CN116706895A - Ultra-short-term wind power prediction method considering deep features and capability constraint - Google Patents

Ultra-short-term wind power prediction method considering deep features and capability constraint Download PDF

Info

Publication number
CN116706895A
CN116706895A CN202310716534.3A CN202310716534A CN116706895A CN 116706895 A CN116706895 A CN 116706895A CN 202310716534 A CN202310716534 A CN 202310716534A CN 116706895 A CN116706895 A CN 116706895A
Authority
CN
China
Prior art keywords
wind power
data
ultra
decomposition
wavelet
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
Application number
CN202310716534.3A
Other languages
Chinese (zh)
Inventor
申艳杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Individual
Original Assignee
Individual
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Individual filed Critical Individual
Priority to CN202310716534.3A priority Critical patent/CN116706895A/en
Publication of CN116706895A publication Critical patent/CN116706895A/en
Pending legal-status Critical Current

Links

Classifications

    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H02J3/004—Generation forecast, e.g. methods or systems for forecasting future energy generation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • 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
    • G06N3/048—Activation functions
    • 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
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H02J3/38—Arrangements for parallely feeding a single network by two or more generators, converters or transformers
    • H02J3/381—Dispersed generators
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Power Engineering (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Wind Motors (AREA)

Abstract

The invention provides an ultra-short-term wind power prediction method considering deep features and capability constraint, which comprises the steps of firstly filling, outlier processing, smoothing processing and wavelet denoising processing on an original wind power data missing value based on a multi-pattern interpolation method, then carrying out normalization processing on the wavelet denoised wind power data, carrying out variation modal decomposition VMD on a normalized data set, extracting feature variables of different modal components, extracting feature variables with importance larger than K by utilizing a random forest algorithm, forming a final sample set, and training an established wind power time sequence prediction model based on a deep neural network. The method can restrict the input and output data to remove abnormal values in the data, can deeply mine effective characteristics in the original wind power sequence, improves the ultra-short-term wind power prediction accuracy, and is beneficial to improving the resource allocation problems of projects such as a power grid, a wind power plant and the like.

