CN115794805B - Method for supplementing measurement data of medium-low voltage distribution network - Google Patents

Method for supplementing measurement data of medium-low voltage distribution network Download PDF

Info

Publication number
CN115794805B
CN115794805B CN202310084891.2A CN202310084891A CN115794805B CN 115794805 B CN115794805 B CN 115794805B CN 202310084891 A CN202310084891 A CN 202310084891A CN 115794805 B CN115794805 B CN 115794805B
Authority
CN
China
Prior art keywords
data
measurement
representing
measurement data
sample
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.)
Active
Application number
CN202310084891.2A
Other languages
Chinese (zh)
Other versions
CN115794805A (en
Inventor
黄旭
丁琪
祖国强
徐智
张春晖
郝子源
魏然
李治
张驰
赵长伟
陆杨
范朕宁
张高磊
魏炜
黄盼
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.)
State Grid Corp of China SGCC
State Grid Tianjin Electric Power Co Ltd
Electric Power Research Institute of State Grid Tianjin Electric Power Co Ltd
Chengdong Power Supply Co of State Grid Tianjin Electric Power Co Ltd
Original Assignee
State Grid Corp of China SGCC
State Grid Tianjin Electric Power Co Ltd
Electric Power Research Institute of State Grid Tianjin Electric Power Co Ltd
Chengdong Power Supply Co of State Grid Tianjin Electric Power Co Ltd
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 State Grid Corp of China SGCC, State Grid Tianjin Electric Power Co Ltd, Electric Power Research Institute of State Grid Tianjin Electric Power Co Ltd, Chengdong Power Supply Co of State Grid Tianjin Electric Power Co Ltd filed Critical State Grid Corp of China SGCC
Priority to CN202310084891.2A priority Critical patent/CN115794805B/en
Publication of CN115794805A publication Critical patent/CN115794805A/en
Application granted granted Critical
Publication of CN115794805B publication Critical patent/CN115794805B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a method for supplementing measurement data of a medium-low voltage distribution network, which classifies original quantity measurement by using K-media clusters, and selects main/standby key quantity measurement from various types as an input sample; on the basis, an input vector matrix and a response vector are constructed, the input vector matrix and the response vector are used as the input of an LSTM model, and the LSTM model is trained, so that measurement alignment models oriented to different measurement types and moments are obtained; and according to the deficiency conditions of the primary/standby key quantity measurement at the preamble time and the current time, respectively supplementing and aligning the primary/standby key quantity measurement, and supplementing and aligning other deficiency quantity measurement in various types based on average errors. According to the invention, the scale of the LSTM neural network model is greatly reduced by selecting the main/standby key quantity measurement, the compensation of different types of quantity measurement can be realized, the observability of the power distribution network is improved, and the method is suitable for the power distribution network with higher measurement data loss rate.

