CN110110848A - A kind of combination forecasting construction method and device - Google Patents
A kind of combination forecasting construction method and device Download PDFInfo
- Publication number
- CN110110848A CN110110848A CN201910367539.3A CN201910367539A CN110110848A CN 110110848 A CN110110848 A CN 110110848A CN 201910367539 A CN201910367539 A CN 201910367539A CN 110110848 A CN110110848 A CN 110110848A
- Authority
- CN
- China
- Prior art keywords
- training
- deep learning
- model
- learning model
- cross validation
- 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
- 238000010276 construction Methods 0.000 title claims abstract description 20
- 238000012549 training Methods 0.000 claims abstract description 153
- 238000012360 testing method Methods 0.000 claims abstract description 87
- 238000013136 deep learning model Methods 0.000 claims abstract description 83
- 238000002790 cross-validation Methods 0.000 claims abstract description 70
- 239000013598 vector Substances 0.000 claims abstract description 57
- 238000000034 method Methods 0.000 claims abstract description 21
- 238000011156 evaluation Methods 0.000 claims abstract description 10
- 238000013527 convolutional neural network Methods 0.000 claims description 26
- 239000011159 matrix material Substances 0.000 claims description 13
- 238000006243 chemical reaction Methods 0.000 claims description 11
- 238000005096 rolling process Methods 0.000 claims description 5
- 210000002569 neuron Anatomy 0.000 description 8
- 230000004913 activation Effects 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 6
- 238000003062 neural network model Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 230000003044 adaptive effect Effects 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000013277 forecasting method Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000002779 inactivation Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 210000004218 nerve net Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000000452 restraining effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention provides a kind of combination forecasting construction method and device, method includes: to carry out K folding cross validation to basic model on training set, obtains the predicted value of training set；After each folding cross validation, test set is predicted with the folding cross validation resulting basic model, obtains test set in the predicted value of the folding cross validation；Using the predicted value of training set as one-dimensional training vector；And test set is calculated in the average value of the predicted value of each folding cross validation, obtain one-dimensional test vector；With one-dimensional training vector training deep learning model, one-dimensional test vector is predicted with training to convergent deep learning model, calculates the index value of training to convergent deep learning model preset evaluation index；Whether judge index value meets pre-set level condition；If meeting, basic model and training to the convergent deep learning model after cross validation is rolled over K constitute combination forecasting.Using the embodiment of the present invention, the precision of prediction of combination forecasting is improved.
Description
Technical field
The present invention relates to composite prediction technology field more particularly to a kind of combination forecasting construction methods and device.
Background technique
During real data prediction, since many uncertain factors are interfered, often there is information and mention in Individual forecast model
Insufficient, the not high problem of precision of prediction is taken, combination forecasting becomes one of research hotspot.Currently, combination forecasting is big
Mostly it is the linearly or nonlinearly weighted array of two models, the power method such as generallys use or optimal weights determine that method determines power
Weight, obtains combination forecasting.But in fact, the weight of combination forecasting should be flexibly change rather than immobilize
, and be theoretically proved the combination forecasting methods of optimal weights on precision of prediction not necessarily than etc. power or single model
It is good.
It is therefore desirable to design a kind of new combination forecasting construction method, to overcome the above problem.
Summary of the invention
It is an object of the invention to overcome the defect of the prior art, a kind of combination forecasting construction method and dress are provided
It sets, to realize the precision of prediction for improving combination forecasting.
The present invention is implemented as follows:
In a first aspect, the present invention provides a kind of combination forecasting construction method, which comprises
Initial data obtained is divided into training set and test set；
K is carried out to preset basic model on training set and rolls over cross validation, obtains the predicted value of training set；In each folding
After cross validation, the test set is predicted with the folding cross validation resulting basic model, obtains test set in the folding
The predicted value of cross validation；
Using the predicted value of training set as one-dimensional training vector；And test set is calculated in each predicted value for rolling over cross validation
Average value obtains one-dimensional test vector；
With the preset deep learning model of the one-dimensional training vector training, training is obtained to convergent deep learning mould
Type；One-dimensional test vector is predicted with training to convergent deep learning model, and is based on prediction result, calculates training extremely
The index value of convergent deep learning model preset evaluation index；
Judge whether the index value meets pre-set level condition；If meeting, with K roll over cross validation after basic model and
Training to convergent deep learning model constitutes combination forecasting.
