CN109977098A - Non-stationary time-series data predication method, system, storage medium and computer equipment - Google Patents
Non-stationary time-series data predication method, system, storage medium and computer equipment Download PDFInfo
- Publication number
- CN109977098A CN109977098A CN201910173925.9A CN201910173925A CN109977098A CN 109977098 A CN109977098 A CN 109977098A CN 201910173925 A CN201910173925 A CN 201910173925A CN 109977098 A CN109977098 A CN 109977098A
- Authority
- CN
- China
- Prior art keywords
- subsequence
- component
- prediction
- prediction model
- obtains
- 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
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention relates to a kind of Non-stationary time-series data predication methods, comprising: obtains and decomposes original temporal data, obtain three subsequences with different law characteristics；Select matched prediction model respectively for subsequence, sub-sequences carry out analysis prediction respectively by matched prediction model, obtain the prediction result of the subsequence；The prediction result of all subsequences is integrated, the prediction result of original temporal data is obtained.The present invention obtains the prediction result of original sequence data by way of decomposing original temporal data, predicting subsequence and fusant sequence prediction result respectively, the law characteristic of each subsequence is comprehensively considered, forecasting accuracy is high, and can realize the processing and prediction to any Non-stationary time-series data.The present invention also provides a kind of Non-stationary time-series data forecasting system, storage medium and computer equipments.
Description
Technical field
The present invention relates to time series forecasting technical field more particularly to a kind of Non-stationary time-series data predication method, it is
System, storage medium and computer equipment.
Background technique
With the fast development of the commonly used and computer technology of sensor network and handheld mobile device etc., Ren Menke
To obtain a large amount of time series datas.These time series datas dynamic change at any time, the overwhelming majority be all it is non-smoothly, contain noise
, there is particularity and complexity largely.How implicit time series pattern is excavated from these time series datas, and
These modes are analyzed, to extract valuable information and for really to timing driving and Predicting Technique
A great challenge.
Time series forecasting technology is widely used in the uncertain and further support decision-making level of processing, special
It is not in the field for being related to time measure, such as finance, meteorological and communications and transportation.The sequence that can be often encountered in real life,
Especially reflection society, economic phenomenon sequence, it is most of and unstable, if can Accurate Prediction these non-stationary series, can be with
Control and directive function well are played to society, expanding economy.Therefore the modeling and prediction of nonstationary time series are studied
Method has critically important realistic meaning.
It is the prediction technique based on statistical probability that time series forecasting technology, which includes: the first kind, in the prior art, and this method exists
Time series have linear and stationarity hypothesis or under the conditions of could obtain preferable prediction result, but to nonlinear
Data modeling effect is undesirable.Second class is prediction technique neural network based, and such method can be to the non-linear of complexity
Time series data is modeled, but is constrained to the size of data volume, if data volume is not big enough, is modeled still not accurate enough.In addition,
If time series data contains noise, convergence when model training will affect, easily cause the over-fitting of model.
Summary of the invention
The technical problem to be solved by the present invention is to aiming at the problems existing in the prior art, provide a kind of Non-stationary time-series
Data predication method, system, storage medium and computer equipment.
The technical scheme to solve the above technical problems is that a kind of Non-stationary time-series data predication method, comprising:
S1 obtains original temporal data, and pre-processes to the original temporal data；
S2 decomposes the original temporal data, obtains three subsequences with different law characteristics；
S3 selects matched prediction model for the subsequence, by the matched prediction model respectively to institute respectively
It states subsequence and carries out analysis prediction, obtain the prediction result of the subsequence；
S4 integrates the prediction result of all subsequences, obtains the prediction result of the original temporal data.
In order to solve the above technical problems, the present invention also provides a kind of Non-stationary time-series data forecasting systems, comprising:
Preprocessing module is pre-processed for obtaining original temporal data, and to the original temporal data；
Decomposing module obtains three subsequences with different law characteristics for decomposing the original temporal data；
Prediction module passes through the matched prediction mould for selecting matched prediction model respectively for the subsequence
Type carries out analysis prediction to the subsequence respectively, obtains the prediction result of the subsequence；
Fusion Module obtains the original temporal data for integrating the prediction result of all subsequences
Prediction result.
In order to solve the above technical problems, the present invention also provides a kind of computer readable storage medium, including instruction, when described
When instruction is run on computers, the computer is made to execute the above method.
In order to solve the above technical problems, the present invention also provides a kind of computer equipment, including memory, processor and storage
Computer program that is on the memory and can running on the processor, the processor execute real when described program
The existing above method.
The beneficial effects of the present invention are: the present invention decomposes original temporal data, obtaining three has different law characteristics
Subsequence, choose matching prediction model respectively according to the different law characteristic of each subsequence, pass through the prediction model of selection point
Other sub-sequences are predicted, the prediction result of each subsequence is obtained, and are then integrated the prediction result of subsequence, to obtain
The prediction result of original sequence data, the present invention is by decomposing original temporal data, predicting subsequence and fusant sequence respectively
The mode of column prediction result obtains the prediction result of original sequence data, has comprehensively considered the law characteristic of each subsequence, prediction
Accuracy is high, and can realize the processing and prediction to any Non-stationary time-series data.