Description

Method for supplementing measurement data of medium-low voltage distribution network
Technical Field
The invention belongs to the technical field of distribution network measurement data processing, and particularly relates to a medium-low voltage distribution network measurement data filling method.
Background
Along with the continuous increase of the scale of the power system, the measurement data of the power system shows a rapid growth trend, however, the problems of data deletion may occur in the processes of acquisition, measurement, transmission, storage and the like of massive data, especially the quality and the deletion condition of the measurement data of the power distribution network are worse than those of the power transmission network, the problem that observability of the distribution network is seriously affected and the safety of the power transmission network is threatened is not solved, meanwhile, the number of measurement devices is huge, the calculation resources occupied by complete training are huge, the measurement types are various (current, voltage and power), the change rule of the measurement of different types along with time is different, and how to effectively complement the measurement data of the distribution network is a problem to be solved. The traditional measurement data alignment method is mainly used for predicting single measurement or small-scale measurement, but the calculation accuracy and calculation speed are difficult to meet the requirement of large-scale measurement data alignment.
Therefore, it is necessary to provide a new method for supplementing measurement data of a medium-low voltage distribution network to solve the above-mentioned technical problems.
Disclosure of Invention
The present invention is directed to a method for supplementing measurement data of a medium-low voltage distribution network to solve the above-mentioned problems.
The invention realizes the above purpose through the following technical scheme:
A method for supplementing measurement data of a medium-low voltage distribution network comprises the following steps:
obtaining historical sample measurement data, and processing the sample measurement data to obtain a sample input vector matrix and a sample response vector;
constructing a measurement alignment model, inputting the sample input vector matrix and the sample response vector into the measurement alignment model for training to obtain a trained measurement alignment model, and optimizing the trained measurement alignment model to obtain a final measurement alignment model;
acquiring measurement data at the current moment, and processing the measurement data to obtain a current input vector matrix and a current response vector;
inputting the current input vector matrix and the current response vector into the final measurement alignment model to obtain a measurement predicted value;
performing filling processing on the measurement data at the current moment based on the measurement predicted value to fill in the measurement data of the medium-low voltage distribution network;
obtaining historical sample measurement data, and processing the sample measurement data to obtain a sample input vector matrix and a sample response vector, wherein the specific process comprises the following steps of:
acquiring historical sample measurement data;
Classifying the sample measurement data based on a K-media clustering method, distinguishing main/standby key measurement data and non-key measurement data in various data, and calculating average errors between the non-key measurement data and the main/standby key measurement data in the same type of data;
based on the n+1 consecutive values in the preset sample amount measurement data
Figure SMS_2
Construction of +_for each of the primary and backup key quantity measurement data>
Figure SMS_6
Sample input vector matrix of dimensions
Figure SMS_10
, wherein ,/>
Figure SMS_4
Vector representing the constitution of the sample size measurement data, +.>
Figure SMS_8
Measurement data value representing a first time instant +.>
Figure SMS_12
Representing the measured data value at the second instant,
Figure SMS_14
measurement data value representing a third time instant +.>
Figure SMS_1
Measurement data value representing time n-1, < >>
Figure SMS_5
A measurement data value representing an nth time; />
Figure SMS_9
Representing and vector->
Figure SMS_13
N-dimensional vector formed by time series characteristics of one-to-one correspondence of n measured data values, vector ∈>
Figure SMS_3
Representation and vector->
Figure SMS_7
The n measured data values are corresponding to each other in one-to-one mode and are related to the judgment of workdays, and the vector is +.>
Figure SMS_11
Representing the type of the measurement object;
constructing a sample response vector based on a sample input vector matrix
Figure SMS_15
, wherein
Figure SMS_16
； wherein ,b 1 representing the first data value in the sample response vector, b 2 Representing the second data value in the sample response vector,b 3 representing the third data value in the sample response vector,b n-1 representing the n-1 data value in the sample response vector,b n representing an nth data value in the sample response vector;
classifying the sample measurement data based on a K-media clustering method, distinguishing main/standby key measurement data and non-key measurement data in various types of data, and calculating average errors between the non-key measurement data and the main/standby key measurement data in the same type of data, wherein the specific process is as follows:
setting sample quantity measurement data as X (N.times.M), wherein N is the number of data samples, M is the feature dimension of each data, and the given cluster number is K to obtain K cluster centers;
among the sample amount measurement data, K sample data are randomly selected
Figure SMS_17
As an initial cluster center, wherein ∈>
Figure SMS_18
Representing the first sample data, < >>
Figure SMS_19
Representing the second sample data,/for example>
Figure SMS_20
A third sample of data is represented and is displayed,
Figure SMS_21
represents the kth sample data;
calculating the remaining N-K sample data
Figure SMS_22
Euclidean distance to K cluster centers, wherein, < ->
Figure SMS_23
Representing the first sample data of the remaining sample data, and (2)>
Figure SMS_24
Representing the second sample data of the remaining sample data,/- >
Figure SMS_25
Representing the third sample data of the remaining sample data,/->
Figure SMS_26
Representing the N-K sample data in the rest sample data, dividing the rest sample data under the corresponding class clusters according to the Euclidean distance minimum value to obtain a clustering result, and realizing cluster updating; the Euclidean distance calculation formula is as follows:
Figure SMS_27
（1）；/>
wherein ,
Figure SMS_28
representing sample data->
Figure SMS_29
The first element of (2)>
Figure SMS_30
Representing sample data->
Figure SMS_31
The first element of (a);
traversing all sample points in various clusters, updating the cluster center point by taking the minimum sum of Euclidean distances from all other points in the clusters to the center point as an objective function, wherein the objective function formula is as follows:
Figure SMS_32
（2）；
wherein ,
Figure SMS_33
indicating Euclidean distance from the jth sample point to the 1 st cluster center, +.>
Figure SMS_34
Indicating Euclidean distance from the jth sample point to the 2 nd cluster center, +.>
Figure SMS_35
Representing Euclidean distance from the jth sample point to the kth cluster center;
repeating the processes of cluster updating and cluster center point updating, iterating until all cluster center points and cluster results do not change any more or reach the preset maximum iteration times, and ending the clustering;
setting K clustering center points
Figure SMS_38
Namely, the main critical measurement data are respectively calculated as the average error between the non-critical measurement data and the main critical measurement data in the K classes >
Figure SMS_40
Let the number of non-critical quantity measurement data in each class be +.>
Figure SMS_46
The non-critical measurement data in class i is +.>
Figure SMS_39
, wherein />
Figure SMS_43
Representing the first non-critical measurement data in class i,/I>
Figure SMS_44
Representing second non-critical measurement data in class i,/I>
Figure SMS_47
Representing third non-critical measurement data in class i,/I>
Figure SMS_36
Represents the%>
Figure SMS_41
Measuring data by non-key quantity; />
Figure SMS_45
Mean error vector representing the first class, +.>
Figure SMS_48
Mean error vector representing the second class, +.>
Figure SMS_37
Mean error vector representing class III, -)>
Figure SMS_42
The average error vector of the K-th class is represented, and the average error calculation formula is as follows:
Figure SMS_49
（3）；
wherein ,
Figure SMS_50
representation->
Figure SMS_51
The j-th element in the vector,>
Figure SMS_52
representing major key quantity measurement data->
Figure SMS_53
The kth element of (a)>
Figure SMS_54
Representing non-critical quantity data->
Figure SMS_55
The kth element of (a);
removing the K main key quantity measurement data from various clusters, and respectively searching the clustering center points of the rest data samples in the various clusters again to obtain new K clustering center points
Figure SMS_66
Namely, standby key quantity measurement data, wherein ∈>
Figure SMS_58
Representing the first cluster center of the new K cluster centers,/>
Figure SMS_62
Representing the second cluster center point of the new K cluster centers,/for>
Figure SMS_59
Representing the third cluster center of the new K cluster centers,/ >
Figure SMS_63
Representing the Kth cluster center point in the new K cluster center points, and respectively calculating non-key quantity measurement data in K classesNew mean error between critical quantity data for standby +.>
Figure SMS_64
Let the number of non-critical quantity measurement data in each class be +.>
Figure SMS_68
The non-critical measurement data in class i is
Figure SMS_65
, wherein />
Figure SMS_70
Representing the first non-critical measurement data in class i,/I>
Figure SMS_56
Representing second non-critical measurement data in class i,/I>
Figure SMS_61
Representing third non-critical measurement data in class i,
Figure SMS_67
represents the%>
Figure SMS_71
Measuring data by non-key quantity; />
Figure SMS_69
Representing a new average error vector of the first type,
Figure SMS_72
new average error vector representing the second class, < >>
Figure SMS_57
Representing a third class of new average error vectors, and (2)>
Figure SMS_60
A new average error vector representing class K, the new average error calculation formula is as follows:
Figure SMS_73
（4）；
wherein ,
Figure SMS_74
representation->
Figure SMS_75
The j-th element in the vector,>
Figure SMS_76
representing spare critical quantity measurement data->
Figure SMS_77
The kth element of (a)>
Figure SMS_78
Representing non-critical quantity data->
