CN112990593A - Transformer fault diagnosis and state prediction method based on CSO-ANN-EL algorithm - Google Patents

Transformer fault diagnosis and state prediction method based on CSO-ANN-EL algorithm Download PDF

Info

Publication number
CN112990593A
CN112990593A CN202110333795.8A CN202110333795A CN112990593A CN 112990593 A CN112990593 A CN 112990593A CN 202110333795 A CN202110333795 A CN 202110333795A CN 112990593 A CN112990593 A CN 112990593A
Authority
CN
China
Prior art keywords
dga
model
ann
cso
training
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
CN202110333795.8A
Other languages
Chinese (zh)
Inventor
范竞敏
曹云飞
曾伟良
孟安波
殷豪
王裕
周永旺
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.)
Guangdong University of Technology
Original Assignee
Guangdong University of Technology
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 Guangdong University of Technology filed Critical Guangdong University of Technology
Priority to CN202110333795.8A priority Critical patent/CN112990593A/en
Publication of CN112990593A publication Critical patent/CN112990593A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • 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
    • G06F18/2148—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the process organisation or structure, e.g. boosting cascade
    • 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
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/20—Administration of product repair or maintenance
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Strategic Management (AREA)
  • Marketing (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • General Business, Economics & Management (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Biophysics (AREA)
  • Tourism & Hospitality (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Primary Health Care (AREA)
  • Water Supply & Treatment (AREA)
  • Physiology (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a transformer fault diagnosis and state prediction method based on a CSO-ANN-EL algorithm in order to overcome the defects of unsatisfactory training effect and low prediction precision when a traditional artificial neural network is applied to transformer fault diagnosis, historical DGA sample data is collected, self-sampling is carried out after standard pretreatment is carried out on the historical DGA sample data, a training sub-sample is created, a plurality of base classifiers are trained by utilizing the training sub-sample, and an EL model is constructed; obtaining a DGA data set with a time sequence and carrying out standardized preprocessing; constructing an ANN model, and optimizing parameters of the ANN model by adopting a CSO algorithm to obtain a CSO-ANN algorithm model; inputting the DGA data set into a CSO-ANN algorithm model for training and outputting a DGA data prediction result; and inputting the DGA data prediction result and the output result of the base classifier into the trained EL model to diagnose the transformer fault to obtain a transformer fault diagnosis result.