Detailed description of the invention
Fig. 1 is Non-stationary time-series data predication method schematic flow chart provided in an embodiment of the present invention；
Fig. 2 is Non-stationary time-series data predication method schematic flow chart provided in an embodiment of the present invention；
Fig. 3 is the basic block diagram of Dense network provided in an embodiment of the present invention；
Fig. 4 is the cellular construction figure of GRU network provided in an embodiment of the present invention；
Fig. 5 is the observation data of temperature provided in an embodiment of the present invention；
Fig. 6 is exploded view of the STL provided in an embodiment of the present invention to temperature-time sequence；
Fig. 7 is the prediction result provided in an embodiment of the present invention to trend component；
Fig. 8 is the prediction result provided in an embodiment of the present invention to remainder component；
Fig. 9 is the prediction result provided in an embodiment of the present invention to periodic component；
Figure 10 is the fusion forecasting result provided in an embodiment of the present invention to temperature；
Figure 11 is Non-stationary time-series data forecasting system schematic block diagram provided in an embodiment of the present invention.
Specific embodiment
The principle and features of the present invention will be described below with reference to the accompanying drawings, and the given examples are served only to explain the present invention, and
It is non-to be used to limit the scope of the invention.
Fig. 1 and Fig. 2 gives a kind of Non-stationary time-series data predication method schematic flow provided in an embodiment of the present invention
Figure.As depicted in figs. 1 and 2, which includes:
S1 obtains original temporal data, and pre-processes to the original temporal data；
S2 decomposes the original temporal data, obtains three subsequences with different law characteristics；
S3 selects matched prediction model for the subsequence, by the matched prediction model respectively to institute respectively
It states subsequence and carries out analysis prediction, obtain the prediction result of the subsequence；
S4 integrates the prediction result of all subsequences, obtains the prediction result of the original temporal data.
In above-described embodiment, by decomposing original temporal data, three subsequences with different law characteristics are obtained,
It chooses matching prediction model respectively according to the different law characteristic of each subsequence, sub-sequences is distinguished by the prediction model of selection
It is predicted, obtains the prediction result of each subsequence, then integrated the prediction result of subsequence, to obtain original series number
According to prediction result, the present invention is by decomposing original temporal data, predicting subsequence and fusant sequence prediction result respectively
Mode obtain the prediction result of original sequence data, comprehensively considered the law characteristic of each subsequence, forecasting accuracy is high, and
The processing and prediction to any Non-stationary time-series data can be achieved.In the embodiment, arbitrary complex time sequence data can be decomposed
For three subsequences, these subsequences all respectively have an evident regularity feature, and readily selected suitable model is modeled；To with not
Subsequence with rule characteristic separately designs different prediction models, finally combines the advantage of different models, improves final
Prediction effect.
Preferably, original temporal data are obtained, and the original temporal data are pre-processed, comprising: are obtained original
Time series data encloses corresponding timestamp for the original temporal data.
Preferably, described to decompose the original temporal data, obtain three subsequences with different law characteristics, packet
It includes:
The original temporal data are decomposed using STL, obtain trend component subsequence, periodic component subsequence and
Remainder component subsequence.
In above-described embodiment, the original temporal data are decomposed using STL, obtain trend component subsequence, week
Phase component subsequence and remainder component subsequence.Original temporal data can be divided into three by STL decomposition method has different rule
The component of feature is restrained, three components are trend component Trend, periodic component Seasonal and remainder components R esidual respectively.
In the embodiment, solve interfering with each other between heterogeneity inside Non-Stationary Time Series using STL decomposition method
Former time series data has been resolved into the subsequence with different regular characteristics by problem.Different subsequences are analyzed, are understood every
The global regularity feature of a subsequence, respectively different prediction network carry out multi-step prediction.Non-stationary has been arrived in this method study
The development model of time series data, while having also contemplated the long-rang dependence of data.
