lance
In the case where shield, it can be combined in any appropriate way.In order to avoid unnecessary repetition, the disclosure to it is various can
No further explanation will be given for the combination of energy.
In addition, any combination can also be carried out between a variety of different embodiments of the disclosure, as long as it is without prejudice to originally
Disclosed thought equally should be considered as disclosure disclosure of that.
Claims (10)
1. a kind of Time Series Forecasting Methods, which is characterized in that the described method includes:
The time series data at current time is separately input into multiple decision models corresponding with index to be predicted, is directed to
Multiple first prediction results of the index to be predicted of subsequent time;
The multiple first prediction result is merged, determines second of the index to be predicted for the subsequent time
Prediction result.
2. the method according to claim 1, wherein multiple decision models corresponding with the index to be predicted are
It constructs in the following way:
Obtain historical time sequence data;
According to the historical time sequence data and the index to be predicted, multiple training datasets are constructed;
The decision model is constructed for training dataset described in each.
3. according to the method described in claim 2, it is characterized in that, the historical time sequence data includes a variety of pre-set levels
In corresponding index value of multiple historical junctures, the index to be predicted is one of described a variety of pre-set levels;
It is described according to the historical time sequence data and the index to be predicted, construct multiple training datasets, comprising:
Circulation executes and chooses at least one from the historical juncture in the multiple historical juncture in addition to a nearest historical juncture
A historical juncture, and a variety of pre-set levels are chosen, for selected each historical juncture, according to selected a variety of default fingers
The historical juncture corresponding index value and the index to be predicted are marked on corresponding to the subsequent time of the historical juncture
Index value, construct a training dataset operation, until the sum of the training dataset constructed reaches default
Until quantity, the preset quantity is the positive integer greater than 1.
4. according to the method described in claim 3, it is characterized in that, the historical time sequence data got is X=[x1,
x2, x3..., xt], wherein xi=[xi,r1, xi,r2..., xi,rn]T, xiIndicate n kind pre-set level i-th of historical juncture respectively
Corresponding index value, wherein n is the positive integer greater than 1;
The circulation execute chosen from the historical juncture in the multiple historical juncture in addition to a nearest historical juncture to
Few historical juncture, and choose a variety of pre-set levels, for selected each historical juncture, according to selected a variety of pre-
If index the historical juncture corresponding index value and the index to be predicted the historical juncture subsequent time institute
Corresponding index value constructs the operation of a training dataset, until the sum of the training dataset constructed reaches
Until preset quantity, comprising:
By each moment of first historical juncture into the t-1 historical juncture corresponding to the historical time sequence data
It is successively used as the target histories moment, and is directed to each target histories moment, is selected from the historical time sequence data
A variety of pre-set levels are in target histories moment corresponding index value, and by selected index value and when the history
Between next historical juncture corresponding index value group of the index to be predicted described in sequence data at the target histories moment become
Target histories moment corresponding column vector described in q-th of training dataset, the initial value of q are 1;
If q is less than the preset quantity, after q=q+1, return will be first corresponding to the historical time sequence data
Each moment of the historical juncture into the t-1 historical juncture is successively used as the target histories moment, and is directed to each target
Historical juncture selects a variety of pre-set levels corresponding at the target histories moment from the historical time sequence data
Index value, and by index to be predicted described in selected index value and the historical time sequence data in the target histories
Corresponding index value group of next historical juncture at moment becomes target histories moment corresponding column described in q-th of training dataset
The step of vector, until q is equal to the preset quantity.
5. the method according to claim 3 or 4, which is characterized in that selected a variety of pre-set levels are with described to pre-
Surveying index, there are the pre-set levels of correlativity.
6. first prediction is tied the method according to claim 1, wherein the decision model is decision tree
Fruit includes the first prediction mean value and the first prediction variance, and second prediction result includes the second prediction mean value and the second prediction side
Difference;
It is described to merge the multiple first prediction result, determine the index to be predicted for being directed to the subsequent time
Second prediction result, comprising:
Each first prediction result is successively regard as the first prediction result of target, for each prediction of target first knot
Fruit performs the following operations:
Resulting second prediction side is merged according to the first prediction variance for including in first prediction result of target and last time
Difference determines the weight of the corresponding decision model of first prediction result of target, wherein in initial fusion, described second is pre-
Survey variance is ∞;
According to last time merges resulting second prediction mean value, the weight and the first prediction result of the target include the
One prediction mean value determines the resulting second prediction mean value of this fusion, wherein in initial fusion, the second prediction mean value
It is 1;
The the first prediction variance and weight for including according to first prediction result of target determines this fusion resulting the
Two prediction variances.
7. according to the method described in claim 6, it is characterized in that,
It is described pre- according to the include in first prediction result of target first prediction variance and last time fusion resulting second
Variance is surveyed, determines the weight of the corresponding decision model of first prediction result of target, comprising:
(1) determines the weight according to the following formula:
Wherein, knowFor the weight, vnowFor the first prediction variance that first prediction result of target includes, v (s) is last time
Merge resulting second prediction variance;
It is described to include according to the resulting second prediction mean value of last time fusion, the weight and the first prediction result of the target
First prediction mean value, determine this fusion it is resulting second prediction mean value, comprising:
(2) determine the resulting second prediction mean value of this fusion according to the following formula:
P (s+1)=know*pnow+(1-know)*p(s) (2)
Wherein, knowFor the weight, pnowFor the first prediction mean value that first prediction result of target includes, p (s) is last time
Resulting second prediction mean value is merged, p (s+1) is the resulting second prediction mean value of this fusion;
The first prediction variance and the weight for including according to first prediction result of target, determines this fusion gained
Second prediction variance, comprising:
(3) determine the resulting second prediction variance of this fusion according to the following formula:
V (s+1)=know*vnow (3)
Wherein, knowFor the weight, vnowFor the first prediction variance that first prediction result of target includes, v (s+1) is this
The secondary resulting second prediction variance of fusion.