Optionally, deep learning model be convolutional neural networks model, including input layer, for single number to be converted to
The hidden layer of column vector, for by hidden layer convert resulting column vector be converted to the conversion layer of matrix, convolutional layer, pond layer,
Full articulamentum and output layer.
Optionally, the index value includes the RMSE of training set and the RMSE of test set, judges whether the index value accords with
Close pre-set level condition, comprising:
If difference of the RMSE of test set less than the RMSE of the RMSE and test set of the first preset value and training set is less than second
Preset value, then Judging index value meets pre-set level condition；Otherwise, it is determined that index value does not meet pre-set level condition.
Optionally, the basic model is XGBoost model.
Optionally, if index value does not meet pre-set level condition, the method also includes:
The parameter of percentage regulation learning model is re-executed with the one-dimensional training vector training deep learning model.
Optionally, preset convolutional neural networks model has multiple, and the parameter of each convolutional neural networks model is different；Judgement
Whether the index value meets pre-set level condition, if meeting, with the basic model after K folding cross validation and trains to convergent
Deep learning model constitutes combination forecasting, comprising:
Whether the index value of training of judgement to convergent each convolutional neural networks model meets pre-set level condition；
From each convolutional neural networks model that index value meets pre-set level condition, the optimal convolution mind of selective goal value
Through network model as target convolution neural network model, basic model and target convolution nerve net after cross validation is rolled over K
Network model constitutes combination forecasting.
Second aspect, the present invention provide a kind of combination forecasting construction device, and described device includes:
First obtains module, for initial data obtained to be divided into training set and test set；
Cross validation module rolls over cross validation for carrying out K to preset basic model on training set, obtains training set
Predicted value；After each folding cross validation, the test set is predicted with the folding cross validation resulting basic model,
Test set is obtained in the predicted value of the folding cross validation；
Second obtains module, for using the predicted value of training set as one-dimensional training vector；And test set is calculated in each folding
The average value of the predicted value of cross validation obtains one-dimensional test vector；
Training module obtains training to convergence for training preset deep learning model with the one-dimensional training vector
Deep learning model；One-dimensional test vector is predicted with training to convergent deep learning model, and based on prediction knot
Fruit calculates the index value of training to convergent deep learning model preset evaluation index；
Judgment module, for judging whether the index value meets pre-set level condition；If meeting, cross validation is rolled over K
Basic model and training to convergent deep learning model afterwards constitutes combination forecasting.
Optionally, deep learning model be convolutional neural networks model, including input layer, for single number to be converted to
The hidden layer of column vector, for by hidden layer convert resulting column vector be converted to the conversion layer of matrix, convolutional layer, pond layer,
Full articulamentum and output layer.
Optionally, the index value includes the RMSE of training set and the RMSE of test set, described in the judgment module judgement
Whether index value meets pre-set level condition, specifically:
If difference of the RMSE of test set less than the RMSE of the RMSE and test set of the first preset value and training set is less than second
Preset value, then Judging index value meets pre-set level condition；Otherwise, it is determined that index value does not meet pre-set level condition.
Optionally, the basic model is XGBoost model.
Optionally, described device further includes adjustment module, is used for:
When the judging result of the judgment module is no, the parameter of percentage regulation learning model is re-executed described in use
One-dimensional training vector training deep learning model.
Optionally, preset deep learning model has multiple, and the parameter of each deep learning model is different；The judgment module
It is specifically used for:
Whether the index value of training of judgement to convergent each deep learning model meets pre-set level condition；
From each deep learning model that index value meets pre-set level condition, the optimal deep learning mould of selective goal value
For type as target deep learning model, it is pre- that basic model and target deep learning model after cross validation is rolled over K constitute combination
Survey model.
The invention has the following advantages: first completing K folding cross validation instruction by basic model using the embodiment of the present invention
It after white silk, then is trained with the predicted value exported by basic model and obtains deep learning model, realize not only independent training pattern, but also protect
The relevance between model has been held, model construction speed, and the deep learning model for constructing combination forecasting are improved
Index value meet pre-set level condition, to ensure that the precision of combination forecasting and generalization ability are met the requirements, improve
The precision of prediction of combination forecasting.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
Other attached drawings are obtained according to these attached drawings.
Fig. 1 is a kind of flow diagram of combination forecasting construction method provided in an embodiment of the present invention；
Fig. 2 is a kind of structural schematic diagram of combination forecasting provided in an embodiment of the present invention；
Fig. 3 is the prediction result figure that data prediction is carried out using combination forecasting provided in an embodiment of the present invention；
Fig. 4 is the evaluation index of present invention example provides in real time combination forecasting and existing single XGBoost model
Comparison diagram；
Fig. 5 is a kind of structural schematic diagram of combination forecasting construction device provided in an embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts all other
Embodiment shall fall within the protection scope of the present invention.