Description

Ultra-short-term wind power prediction method considering deep features and capability constraint
Technical Field
The invention relates to the field of wind power prediction, in particular to an ultra-short-term wind power prediction method considering deep features and capability constraint.
Background
Wind energy has the characteristics of fluctuation, randomness, intermittence and the like, and can cause the problems of large fluctuation of wind power generation power, unstable output power and the like, thereby affecting the safe operation of a power grid. The wind power change trend of the wind power of the wind turbine and the wind power plant can be obtained in advance by the wind power prediction of the wind turbine and the wind power plant, and the problems of high fluctuation of generated power, unstable output power and the like can be solved in advance. Therefore, research on wind power prediction technology improves the accuracy of wind power prediction, and has important significance.
The ultra-short term wind power prediction method is mainly divided into a digital statistical method and a neural network method. The mathematical statistics method is mainly to conduct statistical prediction by analyzing rules existing among data. Because wind power data has randomness, the wind power prediction accuracy adopting a mathematical statistical method is difficult to meet the actual needs. The artificial neural network method has good fitting effect on nonlinear time sequence data, so that the neural network method is more advantageous than a mathematical statistics method in the aspect of ultra-short-term wind power prediction.
In order to obtain a higher-precision ultra-short-term wind power prediction model, deeper features, such as spectrum features and the like, need to be mined from an original wind power sequence, and a wind power model considering the deep features is constructed by extracting the features with strong correlation and adopting a neural network method.
Disclosure of Invention
Aiming at the defects existing in the prior art, the invention aims to provide an ultra-short-term wind power prediction method taking deep features and capability constraints into consideration, and provides a mixed wind power prediction method of variation modal decomposition-feature extraction-random forest-deep neural network, which has higher precision in ultra-short-term wind power prediction.
In order to achieve the above object, the present invention provides an ultra-short term wind power prediction method considering deep features and capability constraints, comprising:
step 1: time L is extracted in time sequence with sampling period delta t 1 Wind power original data in the fan to construct a fan operation original data set YT 0 ＝{(Y(t n ) 0 ) N=1, 2, & N; wherein t is n Represents the time corresponding to the nth sampling point, N is the total number of sampling points, t n+1 -t n ＝Δt，L 1 ＝(N-1)Δt，Y(t n ) 0 At t n Raw data of wind power at moment;
step 2: preprocessing the original data to construct a sample set; comprising the following steps:
step 2.1: original data set YT based on multiple interpolation method 0 Filling the missing value of the wind power data; the multi-mode interpolation method is an algorithm for carrying out data interpolation according to the type of the missing value, and the formula is shown in the formula (1);
wherein, the continuous missing wind power data refers to data of missing wind power values at more than 2 continuous moments; the discontinuous missing wind power data means wind power data in which only middle time missing data in adjacent 3 times. Obtaining the fan operation data set as YT 1 ＝{(Y(t n ) 1 )|n＝1,2,...,N}，YT 1 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 1 Is the t th processed based on the multi-mode interpolation method n Wind power data at the moment;
step 2.2: carrying out comprehensive outlier processing on wind power data processed based on a multi-mode interpolation method; the method comprises the following steps:
step 2.2.1: if the wind power data Y (t n ) 1 Is larger than the maximum rated power value Y (t) n ) emax Processing the data, wherein the formula is shown as (2);
Y(t n ) 2 ＝{Y(t n ) emax |if(Y(t n ) 1 ＞Y(t n ) emax } (2)
wherein ,Y(tn ) emax Is the maximum rated power value of the fan; obtaining the fan operation data set as YT 2 ＝{(Y(t n ) 2 )|n＝1,2,...,N}，YT 2 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 2 Is t after being processed based on LDIAOM outlier n Wind power data at the moment;
step 2.2.2: pair Y (t) n ) 2 Performing outlier processing based on random standard deviation to obtain a fan operation data set YT 3 ＝{(Y(t n ) 3 )|n＝1,2,...,N}，YT 3 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 3 Post-processing t for outlier based on random standard deviation n Wind power data at the moment; outlier processing based on random standard deviation comprises the following specific steps:
step 2.2.2.1: computing wind power dataset YT 2 Arithmetic mean of (2)The calculation formula is shown as (3):
wherein N is YT 2 Wind power number in wind power data set:
step 2.2.2.2: obtainingThereafter, the standard deviation σ is calculated according to equation (4):
step 2.2.2.3: determining a wind power standard interval Ω K The method comprises the following steps:
wherein k=1, 2, N;
step 2.2.2.4: if the wind power valueY(t n ) 2 Not belonging to standard interval omega K Y (t) n ) 2 Replaced by Y (t) n ) 3 ；
Step 2.3: smoothing the wind power data processed by the abnormal value to obtain a fan operation data set YT 4 ＝{(Y(t n ) 4 )|n＝1,2,…,N}，YT 4 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the The smoothing calculation formula is shown as (7), wherein Y (t) n ) 4 For t after smoothing n Wind power data at the moment;