Figure SMS_79
Is the kth element in (c).
As a further optimization scheme of the invention, the time series characteristics comprise year, quarter, month, day, time and minute; the judgment about the working day specifically comprises a working day 0 representation and a non-working day 1 representation; the measured object types comprise an electric current amount, a voltage amount and a power amount, wherein the electric current amount is represented by 0, the voltage amount is represented by 1, and the power amount is represented by 2.
As a further optimization scheme of the invention, the measurement complement model is an LSTM model, the LSTM model comprises an input gate, a forgetting gate and an output gate, and a double-layer structure is adopted, and the formula is as follows:
Figure SMS_80
（5）；
Figure SMS_81
（6）；
Figure SMS_82
（7）；
Figure SMS_83
（8）；
Figure SMS_84
（9）；
Figure SMS_85
（10）；/>
in the formulas (5) to (10),
Figure SMS_87
indicating the state of the input door at the current moment +.>
Figure SMS_92
Indicating the state of forgetting the door at the current moment, +.>
Figure SMS_95
Indicating the current output door state +.>
Figure SMS_88
Representing the state of the LSTM model at the current moment, +.>
Figure SMS_91
Representing the state of the LSTM model at the previous moment, < + >>
Figure SMS_93
The candidate state of the current moment of the LSTM model is expressed as a pair of the current moment of the LSTM model +.>
Figure SMS_96
and />
Figure SMS_86
For calculating the current cell state +.>
Figure SMS_90
，/>
Figure SMS_94
Input representing the current moment of the LSTM model, +.>
Figure SMS_97
For input to the input gate->
Figure SMS_99
Weight of->
Figure SMS_104
Hidden layer to input gate for previous time>
Figure SMS_108
Weight of->
Figure SMS_111
For input to forget gate->
Figure SMS_109
Weight of->
Figure SMS_112
Hidden layer to forget door for the previous moment>
Figure SMS_113
Weight of->
Figure SMS_114
For input to the output gate->
Figure SMS_89
Weight of->
Figure SMS_102
Hidden layer to output gate for the previous time>
Figure SMS_106
Weight of->
Figure SMS_110
For input of
Figure SMS_98
Weight in feature extraction process, +.>
Figure SMS_101
Implicit layer for the previous moment->
Figure SMS_105
Weight in feature extraction process, +.>
Figure SMS_107
and />
Figure SMS_100
The symbols +. >
Figure SMS_103
Representing the Hadamard product.
As a further optimization scheme of the invention, the sample input vector matrix and the sample response vector are input into the measurement alignment model for training, and the measurement alignment model after training is obtained, which comprises the following specific processes:
vector in sample input vector matrix by Symlet wavelet function
Figure SMS_115
Denoising the data in the sample input vector matrix after denoising and all the data in the sample response vector after denoising are normalized to data with zero mean and unit variance, so as to obtain the normalized sample input vector matrix and the normalized sample response vector;
dividing the normalized sample input vector matrix and the normalized sample response vector into a training set, a verification set and a test set according to the proportion;
and inputting the standardized sample input vector matrix and the standardized sample response vector in the training set into the measurement alignment model for training, inputting the standardized sample input vector matrix and the standardized sample response vector in the verification set into the measurement alignment model, and correcting parameters of the measurement alignment model to obtain the measurement alignment model after training.
As a further optimization scheme of the invention, the parameters of the measurement alignment model comprise weight and bias values.
As a further optimization scheme of the invention, the measurement alignment model is trained by adopting an Adam algorithm, and a weight updating formula is as follows:
Figure SMS_116
（11）；
Figure SMS_117
（12）；/>
Figure SMS_118
（13）；
in the formulae (11) - (13),
Figure SMS_120
and />
Figure SMS_124
For the network weight parameter to be updated in adjacent time steps,/->
Figure SMS_126
For smooth parameters +.>
Figure SMS_121
For learning rate->
Figure SMS_123
and />
Figure SMS_128
Exponential decay rate estimated for first and second moments, respectively,/->
Figure SMS_129
、/>
Figure SMS_119
Deviation correction values for the first and second moment estimates, respectively; />
Figure SMS_125
Representing a first moment estimation of the gradient when the time step is t-1;
Figure SMS_127
representing the gradient at time step t; />
Figure SMS_130
Representing a second moment estimate of the gradient at time step t-1; />
Figure SMS_122
Representing the square of the gradient at time step t;
setting a root mean square error function (RMSE) as a loss function trained by the measurement and alignment model, wherein the formula is as follows:
Figure SMS_131
（14）；
in the formula ,
Figure SMS_132
for the total number of measurement data to be predicted, +.>
Figure SMS_133
For measuring the true value of the data to be predicted,
Figure SMS_134
predicted values of measurement data output by the measurement patch model.
As a further optimization scheme of the invention, the post-training measurement and alignment model is optimized to obtain a final measurement and alignment model, and the specific process is as follows:
And optimizing the super parameters of the trained measurement alignment model by adopting a Bayesian optimization method to obtain optimized network parameters, reconstructing the trained measurement alignment model based on the optimized network parameters, and obtaining the final measurement alignment model.
As a further optimization scheme of the invention, the super parameters of the post-training measurement and alignment model comprise iteration times, hidden layer numbers, neuron numbers of each layer and learning rate.
As a further optimization scheme of the invention, the super-parameters of the measurement and complement model after training are optimized by adopting a Bayesian optimization method, so as to obtain optimized network parameters, and the specific process is as follows:
setting objective functions of a Bayesian framework
Figure SMS_135
Independent variable->
Figure SMS_136
Representing the super-parameters;
selection of
Figure SMS_137
Calculating objective function of each observation point>
Figure SMS_138
The value at the observation point, namely the observation value of a preset observation model;
setting up
Figure SMS_139
Based on the observation +.>
Figure SMS_140
Estimating to obtain the objective function->
Figure SMS_141
Functional distribution of->
Figure SMS_142
A minimum value of the target value;
setting current observation data
Figure SMS_143
A base (B)At the current observation dataDCalculating a preset acquisition function and determining the next observation point +.>
Figure SMS_144
Calculate the +.>
Figure SMS_145
Acquisition function value->
Figure SMS_146
Setting- >
Figure SMS_147
Updating a preset probability agent model;
repeating the above steps until the target value reaches the preset maximum observation times P to obtain the optimized network parameters
Figure SMS_148
。
As a further optimization scheme of the invention, the probability agent model is a Gaussian process regression model, and the Gaussian process regression model obeys k-dimensional normal distribution and has the following formula:
Figure SMS_149
（15）；
wherein ,
Figure SMS_150
representing an n-dimensional vector, ">
Figure SMS_151
As a mean function>
Figure SMS_152
As a covariance function.
As a further optimization of the invention, the expected improvement function is adopted as the acquisition function and the next observation point is determined
Figure SMS_153
The formula is as follows:
Figure SMS_154
（16）；
Figure SMS_155
（17）；
in the formulae (16) to (17),
Figure SMS_156
is the position observed in step i, +.>
Figure SMS_157
Is the posterior mean of the agent at time s+1; />
Figure SMS_158
Representing the observation position when the current objective function is maximized; />
Figure SMS_159
Maximum value of the current objective function; the argmax (f (u)) function is an argument u that maximizes the value of f (u); the max (f (u)) function is the maximum value of f (u); d represents the current observation data set; e (f (u)) functions are expected for f (u).
As a further optimization scheme of the invention, measuring data at the current moment is obtained, the measuring data is processed to obtain a current input vector matrix and a current response vector, and the specific process is as follows:
Acquiring measurement data at the current moment;
classifying the measured data at the current moment based on a K-media clustering method, distinguishing main/standby key measured data at the current moment and non-key measured data at the current moment in various data, and calculating average errors between the non-key measured data at the current moment and the main/standby key measured data at the current moment in the same type of data;
judging whether the current main key measurement data is missing, if not, supplementing the missing current non-key measurement data based on the average error between the current non-key measurement data and the current main key measurement data; if yes, comparing the main key measurement data at the current moment with the missing degree of the data of the front period of the standby key measurement data at the current moment, and taking the main/standby key measurement data at the current moment with lower missing degree as the measurement data to be predicted;
selecting data of the first m times of the waiting predicting time of the waiting predicting quantity data to construct an m-dimensional vector
Figure SMS_160
And constructs +.>
Figure SMS_161
Current input vector matrix of dimension->
Figure SMS_162
, wherein ,c 1 the first vector representing the first m moments, c 2 The second vector representing the first m moments,c 3 the third vector representing the first m moments,c m-1 the m-1 st vector representing the first m moments,c m an mth vector representing the first m times;y 2 、y 3 、y 4 、y 5 、y 6 、y 7 respectively represent and vectory 1 M-dimensional vector formed by time sequence features corresponding to m measured data values one by oney 8 Representation and vectory 1 The m measurement data values are in one-to-one correspondence with the judgment of working days and the vectory 9 Representing the type of the measurement object;
construction of the current response vector
Figure SMS_163
, wherein ,
Figure SMS_164
，/>
Figure SMS_165
for the measurement value at time t, set +.>
Figure SMS_166
。/>
As a further optimization scheme of the present invention, the current input vector matrix and the current response vector are input into the final measurement alignment model to obtain a measurement predicted value, and the specific process is as follows:
the current input vector matrix and the current response vector are used as input parameters to be input into a final measurement and alignment model, and the output quantity measurement predicted value is
Figure SMS_167
, wherein ,/>
Figure SMS_168
Measurement prediction value representing the first moment, < +.>
Figure SMS_169
Measurement prediction value representing the second moment, < + >>
Figure SMS_170
Indicating the measurement prediction value at the third time,
Figure SMS_171
measurement prediction value indicating m-1 th time,/->
Figure SMS_172
The measurement predicted value at the time t is obtained.
As a further optimization scheme of the invention, the method for carrying out the filling processing on the measurement data at the current moment based on the measurement predicted value comprises the following specific processes:
And based on the measurement predicted value, supplementing the missing measurement data according to the average error between the non-key measurement data and the main/standby key measurement data in the same class until all measurement data in all classes are supplemented.
The invention has the beneficial effects that:
the invention effectively utilizes the relevance of each measured data, thereby reducing the scale of the problem to be solved, improving the efficiency of large-scale measured data filling, adding time sequence characteristics, data types and other influencing factors in the LSTM model, and improving the accuracy of data prediction.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a flow chart of a model training phase of the present invention;
FIG. 3 is a flow chart of the measurement patch phase of the present invention;
FIG. 4 is an overall block diagram of an LSTM model employed in the present invention;
FIG. 5 is an internal structure diagram of the LSTM model employed in the present invention;
FIG. 6 is a flowchart of a K-media clustering algorithm employed by the present invention.
Detailed Description
The following detailed description of the present application is provided in conjunction with the accompanying drawings, and it is to be understood that the following detailed description is merely illustrative of the application and is not to be construed as limiting the scope of the application, since numerous insubstantial modifications and adaptations of the application will be to those skilled in the art in light of the foregoing disclosure.
As shown in fig. 1, a method for supplementing measurement data of a medium-low voltage distribution network includes the following steps:
s1: obtaining historical sample measurement data, and processing the sample measurement data to obtain a sample input vector matrix and a sample response vector;
s2: constructing a measurement alignment model, inputting the sample input vector matrix and the sample response vector into the measurement alignment model for training to obtain a trained measurement alignment model, and optimizing the trained measurement alignment model to obtain a final measurement alignment model;
s3: acquiring measurement data at the current moment, and processing the measurement data to obtain a current input vector matrix and a current response vector;
s4: inputting the current input vector matrix and the current response vector into the final measurement alignment model to obtain a measurement predicted value;
s5: and carrying out filling processing on the measurement data at the current moment based on the measurement predicted value so as to fill in the measurement data of the medium-low voltage distribution network.
In this embodiment, the method specifically includes the following steps:
as shown in fig. 2, model training phase:
step 1: acquiring historical sample measurement data, namely acquiring existing measurement data, classifying the existing measurement data by adopting a K-media clustering method, selecting main/standby key measurement from various types, and recording average errors between other measurement and main/standby key measurement in the same type, wherein the other measurement data is non-key measurement data;
Step 2: for each of the primary and backup key measurements, based on n+1 values in succession in the measurement
Figure SMS_174
) Structure->
Figure SMS_180
Dimensional input vector matrix
Figure SMS_181
, wherein
Figure SMS_175
For a vector of measurement data, each element represents a measurement data value at a moment,/-for the measurement data>
Figure SMS_176
Respectively is->
Figure SMS_178
N-dimensional vector formed by time series characteristics (year, quarter, month, day, time and minute) of one-to-one correspondence of n measured data values in vector, ++>
Figure SMS_179
Vector representation and +.>
Figure SMS_173
The n measured data values in the vector are in one-to-one correspondence with the judgment about the working day (working day 0 representation and non-working day 1 representation),>
Figure SMS_177
vector represents the type of object measured (current amount is represented by 0, voltage amount is represented by 1, power amount is represented by 2);
step 3: constructing response vectors
Figure SMS_182
, wherein />
Figure SMS_183
，
Figure SMS_184
I.e. predicting the value of the sequence at a future time step, assigning a response vector as a training sequence with the value shifted by one time step, at each time step of the input sequence, the LSTM network learns the value of the predicted next time step;
step 4: establishing a measurement complement model, and adopting an LSTM model structure, wherein the structure mainly comprises an input door (controlling how much information of a candidate state at the current moment needs to be saved), a forgetting door (controlling how much information of an internal state at the previous moment needs to be forgotten) and an output door (controlling how much information of the internal state at the current moment needs to be output to an external state); setting an LSTM model training mode, adopting an Adam optimization algorithm, wherein a loss function is a root mean square error function (RMSE), the learning rate is 0.005, input parameters are a multidimensional input vector matrix constructed in the step 2 after pretreatment and a response vector constructed in the step 3, training model parameters, and each main key quantity measurement and each standby key quantity measurement are involved in training to obtain measurement alignment models for different measurement types and moments;
Step 5: and (3) optimizing the super parameters (iteration times, hidden layer numbers, neuron numbers of each layer, learning rate and the like) of the established LSTM model by using a Bayesian optimization (Bayesian optimization) method to obtain optimized network parameters, and reconstructing the LSTM model for prediction by using the optimized parameters.
As shown in fig. 3, the measurement replenishment phase:
step 1: selecting quantity measurement to be predicted according to the missing condition of main key quantity measurement data at the current moment (t moment), if the quantity measurement is not missing, based on the main key quantity measurement, supplementing the quantity measurement with other missing quantity measurement according to the average error between the other quantity measurement and the main key quantity measurement in the class, jumping to the step 6, and if the quantity measurement is missing, comparing the quality (missing degree) of the data in the preamble period of the main key quantity measurement and the spare key quantity measurement, wherein the quality is good as the quantity measurement to be predicted;
step 2: based on the selected measurement to be predicted, the data of the first m times of the measurement to be predicted time (t time) are selected to construct an m-dimensional vector
Figure SMS_185
And based on the m data constructs
Figure SMS_186
Dimension input vector matrix +.>
Figure SMS_187
The construction method is the same as the model training stage step 2;
step 3: constructing response vectors
Figure SMS_188
, wherein />
Figure SMS_189
，
Figure SMS_190
，/>
Figure SMS_191
Measurement of t time (absence) is given by ∈K>
Figure SMS_192
；
Step 4: preprocessing an input vector matrix and a response vector, taking the preprocessed input vector matrix and the preprocessed response vector as input parameters, predicting based on a trained measurement alignment model for the measurement type, and outputting the output result as
Figure SMS_193
Inverse normalized +.>
Figure SMS_194
The measurement predicted value at the time t is obtained;
step 5: based on the main/standby key quantity measurement predicted value, compensating the measurement of other missing quantity according to the average error between the other quantity measurement in the same class and the main/standby key quantity measurement;
step 6: the above steps are repeated for each class of measurement until all of the measurement data in all classes are filled.
In this embodiment, as shown in fig. 6, the specific steps of clustering by using the K-media clustering method include:
step 1: setting an input data sample as X (N.times.M), wherein N is the number of data samples, M is the feature dimension of each data, and the given cluster number is K, namely gathering N data into K types;
step 2: among the original data samples, K samples are randomly selected
Figure SMS_195
As an initial cluster center;
step 3: cluster updating: calculating the remaining N-K samples
Figure SMS_196
And dividing the rest data into corresponding class clusters according to the Euclidean distance between the rest data and the K clustering centers to obtain a clustering result. The Euclidean distance calculation formula is as follows:
Figure SMS_197
（1）；
wherein ,
Figure SMS_198
representing sample data->
Figure SMS_199
The first element of (2)>
Figure SMS_200
Representing sample data->
Figure SMS_201
The first element of (a);
Figure SMS_202
and />
Figure SMS_203
All are M-dimensional sample data vectors, and each element in the vector represents measurement data at a certain moment. />
Step 4: cluster center point update: traversing all sample points in various clusters, updating the cluster center point by taking the minimum sum of Euclidean distances from all other points (T-1) in the clusters to the center point as an objective function, wherein the formula of the objective function is as follows:
Figure SMS_204
（2）；
providing T sample points in a certain class of clusters, respectively taking each sample point as a cluster center point, calculating the sum of Euclidean distances from all other points in the clusters to the center point, wherein,
Figure SMS_205
representing the first sample point as the cluster center point, the sum of the Euclidean distances of all other points in the cluster to the center point, and so on, +.>
Figure SMS_206
Representing the euclidean distance from the jth sample point to the 1 st sample point (center point),/>
Figure SMS_207
indicating Euclidean distance from the jth sample point to the 2 nd cluster center, +.>
Figure SMS_208
Representing Euclidean distance from the jth sample point to the T clustering center; and so on, wherein
Figure SMS_209
Are all 0;