It should be noted that combination forecasting construction method provided by the present invention can be applied to electronic equipment,
In, in a particular application, which can be computer, PC, plate, mobile phone etc., this is all reasonable.
Referring to Fig. 1, the embodiment of the present invention provides a kind of combination forecasting construction method, and method includes the following steps:
S101, initial data obtained is divided into training set and test set；
It can be by downloading or being obtained from the modes such as webpage capture or the data for obtaining other equipment acquisition original from website
Data, initial data can be any kind of data, for example, certain website day pageview, the download time of certain file, certain scape
The day visit amount in area etc..Initial data can be divided into training set and test set according to a certain percentage, for example, at random will be original
The 70% of data is used as training set, and 30% is used as test set.
S102, K folding cross validation is carried out to preset basic model on training set, obtains the predicted value of training set；?
After each folding cross validation, the test set is predicted with the folding cross validation resulting basic model, obtains test set
In the predicted value of the folding cross validation；
Basic model can be supporting vector machine model or XGBoost model or LASSO model etc., due to XGBoost
The calculating speed of model is fast, can comparatively fast complete K folding cross validation training, therefore preferably, basic model can be XGBoost mould
Type makees K folding cross validation using XGBoost model on training set, may include:
Training set is divided into K one's share of expenses for a joint undertaking collection, taking out a copy of it every time is test subset, remaining K-1 parts is training subset,
GridSearchCV (grid search) method pair can be used in the training process using training subset training XGBoost model
XGBoost model carries out tune ginseng, can will when XGBoost model is minimum or after be less than preset threshold in the RSME for testing subset
Model at this time is the training result of this folding cross validation, can be with trained after XGBoost model training is good
XGBoost model is tested in original test set and this folding cross validation respectively and tests subset, obtains surveying in this folding cross validation
The predicted value of swab collection and the predicted value of test set, after K rolls over cross validation, each subset of training set is intersected in certain folding respectively
Verifying is as test subset, and therefore, the predicted value of all test subsets just constitutes the predicted value of entire training set.K can root
It is previously set according to demand, such as can be 5/10/20 etc..
For example, K takes 10, then training set is divided into 10 one's share of expenses for a joint undertaking collection, every folding cross validation takes out a copy of it successively to survey
Swab collection, remaining 9 parts are training subset, after 10 folding cross validations have been trained, so that it may the predicted value of 10 parts of test subsets is obtained,
With regard to constituting the predicted value of entire training set；After every folding cross validation trains XGBoost model, XGBoost model can be used
Entire test set is tested, a predicted value of test set is obtained, ten parts of predicted values of last available test set.
S103, using the predicted value of training set as one-dimensional training vector；And test set is calculated in the pre- of each folding cross validation
The average value of measured value obtains one-dimensional test vector；
Training set may include multiple training samples, and the predicted value of training set includes the predicted value of each training sample, and one
Dimension training vector can be the column vector being made of the predicted value of each training sample.
Assuming that K folding cross validation be 5 folding cross validations, test set it is each folding cross validation predicted value be respectively a1, a2,
A3, a4 and a5 then take the average value of a1, a2, a3, a4 and a5 as one-dimensional test vector.One-dimensional test vector can be by surveying
Examination concentrates each test sample to constitute in the average value of the predicted value of each folding cross validation.
S104, preset deep learning model is trained with the one-dimensional training vector, obtains training to convergent depth
Practise model；One-dimensional test vector is predicted with training to convergent deep learning model, and is based on prediction result, calculates instruction
Practice to the index value of convergent deep learning model preset evaluation index；
Deep learning model can remember mould for convolutional neural networks model (CNN), recurrent neural networks model, shot and long term
Type (LSTM) etc., it is preferable that can be convolutional neural networks model, may include input layer, for converting single number
For the hidden layer of column vector, for hidden layer to be converted the conversion layer, convolutional layer, Chi Hua that resulting column vector is converted to matrix
Layer, full articulamentum and output layer.
The each network layer of convolutional neural networks is all made of neuron, each neuron may include weight weight and
Bias bias；It can receive one layer of input, by the received input of institute with multiplied by weight along with biasing is as output, separately
Outside, it can also be exported using a nonlinear function (such as ReLu activation primitive).