step 2.4: for the wind power data Y (t n ) 4 Performing wavelet denoising treatment; comprising the following steps:
step 2.4.1: pair Y (t) n ) 4 Performing wavelet decomposition to obtain decomposition signal with M as decomposition scale
Step 2.4.2: the approximate coefficient and the detail coefficient are obtained after wavelet decomposition, the approximate coefficient mainly represents the real signal, and the detail coefficient mainly comprises the noise signal, so that the detail coefficient needs to be processed by adopting a soft threshold function to reduce the noise signal, and the soft threshold formula is as follows:
where w represents the wavelet coefficient, thr represents a given threshold, w thr Representing the wavelet coefficient after threshold processing;
finally, carrying out wavelet according to the denoised wavelet coefficientWave reconstruction to obtain a denoised signal YT 5 ＝{(Y(t n ) 5 )|n＝1,2,...,N}，YT 5 ∈R N×1 Wherein Y (t) n ) 5 T after denoising for wavelet n Wind power data at the moment;
step 2.5: normalizing the wind power data after wavelet denoising to obtain a normalized data set YT 6 ＝{(Y(t n ) 6 ) N=1, 2, …, N }, the calculation formula is shown as (9);
wherein ,Y(tn ) 6 Is t < th > after normalization n Wind power data; y (t) n ) 5 max Is the maximum value in the wind power data after wavelet denoising, Y (t n ) 5 min The minimum value in the wind power data after wavelet denoising is set;
step 2.6: for normalized data set YT 6 Performing Variational Modal Decomposition (VMD) to obtain H modal components a=1, 2, and extracting characteristic variables of different modal components;
step 2.6.1: para YT 6 VMD is performed, the decomposition steps are as follows:
step 2.6.1.1: converting an intrinsic mode function into an amplitude modulation frequency modulation signal u with limited bandwidth a Taking the minimum sum of bandwidths estimated by the intrinsic mode functions as a constraint condition of decomposition, and calculating a formula as shown in (10);
wherein ,ua Is the a-th modal component signal after VMD decomposition, which is convolution, omega a Is the center frequency of the a-th modal component signal,expressed in the formula of bracketsT is derived, t represents time, delta t Is a dirac function;
step 2.6.1.2: converting the formula (10) into an unconstrained problem by introducing a quadratic penalty term and a Lagrangian multiplier beta, wherein the calculation formula is shown as (11);
step 2.6.1.3: using an alternate multiplier direction method to model the function u a Center frequency omega a Iteratively updating Lagrangian multiplier beta, wherein the calculation formulas are shown as (12), (13) and (14);
wherein ,for the wiener filtering result of the residual component, +.>The center frequency of the power spectrum of the modal function is represented by tau, which is an update factor;
step 2.6.1.4: repeating the steps 2.6.1.1, 2.6.1.2 and 2.6.1.3 when the conditions are satisfiedWhen the iteration is stopped, the convergence is considered, wherein f is a threshold value;
will correlate sample set YT 6 Is equal to the number of samples Y (t) n ) 6 Decomposing to obtainH modal components with limited bandwidth and forming a sample set as Represents the t n The modal component of the a-th limited bandwidth of the moment, DATA ε R N×H ；
Step 2.6.2: extracting characteristic variables of different modal components;
the characteristic variables include: the maximum value (Max), the minimum value (Min), and (Sum), the average value (Mean), the Range (Range), and the feature extraction are performed on the sample set DATA of the decomposed wind power, and the sample set is formed as follows: DATA 2= { (Max (t) n ),Min(t n ),Sum(t n ),Mean(t n ),Range(t n ))|n＝1,2,...,N}，DATA2∈R N×5 ， Represents t n Maximum value of H modal components at time, < -> Represents t n The minimum of the H modal components at the time instant, represents t n The H modal components add at the moment +.> Represents t n Average value of H modal components at time, range (t n )＝{Max(t n )-Min(t n )|n＝1,2,...,N}，(Max(t n )-Min(t n ) (d) represents t n The difference between the maximum value and the minimum value in the H modal components at the moment;
step 2.7: calculating the importance of a characteristic variable in DATA2 by using a random forest algorithm, selecting the characteristic variable according to the characteristic importance, extracting the characteristic variable with the importance larger than K, wherein the characteristic importance calculation is an embedded function of the random forest algorithm, and taking the variable of the out-of-bag (OOB) DATA classification accuracy as an evaluation criterion; the feature importance is calculated as follows;
step 2.7.1: the number of samples was n=1, 2,.. initializing n=1, creating a decision tree L n ；
Step 2.7.2: for decision tree L n Training, and calculating classification accuracy of n pieces of data outside the bag
Step 2.7.3: for features W in the out-of-bag dataset n1 Disturbance and calculation of classification accuracy
Step 2.7.4: by characteristics W n1 Is related to the magnitude of the correlation I n1 As shown in the formula;
step 2.7.5: through I n1 The importance ranking is carried out on the calculated correlation magnitude;
according to the feature importance ranking, G feature variables with the greatest importance are selected as extracted feature variables to form a sample set DATA3 of the extracted feature variables with strong correlation, DATA3 epsilon R N×G ；