8. a kind of time series forecasting device, which is characterized in that described device includes:
Module is obtained, for the time series data at current time to be separately input into multiple decisions corresponding with index to be predicted
Model obtains multiple first prediction results of the index to be predicted for subsequent time;
Fusion Module, for the multiple first prediction result to be merged, determine for the subsequent time it is described to
Second prediction result of prediction index.
9. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is held by processor
The step of any one of claim 1-7 the method is realized when row.
10. a kind of electronic equipment characterized by comprising
Memory is stored thereon with computer program;
Processor, for executing the computer program in the memory, to realize described in any one of claim 1-7
The step of method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910152919.5A CN109993205A (en) | 2019-02-28 | 2019-02-28 | Time Series Forecasting Methods, device, readable storage medium storing program for executing and electronic equipment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910152919.5A CN109993205A (en) | 2019-02-28 | 2019-02-28 | Time Series Forecasting Methods, device, readable storage medium storing program for executing and electronic equipment |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109993205A true CN109993205A (en) | 2019-07-09 |
Family
ID=67129484
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910152919.5A Pending CN109993205A (en) | 2019-02-28 | 2019-02-28 | Time Series Forecasting Methods, device, readable storage medium storing program for executing and electronic equipment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109993205A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114338416A (en) * | 2020-09-29 | 2022-04-12 | 中国移动通信有限公司研究院 | Space-time multi-index prediction method and device and storage medium |
CN116595356A (en) * | 2023-07-17 | 2023-08-15 | 腾讯科技(深圳)有限公司 | Time sequence signal prediction method and device, electronic equipment and storage medium |
CN117408364A (en) * | 2023-08-08 | 2024-01-16 | 湖北泰跃卫星技术发展股份有限公司 | Crop disease and pest index prediction method and device based on model fusion |
WO2024021630A1 (en) * | 2022-07-27 | 2024-02-01 | 京东城市(北京)数字科技有限公司 | Method and apparatus for calculating indicator data |
-
2019
- 2019-02-28 CN CN201910152919.5A patent/CN109993205A/en active Pending
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114338416A (en) * | 2020-09-29 | 2022-04-12 | 中国移动通信有限公司研究院 | Space-time multi-index prediction method and device and storage medium |
CN114338416B (en) * | 2020-09-29 | 2023-04-07 | 中国移动通信有限公司研究院 | Space-time multi-index prediction method and device and storage medium |
WO2024021630A1 (en) * | 2022-07-27 | 2024-02-01 | 京东城市(北京)数字科技有限公司 | Method and apparatus for calculating indicator data |
CN116595356A (en) * | 2023-07-17 | 2023-08-15 | 腾讯科技(深圳)有限公司 | Time sequence signal prediction method and device, electronic equipment and storage medium |
CN116595356B (en) * | 2023-07-17 | 2023-09-22 | 腾讯科技(深圳)有限公司 | Time sequence signal prediction method and device, electronic equipment and storage medium |
CN117408364A (en) * | 2023-08-08 | 2024-01-16 | 湖北泰跃卫星技术发展股份有限公司 | Crop disease and pest index prediction method and device based on model fusion |
CN117408364B (en) * | 2023-08-08 | 2024-05-24 | 湖北泰跃卫星技术发展股份有限公司 | Crop disease and pest index prediction method and device based on model fusion |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109993205A (en) | Time Series Forecasting Methods, device, readable storage medium storing program for executing and electronic equipment | |
Yu et al. | Multi-agent adversarial inverse reinforcement learning | |
Shen et al. | Mathematical modeling and multi-objective evolutionary algorithms applied to dynamic flexible job shop scheduling problems | |
CN109870903B (en) | Parameter optimization method, device and non-transitory computer readable medium | |
CN107590603B (en) | Based on the dispatching method and system for improving change neighborhood search and differential evolution algorithm | |
CN109299142A (en) | A kind of convolutional neural networks search structure method and system based on evolution algorithm | |
Peng et al. | Deep reinforcement learning approach for capacitated supply chain optimization under demand uncertainty | |
Busa-Fekete et al. | Multi-objective bandits: Optimizing the generalized gini index | |
CN110163368A (en) | Deep learning model training method, apparatus and system based on mixed-precision | |
CN106296669A (en) | A kind of image quality evaluating method and device | |
CN110118566B (en) | Robot operation path generation method and device | |
Koulinas et al. | Construction resource allocation and leveling using a threshold accepting–based hyperheuristic algorithm | |
JP2005527034A (en) | A dynamic method for solving large-scale discrete and continuous optimization problems. | |
CN113361680A (en) | Neural network architecture searching method, device, equipment and medium | |
CA2436352A1 (en) | Process and system for developing a predictive model | |
CN106610867B (en) | On-chip network task scheduling method and device | |
CN107392307A (en) | The Forecasting Methodology of parallelization time series data | |
CN107306207A (en) | Calculated and multiple target intensified learning service combining method with reference to Skyline | |
CN108304925A (en) | A kind of pond computing device and method | |
Guo et al. | Dynamic Fuzzy Logic Control of Genetic Algorithm Probabilities. | |
CN110009181A (en) | Distribution network transform measure and mistake load figureofmerit relevance method for digging and device | |
Brum et al. | Automatic design of heuristics for minimizing the makespan in permutation flow shops | |
CN109543879A (en) | Load forecasting method and device neural network based | |
CN110222824A (en) | Intelligent algorithm model is autonomously generated and evolvement method, system, device | |
CN109146060A (en) | A kind of method and device based on convolutional neural networks processing data |
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: 20190709 |