step 5: repeating the processes of cluster updating and cluster center point updating, iterating until all cluster center points and cluster results do not change any more or reach the maximum iteration times designated in advance, and ending the clustering;
Step 6: the obtained K clustering center points
Figure SMS_210
Namely, the main key quantity measurement is respectively calculated as the measurement of other quantities in K classes +.>
Figure SMS_211
Average error from primary key measurement
Figure SMS_212
And recording, wherein the average error calculation formula is as follows:
Figure SMS_213
（3）；
wherein ,
Figure SMS_215
representation->
Figure SMS_217
The j-th element in the vector,>
Figure SMS_219
representing major key quantity measurement data->
Figure SMS_216
The kth element of (a)>
Figure SMS_218
Representing non-critical quantity data->
Figure SMS_220
The kth element of (a); />
Figure SMS_221
and />
Figure SMS_214
The data are M-dimensional sample data vectors, and each element in the vectors represents measurement data at a certain moment;
step 7: removing the selected K main key quantity measurements from various clusters, and searching cluster center points for the rest data samples in the various clusters;
step 8: the obtained new K clustering center points
Figure SMS_222
Namely, the standby key quantity measurement is respectively calculated as the other quantity measurement in the K class>
Figure SMS_223
Average error between the measurement of the spare key quantity +.>
Figure SMS_224
And recording, wherein the average error calculation formula is as follows:
Figure SMS_225
（4）；
wherein ,
Figure SMS_228
representation->
Figure SMS_230
The j-th element in the vector,>
Figure SMS_232
representing spare critical quantity measurement data->
Figure SMS_227
The kth element of (a)>
Figure SMS_229
Representing non-critical quantity data->
Figure SMS_231
The kth element of (a); />
Figure SMS_233
and />
Figure SMS_226
All are M-dimensional sample data vectors, and each element in the vector represents measurement data at a certain moment.
In this embodiment, a measurement alignment model is constructed, the sample input vector matrix and the sample response vector are input into the measurement alignment model to train, a trained measurement alignment model is obtained, the trained measurement alignment model is optimized, and a final measurement alignment model is obtained, and the method specifically includes the following steps:
as shown in fig. 5, the measurement and alignment model adopts an LSTM model, the LSTM model adopts a double-layer structure, the number of neurons in an hidden layer is 96×3, and a single LSTM structure is composed of an input gate, a forgetting gate and an output gate, and the formula is as follows:
Figure SMS_234
（5）；
Figure SMS_235
（6）；
Figure SMS_236
（7）；
Figure SMS_237
（8）；
Figure SMS_238
（9）；
Figure SMS_239
（10）；
in the formula ,
Figure SMS_255
indicating the state of the input door at the current moment +.>
Figure SMS_258
Indicating the state of forgetting the door at the current moment, +.>
Figure SMS_262
Indicating the current output door state +.>
Figure SMS_242
Representing the state of the LSTM at the current moment, +.>
Figure SMS_244
Indicating the state of the LSTM immediately preceding, +.>
Figure SMS_248
For the candidate state of the LSTM current time, the LSTM current time pair is expressed>
Figure SMS_252
and />
Figure SMS_257
For calculating the current cell state +.>
Figure SMS_259
，/>
Figure SMS_264
Input representing the current time of LSTM, +.>
Figure SMS_266
For input to the input gate->
Figure SMS_263
Weight of->
Figure SMS_265
Hidden layer to input gate for previous time>
Figure SMS_267
Weight of->
Figure SMS_268
For input to forget gate->
Figure SMS_243
Weight of->
Figure SMS_247
Hidden layer to forget door for the previous moment>
Figure SMS_249
Weight of- >
Figure SMS_254
For input to the output gate->
Figure SMS_240
Weight of->
Figure SMS_245
Hidden layer to output gate for the previous time>
Figure SMS_251
Weight of->
Figure SMS_253
For input +.>
Figure SMS_241
Weight in feature extraction process, +.>
Figure SMS_246
Implicit layer for the previous moment->
Figure SMS_250
Weight in feature extraction process, +.>
Figure SMS_261
and />
Figure SMS_256
The symbols +.>
Figure SMS_260
Representing the Hadamard product.
Preprocessing an input vector matrix constructed in a model training stage step 2 and a response vector constructed in a model training stage step 3, and firstly selecting a Symlet wavelet function to construct a vector for measuring data in the input vector matrix
Figure SMS_269
Denoising the data in the input vector matrix and the response vector, and normalizing all the data in the input vector matrix and the response vector into data with zero mean and unit variance;
the input vector matrix and the response vector are divided into a training set, a verification set and a test set according to the proportion of 8:1:1, wherein the training set and the verification set are used for training a model and determining parameters (a weight value W, U and a bias value b), and the test set is used for checking the generalization capability of the model.
And taking the input vector matrix and the response vector in the training set as the LSTM input, finally obtaining output through the full connection layer, and taking the input vector matrix and the response vector in the verification set as the LSTM input to correct parameters in the LSTM model to obtain a trained LSTM model, namely a measurement and alignment model after training.
Training the LSTM layer model by adopting an Adam algorithm, wherein the weight updating formula is as follows:
Figure SMS_270
（11）；
Figure SMS_271
（12）；
Figure SMS_272
（13）；
in the formula ,
Figure SMS_275
and />
Figure SMS_278
For the network weight parameter to be updated in adjacent time steps,/->
Figure SMS_280
In order to smooth the parameters of the image,
Figure SMS_273
for learning rate->
Figure SMS_276
and />
Figure SMS_279
Exponential decay rate estimated for first and second moments, respectively,/->
Figure SMS_282
Deviation correction values for the first and second moment estimates, respectively; />
Figure SMS_274
Representing a first moment estimation of the gradient when the time step is t-1; />
Figure SMS_277
Representing the gradient at time step t; />
Figure SMS_281
Representing a second moment estimate of the gradient at time step t-1; />
Figure SMS_283
Representing the square of the gradient at time step t;
a root mean square error function (root mean square erro, RMSE) is defined as a model trained loss function, whose formula is:
Figure SMS_284
（14）；
in the formula ,
Figure SMS_285
for the total number of measurement data to be predicted, +.>
Figure SMS_286
For measuring the true value of the data to be predicted,
Figure SMS_287
predicted values of measurement data output for the LSTM model.
In this embodiment, the bayesian optimization (Bayesian optimization) method involved in the model training stage step 5 is as follows:
(1) Bayesian optimization framework
Step 1: independent variable
Figure SMS_288
For hyper-parametric space, model-trained loss function (RMSE) is used as the objective function of bayesian framework +.>
Figure SMS_289
；
Step 2: selection of
Figure SMS_290
Calculating the ∈10 in each observation point>
Figure SMS_291
The values at these points are then used to determine the observations of the observation model;
step 3: order the
Figure SMS_292
；
Step 4: estimating the function from finite observations, this assumption being called a priori assumption in Bayesian optimization, by which the estimation is made
Figure SMS_293
(function distribution) minimum value of target value on (function distribution);
step 5: based on current observation data
Figure SMS_294
Calculating an acquisition function and determining the next observation point +.>
Figure SMS_295
The function value at the next observation point is calculated: />
Figure SMS_296
And let->
Figure SMS_297
Updating the probability agent model;
step 6: repeating the steps 4 and 5 until the target value on the assumed distribution reaches a preset standard or reaches a preset maximum observation frequency P;
step 7: output of
Figure SMS_298
Corresponding y, < >>
Figure SMS_299
And the optimized super parameter is obtained.
(2) The tool (probability agent model) for estimating the function distribution used in the bayesian optimization framework step 4 is as follows:
the probability proxy model comprises a priori probability model
Figure SMS_300
And observation model->
Figure SMS_301
Updating the probability agent model, i.e. obtaining the posterior probability distribution comprising more data information according to the formula +.>
Figure SMS_302
The calculation formula is as follows:
Figure SMS_303
（15）；
in the formula ,
Figure SMS_304
for the objective function, D represents the observed set, +.>
Figure SMS_305
For likelihood distribution of observations +. >
Figure SMS_306
Representation->
Figure SMS_307
Is>
Figure SMS_308
Representation->
Figure SMS_309
Posterior probability distribution of (c).
Gaussian process regression is adopted as a probability agent model, and the distribution of an objective function is estimated according to a few observation points
Figure SMS_310
(include->
Figure SMS_311
The value of each point and the confidence level corresponding to the point), wherein the Gaussian process regression obeys the k-dimensional normal distribution:
Figure SMS_312
（16）；
wherein ,
Figure SMS_313
representing an n-dimensional vector, ">
Figure SMS_314
As a mean function>
Figure SMS_315
As a covariance function.
(3) The acquisition function in the bayesian optimization framework step 5 is as follows:
acquisition function measurement observation point pair fitting
Figure SMS_316
The generated influence is selected and the point with the largest influence is selected to execute the next observation, the expected lifting EI (Expected Improvement) is adopted as the acquisition function and the next observation point +.>
Figure SMS_317
I.e. if a certain point can be the current maximum +.>
Figure SMS_318
If the maximum expected lifting is brought, the point is selected as the next observation point, and the calculation formula is as follows:
Figure SMS_319
（17）；
Figure SMS_320
（18）；
in the formulae (16) to (17),
Figure SMS_321
is the position observed in step i, +.>
Figure SMS_322
Is the posterior mean of the agent at time s+1; />
Figure SMS_323
Representing the observation position when the current objective function is maximized; />
Figure SMS_324
Maximum value of the current objective function; the argmax (f (u)) function is an argument u that maximizes the value of f (u); the max (f (u)) function is the maximum value of f (u); d represents the current observation data set; e (f (u)) functions are expected for f (u).
The foregoing examples illustrate only a few embodiments of the invention and are described in detail herein without thereby limiting the scope of the invention. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the invention, which are all within the scope of the invention.