Preferably, described that the original temporal data are decomposed using STL, obtain trend component subsequence, period
Component subsequence and remainder component subsequence, comprising:
S2.1, by the original temporal data Y at current timetIt is decomposed into trend component Tt, periodic term component StWith remainder component
RtThe sum of, it is trend component TtAssign initial value Tt (0), generally setting Tt (0)=0；
The original temporal data at current time are subtracted the last round of trend component for recycling and obtaining, when obtaining first by S2.2
Between sequence Yt-Tt (k), wherein k is the number of iterations；Wherein if original temporal data YtIt is missing from some specific points, that
, the point for going the sequence of trending to lack at this can similarly lack；
S2.3, to the first time sequence Yt-Tt (k)The recurrence of LOESS local polynomial fitting is carried out, when calculating each
Between the smooth value put, sharpening result combines to obtain temporary period subsequenceT=-n(p)+1,...,-N+n(p), length
For N+2 × np, wherein npFor the sample number of a cycle, this step need to select LOESS and return smoothing parameter ns；
S2.4, to the temporary period subsequenceSuccessively carrying out length is respectively np、npSliding three times with 3 is flat
, LOESS recurrence is carried out again, obtains the second time series that length is NT=1 ..., N；Removal is periodically poor
It is different, it is equivalent to the small throughput of extracting cycle subsequence, this step need to select LOESS and return smoothing parameter nl；
S2.5 utilizes the temporary period subsequenceSubtract second time seriesObtain the period point
+ 1 iteration result of kth of amount
S2.6 utilizes original temporal data Yt+ 1 iteration result of kth for subtracting periodic component, obtains third time seriesIf original temporal data YtIt is missing from specific point, then going periodic sequence same scarce in this point
It loses；
S2.7, to the third time seriesLOESS recurrence is carried out, the K+1 times for obtaining trend component changes
For result Tt (k+1), this process need to select LOESS recurrence smoothing parameter nt；
Whether S2.8, the trend component judged restrain with periodic component, if convergence, the trend component of time series
For Tt=Tt (k+1), periodic term component isRemainder component is Rt=Yt-Tt-St；If not restraining, return step
S2.2 re-starts circulation, until trend component and periodic component convergence.
It should be noted that STL is made of two circulative metabolisms, interior loop nesting is in outer circulation.The interior every operation of circulation
Once, periodic component and trend component will be updated once, and complete decomposes runs all by niA such interior circulation
Process composition.Each outside the circular channels is made of interior circulation, and robust weight can be calculated by outer circulation, these power
It can be used in next interior circulation again, for reducing abnormal behaviour of short duration in trend component and periodic component.It is outer for the first time to follow
The robust weight of ring setting is equal to 1, carries out noSecondary outer circulation.STL can be by the data Y of arbitrary sequence data at a certain momenttIt decomposes
For trend component Tt, periodic component StWith remainder components RtThe sum of, as shown in formula (1):
Yt=Tt+St+RtT=1,2 ..., N (1)
The circulation interior each time of STL can all first pass through smooth update cycle component, followed by carry out to trend component smooth
And update.
In the embodiment, interior circulation and outer circulation are devised, makes algorithm that there is enough robustness.Particularly, if niFoot
Enough big, then at the end of interior circulation, trend component has been restrained with periodic component；If without apparent exceptional value in time series data,
It can be by noIt is set as 0.npIt is the quantity of point of observation in each period, if the period of time series is one day, counts by the hour, then
np=24.nlUsually it can be assumed that for more than or equal to npMinimum odd number.nsIt is generally set to odd number, while nsAt least 7,
With nsIncrease, each period subsequence is smoothened.ntAlso it is generally set to odd number.
Preferably, described to select matched prediction model respectively for the subsequence, pass through the matched prediction model
Analysis prediction is carried out to the subsequence respectively, obtains the prediction result of the subsequence, comprising:
S3.1 constructs Dense Network Prediction Model, predicts the trend by Dense Network Prediction Model analysis
Component subsequence obtains the predicted value of trend component subsequence；
S3.2 constructs GRU Network Prediction Model, predicts the periodic component by GRU Network Prediction Model analysis
Subsequence and remainder component subsequence, obtain the predicted value of periodic component subsequence and remainder component subsequence.
It should be noted that before carrying out analysis prediction by prediction model, it is thus necessary to determine that the time step of prediction model, when
Between step-length can arbitrarily set, in the embodiment of the present invention, time step be can be set to 24 hours.
In above-described embodiment, trend component Trend has apparent trend development mode, and data are relatively simple, therefore adopt
It is predicted with the Dense network of shallow-layer.And periodic component Seasonal is divided with the remainder with Complex Noise
Residual is measured, devises a GRU network to analyze it and predict.GRU network be timing for data into
Row modeling deep learning network, can preferably learn the fluctuation and uncertainty of remainder components R esidual.The embodiment
In, select suitable prediction model to be analyzed and predicted respectively the subsequence with different law characteristics, simple trend
Component selects shallow-layer Dense network to be predicted, this not only simplifies overall model structure, also saves and calculates the time.Period
Component and remainder component are predicted with GRU network, with the long Time-Dependent feature in acquisition time sequence data.Pass through this point
The prediction accuracy of each subsequence can be improved in cloth prediction technique.
Decomposing the trend component that obtains through STL decomposition method is smoother curve, this curve can regard as by
Local linear fitting as a result, describing the basic situation of change of former data on a timeline, data variation is gentle, relatively simple
It is single.Therefore, for trend component Trend, relatively simple model structure is had chosen to analyze it and predict.Dense
Network structure is simple, there is good data approximation ability and mature training method.The basic structure of Dense network such as Fig. 3 institute
Show.In Dense Network Prediction Model, the relationship of output and input can be indicated with formula (2).
Wherein, b1With ωij(i=0,1,2 ..., I；J=0,1,2 ..., J) it is respectively input layer to first hidden layer
Biasing and connection weight, b2With ujk(j=0,1,2 ..., J；K=0,1,2 ..., K) it is respectively between two hidden layers
Biasing and connection weight, b3With vkl(k=0,1,2 ..., K；L=0,1,2 ..., L) it is respectively hidden layer to the inclined of output layer
It sets and connection weight, I, L is input layer and output layer number of nodes respectively, J, K are two hidden layer number of nodes, g1(·),g2
(·),g3() is activation primitive, xiAnd ylIt is i-th of input and first of output respectively.