Since the convolution of convolutional neural networks model, pond operation object are usually three-dimensional data, and one-dimensional training vector
Only one dimension, therefore one-dimensional training vector can be become into two-dimensional matrix by hidden layer and conversion layer, with the line number of matrix
With columns respectively as a dimension, along with the port number of default constitutes three-dimensional data, and then can by convolutional layer,
ReLu activation primitive and pond layer carry out 2 convolution, pondization operates, subsequently into full articulamentum, by activation primitive and
After Dropout (random inactivation) operation, it is passed to output layer, completes the primary training of CNN network.
Dropout operation refers in the training process of convolutional neural networks model, by neuron according to certain probability
It is temporarily abandoned from network, over-fitting can be effectively prevented by Dropout operation.Dropout is carried out in full articulamentum
Operation, can be by after certain neuron random drops in full articulamentum, then the calculation result is sent to output layers.
In a kind of implementation, the structure of convolutional neural networks model as shown in Fig. 2, each network layer successively are as follows: input layer,
Hidden layer, conversion layer, convolutional layer, pond layer, convolutional layer, pond layer, full articulamentum and output layer.
For example, as shown in Fig. 2, each sample of training set and test set is 42 dimensional vectors, after K rolls over cross validation
The predicted value of basic model output is single number, and single number is input to the input layer of convolutional neural networks model, through inputting
Layer enters hidden layer, and hidden layer may include multiple neurons (such as 49), and each neuron of hidden layer is received by institute
Input adds the output itself biased as the neuron with itself multiplied by weight, so as to obtain by each of hidden layer
The column vector that the output of neuron is constituted, realizes single number being converted to column vector；And then column vector enters conversion layer, passes through
Conversion layer is converted to matrix, and the size of matrix can be previously set, and the element in matrix can be the element in column vector；In turn
It with matrix size (such as 7*7) multiplied by default channel number (such as 1), obtains three-dimensional data (such as 7*7*1), and then three-dimensional data
It can be calculated by the convolution operation and ReLu activation primitive of convolutional layer 1, so that port number becomes 10, three-dimensional data is expressed as 7*
7*10, and then the pondization for entering pond layer 1 operates, by matrix size boil down to 6*6, three-dimensional data is expressed as 6*6*10, then passes through
2 convolution of convolutional layer 2 and pond layer, the calculating of ReLu activation primitive and Chi Huahou are crossed, three-dimensional data is expressed as 5*5*20, and then enters
Full articulamentum is operated using ReLu activation primitive and Dropout, is passed to output layer, obtains the output of convolutional neural networks model
Single predicted value.
It, can be by the network application of convolution mind in numeric type regression forecasting, prediction technique by addition hidden layer and conversion layer
Actual prediction problem with versatility, suitable for Various types of data.
Preset evaluation index can be previously set according to demand, such as may include RMSE, alternatively, in other embodiments
In, or mean absolute error (MAE), mean square error (MSE) etc..Can in prediction result each test sample it is pre-
Measured value and actual value calculate the index value of training to convergent deep learning model preset evaluation index.
S105, judge whether the index value meets pre-set level condition；Basis if meeting, after cross validation is rolled over K
Model and training to convergent deep learning model constitute combination forecasting.
Index value may include the RMSE (Root Mean Squard Error, root-mean-square error) and test set of training set
RMSE, alternatively, in other embodiments, or the mean absolute error (MAE) of training set and/or test set,
Square error (MSE) etc..
Judge whether the index value meets pre-set level condition, comprising:
If difference of the RMSE of test set less than the RMSE of the RMSE and test set of the first preset value and training set is less than second
Preset value, then Judging index value meets pre-set level condition；Otherwise, it is determined that index value does not meet pre-set level condition.
First preset value and the second preset value can be previously set, and the two may be the same or different.For example, can be with
Respectively 0.1 and 0.2..
The RMSE of test set is less than the first preset value, it is believed that training to convergent deep learning model precision compared with
It is high；The difference of the RMSE of the RMSE and test set of training set is less than the second preset value, it is believed that training to convergent deep learning
The generalization ability of model is stronger, therefore, if index value meets pre-set level condition, it is believed that training to convergent deep learning
The precision of model is higher and generalization ability is stronger, ensure that the precision and generalization ability of combination forecasting.
Alternatively, judging whether the index value meets pre-set level condition in other implementations, if may include:
The mean absolute error of test set is less than the first preset value, then Judging index value meets pre-set level condition；Otherwise, it is determined that index
Value does not meet pre-set level condition.