Claims (14)

1. The method for supplementing measurement data of the medium-low voltage distribution network is characterized by comprising the following steps of:
obtaining historical sample measurement data, and processing the sample measurement data to obtain a sample input vector matrix and a sample response vector;
constructing a measurement alignment model, inputting the sample input vector matrix and the sample response vector into the measurement alignment model for training to obtain a trained measurement alignment model, and optimizing the trained measurement alignment model to obtain a final measurement alignment model;
acquiring measurement data at the current moment, and processing the measurement data to obtain a current input vector matrix and a current response vector;
inputting the current input vector matrix and the current response vector into the final measurement alignment model to obtain a measurement predicted value;
Performing filling processing on the measurement data at the current moment based on the measurement predicted value to fill in the measurement data of the medium-low voltage distribution network;
obtaining historical sample measurement data, and processing the sample measurement data to obtain a sample input vector matrix and a sample response vector, wherein the specific process comprises the following steps of:
acquiring historical sample measurement data;
classifying the sample measurement data based on a K-media clustering method, distinguishing main/standby key measurement data and non-key measurement data in various data, and calculating average errors between the non-key measurement data and the main/standby key measurement data in the same type of data;
based on the n+1 consecutive values in the preset sample amount measurement data
Figure QLYQS_1
Construction of +_for each of the primary and backup key quantity measurement data>
Figure QLYQS_8
Sample input vector matrix of dimensions
Figure QLYQS_11
, wherein ,/>
Figure QLYQS_3
Vector representing the constitution of the sample size measurement data, +.>
Figure QLYQS_7
Measurement data value representing a first time instant +.>
Figure QLYQS_10
Representing the measured data value at the second instant,
Figure QLYQS_14
measurement data value representing a third time instant +.>
Figure QLYQS_2
Measurement data value representing time n-1, < >>
Figure QLYQS_6
A measurement data value representing an nth time; / >
Figure QLYQS_12
Representing and vector->
Figure QLYQS_13
N-dimensional vector formed by time series characteristics of one-to-one correspondence of n measured data values, vector ∈>
Figure QLYQS_4
Representation and vector->
Figure QLYQS_5
The n measured data values are corresponding to each other in one-to-one mode and are related to the judgment of workdays, and the vector is +.>
Figure QLYQS_9
Representing the type of the measurement object;
constructing a sample response vector based on a sample input vector matrix
Figure QLYQS_15
, wherein
Figure QLYQS_16
； wherein ,b 1 representing the first data value in the sample response vector,b 2 representing the second data value in the sample response vector,b 3 representing the third data value in the sample response vector,b n-1 representing the n-1 data value in the sample response vector,b n representing an nth data value in the sample response vector;
classifying the sample measurement data based on a K-media clustering method, distinguishing main/standby key measurement data and non-key measurement data in various types of data, and calculating average errors between the non-key measurement data and the main/standby key measurement data in the same type of data, wherein the specific process is as follows:
setting sample quantity measurement data as X (N.times.M), wherein N is the number of data samples, M is the feature dimension of each data, and the given cluster number is K to obtain K cluster centers;
among the sample amount measurement data, K sample data are randomly selected
Figure QLYQS_17
As an initial cluster center, wherein ∈>
Figure QLYQS_18
Representing the first sample data, < >>
Figure QLYQS_19
Representing the second sample data,/for example>
Figure QLYQS_20
Representing the third sample data, ++>
Figure QLYQS_21
Represents the kth sample data;
calculating the remaining N-K sample data
Figure QLYQS_22
Euclidean distance to K cluster centers, wherein, < ->
Figure QLYQS_23
Representing the first sample data of the remaining sample data, and (2)>
Figure QLYQS_24
Representing the second sample data of the remaining sample data,/->
Figure QLYQS_25
Representing the third sample data of the remaining sample data,/->
Figure QLYQS_26
Representing the N-K sample data in the rest sample data, dividing the rest sample data under the corresponding class clusters according to the Euclidean distance minimum value to obtain a clustering result, and realizing cluster updating; the Euclidean distance calculation formula is as follows:
Figure QLYQS_27
（1）；
wherein ,
Figure QLYQS_28
representing sample data->
Figure QLYQS_29
The first element of (2)>
Figure QLYQS_30
Representing sample data->
Figure QLYQS_31
The first element of (a);
traversing all sample points in various clusters, updating the cluster center point by taking the minimum sum of Euclidean distances from all other points in the clusters to the center point as an objective function, wherein the objective function formula is as follows:
Figure QLYQS_32
（2）；
wherein ,
Figure QLYQS_33
indicating Euclidean distance from the jth sample point to the 1 st cluster center, +.>
Figure QLYQS_34
Indicating Euclidean distance from the jth sample point to the 2 nd cluster center, +. >
Figure QLYQS_35
Represents the jth sample pointEuclidean distance to the kth cluster center;
repeating the processes of cluster updating and cluster center point updating, iterating until all cluster center points and cluster results do not change any more or reach the preset maximum iteration times, and ending the clustering;
setting K clustering center points
Figure QLYQS_37
Namely, the main critical measurement data are respectively calculated as the average error between the non-critical measurement data and the main critical measurement data in the K classes>
Figure QLYQS_42
Let the number of non-critical quantity measurement data in each class be +.>
Figure QLYQS_46
The non-critical measurement data in class i is +.>
Figure QLYQS_38
, wherein />
Figure QLYQS_43
Representing the first non-critical measurement data in class i,/I>
Figure QLYQS_45
Representing second non-critical measurement data in class i,/I>
Figure QLYQS_48
Representing third non-critical measurement data in class i,/I>
Figure QLYQS_36
Represents the%>
Figure QLYQS_40
Measuring data by non-key quantity; />
Figure QLYQS_44
Mean error vector representing the first class, +.>
Figure QLYQS_47
Mean error vector representing the second class, +.>
Figure QLYQS_39
Mean error vector representing class III, -)>
Figure QLYQS_41
The average error vector of the K-th class is represented, and the average error calculation formula is as follows:
Figure QLYQS_49
（3）；
wherein ,
Figure QLYQS_50
representation->
Figure QLYQS_51
The j-th element in the vector,>
Figure QLYQS_52
representing major key quantity measurement data->
Figure QLYQS_53
The (c) is a group of elements,
Figure QLYQS_54
representing non-critical quantity data- >
Figure QLYQS_55
The kth element of (a);
removing the K main key quantity measurement data from various clusters, and respectively searching the clustering center points of the rest data samples in the various clusters again to obtain new K clustering center points
Figure QLYQS_71
Namely, standby key quantity measurement data, wherein ∈>
Figure QLYQS_58
Representing the first cluster center of the new K cluster centers,/>
Figure QLYQS_64
Representing the second cluster center point of the new K cluster centers,/for>
Figure QLYQS_59
Representing the third cluster center of the new K cluster centers,/>
Figure QLYQS_60
Representing the Kth cluster center point in the K new cluster centers, and respectively calculating new average error between the non-key measured data and the standby key measured data in the K classes +.>
Figure QLYQS_63
The number of non-critical quantity measurement data in each class is respectively set as
Figure QLYQS_67
The non-critical measurement data in class i is +.>
Figure QLYQS_66
, wherein />
Figure QLYQS_70
Representing the first non-critical measurement data in class i,/I>
Figure QLYQS_56
Representing second non-critical measurement data in class i,/I>
Figure QLYQS_61
Representing third non-critical measurement data in class i,/I>
Figure QLYQS_65
Represents the%>
Figure QLYQS_68
Measuring data by non-key quantity; />
Figure QLYQS_69
New mean error vector representing the first class, < ->
Figure QLYQS_72
Representing a new average error vector of the second class,
Figure QLYQS_57
representing a third class of new average error vectors, and (2) >
Figure QLYQS_62
A new average error vector representing class K, the new average error calculation formula is as follows:
Figure QLYQS_73
（4）；/>
wherein ,
Figure QLYQS_74
representation->
Figure QLYQS_75
The j-th element in the vector,>
Figure QLYQS_76
representing spare critical quantity measurement data->
Figure QLYQS_77
The kth element of (a)>
Figure QLYQS_78
Representing non-critical quantity data->
Figure QLYQS_79
Is the kth element in (c).
2. The method for supplementing measurement data to a medium-low voltage distribution network according to claim 1, wherein the method comprises the steps of: the time sequence features comprise year, quarter, month, day, time and minute; the judgment about the working day specifically comprises a working day 0 representation and a non-working day 1 representation; the measured object types comprise an electric current amount, a voltage amount and a power amount, wherein the electric current amount is represented by 0, the voltage amount is represented by 1, and the power amount is represented by 2.
3. The method for supplementing measurement data to a medium-low voltage distribution network according to claim 1, wherein the method comprises the steps of: the measuring and complementing model is an LSTM model, the LSTM model comprises an input door, a forgetting door and an output door, a double-layer structure is adopted, and the formula is as follows:
Figure QLYQS_80
（5）；
Figure QLYQS_81
（6）；
Figure QLYQS_82
（7）；
Figure QLYQS_83
（8）；
Figure QLYQS_84
（9）；
Figure QLYQS_85
（10）；
in the formulas (5) to (10),
Figure QLYQS_101
indicating the state of the input door at the current moment +.>
Figure QLYQS_107
Indicating that the door state is forgotten at the current moment,
Figure QLYQS_109
indicating the current output door state +.>
Figure QLYQS_87
Representing the state of the LSTM model at the current moment, +. >
Figure QLYQS_90
Representing the state of the LSTM model at the previous moment, < + >>
Figure QLYQS_94
The candidate state of the current moment of the LSTM model is expressed as a pair of the current moment of the LSTM model +.>
Figure QLYQS_98
and />
Figure QLYQS_89
For calculating the current cell state +.>
Figure QLYQS_93
，/>
Figure QLYQS_95
Input representing the current moment of the LSTM model, +.>
Figure QLYQS_99
For input to the input gate->
Figure QLYQS_102
Weight of->
Figure QLYQS_110
Hidden layer to input gate for previous time>
Figure QLYQS_112
Weight of->
Figure QLYQS_114
For input to forget gate->
Figure QLYQS_105
Weight of->
Figure QLYQS_108
Hidden layer to forget door for the previous moment>
Figure QLYQS_111
Weight of->
Figure QLYQS_113
For input to the output gate->
Figure QLYQS_86