The present invention rule of thumb chooses the number of plies and neuron number of network with many experiments.Specifically: Dense net
Network is made of input layer, two hidden layers and output layer.Wherein, the neuron number of input layer and output layer is by real data
Dimension decision is output and input, is disposed as 24 in embodiments of the present invention, the neuron number of two intermediate hidden layers is equal
It is 512.
Preferably, the building Dense Network Prediction Model analyzes prediction institute by the Dense Network Prediction Model
Trend component subsequence is stated, the predicted value of trend component subsequence is obtained, comprising:
S3.1.1 designs Dense Network Prediction Model, and Dense Network Prediction Model is by input layer, two hidden layers and defeated
Layer forms out；The neuron number of its input layer and output layer is determined by the dimension that outputs and inputs of real data；
S3.1.2 divides the training set, verifying collection and test set of trend component subsequence；
Trend component subsequence is normalized in S3.1.3；
Trend component subsequence is processed into the data for meeting Dense Network Prediction Model input and output requirement by S3.1.4
Format；It is every time (1,24) into the dimension of Dense network, output dimension is also (1,24) in the embodiment；
S3.1.5, it is least random number that all connection weights and threshold value, which are arranged, using ReLU function as Dense network
Activation primitive optimizes network objectives function using Adam optimizer；
S3.1.6 obtains the Dense Network Prediction Model for meeting error requirements, to Dense by network training and test
Anti-normalization processing was made in the output of Network Prediction Model, obtained the predicted value of trend component subsequence.
In above-described embodiment, normalization pretreatment is done to trend component, and design Dense network model, to trend component
Subsequence is trained and tests, and obtains the prediction result of network, and doing anti-normalization processing to neural network forecast result can be realized
The prediction of trend component, as shown in Figure 7.The Accurate Prediction of trend component can effectively improve the prediction of original time series.
Periodic component has periodicity and constancy on a timeline, therefore it can be used for predicting future time sequence
Numerical value.Different from trend component and periodic component, irregular remainder component contains the most information of former data, including original
Non-stationary, the uncertain and strong fluctuation of data, for this sequence, conventional network is difficult to carry out accurate modeling to it.
The embodiment of the present invention is by one GRU network of design come the prediction of process cycle component Seasonal and remainder components R esidual
Problem, the basic network cellular construction of GRU network is as shown in figure 4, the output of each GRU unit is calculated by formula (3).
Wherein, zt,rt,htGRU network is respectively represented in the update door of moment t, resetting door, candidate activation value and is worked as
Preceding activation value；Uz,Ur,Uh,Wz,Wr,WhFor weight matrix, bz,br,bhFor bias term, ht-1It is the activation value of last moment, o table
Show that matrix element is corresponding to be multiplied, σ and tanh are activation primitive.
Preferably, the building GRU Network Prediction Model predicts the week by GRU Network Prediction Model analysis
Phase component subsequence and remainder component subsequence, obtain the predicted value of periodic component subsequence and remainder component subsequence, comprising:
S3.2.1 designs GRU Network Prediction Model, and GRU Network Prediction Model is formed by two GRU layers, neuron
Number is determined by the dimension that outputs and inputs of real data；
S3.2.2 divides the training set, verifying collection and test set of periodic component subsequence and remainder component subsequence；
S3.2.3 makees normalized to periodic component subsequence and remainder component subsequence；
Periodic component subsequence and remainder component subsequence are processed into and meet the input of GRU Network Prediction Model by S3.2.4
Export desired data format；The input layer dimension of GRU network requirement is three-dimensional, and data to be processed are in the embodiment
One-dimensional, it is every time (1,24,1) into the dimension of GRU network, output dimension is (1,24) therefore；
S3.2.5 carries out the initialization of GRU deep neural network (initialization of such as weight) using the default parameters of Keras,
Using tanh function as the activation primitive of GRU Network Prediction Model, optimize network objectives function using Adam optimizer；
S3.2.6 obtains the GRU Network Prediction Model for meeting error requirements, to GRU network by network training and test
Anti-normalization processing is made in the output of prediction model, obtains the predicted value of periodic component subsequence and remainder component subsequence.
In above-described embodiment, normalization pretreatment is done to periodic component and remainder component, and separately design two GRU networks
Model is trained and tests to periodic component and remainder component subsequence, obtain the prediction result of network, to neural network forecast knot
Fruit, which is cooked anti-normalization processing, can be realized the prediction of periodic component and remainder component, as Figure 8-9.Periodic component and remainder point
The Accurate Prediction of amount can effectively improve the prediction of original time series.
Preferentially, the prediction result by all subsequences is integrated, and obtains the original temporal data
Prediction result, comprising: the prediction result of all subsequences is integrated by linear fusion mode, is obtained described original
The prediction result of time series data.