Using the embodiment of the present invention, after first completing K folding cross validation training by basic model, then exported with by basic model
Predicted value training obtain deep learning model, realize not only independent training pattern, but also maintain the relevance between model, mention
High model construction speed, and the index value of the deep learning model for constructing combination forecasting meets pre-set level item
Part improves the predictability of combination forecasting to ensure that the precision of combination forecasting and generalization ability are met the requirements
Energy.
If index value does not meet pre-set level condition, the method can also include:
The parameter of percentage regulation learning model is re-executed with the one-dimensional training vector training deep learning model.
Can using Adam (Adaptive Moment Estimation, adaptive moment estimation method) algorithm or
The parameter of the optimization algorithms percentage regulation learning model such as Adagrad method or stochastic gradient descent (SDG), and then again to adjustment
Deep learning model after parameter is trained.
In another implementation, multiple deep learning models can be preset, the parameter of each deep learning model is different；Sentence
Whether the index value of breaking meets pre-set level condition, if meeting, basic model and training after cross validation is rolled over K are extremely restrained
Deep learning model constitute combination forecasting, comprising:
Whether the index value of training of judgement to convergent each deep learning model meets pre-set level condition；
From each deep learning model that index value meets pre-set level condition, the optimal deep learning mould of selective goal value
For type as target deep learning model, it is pre- that basic model and target deep learning model after cross validation is rolled over K constitute combination
Survey model.
Such as multiple convolutional neural networks models are preset, it is the convolution kernel size of each convolutional neural networks model, port number, complete
The parameters such as articulamentum dimension, Dropout probability value can be different, are trained for each model, obtain training to convergence
Each convolutional neural networks model, from each convolutional neural networks model that index value meets pre-set level condition, selective goal
It is worth optimal convolutional neural networks model as target convolution neural network model, is commented for example, carrying out prediction model using RMSE
Valence, choosing the model that test set RMSE value is small and the difference of training set, test set RMSE is small is target convolution neural network model,
And then basic model and target convolution neural network model after cross validation are rolled over K and constitutes combination forecasting.It is specific to choose
The principle of model can be with are as follows: the model for first selecting test set RMSE value small into convergent each convolutional neural networks model from training,
The model for selecting the difference of training set and test set RMSE small from the model selected again is as target convolution neural network model.
If presetting a deep learning model, the parameter of model can be readjusted simultaneously in each training to after restraining
Re -training, and then obtain being trained under each parameter to convergent deep learning model, and the therefrom optimal depth of selective goal value
As target deep learning model, basic model and target deep learning model after cross validation is rolled over K are constituted learning model
Combination forecasting.
Training set and test set are predicted using the combination forecasting that the present invention constructs, it is available such as Fig. 3 institute
The prediction result shown compares combination forecasting of the invention and existing single XGBoost model, obtains comparison knot
Fruit is as shown in Figure 4, the results showed that XGBoost-CNN combination forecasting is on generalization ability and precision all than single XGBoost
Model is good.
As it can be seen that, since XGBoost model calculating speed is fast, can comparatively fast complete K folding cross validation using the embodiment of the present invention
Training, therefore after first completing K folding cross validation training by XGBoost model, then instructed with the predicted value that is exported by XGBoost model
Deep learning model is got, improves model construction speed, and constructed combination forecasting is in generalization ability and precision
It is upper to be higher than Individual forecast model, improve the estimated performance of combination forecasting.
Corresponding with above-mentioned embodiment of the method, the embodiment of the present invention also provides a kind of combination forecasting construction device.
Referring to Fig. 5, Fig. 5 is a kind of structural representation of combination forecasting construction device provided by the embodiment of the present invention
Figure, device include:
First obtains module 201, for initial data obtained to be divided into training set and test set；
Cross validation module 202 is rolled over cross validation for carrying out K to preset basic model on training set, is instructed
Practice the predicted value of collection；After each folding cross validation, the test set is carried out with the folding cross validation resulting basic model
Prediction, obtains test set in the predicted value of the folding cross validation；
Second obtains module 203, for using the predicted value of training set as one-dimensional training vector；And test set is calculated each
The average value for rolling over the predicted value of cross validation, obtains one-dimensional test vector；
Training module 204 obtains training to receipts for training preset deep learning model with the one-dimensional training vector
The deep learning model held back；One-dimensional test vector is predicted with training to convergent deep learning model, and based on prediction
As a result, calculating the index value of training to convergent deep learning model preset evaluation index；
Judgment module 205, for judging whether the index value meets pre-set level condition；If meeting, is intersected with K folding and tested
Basic model and training to convergent deep learning model after card constitutes combination forecasting.