Weight of->
Figure QLYQS_92
Hidden layer to output gate for the previous time>
Figure QLYQS_97
Weight of->
Figure QLYQS_103
For input of
Figure QLYQS_88
Weight in feature extraction process, +.>
Figure QLYQS_91
Implicit layer for the previous moment->
Figure QLYQS_96
Weight in feature extraction process, +.>
Figure QLYQS_100
and />
Figure QLYQS_104
The symbols +.>
Figure QLYQS_106
Representing the Hadamard product.
4. The method for supplementing measurement data to a medium-low voltage distribution network according to claim 1, wherein the method comprises the steps of: inputting the sample input vector matrix and the sample response vector into the measurement alignment model for training to obtain a measurement alignment model after training, wherein the specific process is as follows:
vector in sample input vector matrix by Symlet wavelet function
Figure QLYQS_115
Denoising the data in the sample input vector matrix after denoising and all the data in the sample response vector after denoising are normalized to data with zero mean and unit variance, so as to obtain the normalized sample input vector matrix and the normalized sample response vector;
dividing the normalized sample input vector matrix and the normalized sample response vector into a training set, a verification set and a test set according to the proportion;
and inputting the standardized sample input vector matrix and the standardized sample response vector in the training set into the measurement alignment model for training, inputting the standardized sample input vector matrix and the standardized sample response vector in the verification set into the measurement alignment model, and correcting parameters of the measurement alignment model to obtain the measurement alignment model after training.
5. The method for supplementing measurement data to a medium-low voltage distribution network according to claim 4, wherein the method comprises the steps of: the parameters of the measurement complement model comprise weights and bias values.
6. The method for measuring and data supplementing of a medium-low voltage distribution network according to claim 4, wherein the measuring and supplementing model is trained by Adam algorithm, and the weight updating formula is as follows:
Figure QLYQS_116
（11）；
Figure QLYQS_117
（12）；
Figure QLYQS_118
（13）；
In the formulae (11) - (13),
Figure QLYQS_120
and />
Figure QLYQS_125
For the network weight parameter to be updated in adjacent time steps,/->
Figure QLYQS_128
For smooth parameters +.>
Figure QLYQS_122
For learning rate->
Figure QLYQS_124
and />
Figure QLYQS_127
Exponential decay rate estimated for first and second moments, respectively,/->
Figure QLYQS_130
、/>
Figure QLYQS_119
Deviation correction values for the first and second moment estimates, respectively; />
Figure QLYQS_123
Representing a first moment estimation of the gradient when the time step is t-1;
Figure QLYQS_126
representing the gradient at time step t; />
Figure QLYQS_129
Representing a second moment estimate of the gradient at time step t-1; />
Figure QLYQS_121
Representing the square of the gradient at time step t;
setting a root mean square error function (RMSE) as a loss function trained by the measurement and alignment model, wherein the formula is as follows:
Figure QLYQS_131
（14）；
in the formula ,
Figure QLYQS_132
for the total number of measurement data to be predicted, +.>
Figure QLYQS_133
For the measurement data true value to be predicted, < +.>
Figure QLYQS_134
Predicted values of measurement data output by the measurement patch model. />
7. The method for compensating measurement data of a medium-low voltage distribution network according to claim 6, wherein the method is characterized by optimizing the trained measurement compensation model to obtain a final measurement compensation model, and comprises the following specific steps:
and optimizing the super parameters of the trained measurement alignment model by adopting a Bayesian optimization method to obtain optimized network parameters, reconstructing the trained measurement alignment model based on the optimized network parameters, and obtaining the final measurement alignment model.
8. The method for measuring and data supplementing in a medium-low voltage distribution network according to claim 7, wherein the super parameters of the trained measuring and supplementing model comprise iteration times, hidden layer numbers, neuron numbers of each layer and learning rate.
9. The method for supplementing measurement data of a medium-low voltage distribution network according to claim 7, wherein the super parameters of the trained measurement and supplementation model are optimized by adopting a Bayesian optimization method, and the optimized network parameters are obtained by the following specific processes:
setting objective functions of a Bayesian framework
Figure QLYQS_135
Independent variable->
Figure QLYQS_136
Representing the super-parameters;
selection of
Figure QLYQS_137
Calculating objective function of each observation point>
Figure QLYQS_138
The value at the observation point, namely the observation value of a preset observation model;
setting up
Figure QLYQS_139
Based on the observation +.>
Figure QLYQS_140
Estimating to obtain the objective function->
Figure QLYQS_141
Is a function distribution of (2)
Figure QLYQS_142
A minimum value of the target value;
setting current observation data
Figure QLYQS_143
Based on current observation dataDCalculating a preset acquisition function and determining the next observation point +.>
Figure QLYQS_144
Calculate the +.>
Figure QLYQS_145
Acquisition function value->
Figure QLYQS_146
Setting up
Figure QLYQS_147
Updating a preset probability agent model;
repeating the above steps until the target value reaches the preset maximum observation times P to obtain the optimized network parameters
Figure QLYQS_148
。
10. The method for supplementing measurement data of a medium-low voltage distribution network according to claim 9, wherein the probability agent model is a gaussian process regression model, and the gaussian process regression model obeys k-dimensional normal distribution and has the following formula:
Figure QLYQS_149
（15）；
wherein ,
Figure QLYQS_150
representing an n-dimensional vector, ">
Figure QLYQS_151
As a mean function>
Figure QLYQS_152
As a covariance function.
11. The method of claim 9, wherein the expected improvement function is used as an acquisition function and a next observation point is determined
Figure QLYQS_153
The formula is as follows:
Figure QLYQS_154
（16）；
Figure QLYQS_155
（17）；
in the formulae (16) to (17),
Figure QLYQS_156
is the position observed in step i, +.>
Figure QLYQS_157
Is the posterior mean of the agent at time s+1; />
Figure QLYQS_158
Representing the observation position when the current objective function is maximized; />
Figure QLYQS_159
Maximum value of the current objective function; the argmax (f (u)) function is an argument u that maximizes the value of f (u); the max (f (u)) function is the maximum value of f (u); d represents the current observation data set; e (f (u)) functions are expected for f (u).
12. The method for supplementing measurement data of a medium-low voltage distribution network according to claim 1, wherein the method is characterized by obtaining measurement data at the current moment, and processing the measurement data to obtain a current input vector matrix and a current response vector, and comprises the following specific steps:
Acquiring measurement data at the current moment;
classifying the measured data at the current moment based on a K-media clustering method, distinguishing main/standby key measured data at the current moment and non-key measured data at the current moment in various data, and calculating average errors between the non-key measured data at the current moment and the main/standby key measured data at the current moment in the same type of data;
judging whether the current main key measurement data is missing, if not, supplementing the missing current non-key measurement data based on the average error between the current non-key measurement data and the current main key measurement data; if yes, comparing the main key measurement data at the current moment with the missing degree of the data of the front period of the standby key measurement data at the current moment, and taking the main/standby key measurement data at the current moment with lower missing degree as the measurement data to be predicted;
selecting data of the first m times of the waiting predicting time of the waiting predicting quantity data to construct an m-dimensional vector
Figure QLYQS_160
And constructs +.>
Figure QLYQS_161
Current input vector matrix of dimension->
Figure QLYQS_162
, wherein ,c 1 the first vector representing the first m moments, c 2 The second vector representing the first m moments,c 3 the third vector representing the first m moments,c m-1 the m-1 st vector representing the first m moments,c m an mth vector representing the first m times;y 2 、y 3 、y 4 、y 5 、y 6 、y 7 respectively represent and vectory 1 M-dimensional vector formed by time sequence features corresponding to m measured data values one by oney 8 Representation and vectory 1 The m measurement data values are in one-to-one correspondence with the judgment of working days and the vectory 9 Representing the type of the measurement object;
construction of the current response vector
Figure QLYQS_163
, wherein ,
Figure QLYQS_164
，/>
Figure QLYQS_165
for the measurement value at time t, set +.>
Figure QLYQS_166
。
13. The method for supplementing measurement data of a medium-low voltage distribution network according to claim 12, wherein the method is characterized in that the current input vector matrix and the current response vector are input into the final measurement supplementing model to obtain a measurement predicted value, and comprises the following steps:
the current input vector matrix and the current response vector are used as input parameters to be input into a final measurement and alignment model, and the output quantity measurement predicted value is
Figure QLYQS_167
, wherein ,/>
Figure QLYQS_168
Measurement prediction value representing the first moment, < +.>
Figure QLYQS_169
Measurement prediction value representing the second moment, < + >>
Figure QLYQS_170
Indicating the measurement prediction value at the third time,
Figure QLYQS_171
measurement prediction value indicating m-1 th time,/->
Figure QLYQS_172
The measurement predicted value at the time t is obtained.
14. The method for supplementing measurement data of a medium-low voltage distribution network according to claim 1, wherein the method for supplementing the measurement data at the current moment based on the measurement predicted value comprises the following steps:
and based on the measurement predicted value, supplementing the missing measurement data according to the average error between the non-key measurement data and the main/standby key measurement data in the same class until all measurement data in all classes are supplemented.
CN202310084891.2A 2023-02-03 2023-02-03 Method for supplementing measurement data of medium-low voltage distribution network Active CN115794805B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310084891.2A CN115794805B (en) 2023-02-03 2023-02-03 Method for supplementing measurement data of medium-low voltage distribution network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310084891.2A CN115794805B (en) 2023-02-03 2023-02-03 Method for supplementing measurement data of medium-low voltage distribution network