Step 2.8: the selected feature set DATA3, the decomposed wind power DATA DATA and the normalized wind power DATA YT 6 Is denoted as DATA4 = { (YT) as a sample input set 6 ,DATA,DATA3)}，DATA4∈R N ×(1+H+G) ，YT out ＝{(Y(t n+1 ) 6 ) I n=1, 2, …, N } is the sample output set, YT out ∈R N×1 ；
Step 3: establishing a wind power time sequence prediction model based on a deep neural network DNN, taking 70% of DATA before DATA4 as input of a training set and YT out The first 70% of DATA is used as the output of the training set to form the training set, the remaining 30% of DATA of DATA4 is used as the input of the test set, and the training set is used for training the prediction model; according to the DNN structure, the method comprises the steps of dividing the DNN structure into an input layer, J hidden layers, r=1, 2, and J, and an output layer, wherein in the DNN modeling process, each hidden layer obtains an input value from the previous hidden layer, then nonlinear processing is carried out on an input variable through an activation function of the layer, finally, the processed data is used as output of the layer to be transmitted to the next hidden layer until the last hidden layer is finished, and a prediction result is output. The predictive calculation result formulas are shown as (21) and (22);
I r ＝g(w r ·YT 6 +b r ) (21)
wherein ,Ir Output matrix representing the r hidden layer, w r Is the weight parameter of the r hidden layer, b r Is the threshold parameter of the r-th hidden layer, g (x) represents the activation function Sigmoid;
YT 7 ＝g(w J ·I J-1 +b J ) (22)
wherein ,wJ Is the weight parameter between the last hidden layer and the output layer, b J Is the threshold parameter between the last hidden layer and the output layer, I J-1 Is the output matrix of the penultimate hidden layer;
finally, the wind power prediction data set is YT 7 ＝{(Y(t n ) 7 )|n＝1,2,…,N}，YT 7 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 7 Is t after inverse normalization n Predicted wind power data;
step 4: for the obtained actual wind power data, outputting a predicted value of wind power by using a trained prediction model, if the outputted predicted value of wind power is larger than the maximum rated power value Y (t) n ) emax Then the predicted value needs to be processed by an abnormal value based on LDIAOM to obtain a final wind power data set YT 8 ＝{(Y(t n ) 8 )|n＝1,2,…,N}，YT 8 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 8 Is the t after data processing n And wind power data.
The invention has the beneficial effects that:
according to the ultra-short-term wind power prediction method considering deep features and capability constraint, not only can abnormal values in input and output data be removed, but also effective features in an original wind power sequence can be deeply mined, and the ultra-short-term wind power prediction precision is improved.
Drawings
FIG. 1 is a graph of the original wind power curve according to the present invention;
FIG. 2 is a graph showing wind power curves after anomaly rejection and interpolation in the present invention;
FIG. 3 is a graph showing the decomposition results of VMD according to the present invention;
FIG. 4 is a schematic diagram of a DNN structure network according to the present invention;
FIG. 5 is a flow chart of an ultra-short term wind power prediction method taking deep features and capability constraints into consideration.
Detailed Description
The invention will be further described with reference to the accompanying drawings and examples of specific embodiments.
As shown in FIG. 5, the present invention provides an ultra-short term wind power prediction method taking deep features and capability constraints into account, comprising:
step 1: time L is extracted in time sequence with sampling period delta t 1 Wind power original data in the fan to construct a fan operation original data set YT 0 ＝{(Y(t n ) 0 ) N=1, 2, & N; wherein t is n Represents the time corresponding to the nth sampling point, N is the total number of sampling points, t n+1 -t n ＝Δt，L 1 ＝(N-1)Δt，Y(t n ) 0 At t n Raw data of wind power at moment;
step 2: preprocessing the original data to construct a sample set; comprising the following steps:
step 2.1: original data set YT based on multiple interpolation method 0 Filling the missing value of the wind power data; the multi-pattern interpolation method is an algorithm for performing data interpolation according to the missing value type, and as shown in fig. 2, the data is filled by the multi-pattern interpolation method, and as shown in fig. 1, the data is not filled by the interpolation method. The formula is shown as (1);
wherein, the continuous missing wind power data refers to data of missing wind power values at more than 2 continuous moments; the discontinuous missing wind power data means wind power data in which only middle time missing data in adjacent 3 times. Obtaining the fan operation data set as YT 1 ＝{(Y(t n ) 1 )|n＝1,2,...,N}，YT 1 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 1 Is the t th processed based on the multi-mode interpolation method n Wind power data at the moment;
step 2.2: carrying out comprehensive outlier processing on wind power data processed based on a multi-mode interpolation method; the method comprises the following steps:
step 2.2.1: if the wind power data Y (t n ) 1 Is larger than the maximum rated power value Y (t) n ) emax Processing the data, wherein the formula is shown as (2);
Y(t n ) 2 ＝{Y(t n ) emax |if(Y(t n ) 1 ＞Y(t n ) emax } (2)
wherein ,Y(tn ) emax Is the maximum rated power value of the fan; obtaining the fan operation data set as YT 2 ＝{(Y( t n) 2 )|n＝1,2,...,N}，YT 2 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 2 Is t after being processed based on LDIAOM outlier n Wind power data at the moment;