In above-described embodiment, by the prediction knot of trend component, periodic component and remainder component by way of linear fusion
Fruit is directly added, and the prediction to original temporal data can be realized, the results are shown in Figure 10, the results showed that, first divided by this
The time series forecasting mode that solution merges again effectively improves the prediction effect of Non-stationary time-series data.
Below by taking the Non-stationary time-series data development trend of following 24 time steps of this index of predicted temperature as an example, tool
Body includes the following steps:
The crucial meteorological element data set that data source used in the embodiment of the present invention is recorded in Beijing meteorological observatory.Utilize pass
Key meteorological element (i.e. temperature) verifies fusion forecasting model, select 2400 groups of measured data of experiment to the method for the present invention into
Row simulation calculation, the time step of data is 1h in experiment, and the overall condition of original temporal data is as shown in Figure 5.In view of protecting
Similar statistical nature is held, the data of selection preceding 80% are as training set, and 13% data are as verifying collection, remaining 7% conduct
Test set.Three components that data obtain after STL is decomposed are original temporal respectively from top to bottom as shown in fig. 6, in attached drawing
Data, trend component, periodic component and remainder component；Three components are normalized, are then separately processed in pairs
The input/output format for answering network to need, learning time step-length and predicted time step-length are for 24 hours, i.e., with preceding data letter for 24 hours
Value after breath prediction for 24 hours.Verifying iteration 100 times is respectively trained in three components, carries out at renormalization to last prediction result
Reason, the prediction result for obtaining three components is as Figure 7-9, and three prediction results are directly added, and obtains the fused side of temperature
The final result of method prediction, as shown in Figure 10.The root-mean-square error of prediction is 2.6251, testing time 0.018s, prediction knot
Fruit illustrates that the distributed fusion forecasting method of decomposing for the Non-stationary time-series data that the present invention designs has practicability and timeliness.
Above in association with Fig. 1 to Figure 10, the Non-stationary time-series data prediction provided according to embodiments of the present invention is described in detail
Method.The Non-stationary time-series data forecasting system provided below with reference to Figure 11, the present invention is described in detail embodiment.
As shown in figure 11, which includes and processing module, decomposing module, prediction module and Fusion Module.Pre-process mould
Block pre-processes the original temporal data for obtaining original temporal data；Decomposing module is for decomposing the original
Beginning time series data obtains three subsequences with different law characteristics；Prediction module is used to select respectively for the subsequence
Matched prediction model carries out analysis prediction to the subsequence respectively by the matched prediction model, obtains the son
The prediction result of sequence；Fusion Module is for integrating the prediction result of all subsequences, when obtaining described original
The prediction result of ordinal number evidence.
In above-described embodiment, by decomposing original temporal data, three subsequences with different law characteristics are obtained,
It chooses matching prediction model respectively according to the different law characteristic of each subsequence, sub-sequences is distinguished by the prediction model of selection
It is predicted, obtains the prediction result of each subsequence, then integrated the prediction result of subsequence, to obtain original series number
According to prediction result, the present invention is by decomposing original temporal data, predicting subsequence and fusant sequence prediction result respectively
Mode obtain the prediction result of original sequence data, comprehensively considered the law characteristic of each subsequence, forecasting accuracy is high, and
The processing and prediction to any Non-stationary time-series data can be achieved.In the embodiment, arbitrary complex time sequence data can be decomposed
For three subsequences, these subsequences all respectively have an evident regularity feature, and readily selected suitable model is modeled；To with not
Subsequence with rule characteristic separately designs different prediction models, finally combines the advantage of different models, improves final
Prediction effect.
In the embodiment of the present invention, each module of Non-stationary time-series data forecasting system can realize method institute in above-described embodiment
The correspondence effect of the repertoire of description, specific implementation and acquirement is corresponding with method part, and details are not described herein again.
The embodiment of the present invention also provides a kind of computer readable storage medium, including instruction, when described instruction is in computer
When upper operation, the computer is made to execute the method in above-described embodiment.
The embodiment of the present invention also provides a kind of computer equipment, including memory, processor and is stored in the memory
On and the computer program that can run on the processor, the processor realize above-described embodiment when executing described program
In method.
It is apparent to those skilled in the art that for convenience of description and succinctly, the dress of foregoing description
The specific work process with unit is set, can refer to corresponding processes in the foregoing method embodiment, details are not described herein.
In several embodiments provided herein, it should be understood that disclosed device and method can pass through it
Its mode is realized.For example, the apparatus embodiments described above are merely exemplary, for example, the division of unit, only
A kind of logical function partition, there may be another division manner in actual implementation, for example, multiple units or components can combine or
Person is desirably integrated into another system, or some features can be ignored or not executed.
Unit may or may not be physically separated as illustrated by the separation member, shown as a unit
Component may or may not be physical unit, it can and it is in one place, or may be distributed over multiple networks
On unit.It can select some or all of unit therein according to the actual needs to realize the mesh of the embodiment of the present invention
's.