Using the embodiment of the present invention, after first completing K folding cross validation training by basic model, then exported with by basic model
Predicted value training obtain deep learning model, realize not only independent training pattern, but also maintain the relevance between model, mention
High model construction speed, and the index value of the deep learning model for constructing combination forecasting meets pre-set level item
Part improves the prediction essence of combination forecasting to ensure that the precision of combination forecasting and generalization ability are met the requirements
Degree.
Optionally, deep learning model be convolutional neural networks model, including input layer, for single number to be converted to
The hidden layer of column vector, for by hidden layer convert resulting column vector be converted to the conversion layer of matrix, convolutional layer, pond layer,
Full articulamentum and output layer.
Optionally, the index value includes the RMSE of training set and the RMSE of test set, described in the judgment module judgement
Whether index value meets pre-set level condition, specifically:
If difference of the RMSE of test set less than the RMSE of the RMSE and test set of the first preset value and training set is less than second
Preset value, then Judging index value meets pre-set level condition；Otherwise, it is determined that index value does not meet pre-set level condition.
Optionally, the basic model is XGBoost model.
Optionally, described device further includes adjustment module, is used for:
When the judging result of the judgment module is no, the parameter of percentage regulation learning model is re-executed described in use
One-dimensional training vector training deep learning model.
Optionally, preset deep learning model has multiple, and the parameter of each deep learning model is different；The judgment module
It is specifically used for:
Whether the index value of training of judgement to convergent each deep learning model meets pre-set level condition；
From each deep learning model that index value meets pre-set level condition, the optimal deep learning mould of selective goal value
For type as target deep learning model, it is pre- that basic model and target deep learning model after cross validation is rolled over K constitute combination
Survey model.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Within mind and 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 combination forecasting construction method, which is characterized in that the described method includes:
Initial data obtained is divided into training set and test set；
K is carried out to preset basic model on training set and rolls over cross validation, obtains the predicted value of training set；Intersect in each folding
After verifying, the test set is predicted with the folding cross validation resulting basic model, test set is obtained and intersects in the folding
The predicted value of verifying；
Using the predicted value of training set as one-dimensional training vector；And calculate test set being averaged in each predicted value for rolling over cross validation
Value, obtains one-dimensional test vector；
With the preset deep learning model of the one-dimensional training vector training, training is obtained to convergent deep learning model；With
Training to convergent deep learning model predicts one-dimensional test vector, and is based on prediction result, calculates training and extremely restrains
Deep learning model preset evaluation index index value；
Judge whether the index value meets pre-set level condition；Basic model and training if meeting, after cross validation is rolled over K
Combination forecasting is constituted to convergent deep learning model.
2. the method according to claim 1, wherein deep learning model be convolutional neural networks model, including
Input layer, the hidden layer for single number to be converted to column vector are used to converting hidden layer into resulting column vector and are converted to
Conversion layer, convolutional layer, pond layer, full articulamentum and the output layer of matrix.
3. the method according to claim 1, wherein the index value includes the RMSE and test set of training set
RMSE, judges whether the index value meets pre-set level condition, comprising:
If difference of the RMSE of test set less than the RMSE of the RMSE and test set of the first preset value and training set is default less than second
Value, then Judging index value meets pre-set level condition；Otherwise, it is determined that index value does not meet pre-set level condition.
4. method according to claim 1-3, which is characterized in that the basic model is XGBoost model.
5. the method according to claim 1, wherein if index value does not meet pre-set level condition, the method
Further include:
The parameter of percentage regulation learning model is re-executed with the one-dimensional training vector training deep learning model.
6. method according to claim 1 or 2, which is characterized in that preset deep learning model has multiple, each depth
The parameter for practising model is different；Judge whether the index value meets pre-set level condition, if meeting, after rolling over cross validation with K
Basic model and training to convergent deep learning model constitute combination forecasting, comprising:
Whether the index value of training of judgement to convergent each deep learning model meets pre-set level condition；
From each deep learning model that index value meets pre-set level condition, the optimal deep learning model of selective goal value is made
For target deep learning model, basic model and target deep learning model after cross validation is rolled over K constitute combined prediction mould
Type.