step 2.2.2: pair Y (t) n ) 2 Performing outlier processing based on random standard deviation to obtain a fan operation data set YT 3 ＝{(Y(t n ) 3 )|n＝1,2,...,N}，YT 3 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 3 Post-processing t for outlier based on random standard deviation n Wind power data at the moment; the outlier processing based on the random standard deviation specifically comprises the following steps:
step 2.2.2.1: computing wind power dataset YT 2 Arithmetic mean of (2)The calculation formula is shown as (3):
wherein N is YT 2 Wind power number in wind power data set:
step 2.2.2.2: obtainingThereafter, the standard deviation σ is calculated according to equation (4):
step 2.2.2.3: determining a wind power standard interval Ω K The method comprises the following steps:
wherein k=1, 2, N;
step 2.2.2.4: if the wind power value Y (t) n ) 2 Not belonging to standard interval omega K Y (t) n ) 2 Substitution Y (t) n ) 3 ；
Step 2.3: smoothing the wind power data processed by the abnormal value to obtain a fan operation data set YT 4 ＝{(Y(t n ) 4 )|n＝1,2,…,N}，YT 4 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the The smoothing calculation formula is shown as (7), wherein Y (t) n ) 4 For t after smoothing n Wind power data at the moment;
step 2.4: for the wind power data Y (t n ) 4 Performing wavelet denoising treatment;
step 2.4.1: pair Y (t) n ) 4 Wavelet decomposition is performed to obtain a decomposition signal { Y (t) n ) 4 L ,L＝1，2，...，M}；
Step 2.4.2: the method comprises the steps of obtaining a generated approximate coefficient and a detail coefficient after wavelet decomposition, wherein the approximate coefficient mainly represents a real signal, and the detail coefficient mainly comprises a noise signal, so that the detail coefficient needs to be processed by adopting a soft threshold function for noise signal reduction, and the soft threshold formula is as follows;
where w represents the wavelet coefficient, thr represents a given threshold, w thr Representing the wavelet coefficient after threshold processing;
finally, carrying out wavelet reconstruction according to the denoised wavelet coefficient to obtain a denoised signal YT 5 ＝{(Y(t n ) 5 )|n＝1,2,...,N}，YT 5 ∈R N×1 Wherein Y (t) n ) 5 T after denoising for wavelet n Wind power data at the moment;
step 2.5: normalizing the wind power data after wavelet denoising to obtain a normalized data set YT 6 ＝{(Y(t n ) 6 ) N=1, 2,..n }, the calculation formula is shown as (9);
wherein ,Y(tn ) 6 Is t < th > after normalization n Wind power data; y (t) n ) 5 max Is the maximum value in the wind power data after wavelet denoising, Y (t n ) 5 min The minimum value in the wind power data after wavelet denoising is set;
step 2.6: para YT 6 Performing Variable Modal Decomposition (VMD), setting the decomposition layer number to 6 layers as shown in fig. 3, obtaining 6 modal components a=1, 2, 6, and extracting characteristic variables of different modal components;
step 2.6.1: para YT 6 Performing VMD, and decomposing as follows;
step 2.6.1.1: converting an intrinsic mode function into an amplitude modulation frequency modulation signal u with limited bandwidth a Taking the minimum sum of the estimated bandwidths of the natural mode functions as a constraint condition of decomposition, wherein a calculation formula is shown as (10);
wherein ,ua Is the a-th modal component signal after VMD decomposition, which is convolution, omega a Is the center frequency of the a-th modal component signal,represent deriving t in bracket formula, t represents time, delta t Is a dirac function;
step 2.6.1.2: converting the formula (10) into an unconstrained problem by introducing a quadratic penalty term and a Lagrangian multiplier beta, wherein the calculation formula is shown as (11);
step 2.6.1.3: using an alternate multiplier direction method to model the function u a Center frequency omega a Iteratively updating Lagrangian multiplier beta, wherein the calculation formulas are shown as (12), (13) and (14);
wherein ,for the wiener filtering result of the residual component, +.>Is the center frequency of the power spectrum of the mode function, and τ is the update factor.
Step 2.6.1.4: repeating the steps 2.6.1.1, 2.6.1.2 and 2.6.1.3 when the conditions are satisfiedWhen (f is a threshold value), the convergence is considered to stop iteration;
will correlate the sampleSet YT 6 Is equal to the number of samples Y (t) n ) 6 Decomposing to obtain H modal components with limited bandwidth and forming a sample set as Represents the t n The modal component of the a-th limited bandwidth of the moment, DATA ε R N×H ；
Step 2.6.2: extracting characteristic variables of different modal components;
the characteristic variables include: the maximum value (Max), the minimum value (Min), and (Sum), the average value (Mean), the Range (Range), and the feature extraction are performed on the sample set DATA of the decomposed wind power, and the sample set is formed as follows: DATA 2= { (Max (t) n ),Min(t n ),Sum(t n ),Mean(t n ),Range(t n ))|n＝1,2,...,N}，DATA2∈R N×5 ， Represents t n Maximum value of H modal components at time, < -> Represents t n The minimum of the H modal components at the time instant, represents t n The H modal components add at the moment +.> Represents t n Average value of H modal components at time, range (t n )＝{Max(t n )-Min(t n )|n＝1,2,...,N}，(Max(t n )-Min(t n ) (d) represents t n The difference between the maximum value and the minimum value in the H modal components at the moment;