It, can also be in addition, the functional units in various embodiments of the present invention may be integrated into one processing unit
It is that each unit physically exists alone, is also possible to two or more units and is integrated in one unit.It is above-mentioned integrated
Unit both can take the form of hardware realization, can also realize in the form of software functional units.
It, can if integrated unit is realized in the form of SFU software functional unit and when sold or used as an independent product
To be stored in a computer readable storage medium.Based on this understanding, technical solution of the present invention substantially or
Say that all or part of the part that contributes to existing technology or the technical solution can embody in the form of software products
Out, which is stored in a storage medium, including some instructions are used so that a computer equipment
(can be personal computer, server or the network equipment etc.) executes all or part of each embodiment method of the present invention
Step.And storage medium above-mentioned include: USB flash disk, it is mobile hard disk, read-only memory (ROM, Read-Only Memory), random
Access various Jie that can store program code such as memory (RAM, Random Access Memory), magnetic or disk
Matter.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (10)
1. a kind of Non-stationary time-series data predication method characterized by comprising
S1 obtains original temporal data, and pre-processes to the original temporal data；
S2 decomposes the original temporal data, obtains three subsequences with different law characteristics；
S3 selects matched prediction model for the subsequence, by the matched prediction model respectively to the son respectively
Sequence carries out analysis prediction, obtains the prediction result of the subsequence；
S4 integrates the prediction result of all subsequences, obtains the prediction result of the original temporal data.
2. obtaining three tools the method according to claim 1, wherein described decompose the original temporal data
There is the subsequence of different law characteristics, comprising:
The original temporal data are decomposed using STL, obtain trend component subsequence, periodic component subsequence and remainder
Component subsequence.
3. according to the method described in claim 2, it is characterized in that, described divide the original temporal data using STL
Solution, obtains trend component subsequence, periodic component subsequence and remainder component subsequence, comprising:
S2.1, by the original temporal data Y at current timetIt is decomposed into trend component Tt, periodic term component StWith remainder components RtIt
With, be trend component Tt (k)Assign initial value Tt (0)；
The original temporal data at current time are subtracted the last round of trend component for recycling and obtaining, obtain first time sequence by S2.2
Arrange Yt-Tt (k), wherein k is the number of iterations；
S2.3, to the first time sequence Yt-Tt (k)LOESS recurrence is carried out, calculates the smooth value at each time point, smoothly
As a result combination obtains temporary period subsequenceT=-n(p)+1,...,-N+n(p), the length is N+2 × np, wherein npFor
The sample number of a cycle；
S2.4, to the temporary period subsequenceSuccessively carrying out length is respectively np、npWith 3 sliding average three times, then
Secondary progress LOESS recurrence obtains the second time series that length is NT=1 ..., N；
S2.5 utilizes the temporary period subsequenceSubtract second time seriesObtain the kth of periodic component
+ 1 iteration result
S2.6 utilizes original temporal data Yt+ 1 iteration result of kth for subtracting periodic component, obtains third time series
S2.7, to the third time seriesLOESS recurrence is carried out ,+1 iteration result of kth of trend component is obtained
Tt (k+1)；
Whether S2.8, the trend component judged restrain with periodic component, if convergence, the trend component of original time series
For Tt=Tt (k+1), periodic component isRemainder component is Rt=Yt-Tt-St；If not restraining, return step S2.2,
Circulation is re-started, until trend component and periodic component convergence.
4. according to the method described in claim 2, it is characterized in that, described select matched prediction mould for the subsequence respectively
Type carries out analysis prediction to the subsequence respectively by the matched prediction model, obtains the prediction knot of the subsequence
Fruit, comprising:
S3.1 constructs Dense Network Prediction Model, predicts the trend component by Dense Network Prediction Model analysis
Subsequence obtains the predicted value of trend component subsequence；
S3.2 constructs GRU Network Prediction Model, predicts the sub- sequence of periodic component by GRU Network Prediction Model analysis
Column and remainder component subsequence, obtain the predicted value of periodic component subsequence and remainder component subsequence.
5. according to the method described in claim 4, it is characterized in that, the building Dense Network Prediction Model, by described
The trend component subsequence is predicted in the analysis of Dense Network Prediction Model, obtains the predicted value of trend component subsequence, comprising:
S3.1.1 designs Dense Network Prediction Model, and Dense Network Prediction Model is by input layer, two hidden layers and output layer
Composition；The neuron number of its input layer and output layer is determined by the dimension that outputs and inputs of real data；
S3.1.2 divides the training set, verifying collection and test set of trend component subsequence；
Trend component subsequence is normalized in S3.1.3；
Trend component subsequence is processed into the data format for meeting Dense Network Prediction Model input and output requirement by S3.1.4；
S3.1.5, it is least random number that all connection weights and threshold value, which are arranged, using ReLU function as the activation letter of Dense network
Number optimizes network objectives function using Adam optimizer；
S3.1.6 obtains the Dense Network Prediction Model for meeting error requirements, to Dense network by network training and test
Anti-normalization processing was made in the output of prediction model, obtained the predicted value of trend component subsequence.