Publications (2)

Publication Number Publication Date
CN115794805A CN115794805A (en) 2023-03-14
CN115794805B true CN115794805B (en) 2023-05-23

Family

ID=85430583

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310084891.2A Active CN115794805B (en) 2023-02-03 2023-02-03 Method for supplementing measurement data of medium-low voltage distribution network

Country Status (1)

Country Link
CN (1) CN115794805B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116452369B (en) * 2023-06-19 2023-08-29 深圳市峰和数智科技有限公司 Real-time data alignment method and device for photovoltaic power station and related equipment

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112418406A (en) * 2020-12-09 2021-02-26 甘肃靖远航天风力发电有限公司 Wind power tower inclination angle missing data supplementing method based on SSA-LSTM model

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112418406A (en) * 2020-12-09 2021-02-26 甘肃靖远航天风力发电有限公司 Wind power tower inclination angle missing data supplementing method based on SSA-LSTM model

Also Published As

Publication number Publication date
CN115794805A (en) 2023-03-14

Similar Documents

Publication Publication Date Title
CN111563706A (en) Multivariable logistics freight volume prediction method based on LSTM network
US20190340533A1 (en) Systems and methods for preparing data for use by machine learning algorithms
CN111401599B (en) Water level prediction method based on similarity search and LSTM neural network
CN113298230B (en) Prediction method based on unbalanced data set generated against network
CN113159048A (en) Weak supervision semantic segmentation method based on deep learning
CN109787821B (en) Intelligent prediction method for large-scale mobile client traffic consumption
CN115794805B (en) Method for supplementing measurement data of medium-low voltage distribution network
CN111178585A (en) Fault reporting amount prediction method based on multi-algorithm model fusion
US20230342606A1 (en) Training method and apparatus for graph neural network
CN116187835A (en) Data-driven-based method and system for estimating theoretical line loss interval of transformer area
CN114330934A (en) Model parameter self-adaptive GRU new energy short-term power generation power prediction method
CN114609994A (en) Fault diagnosis method and device based on multi-granularity regularization rebalance incremental learning
US20230281363A1 (en) Optimal materials and devices design using artificial intelligence
CN112036598A (en) Charging pile use information prediction method based on multi-information coupling
CN111882114A (en) Short-term traffic flow prediction model construction method and prediction method
Anh et al. Effect of gradient descent optimizers and dropout technique on deep learning LSTM performance in rainfall-runoff modeling
CN117458480A (en) Photovoltaic power generation power short-term prediction method and system based on improved LOF
CN113033898A (en) Electrical load prediction method and system based on K-means clustering and BI-LSTM neural network
CN116245259B (en) Photovoltaic power generation prediction method and device based on depth feature selection and electronic equipment
CN117113086A (en) Energy storage unit load prediction method, system, electronic equipment and medium
CN116303786B (en) Block chain financial big data management system based on multidimensional data fusion algorithm
CN115907228A (en) Short-term power load prediction analysis method based on PSO-LSSVM
CN115619028A (en) Clustering algorithm fusion-based power load accurate prediction method
CN109033678A (en) A kind of aircraft near-optimal design method generated based on virtual sample
CN115630979A (en) Day-ahead electricity price prediction method and device, storage medium and computer equipment

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
GR01 Patent grant
GR01 Patent grant