step 2.7: the importance of the characteristic variable in the DATA2 is calculated by utilizing a random forest algorithm, the characteristic variable is selected according to the characteristic importance, the characteristic variable with the importance larger than 0.4 is extracted, the characteristic importance calculation is an embedded function of the random forest algorithm, and the variable of the out-of-bag (OOB) DATA classification accuracy is taken as an evaluation criterion. The feature importance is calculated as follows;
step 2.7.1: the number of samples was n=1, 2,.. initializing n=1, creating a decision tree L n ；
Step 2.7.2: for decision tree L n Training, and calculating classification accuracy of n pieces of data outside the bag
Step 2.7.3: for features W in the out-of-bag dataset n1 Disturbance and calculation of classification accuracy
Step 2.7.4: by characteristics W n1 Is related to the magnitude of the correlation I n1 As shown in the formula;
step 2.7.5: through I n1 The calculated correlation magnitudes are ranked for importance as shown in table 1;
TABLE 1 characteristic variable correlation comparison Table
Characteristic variable Correlation size
Maximum value (Max) 0.212
Minimum value (Min) 0.012
and (Sum) 0.409
Average (Mean) 0.412
Extremely poor (Range) 0.312
And selecting the characteristic variable with the correlation larger than 0.4 as the extracted characteristic variable according to the characteristic correlation. Therefore, the Sum and average are selected as extracted feature variables, and a sample set of feature variables forming a strong correlation after extraction is data3= { (Sum (t) n ),Mean(t n ))|n＝1,2,...,N}，DATA3∈R N×2 ；
Step 2.8: the selected feature set DATA3 and the decomposed wind power DATA DATA and the normalized wind power DATA YT 6 The composition is noted as a sample input set as data4= { (YT) 6 ,DATA,DATA3)}，DATA4∈R N×9 YT is to out ＝{(Y(t n+1 ) 6 )|n＝1,2,...,N}，YT out ∈R N×1 Is a sample output set;
step 3: establishing a baseIn the wind power time series prediction model of the deep neural network DNN, as shown in FIG. 4, 70% of the DATA before DATA4 is used as the input of the training set and YT out The first 70% of DATA is used as the output of the training set to form the training set, the remaining 30% of DATA of DATA4 is used as the input of the test set, and the training set is used for training the prediction model; according to the DNN structure, the method comprises the steps of dividing the DNN structure into an input layer, J hidden layers, r=1, 2, and J, and an output layer, wherein in the DNN modeling process, each hidden layer obtains an input value from the previous hidden layer, then nonlinear processing is carried out on an input variable through an activation function of the layer, finally, the processed data is used as output of the layer to be transmitted to the next hidden layer until the last hidden layer is finished, and a prediction result is output. The predictive calculation result formulas are shown as (21) and (22);
I r ＝g(w r ·YT 6 +b r ) (21)
wherein ,Ir Output matrix representing the r hidden layer, w r Is the weight parameter of the r hidden layer, b r Is the threshold parameter of the r-th hidden layer, g (x) represents the activation function Sigmoid;
YT 7 ＝g(w J ·I J-1 +b J ) (22)
wherein ,wJ Is the weight parameter between the last hidden layer and the output layer, b J Is the threshold parameter between the last hidden layer and the output layer, I J-1 Is the output matrix of the penultimate hidden layer;
finally, the wind power prediction data set is YT 7 ＝{(Y(t n ) 7 )|n＝1,2,...,N}，YT 7 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 7 Is t after inverse normalization n Predicted wind power data;
step 4: for the obtained actual wind power data, outputting a predicted value of wind power by using a trained prediction model, if the outputted predicted value of wind power is larger than the maximum rated power value Y (t) n ) emax Then the predicted value needs to be processed by an abnormal value based on LDIAOM to obtain the final wind power numberFrom the set YT 8 ＝{(Y(t n ) 8 )|n＝1,2,...,N}，YT 8 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 8 Is the t after data processing n Wind power data;
table 2 comparison of ultrashort-term wind power predictive modeling without considering deep features and capability constraints with the results of the inventive method (i.e., ultrashort-term wind power predictive modeling with consideration of deep features and capability constraints)
As can be seen from Table 2, the modeling accuracy obtained by ultra-short term wind power predictive modeling with deep features and capacity constraints is 37.037% improved by R-squared, 27.23% by MSE, 23.87% by MAE, and 37.52% by MAPE, as compared to ultra-short term wind power predictive modeling without deep features and capacity constraints.
The method can restrict the input and output data to remove abnormal values in the data, can deeply mine effective features in the original wind power sequence, and improves the ultra-short-term wind power prediction accuracy. The wind power value obtained by the method can be applied to actual scenes such as power grid dispatching, wind power field management and the like. In power grid dispatching, a dispatcher estimates the power generation amount of a wind power plant according to a wind power predicted value, so as to correspondingly adjust the power generation amount distribution of each power generation unit, solve the problem of unbalanced supply and demand sides caused by wind power randomness and fluctuation, and ensure the stable operation of a power system; in the aspect of wind power plant management, wind power plant operation and maintenance personnel can reasonably arrange the starting and stopping time of a fan according to predicted wind power so as to utilize wind energy to the maximum extent, and meanwhile, overload of a machine set or energy waste is avoided.