Description

Transformer fault diagnosis and state prediction method based on CSO-ANN-EL algorithm
Technical Field
The invention relates to the technical field of transformer fault diagnosis, in particular to a transformer fault diagnosis and state prediction method based on a CSO-ANN-EL algorithm.
Background
The transformer is an important electrical device in a power transmission and transformation system, the running state of the transformer directly affects the safety level of a power system, and the failure of the transformer can cause large-scale power failure and cause immeasurable economic loss. Therefore, the research on the transformer fault diagnosis technology has important significance for improving the accuracy of transformer fault diagnosis, improving the utilization rate of the transformer and improving the operation and maintenance level of the transformer.
The conventional methods for diagnosing the transformer fault comprise a three-ratio method, an improved three-ratio method and the like, and the intelligent algorithm for diagnosing the transformer fault comprises a gray model, a support vector machine and an Artificial Neural Network (ANN) and corresponding improved algorithms. However, the triple ratio method and the improved triple ratio method gradually show the disadvantages of incomplete coding and over-absolute judgment standard in the practical process. The grey theory utilizes an exponential law to establish a time series data model, which is suitable for a monotonous increasing or monotonous decreasing process, but due to the complexity of transformer faults and corresponding gas content of the transformer in oil, errors can be generated when the exponential law is adopted to carry out data fitting. The support vector machine utilizes a structural risk minimization principle rather than an empirical risk minimization principle to have excellent generalization capability on the small sample binary classification problem, but the practicability and effectiveness of the support vector machine are different according to specific applications. When the training samples are sufficient, the use of an Artificial Neural Network (ANN), such as a BP neural network, is one of the most suitable methods for non-linear prediction. However, the traditional artificial neural network has the defects of low convergence rate, easy falling into local minimum and the like in the training process, and the prediction precision is greatly influenced. And because the fault rate of the existing transformer is low, the fault DGA data samples are fewer, and the training effect on the traditional artificial neural network is not ideal.
Disclosure of Invention
The invention provides a transformer fault diagnosis and state prediction method based on a CSO-ANN-EL algorithm, aiming at overcoming the defects of unsatisfactory training effect and low prediction precision when the traditional artificial neural network is applied to transformer fault diagnosis in the prior art.
In order to solve the technical problems, the technical scheme of the invention is as follows:
the transformer fault diagnosis and state prediction method based on the CSO-ANN-EL algorithm comprises the following steps:
collecting historical DGA sample data and carrying out standardized preprocessing;
carrying out self-sampling on DGA sample data after standardized preprocessing, creating a training subsample, training a plurality of base classifiers by using the training subsample, and constructing an integrated classifier to obtain an EL (Ensemble Learning) model; inputting the DGA sample data after the standardization preprocessing into an EL model for training through ensemble learning;
obtaining a DGA data set with a time sequence by a DGA online detection technology and carrying out standardized preprocessing;
setting parameters of an Artificial Neural Network (ANN) and constructing an ANN model; optimizing parameters of the ANN model by adopting a CSO (Crissscross Optimization) algorithm to obtain a CSO-ANN algorithm model; inputting the DGA data set subjected to the standardized preprocessing into a CSO-ANN algorithm model for training and outputting a DGA data prediction result, wherein the prediction result is a short-term prediction result of the fault state of the transformer;
and inputting the DGA data prediction result and the output result of the base classifier into the trained EL model to diagnose the transformer fault to obtain a transformer fault diagnosis result.
Preferably, the step of performing a standardized preprocessing on the DGA sample data or the DGA data set comprises: calculating the volume concentration percentage X of each gas dissolved in the transformer oil in the dataiThe calculation formula is as follows:
Figure BDA0002997366220000021
in the formula (I), the compound is shown in the specification,
Figure BDA0002997366220000022
the gas volume concentration of the i (i ═ 1, 2.., p) th gas is represented, and p represents the number of types of gas in which DGA sample data or DGA data is concentrated and dissolved in the transformer oil.
Preferably, the step of training the plurality of base classifiers by using the training subsamples comprises: dividing the DGA sample data after the standardized preprocessing into a training set and a testing set, carrying out self-sampling from the training set, and creating training subsamples of p gases; and respectively inputting the training subsamples into a base classifier adopting a CART model for training, and performing ensemble learning by adopting a bagging method to construct an ensemble classifier to obtain the EL model.
Preferably, the step of training the EL model by ensemble learning includes: classifying transformer faults of input DGA sample data by adopting p bagged decision trees; and obtaining all outputs from the trained base classifiers, voting the outputs of all the base classifiers through simple weighting and leaf-Bayesian voting, and determining the final diagnosis result output by the EL model.