6. according to the method described in claim 4, it is characterized in that, the building GRU Network Prediction Model, passes through the GRU
The periodic component subsequence and remainder component subsequence are predicted in Network Prediction Model analysis, obtain periodic component subsequence and remaining
The predicted value of item component subsequence, comprising:
S3.2.1 designs GRU Network Prediction Model, and GRU Network Prediction Model is formed by two GRU layers, and neuron number is equal
It is determined by the dimension that outputs and inputs of real data；
S3.2.2 divides the training set, verifying collection and test set of periodic component subsequence and remainder component subsequence；
S3.2.3 makees normalized to periodic component subsequence and remainder component subsequence；
Periodic component subsequence and remainder component subsequence are processed into and meet GRU Network Prediction Model input and output by S3.2.4
It is required that data format；
S3.2.5 carries out the initialization of GRU deep neural network using the default parameters of Keras, using tanh function as GRU network
The activation primitive of prediction model optimizes network objectives function using Adam optimizer；
S3.2.6 obtains the GRU Network Prediction Model for meeting error requirements, to GRU neural network forecast by network training and test
Anti-normalization processing is made in the output of model, obtains the predicted value of periodic component subsequence and remainder component subsequence.
7. method according to any one of claims 1 to 6, which is characterized in that the prediction by all subsequences
As a result it is integrated, obtains the prediction result of the original temporal data, comprising: by linear fusion mode by all sons
The prediction result of sequence is integrated, and the prediction result of the original temporal data is obtained.
8. a kind of Non-stationary time-series data forecasting system characterized by comprising
Preprocessing module is pre-processed for obtaining original temporal data, and to the original temporal data；
Decomposing module obtains three subsequences with different law characteristics for decomposing the original temporal data；
Prediction module passes through the matched prediction model point for selecting matched prediction model respectively for the subsequence
It is other that analysis prediction is carried out to the subsequence, obtain the prediction result of the subsequence；
Fusion Module obtains the pre- of the original temporal data for integrating the prediction result of all subsequences
Survey result.
9. a kind of computer readable storage medium, including instruction, which is characterized in that when described instruction is run on computers,
The computer is set to execute method according to claim 1-7.
10. a kind of computer equipment, including memory, processor and be stored on the memory and can be in the processor
The computer program of upper operation, which is characterized in that the processor is realized when executing described program such as any one of claim 1-7
The method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910173925.9A CN109977098A (en) | 2019-03-08 | 2019-03-08 | Non-stationary time-series data predication method, system, storage medium and computer equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910173925.9A CN109977098A (en) | 2019-03-08 | 2019-03-08 | Non-stationary time-series data predication method, system, storage medium and computer equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109977098A true CN109977098A (en) | 2019-07-05 |
Family
ID=67078210
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910173925.9A Pending CN109977098A (en) | 2019-03-08 | 2019-03-08 | Non-stationary time-series data predication method, system, storage medium and computer equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109977098A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111161538A (en) * | 2020-01-06 | 2020-05-15 | 东南大学 | Short-term traffic flow prediction method based on time series decomposition |
| CN111179595A (en) * | 2020-01-06 | 2020-05-19 | 东南大学 | Traffic flow time sequence decomposition method |
| CN111639814A (en) * | 2020-06-02 | 2020-09-08 | 贝壳技术有限公司 | Method, apparatus, medium, and electronic device for predicting occurrence probability of fluctuating behavior |
| CN111950793A (en) * | 2020-08-17 | 2020-11-17 | 浙江工业大学 | Comprehensive energy system load prediction method considering multivariate load coupling characteristics |
| CN112506752A (en) * | 2020-12-10 | 2021-03-16 | 山东海量信息技术研究院 | Operation and maintenance method, device and equipment based on time series data trend prediction |
| CN112612671A (en) * | 2020-12-16 | 2021-04-06 | 深圳前海微众银行股份有限公司 | System monitoring method, device, equipment and storage medium |
| CN112835771A (en) * | 2019-11-22 | 2021-05-25 | 中国移动通信有限公司研究院 | Traffic prediction method, device, equipment and storage medium |
| CN112929215A (en) * | 2021-02-04 | 2021-06-08 | 博瑞得科技有限公司 | Network flow prediction method, system, computer equipment and storage medium |
| CN112949896A (en) * | 2020-12-22 | 2021-06-11 | 浙江佰安信息技术有限公司 | Time sequence prediction method based on fusion sequence decomposition and space-time convolution |
| WO2021164267A1 (en) * | 2020-02-21 | 2021-08-26 | 平安科技（深圳）有限公司 | Anomaly detection method and apparatus, and terminal device and storage medium |