7. a kind of combination forecasting construction device, which is characterized in that described device includes:
First obtains module, for initial data obtained to be divided into training set and test set；
Cross validation module rolls over cross validation for carrying out K to preset basic model on training set, obtains the pre- of training set
Measured value；After each folding cross validation, the test set is predicted with the folding cross validation resulting basic model, is obtained
Predicted value of the test set in the folding cross validation；
Second obtains module, for using the predicted value of training set as one-dimensional training vector；And it calculates test set and intersects in each folding
The average value of the predicted value of verifying obtains one-dimensional test vector；
Training module obtains training to convergent depth for training preset deep learning model with the one-dimensional training vector
Spend learning model；One-dimensional test vector is predicted with training to convergent deep learning model, and is based on prediction result, meter
Calculate the index value of training to convergent deep learning model preset evaluation index；
Judgment module, for judging whether the index value meets pre-set level condition；If meeting, after rolling over cross validation with K
Basic model and training to convergent deep learning model constitute combination forecasting.
8. device according to claim 7, which is characterized in that deep learning model is convolutional neural networks model, including
Input layer, the hidden layer for single number to be converted to column vector are used to converting hidden layer into resulting column vector and are converted to
Conversion layer, convolutional layer, pond layer, full articulamentum and the output layer of matrix.
9. device according to claim 7, which is characterized in that the index value includes the RMSE and test set of training set
RMSE, the judgment module judge whether the index value meets pre-set level condition, specifically:
If difference of the RMSE of test set less than the RMSE of the RMSE and test set of the first preset value and training set is default less than second
Value, then Judging index value meets pre-set level condition；Otherwise, it is determined that index value does not meet pre-set level condition.
10. according to the described in any item devices of claim 7-9, which is characterized in that the basic model is XGBoost model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910367539.3A CN110110848A (en) | 2019-05-05 | 2019-05-05 | A kind of combination forecasting construction method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910367539.3A CN110110848A (en) | 2019-05-05 | 2019-05-05 | A kind of combination forecasting construction method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110110848A true CN110110848A (en) | 2019-08-09 |
Family
ID=67488143
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910367539.3A Pending CN110110848A (en) | 2019-05-05 | 2019-05-05 | A kind of combination forecasting construction method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110110848A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111008732A (en) * | 2019-11-21 | 2020-04-14 | 中南大学 | Stack model-based fault prediction method and system |
| CN111062486A (en) * | 2019-11-27 | 2020-04-24 | 北京国腾联信科技有限公司 | Method and device for evaluating feature distribution and confidence coefficient of data |
| CN111369039A (en) * | 2020-02-25 | 2020-07-03 | 西安交通大学 | Alloy heat treatment process optimization method based on multi-objective optimization assisted active learning |
| CN111708865A (en) * | 2020-06-18 | 2020-09-25 | 海南大学 | Technology forecasting and patent early warning analysis method based on improved XGboost algorithm |
| CN112380041A (en) * | 2020-11-13 | 2021-02-19 | 重庆金美通信有限责任公司 | Finger control communication equipment fault prediction method based on xgboost |
| CN113222208A (en) * | 2021-03-26 | 2021-08-06 | 辽宁工程技术大学 | Ada-XGboost-based traffic accident prediction system |
| CN113452379A (en) * | 2021-07-16 | 2021-09-28 | 燕山大学 | Section contour dimension reduction model training method and system and data compression method and system |
| CN113570260A (en) * | 2021-07-30 | 2021-10-29 | 北京房江湖科技有限公司 | Task allocation method, computer-readable storage medium and electronic device |
| CN114782775A (en) * | 2022-04-19 | 2022-07-22 | 平安科技（深圳）有限公司 | Method and device for constructing classification model, computer equipment and storage medium |
| CN115392482A (en) * | 2022-08-25 | 2022-11-25 | 上海人工智能创新中心 | Deep learning model selection method, device and equipment |
-
2019
- 2019-05-05 CN CN201910367539.3A patent/CN110110848A/en active Pending
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111008732A (en) * | 2019-11-21 | 2020-04-14 | 中南大学 | Stack model-based fault prediction method and system |
| CN111008732B (en) * | 2019-11-21 | 2023-06-20 | 中南大学 | Stack model-based fault prediction method and system |