As a preferred scheme, the parameters of the artificial neural network ANN include an input mode, an output mode, the number of hidden layers, the number of neurons in hidden layers, a transfer function in hidden layers, the number of output layers, the number of neurons in output layers, and a transfer function in output layers of the artificial neural network ANN.
As a preferred scheme, the steps of setting parameters of the artificial neural network ANN and constructing an ANN model include:
adopting a three-layer BP network model as an ANN model, and inputting a topological structure and a minimum allowable error parameter of the BP network model; the topological structure of the BP network model comprises the number of input layers, the number of hidden layers and the number of output layers; setting a hidden layer activation function and an output layer activation function;
calculating the input alpha of hidden layer neuron in the BP network modelh(k) And output betah(k) And input y to output layer neuronsj(k) And output yoj(k) The calculation formula is as follows:
Figure BDA0002997366220000031
βh(k)＝f1(αh(k))
Figure BDA0002997366220000032
yoj(k)＝f2(yj(k))
in the formula, wihAs a connection weight of the input layer and the hidden layer, bhThreshold for neurons of the hidden layer, whjAs a connection weight of the hidden layer to the output layer, bjA threshold value of each neuron of an output layer is set, wherein h is 1,2, the. x is the number ofi(k) Denotes the ith sample in the kth input sample set x (k), where x (k) ═ x1(k),x2(k),...,xn(k)]；dj(k) Represents the desired output of the kth input sample set x (k); f. of1For the hidden layer activation function, f2Activating a function for an output layer;
defining an error function E, and calculating the connection weight of the hidden layer and the output layer, the threshold value of the output layer, the connection weight of the input layer and the hidden layer and the correction quantity of the threshold value of the hidden layer according to the error function E, wherein the expression formula is as follows:
Figure BDA0002997366220000033
Figure BDA0002997366220000034
Figure BDA0002997366220000035
Figure BDA0002997366220000041
Figure BDA0002997366220000042
where eta is a correction coefficient, deltaj(k) To the sensitivity of the output layer, δh(k) Sensitivity of the hidden layer; f. of1' derivation of the implicit layer activation function, f2' is the derivation of the output layer activation function;
adjusting the corresponding connection weight and threshold according to the correction, and calculating the actual output value d of the BP network model according to the adjusted connection weight and thresholdiAnd the desired output value yiDetermining the optimal positions of an input layer, a hidden layer and an output layer in the BP network model according to the error function fit; the calculation formula of the error function fit is as follows:
Figure BDA0002997366220000043
wherein N is the number of training samples.
As a preferred scheme, the step of optimizing the parameters of the ANN model by using the CSO algorithm includes: setting population scale, particle dimension number, maximum iteration number, transverse crossing probability and longitudinal crossing probability parameters of a CSO algorithm; and calculating the adaptation value of each CSO particle by taking the error function fit as a CSO fitness function to obtain an individual optimal value and a global optimal value of the CSO particles, comparing the individual optimal value and the global optimal value of the CSO particles, and taking the better adaptation value as the current optimal position to adjust the parameters of the BP network model to obtain the CSO-ANN algorithm model.
Preferably, the step of inputting the normalized and preprocessed DGA data set into a CSO-ANN algorithm model for training includes: dividing the DGA data set subjected to the standardized preprocessing into a training set and a testing set, taking parameters optimized by a CSO algorithm as an initial connection weight and a threshold of a CSO-ANN algorithm model, inputting the training set into the CSO-ANN algorithm model for short-term DGA data prediction, evaluating the effectiveness of the DGA data prediction, and optimizing the connection weight and the threshold of the ANN model again when an output error is greater than a preset error precision; when the output error is smaller than the preset error precision, obtaining a CSO-ANN algorithm model which completes training; and inputting the test set into a CSO-ANN algorithm model which finishes training, and outputting a DGA data prediction result.
As a preferable scheme, the effectiveness of DGA data prediction is evaluated by adopting average absolute percentage error, average absolute error and root mean square error.
Preferably, the DGA sample data comprises volume concentrations of hydrogen, methane, ethane, ethylene and acetylene dissolved in transformer oil, and the DGA sample data is provided with a transformer fault type label, wherein the transformer fault type comprises low-energy discharge, high-energy discharge, partial discharge, medium-low temperature overheat and high-temperature overheat; the time-series DGA data set includes the dissolved volume concentrations of hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide in transformer oil.
Compared with the prior art, the technical scheme of the invention has the beneficial effects that: the method is suitable for transformer fault diagnosis of small sample data, can evaluate the state of the transformer in advance, and greatly improves the efficiency of transformer fault diagnosis; and optimizing each weight and threshold of the ANN model by using a CSO algorithm, so that the convergence speed of the neural network is accelerated and the neural network does not fall into local optimum.
Drawings
FIG. 1 is a flow chart of a transformer fault diagnosis and state prediction method based on a CSO-ANN-EL algorithm according to the present invention.
FIG. 2 is a flow chart of a transformer fault diagnosis and state prediction method based on a CSO-ANN-EL algorithm according to the present invention.
FIG. 3 is a block diagram of the CSO-ANN model and EL model of the present invention.
FIG. 4 shows hydrogen (H) in example2) Comparing the predicted value of DGA with the measured value.