| CN113421016A (en) * | 2021-07-09 | 2021-09-21 | 浙江大学 | Resource allocation method, device, equipment and medium based on people number prediction |
| CN117172085A (en) * | 2023-04-17 | 2023-12-05 | 北京市水科学技术研究院 | PCCP broken wire prediction method, device, computer equipment and medium |
-
2019
- 2019-03-08 CN CN201910173925.9A patent/CN109977098A/en active Pending
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112835771A (en) * | 2019-11-22 | 2021-05-25 | 中国移动通信有限公司研究院 | Traffic prediction method, device, equipment and storage medium |
| CN111179595A (en) * | 2020-01-06 | 2020-05-19 | 东南大学 | Traffic flow time sequence decomposition method |
| CN111161538A (en) * | 2020-01-06 | 2020-05-15 | 东南大学 | Short-term traffic flow prediction method based on time series decomposition |
| WO2021164267A1 (en) * | 2020-02-21 | 2021-08-26 | 平安科技（深圳）有限公司 | Anomaly detection method and apparatus, and terminal device and storage medium |
| CN111639814A (en) * | 2020-06-02 | 2020-09-08 | 贝壳技术有限公司 | Method, apparatus, medium, and electronic device for predicting occurrence probability of fluctuating behavior |
| CN111950793A (en) * | 2020-08-17 | 2020-11-17 | 浙江工业大学 | Comprehensive energy system load prediction method considering multivariate load coupling characteristics |
| CN112506752A (en) * | 2020-12-10 | 2021-03-16 | 山东海量信息技术研究院 | Operation and maintenance method, device and equipment based on time series data trend prediction |
| CN112612671A (en) * | 2020-12-16 | 2021-04-06 | 深圳前海微众银行股份有限公司 | System monitoring method, device, equipment and storage medium |
| CN112949896A (en) * | 2020-12-22 | 2021-06-11 | 浙江佰安信息技术有限公司 | Time sequence prediction method based on fusion sequence decomposition and space-time convolution |
| CN112949896B (en) * | 2020-12-22 | 2022-06-17 | 浙江佰安信息技术有限公司 | Time sequence prediction method based on fusion sequence decomposition and space-time convolution |
| CN112929215A (en) * | 2021-02-04 | 2021-06-08 | 博瑞得科技有限公司 | Network flow prediction method, system, computer equipment and storage medium |
| CN112929215B (en) * | 2021-02-04 | 2022-10-21 | 博瑞得科技有限公司 | Network flow prediction method, system, computer equipment and storage medium |
| CN113421016A (en) * | 2021-07-09 | 2021-09-21 | 浙江大学 | Resource allocation method, device, equipment and medium based on people number prediction |
| CN117172085A (en) * | 2023-04-17 | 2023-12-05 | 北京市水科学技术研究院 | PCCP broken wire prediction method, device, computer equipment and medium |
| CN117172085B (en) * | 2023-04-17 | 2024-04-26 | 北京市水科学技术研究院 | PCCP broken wire prediction method, device, computer equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109977098A (en) | Non-stationary time-series data predication method, system, storage medium and computer equipment | |
| Khodayar et al. | Rough deep neural architecture for short-term wind speed forecasting | |
| Chen et al. | Fuzzy forecasting based on two-factors second-order fuzzy-trend logical relationship groups and particle swarm optimization techniques | |
| US7526461B2 (en) | System and method for temporal data mining | |
| Peng et al. | A hybrid forward algorithm for RBF neural network construction | |
| Ramasso et al. | Remaining useful life estimation by classification of predictions based on a neuro-fuzzy system and theory of belief functions | |
| CN113366473A (en) | Method and system for automatic selection of models for time series prediction of data streams | |
| CN112417028A (en) | Wind speed time sequence characteristic mining method and short-term wind power prediction method | |
| JP7020547B2 (en) | Information processing equipment, control methods, and programs | |
| CN113762486A (en) | Method and device for constructing fault diagnosis model of converter valve and computer equipment | |
| CN116629431A (en) | Photovoltaic power generation amount prediction method and device based on variation modal decomposition and ensemble learning | |
| Nalatissifa et al. | Customer decision prediction using deep neural network on telco customer churn data | |
| CN110175637A (en) | Non-stationary time-series data depth prediction technique, system, storage medium and equipment | |
| Hart et al. | An advanced cost estimation methodology for engineering systems | |
| US6910192B2 (en) | Method of robust technology design using rational robust optimization | |
| JP6927409B2 (en) | Information processing equipment, control methods, and programs | |
| Ming | A rough set based hybrid method to feature selection | |
| CN114239948A (en) | Deep traffic flow prediction method, medium and equipment based on time sequence decomposition unit | |
| Zhou et al. | Nested partitions method for assembly sequences merging | |
| Bogaerts et al. | A fast inverse approach for the quantification of set-theoretical uncertainty | |
| Khumprom et al. | A hybrid evolutionary CNN-LSTM model for prognostics of C-MAPSS aircraft dataset | |
| Pradana et al. | Comparison of Multi Layer Perceptron, Random Forest & Logistic Regression on Students Performance Test | |
| CN115577618B (en) | Construction method and prediction method of high-pressure converter valve hall environmental factor prediction model | |
| Eleuteri et al. | Survival analysis and neural networks | |
| Kurogi et al. | Hierarchical clustering of ensemble prediction using LOOCV predictable horizon for chaotic time series |
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 |
Application publication date: 20190705 |
|
| RJ01 | Rejection of invention patent application after publication |