| CN111062486A (en) * | 2019-11-27 | 2020-04-24 | 北京国腾联信科技有限公司 | Method and device for evaluating feature distribution and confidence coefficient of data |
| CN111062486B (en) * | 2019-11-27 | 2023-12-08 | 北京国腾联信科技有限公司 | Method and device for evaluating feature distribution and confidence of data |
| CN111369039A (en) * | 2020-02-25 | 2020-07-03 | 西安交通大学 | Alloy heat treatment process optimization method based on multi-objective optimization assisted active learning |
| CN111369039B (en) * | 2020-02-25 | 2022-06-07 | 西安交通大学 | Alloy heat treatment process optimization method based on multi-objective optimization assisted active learning |
| CN111708865A (en) * | 2020-06-18 | 2020-09-25 | 海南大学 | Technology forecasting and patent early warning analysis method based on improved XGboost algorithm |
| CN111708865B (en) * | 2020-06-18 | 2021-07-09 | 海南大学 | Technology forecasting and patent early warning analysis method based on improved XGboost algorithm |
| CN112380041A (en) * | 2020-11-13 | 2021-02-19 | 重庆金美通信有限责任公司 | Finger control communication equipment fault prediction method based on xgboost |
| CN112380041B (en) * | 2020-11-13 | 2023-11-14 | 重庆金美通信有限责任公司 | Xgboost-based failure prediction method for command communication equipment |
| CN113222208A (en) * | 2021-03-26 | 2021-08-06 | 辽宁工程技术大学 | Ada-XGboost-based traffic accident prediction system |
| CN113452379A (en) * | 2021-07-16 | 2021-09-28 | 燕山大学 | Section contour dimension reduction model training method and system and data compression method and system |
| CN113452379B (en) * | 2021-07-16 | 2022-08-02 | 燕山大学 | Section contour dimension reduction model training method and system and data compression method and system |
| CN113570260A (en) * | 2021-07-30 | 2021-10-29 | 北京房江湖科技有限公司 | Task allocation method, computer-readable storage medium and electronic device |
| CN114782775A (en) * | 2022-04-19 | 2022-07-22 | 平安科技（深圳）有限公司 | Method and device for constructing classification model, computer equipment and storage medium |
| CN115392482A (en) * | 2022-08-25 | 2022-11-25 | 上海人工智能创新中心 | Deep learning model selection method, device and equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110110848A (en) | A kind of combination forecasting construction method and device | |
| Gaur et al. | Real-time wave forecasting using genetic programming | |
| Alencar et al. | Hybrid approach combining SARIMA and neural networks for multi-step ahead wind speed forecasting in Brazil | |
| Yoon et al. | A method to improve the stability and accuracy of ANN-and SVM-based time series models for long-term groundwater level predictions | |
| CN107273429B (en) | A kind of Missing Data Filling method and system based on deep learning | |
| JP6959840B2 (en) | Water level prediction method, water level prediction program and water level prediction device | |
| CN104700321B (en) | A kind of power transmission and transformation equipment state operation trend analysis method | |
| CN107220618A (en) | Method for detecting human face and device, computer-readable recording medium, equipment | |
| Braakmann-Folgmann et al. | Sea level anomaly prediction using recurrent neural networks | |
| CN104899431B (en) | Based on ant colony and swarm of particles into LSSVM fluctuating wind speed Forecasting Methodologies | |
| CN107506938A (en) | A kind of quality of material appraisal procedure based on machine learning | |
| CN107563122A (en) | The method of crime prediction of Recognition with Recurrent Neural Network is locally connected based on interleaving time sequence | |
| CN108346293A (en) | A kind of arithmetic for real-time traffic flow Forecasting Approach for Short-term | |
| CN103942461A (en) | Water quality parameter prediction method based on online sequential extreme learning machine | |
| CN109255469A (en) | Merge the Forecasting Flood method of stack self-encoding encoder and support vector regression | |
| CN107045785A (en) | A kind of method of the short-term traffic flow forecast based on grey ELM neutral nets | |
| CN105933857A (en) | Mobile terminal position prediction method and apparatus | |
| Guan | Wave height prediction based on CNN-LSTM | |
| Awan et al. | Application of artificial neural networks for monsoon rainfall prediction | |
| Li et al. | A method of rainfall runoff forecasting based on deep convolution neural networks | |
| Hoang et al. | Weather prediction based on LSTM model implemented AWS Machine Learning Platform | |
| Reaz et al. | Prototyping of wavelet transform, artificial neural network and fuzzy logic for power quality disturbance classifier | |
| CN117034780B (en) | Multi-scale sub-season precipitation prediction method based on deep learning | |
| Zhong et al. | Airrl: A reinforcement learning approach to urban air quality inference | |
| CN109934845A (en) | Based on the timing behavior capture frame generation method and device from attention network |
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: 20190809 |