FIG. 5 shows methane (CH) of the example4) Comparing the predicted value of DGA with the measured value.
FIG. 6 shows ethylene (C) of the example2H4) Comparing the predicted value of DGA with the measured value.
FIG. 7 shows ethane (C) of the example2H6) Comparing the predicted value of DGA with the measured value.
FIG. 8 shows acetylene (C) of the example2H2) Comparing the predicted value of DGA with the measured value.
FIG. 9 is a graph comparing the predicted value of DGA of carbon monoxide (CO) and the measured value in the examples.
Detailed Description
The drawings are for illustrative purposes only and are not to be construed as limiting the patent;
for the purpose of better illustrating the embodiments, certain features of the drawings may be omitted, enlarged or reduced, and do not represent the size of an actual product;
it will be understood by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.
Example (b):
the present embodiment provides a transformer fault diagnosis and state prediction method based on a CSO-ANN-EL algorithm, which is a flowchart of the transformer fault diagnosis and state prediction method based on the CSO-ANN-EL algorithm of the present embodiment, as shown in fig. 1 to 2.
The transformer fault diagnosis and state prediction method based on the CSO-ANN-EL algorithm provided by the embodiment comprises the following steps:
s1: and collecting historical DGA sample data and carrying out standardized preprocessing.
The historical DGA sample data collected in this embodiment includes hydrogen (H)2) Methane (CH)4) Ethane (C)2H6) Ethylene (C)2H4) Acetylene (C)2H2) Volume concentrations of the five gases; the historical DGA sample data also contains corresponding transformer fault types including low-energy discharge (D)1) High energy discharge (D)2) Partial Discharge (PD), medium and low temperature overheating (C/O)<700℃)(T1) High-temperature superheating (>700℃)(T2) Five states in total; the corresponding output is encoded as: 10000. 01000, 00100, 00010 and 00001.
Since the capacity and voltage level of the transformer differ from each other, the volume concentration of dissolved gas in the transformer oil also differs. In order to eliminate the difference, the present embodiment performs a normalization preprocessing on the input DGA sample data, specifically: calculating the volume concentration percentage X of each gas dissolved in the transformer oil in the dataiThe calculation formula is as follows:
Figure BDA0002997366220000061
in the formula (I), the compound is shown in the specification,
Figure BDA0002997366220000062
represents the gas volume concentration of the i (i ═ 1, 2.., 5) th gas.
S2: and self-sampling the DGA sample data after the standardization preprocessing, creating a training subsample, training a plurality of base classifiers by using the training subsample, and constructing an integrated classifier to obtain the EL model.
In the step, DGA sample data after standardized preprocessing is divided into a training set (accounting for 45 percent of the whole DGA sample data set) and a testing set (accounting for 55 percent of the whole DGA sample data set), and then bootstrap self-sampling is carried out from the training set to create training subsamples of 5 gases; and respectively inputting the training subsamples into a base classifier adopting a CART model for training, and performing ensemble learning by adopting a bagging (bagging) method to construct an ensemble classifier to obtain the EL model.
S3: and inputting the DGA sample data after the standardization preprocessing into an EL model for training through ensemble learning.
Specifically, a test set of DGA sample data is input into an EL model which is trained, and a base classifier adopting a CART model in the EL model respectively outputs transformer fault diagnosis results.
Further, the basis classifiers in this embodiment are bagged decision trees, and the number of the basis classifiers is set to be 175. And after the base classifiers finish classifying the input test set and respectively output transformer fault diagnosis results, determining the transformer fault diagnosis result finally output by the EL model through simple weighting and leaf Bayesian voting on the outputs of all the base classifiers.
S4: and acquiring a DGA data set with a time sequence by using a DGA online detection technology and carrying out standardized preprocessing.
The DGA data set with time series obtained in this example includes hydrogen (H)2) Methane (CH)4) Ethane (C2H)6) Ethylene (C)2H4) Acetylene (C)2H2) Carbon monoxide (CO), gas concentration of dissolved gases in six transformer oils.
In this step, the step of performing normalization preprocessing on the DGA data set having the time series is the same as the step S2.
S5: and setting parameters of the artificial neural network ANN and constructing an ANN model.
In this example, ANN model extractionSix characteristic gases (H) dissolved in transformer oil by using a three-layer BP network model2、CH4、C2H6、C2H4、C2H2CO) was predicted.
Inputting the topological structure and the minimum allowable error parameter of the BP network model; the topological structure of the BP network model comprises the number of input layers, the number of hidden layers and the number of output layers; setting a hidden layer activation function and an output layer activation function;
calculating the input alpha of each neuron of the hidden layer in the BP network modelh(k) And output betah(k) And input y of each neuron of the output layerj(k) And output yoj(k) The calculation formula is as follows:
Figure BDA0002997366220000071
βh(k)＝f1(αh(k))
Figure BDA0002997366220000072
yoj(k)＝f2(yj(k))
in the formula, wihAs a connection weight of the input layer and the hidden layer, bhThreshold for neurons of the hidden layer, whjAs a connection weight of the hidden layer to the output layer, bjA threshold value of each neuron of an output layer is set, wherein h is 1,2, the. x is the number ofi(k) Denotes the ith sample in the kth input sample set x (k), where x (k) ═ x1(k),x2(k),...,xn(k)]；dj(k) Represents the desired output of the kth input sample set x (k); f. of1For the hidden layer activation function, f2Activating a function for an output layer;