Claims (10)

1. An ultra-short term wind power prediction method taking deep features and capability constraints into account, comprising:
step 1: time L is extracted in time sequence with sampling period delta t 1 Wind power source in the interiorInitial data, constructing a fan operation original data set;
step 2: preprocessing the original data to construct a sample set;
step 3: establishing a wind power time sequence prediction model based on a deep neural network DNN, and training the prediction model by using a training set;
step 4: for the obtained actual wind power data, outputting a predicted value of wind power by using a trained prediction model, if the outputted predicted value of wind power is larger than the maximum rated power value Y (t) n ) emax Then, the predicted value needs to be processed by an abnormal value based on LDIAOM, so as to obtain a final wind power data set.
2. The ultra-short term wind power prediction method considering deep features and capability constraints as claimed in claim 1, wherein said step 2 comprises:
step 2.1: original data set YT based on multiple interpolation method 0 Filling the missing value of the wind power data;
step 2.2: carrying out comprehensive outlier processing on wind power data processed based on a multi-mode interpolation method;
step 2.3: smoothing the wind power data after the abnormal value processing;
step 2.4: carrying out wavelet denoising treatment on the smoothed wind power data;
step 2.5: carrying out normalization processing on the wind power data after wavelet denoising;
step 2.6: performing variational modal decomposition VMD on the normalized data set, and extracting characteristic variables of different modal components;
step 2.7: calculating the importance of the feature variable by using a random forest algorithm, selecting the feature variable according to the feature importance, and extracting the feature variable with the importance larger than K;
step 2.8: and taking the combination of the selected characteristic variable, the decomposed wind power data and the normalized wind power data as a sample set.
3. The ultra-short term wind power prediction method considering deep features and capability constraints according to claim 1, wherein the multi-pattern interpolation method in step 2.1 is an algorithm for performing data interpolation according to the missing value type, and the formula is shown in (1);
wherein, the continuous missing wind power data refers to data of missing wind power values at more than 2 continuous moments; the discontinuous missing wind power data refers to wind power data in which only middle time missing data in 3 adjacent time points, and the obtained wind turbine running data set is YT 1 ＝{(Y(t n ) 1 )|n＝1,2,...,N}，YT 1 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 1 Is the t th processed based on the multi-mode interpolation method n Wind power data at time.
4. A method of ultra-short term wind power prediction taking into account deep features and capacity constraints according to claim 3, wherein said step 2.2 comprises:
step 2.2.1: if the wind power data Y (t n ) 1 Is larger than the maximum rated power value Y (t) n ) emax Processing the data, wherein the formula is shown as (2);
Y(t n ) 2 ＝{Y(t n ) emax |if(Y(t n ) 1 ＞Y(t n ) emax } (2)
wherein ,Y(tn ) emax Is the maximum rated power value of the fan; obtaining the fan operation data set as YT 2 ＝{(Y(t n ) 2 )|n＝1,2,...,N}，YT 2 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 2 Is t after being processed based on LDIAOM outlier n Wind power data at the moment;
step 2.2.2: pair Y (t) n ) 2 Performing outlier processing based on random standard deviation to obtain a fan operation data set YT 3 ＝{(Y(t n ) 3 )|n＝1,2,...,N}，YT 3 ∈R N×1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein Y (t) n ) 3 Post-processing t for outlier based on random standard deviation n Time of day wind power data.
5. The ultra-short term wind power prediction method considering deep features and capability constraints according to claim 4, wherein the step 2.2.2 is based on outlier processing of random standard deviation, and the specific steps are as follows:
step 2.2.2.1: computing wind power dataset YT 2 Arithmetic mean of (2)The calculation formula is shown as (3):
wherein N is YT 2 Wind power number in wind power data set:
step 2.2.2.2: obtainingThereafter, the standard deviation σ is calculated according to equation (4):
step 2.2.2.3: determining a wind power standard interval Ω K The method comprises the following steps:
wherein k=1, 2, N;
step 2.2.2.4:if the wind power value Y (t) n ) 2 Not belonging to standard interval omega K Y (t) n ) 2 Replaced by Y (t) n ) 3 ；
6. An ultra-short term wind power prediction method considering deep features and capability constraints as claimed in claim 5, wherein said step 2.4 comprises:
step 2.4.1: for the wind power data Y (t n ) 4 Performing wavelet decomposition to obtain decomposition signal with M as decomposition scale
Step 2.4.2: the approximate coefficient and the detail coefficient are obtained after wavelet decomposition, the approximate coefficient mainly represents the real signal, and the detail coefficient mainly comprises the noise signal, so that the detail coefficient needs to be processed by adopting a soft threshold function to reduce the noise signal, and the soft threshold formula is as follows:
where w represents the wavelet coefficient, thr represents a given threshold, w thr Representing the wavelet coefficient after threshold processing;
finally, carrying out wavelet reconstruction according to the denoised wavelet coefficient to obtain a denoised signal YT 5 ＝{(Y(t n ) 5 )|n＝1,2,...,N}，YT 5 ∈R N×1 Wherein Y (t) n ) 5 T after denoising for wavelet n Time of day wind power data.
7. The ultra-short term wind power prediction method considering deep features and capability constraints as claimed in claim 6, wherein said step 2.6 comprises:
step 2.6.1: for normalized data set YT 6 Performing variation modal decomposition;
step 2.6.2: feature variables of different modal components are extracted.
8. The ultra-short term wind power prediction method considering deep features and capability constraints as claimed in claim 7, wherein said step 2.6.1 is performed on normalized data set YT 6 And performing variable-fraction modal decomposition, wherein the decomposition steps are as follows:
step 2.6.1.1: converting an intrinsic mode function into an amplitude modulation frequency modulation signal u with limited bandwidth a Taking the minimum sum of bandwidths estimated by the intrinsic mode functions as a constraint condition of decomposition, and calculating a formula as shown in (10);
wherein ,ua Is the a-th modal component signal after VMD decomposition, which is convolution, omega a Is the center frequency of the a-th modal component signal,represent deriving t in bracket formula, t represents time, delta t Is a dirac function;
step 2.6.1.2: converting the formula (10) into an unconstrained problem by introducing a quadratic penalty term and a Lagrangian multiplier beta, wherein the calculation formula is shown as (11);
step 2.6.1.3: using an alternate multiplier direction method to model the function u a Center frequency omega a The Lagrangian multiplier beta is iteratively updated, and the calculation formulas are shown as (12), (13) and (14)；
wherein ,for the wiener filtering result of the residual component, +.>The center frequency of the power spectrum of the modal function is represented by tau, which is an update factor;
step 2.6.1.4: repeating the steps 2.6.1.1, 2.6.1.2 and 2.6.1.3 when the conditions are satisfiedWhen the iteration is stopped, the convergence is considered, wherein f is a threshold value;
will correlate sample set YT 6 Is equal to the number of samples Y (t) n ) 6 Decomposing to obtain H modal components with limited bandwidth, and forming a sample set asRepresents the t n The modal component of the a-th limited bandwidth of the moment, DATA ε R N×H 。
9. The ultra-short term wind power prediction method considering deep level features and capability constraints according to claim 7, wherein the feature variables in step 2.6.2 include: maximum Max, minimum Min, sum, average Mean, range.
10. An ultra-short term wind power prediction method taking into account deep features and capability constraints according to claim 2, wherein said step 2.7 comprises:
step 2.7.1: the number of samples was n=1, 2,.. initializing n=1, creating a decision tree L n ；
Step 2.7.2: for decision tree L n Training, and calculating classification accuracy of n pieces of data outside the bag
Step 2.7.3: for features W in the out-of-bag dataset n1 Disturbance and calculation of classification accuracy
Step 2.7.4: by characteristics W n1 Is related to the magnitude of the correlation I n1 As shown in the formula;
step 2.7.5: through I n1 The importance ranking is carried out on the calculated correlation magnitude;
according to the feature importance ranking, G feature variables with the greatest importance are selected as extracted feature variables to form a sample set DATA3 of the extracted feature variables with strong correlation, DATA3 epsilon R N×G 。
CN202310716534.3A 2023-06-16 2023-06-16 Ultra-short-term wind power prediction method considering deep features and capability constraint Pending CN116706895A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310716534.3A CN116706895A (en) 2023-06-16 2023-06-16 Ultra-short-term wind power prediction method considering deep features and capability constraint