defining an error function E, and calculating the connection weight of the hidden layer and the output layer, the threshold value of the output layer, the connection weight of the input layer and the hidden layer and the correction quantity of the threshold value of the hidden layer according to the error function E, wherein the expression formula is as follows:
Figure BDA0002997366220000081
Figure BDA0002997366220000082
Figure BDA0002997366220000083
Figure BDA0002997366220000084
Figure BDA0002997366220000085
where eta is a correction coefficient, deltaj(k) To the sensitivity of the output layer, δh(k) Sensitivity of the hidden layer; f. of1' derivation of the implicit layer activation function, f2' is the derivation of the output layer activation function;
according to the correction quantity, the connection weight w between the hidden layer and the output layer is adjustedhjAnd threshold b of each neuron of output layerjThe connection weight w between the input layer and the hidden layerihPerforming threshold b of each neuron of the hidden layerhAdjusting; calculating the actual output value d of the BP network model according to the adjusted connection weight and the threshold valueiAnd the desired output value yiDetermining the optimal positions of an input layer, a hidden layer and an output layer in the BP network model according to the error function fit; the calculation formula of the error function fit is as follows:
Figure BDA0002997366220000086
wherein N is the number of training samples.
S6: and optimizing parameters of the ANN model by adopting a CSO algorithm to obtain a CSO-ANN algorithm model.
In this step, the step of optimizing the parameters of the ANN model by using the CSO algorithm includes:
setting population scale, particle dimension number, maximum iteration number, transverse crossing probability and longitudinal crossing probability parameters of a CSO algorithm; and calculating the adaptation value of each CSO particle by taking the error function fit as a CSO fitness function to obtain an individual optimal value and a global optimal value of the CSO particles, comparing the individual optimal value and the global optimal value of the CSO particles, and taking the better adaptation value as the current optimal position to adjust the parameters of the BP network model to obtain the CSO-ANN algorithm model.
Specifically, a CSO algorithm is used for optimizing a weight value and a threshold value in a BP neural network serving as an ANN model, and the method comprises the following steps:
establishing mapping between the CSO particles and the weight and the threshold of the BP neural network, namely encoding the weight and the threshold of the neural network into real number vectors to represent individuals in a population, and randomly generating the population of the real number vectors, wherein the optimization object comprises: correction amount to connection weight w between hidden layer and output layerhjAnd threshold b of each neuron of output layerjThe connection weight w between the input layer and the hidden layerihPerforming threshold b of each neuron of the hidden layerhAnd the weight value and the threshold value form an initial population of the CSO algorithm.
(1) Population initialization
Let X be a randomly generated matrix with D columns and M rows that represents a D-dimensional population of M individuals.
(2) Transverse cross operation
Performing random combination on all particles in the population in pairs without repetition, wherein M/2 pairs of combinations are total, and for each pair of combinations, performing transverse crossing on the particles according to the following formula:
Figure BDA0002997366220000091
wherein r is1，r2Is [0,1 ]]Random values evenly distributed over; c. C1，c2Is [ -1,1 [ ]]Random values evenly distributed over; x (i, d), X (j, d) are dimension d of parent particle X (i) and X (j), respectively; MS (Mass Spectrometry)hc(i, d) and MShc(j, d) are X (i), respectively, and X (j) are d-dimension descendants generated by transverse intersection, which are called mediocre solutions;
the transverse crossing result is stored in a mediocre solution matrix MShcCalculating the fitness value of the particle, comparing the fitness value with the fitness values of parent particles X (i), X (j), and keeping the small fitness value in X to participate in the next iteration;
normalizing each dimension of the particles obtained by transverse crossing, then carrying out non-repeated pairwise random pairing on all dimensions of the particles to obtain D/2 pairs, generating a random number rand for any pair of dimensions, and carrying out longitudinal crossing operation on the pair of dimensions if rand < Pv (Pv is usually 0.2-0.8).
(3) Longitudinal cross operation
Unlike transverse crossing, longitudinal crossing operates between different dimensions for all particles. Assuming that the longitudinal crossing is at d of the particle X (i)1And d2Performed dimensionally, they produce intermediate-quality MS after longitudinal crossingvc(i, d) are:
MSvc(i,d)＝r·X(i,d1)+(1-r)·X(i,d2)
i∈N(1,M),d1,d2∈N(1,D)
in the formula: r is [0,1 ]]Random numbers uniformly distributed thereon; MS (Mass Spectrometry)vc(i, d) is the d-th particle of i1And d2The progeny after the vertical crossing of the dimension.
The vertical crossing result is stored in a mediocre solution matrix MSvcIn the method, after the vertical crossing result is subjected to inverse normalization, the fitness value of the particles in the intermediate solution matrix is calculated, the fitness value is compared with the fitness value of the parent particles, the particles with good fitness values are stored in X, and the next iteration is carried out.
Performing competitive operation on filial generations and parent generations generated by transverse crossing and longitudinal crossing to continuously generate new populations; if the new adaptive value is better than the current individual optimum, replacing the current individual optimum with the adaptive value; and if the updated individual optimal value is superior to the current global optimal value, replacing the current global optimal value with the individual optimal value to complete the optimization of the weight and the threshold of the BP neural network.
S7: and inputting the DGA data set subjected to the standardized preprocessing into a CSO-ANN algorithm model for training and outputting a DGA data prediction result.
Dividing the DGA data set subjected to the standardized preprocessing into a training set and a testing set, taking parameters subjected to CSO algorithm optimization as an initial connection weight and a threshold of a CSO-ANN algorithm model, inputting the training set into the CSO-ANN algorithm model for short-term DGA data prediction, evaluating the effectiveness of DGA data prediction, and optimizing the connection weight and the threshold of the ANN model again when an output error is greater than a preset error precision; when the output error is smaller than the preset error precision, obtaining a CSO-ANN algorithm model which completes training; and inputting the test set into a CSO-ANN algorithm model which finishes training, and outputting a DGA data prediction result.
Further, the present embodiment uses the Mean Absolute Percentage Error (MAPE), the Mean Absolute Error (MAE) and the Root Mean Square Error (RMSE) to evaluate the effectiveness of the DGA data prediction, and the calculation formula is as follows:
Figure BDA0002997366220000101
wherein y ishAnd
Figure BDA0002997366220000102
representing the measured and predicted values of gas concentration, Q is the characteristic gas quantity.
S8: and inputting the DGA data prediction result and the output result of the base classifier into the trained EL model to diagnose the transformer fault to obtain a transformer fault diagnosis result.
The CSO-ANN-EL model for DGA data prediction and transformer fault diagnosis constructed by the embodiment is suitable for transformer fault diagnosis of small sample data, can evaluate the state of the transformer in advance, and greatly improves the efficiency of transformer fault diagnosis.
In one implementation, a DGA dataset was constructed using 200 data points acquired at the same monitoring interval (6h) using an online monitoring system. Samples 1-175 are used as training samples, and samples 176-200 are used as testing samples.
In this embodiment, the number of input layer neurons of the ANN model is set to 6, the number of hidden layer neurons is set to 4, and the number of output layer neurons is set to 6.
Through training and testing, the predicted values of 6 gas concentrations of 25 samples in the test set are compared with the actually measured gas concentration values, as shown in fig. 4-9. As can be seen from the graph, the variation trend of the prediction value obtained by the transformer fault diagnosis and state prediction method based on the CSO-ANN-EL algorithm provided in this embodiment is similar to the variation trend of the measured value.
Further, the effectiveness of DGA data prediction was evaluated using Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE), and the results are shown in the table below.
TABLE 1 prediction error results for six gases
Gas (es) H2 C2H2 CH4 C2H4 C2H6 CO
MAPE(％) 3.512 2.011 0.078 1.603 1.711 1.702
MAE(ppm) 8.864 0.827 0.341 0.430 0.150 11.177
RMSE(ppm) 5.606 0.508 0.058 0.269 0.086 7.284
As can be seen from Table 1, the prediction accuracy of MAPE showed six gases was less than 4% and 3.512% maximum, the CSO-ANN-EL algorithm of this example was at H2、CH4、C2H4、C2H6The prediction precision on MAPE is superiorIn PSO-SVM, GM model (S.Fei, M.Wang, Y.Miao, J.Tu, C.Liu.particle swing optimization-based supported vector machines for implementing dispersed gases content in power transfer reactor oil. energy. converters. management. 2009; 50: 1604-. Therefore, the CSO-ANN-EL algorithm provided by the embodiment is more sensitive and accurate to the prediction result of the variation trend of the DGA data.
In one implementation, 200 sets of historical DGA data collected are used, divided into a training data set (having 90 records and accounting for 45% of the total data set) and a test data set (having 110 records and accounting for 55% of the total data set), where the training data set may be smaller than the test data set.
In order to verify the effectiveness of the proposed algorithm, the CSO-ANN-EL algorithm proposed in this embodiment, and five algorithms such as a decision tree, a random forest, an AdaBoost classifier, a gradient boosting tree (GBCT), and a Support Vector Machine (SVM) are simulated. In order to obtain better classification accuracy, a grid search method is used to find the optimal parameters, and the parameters are set as follows: the basis classifiers of the CSO-ANN-EL algorithm, the random forest, the AdaBoost algorithm and the gradient-boosted tree GBCT algorithm proposed in this embodiment are set as decision trees, and the number of the decision trees is set as 175. The maximum sampling rate for each estimator was set to 0.75, using a bootstrap sampling method, where the learning rates of AdaBoost and gradient-boosted tree GBCT were set to 0.1. The classification accuracy of the simulation experiment using the above algorithm is shown in table 2 below.
TABLE 2 Classification accuracy of various Algorithm simulation experiments
Method CSO-ANN-EL Random forest Decision tree GBCT AdaBoost SVM
Training set 100％ 100％ 100％ 100％ 100％ 56.7％
Test set 92.7％ 91.8％ 86.4％ 86.4％ 86.4％ 43.6％
As can be seen from table 2, the transformer fault diagnosis and state prediction method based on the CSO-ANN-EL algorithm provided by the embodiment has better performance than algorithms such as random forest, decision tree, gradient boosting tree GBCT, AdaBoost classifier, support vector machine SVM, and the like.
The same or similar reference numerals correspond to the same or similar parts;
the terms describing positional relationships in the drawings are for illustrative purposes only and are not to be construed as limiting the patent;
it should be understood that the above-described embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. And are neither required nor exhaustive of all embodiments. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.