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310716534.3A CN116706895A (en) 2023-06-16 2023-06-16 Ultra-short-term wind power prediction method considering deep features and capability constraint

Publications (1)

Publication Number Publication Date
CN116706895A true CN116706895A (en) 2023-09-05

Family

ID=87823606

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310716534.3A Pending CN116706895A (en) 2023-06-16 2023-06-16 Ultra-short-term wind power prediction method considering deep features and capability constraint

Country Status (1)

Country Link
CN (1) CN116706895A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117833282A (en) * 2024-03-04 2024-04-05 西安热工研究院有限公司 Frequency modulation method and system for fused salt coupling thermal power generating unit
CN117833282B (en) * 2024-03-04 2024-06-11 西安热工研究院有限公司 Frequency modulation method and system for fused salt coupling thermal power generating unit

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117833282A (en) * 2024-03-04 2024-04-05 西安热工研究院有限公司 Frequency modulation method and system for fused salt coupling thermal power generating unit
CN117833282B (en) * 2024-03-04 2024-06-11 西安热工研究院有限公司 Frequency modulation method and system for fused salt coupling thermal power generating unit

Similar Documents

Publication Publication Date Title
CN107578124B (en) Short-term power load prediction method based on multilayer improved GRU neural network
Li et al. Short-term load-forecasting method based on wavelet decomposition with second-order gray neural network model combined with ADF test
CN108256697B (en) Prediction method for short-term load of power system
CN108022014B (en) Power system load prediction method and system
CN112434848B (en) Nonlinear weighted combination wind power prediction method based on deep belief network
Berrezzek et al. Efficient Wind Speed Forecasting Using Discrete Wavelet Transform and Artificial Neural Networks.
CN111193256A (en) Power load prediction method based on variational modal decomposition and gated cyclic unit
Hu et al. Adaptive confidence boundary modeling of wind turbine power curve using SCADA data and its application
CN110826791A (en) Hybrid wind power prediction method based on long-time and short-time memory neural network
CN111275255A (en) Construction method of concrete dam deformation monitoring and forecasting model
CN105260615B (en) A kind of grain consumption Forecasting Methodology
CN112819238A (en) Short-term wind power prediction method based on chaotic chicken flock optimization algorithm
CN115358461A (en) Natural gas load prediction method, device, equipment and medium
CN115511657A (en) Wind power output and photovoltaic output evaluation method based on combined prediction model
CN115879369A (en) Coal mill fault early warning method based on optimized LightGBM algorithm
CN116995668A (en) Middle-long term power load prediction method and system based on improved seasonal ARIMA model
CN113627685B (en) Wind driven generator power prediction method considering wind power internet load limit
CN112016839A (en) Flood disaster prediction and early warning method based on QR-BC-ELM
CN109921462B (en) New energy consumption capability assessment method and system based on LSTM
Liu et al. A novel interval forecasting system for uncertainty modeling based on multi-input multi-output theory: A case study on modern wind stations
CN116706895A (en) Ultra-short-term wind power prediction method considering deep features and capability constraint
CN105205572A (en) Photovoltaic power generation output forecasting method based on EMD (Empirical Mode Decomposition) and Elman algorithm
CN115034422A (en) Wind power short-term power prediction method and system based on fluctuation identification and error correction
Zhu N days average volume based ARIMA forecasting model for Shanghai metro passenger flow
CN116562454A (en) Manufacturing cost prediction method applied to BIM long-short-time attention mechanism 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