Claims (10)

1. The transformer fault diagnosis and state prediction method based on the CSO-ANN-EL algorithm is characterized by comprising the following steps of:
collecting historical DGA sample data and carrying out standardized preprocessing;
carrying out self-sampling on DGA sample data after standardized preprocessing, creating a training subsample, training a plurality of base classifiers by using the training subsample, and constructing an integrated classifier to obtain an EL model; inputting the DGA sample data after the standardization preprocessing into an EL model for training through ensemble learning;
obtaining a DGA data set with a time sequence by a DGA online detection technology and carrying out standardized preprocessing;
setting parameters of an Artificial Neural Network (ANN) and constructing an ANN model; optimizing parameters of the ANN model by adopting a CSO algorithm to obtain a CSO-ANN algorithm model; inputting the DGA data set subjected to the standardized preprocessing into a CSO-ANN algorithm model for training and outputting a DGA data prediction result;
and inputting the DGA data prediction result and the output result of the base classifier into the trained EL model to diagnose the transformer fault to obtain a transformer fault diagnosis result.
2. The transformer fault diagnosis and condition prediction method according to claim 1, wherein the step of performing a standardized pre-processing on DGA sample data or DGA data set comprises: calculating the volume concentration percentage X of each gas dissolved in the transformer oil in the dataiThe calculation formula is as follows:
Figure FDA0002997366210000011
in the formula (I), the compound is shown in the specification,
Figure FDA0002997366210000012
the gas volume concentration of the i (i ═ 1, 2.., p) th gas is represented, and p represents the number of types of gas in which DGA sample data or DGA data is concentrated and dissolved in the transformer oil.
3. The transformer fault diagnosis and condition prediction method according to claim 2, wherein the step of training the plurality of base classifiers using the training subsamples comprises: dividing the DGA sample data after the standardized preprocessing into a training set and a testing set, carrying out self-sampling from the training set, and creating training subsamples of p gases; and respectively inputting the training subsamples into a base classifier adopting a CART model for training, and performing ensemble learning by adopting a bagging method to construct an ensemble classifier to obtain the EL model.
4. The transformer fault diagnosis and condition prediction method according to claim 3, wherein the step of training the EL model by ensemble learning comprises:
classifying transformer faults of input DGA sample data by adopting p bagged decision trees; and obtaining all outputs from the trained base classifiers, voting the outputs of all the base classifiers through simple weighting and leaf-Bayesian voting, and determining the final diagnosis result output by the EL model.
5. The transformer fault diagnosis and condition prediction method according to claim 1, wherein the parameters of the artificial neural network ANN comprise an input mode, an output mode, a number of hidden layer layers, a number of hidden layer neurons, a hidden layer transfer function, a number of output layer layers, a number of output layer neurons, and an output layer transfer function of the artificial neural network ANN.
6. The transformer fault diagnosis and condition prediction method according to claim 5, wherein the step of setting parameters of an Artificial Neural Network (ANN) and constructing an ANN model comprises:
adopting a three-layer BP network model as an ANN model, and inputting a topological structure and a minimum allowable error parameter of the BP network model; the topological structure of the BP network model comprises the number of input layers, the number of hidden layers and the number of output layers; setting a hidden layer activation function and an output layer activation function;
calculating the input alpha of hidden layer neuron in the BP network modelh(k) And output betah(k) And input y to output layer neuronsj(k) And output yoj(k) The calculation formula is as follows:
Figure FDA0002997366210000021
βh(k)＝f1(αh(k))
Figure FDA0002997366210000022
yoj(k)＝f2(yj(k))
in the formula, wihAs a connection weight of the input layer and the hidden layer, bhThreshold for neurons of the hidden layer, whjAs a connection weight of the hidden layer to the output layer, bjA threshold value of each neuron of an output layer is set, wherein h is 1,2, the. x is the number ofi(k) Denotes the ith sample in the kth input sample set x (k), where x (k) ═ x1(k),x2(k),...,xn(k)]；dj(k) Represents the desired output of the kth input sample set x (k); f. of1For the hidden layer activation function, f2Activating a function for an output layer;
defining an error function E, and calculating the connection weight of the hidden layer and the output layer, the threshold value of the output layer, the connection weight of the input layer and the hidden layer and the correction quantity of the threshold value of the hidden layer according to the error function E, wherein the expression formula is as follows:
Figure FDA0002997366210000023
Figure FDA0002997366210000024
Figure FDA0002997366210000031
Figure FDA0002997366210000032
Figure FDA0002997366210000033
where eta is a correction coefficient, deltaj(k) To the sensitivity of the output layer, δh(k) Sensitivity of the hidden layer; f. of1' derivation of the implicit layer activation function, f2' is the derivation of the output layer activation function;
adjusting the corresponding connection weight and threshold according to the correction, and calculating the actual output value d of the BP network model according to the adjusted connection weight and thresholdiAnd the desired output value yiDetermining the optimal positions of an input layer, a hidden layer and an output layer in the BP network model according to the error function fit; the calculation formula of the error function fit is as follows:
Figure FDA0002997366210000034
wherein N is the number of training samples.
7. The transformer fault diagnosis and condition prediction method according to claim 6, wherein the step of optimizing the parameters of the ANN model using the CSO algorithm comprises:
setting population scale, particle dimension number, maximum iteration number, transverse crossing probability and longitudinal crossing probability parameters of a CSO algorithm; and calculating the adaptation value of each CSO particle by taking the error function fit as a CSO fitness function to obtain an individual optimal value and a global optimal value of the CSO particles, comparing the individual optimal value and the global optimal value of the CSO particles, and taking the better adaptation value as the current optimal position to adjust the parameters of the BP network model to obtain the CSO-ANN algorithm model.
8. The transformer fault diagnosis and condition prediction method of claim 7, wherein the step of inputting the normalized preprocessed DGA data set into a CSO-ANN algorithm model for training comprises:
dividing the DGA data set subjected to the standardized preprocessing into a training set and a testing set, taking parameters optimized by a CSO algorithm as an initial connection weight and a threshold of a CSO-ANN algorithm model, inputting the training set into the CSO-ANN algorithm model for short-term DGA data prediction, evaluating the effectiveness of the DGA data prediction, and optimizing the connection weight and the threshold of the ANN model again when an output error is greater than a preset error precision; when the output error is smaller than the preset error precision, obtaining a CSO-ANN algorithm model which completes training; and inputting the test set into a CSO-ANN algorithm model which finishes training, and outputting a DGA data prediction result.
9. The transformer fault diagnosis and condition prediction method according to claim 8, wherein the effectiveness of DGA data prediction is evaluated using mean absolute percentage error, mean absolute error, root mean square error.
10. The transformer fault diagnosis and condition prediction method according to any one of claims 1 to 9, wherein the DGA sample data includes volume concentrations of hydrogen, methane, ethane, ethylene, and acetylene dissolved in transformer oil, and the DGA sample data is provided with a transformer fault type tag, and the transformer fault types include low energy discharge, high energy discharge, partial discharge, medium and low temperature overheat, and high temperature overheat; the time-series DGA data set includes the dissolved volume concentrations of hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide in transformer oil.
CN202110333795.8A 2021-03-29 2021-03-29 Transformer fault diagnosis and state prediction method based on CSO-ANN-EL algorithm Pending CN112990593A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110333795.8A CN112990593A (en) 2021-03-29 2021-03-29 Transformer fault diagnosis and state prediction method based on CSO-ANN-EL algorithm

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110333795.8A CN112990593A (en) 2021-03-29 2021-03-29 Transformer fault diagnosis and state prediction method based on CSO-ANN-EL algorithm

Publications (1)

Publication Number Publication Date
CN112990593A true CN112990593A (en) 2021-06-18

Family

ID=76337939

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110333795.8A Pending CN112990593A (en) 2021-03-29 2021-03-29 Transformer fault diagnosis and state prediction method based on CSO-ANN-EL algorithm

Country Status (1)

Country Link
CN (1) CN112990593A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113985207A (en) * 2021-10-28 2022-01-28 国网北京市电力公司 Method, system and device for monitoring faults of power grid operation equipment and storage medium
CN114109737A (en) * 2021-07-01 2022-03-01 国电电力宁夏新能源开发有限公司 Wind turbine generator hydraulic station state diagnosis system and method
CN114878674A (en) * 2022-04-28 2022-08-09 国电南瑞南京控制***有限公司 Transformer winding defect diagnosis method based on comprehensive characteristics of winding stress and magnetic leakage parameter of fusion algorithm

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110516757A (en) * 2019-09-02 2019-11-29 广东工业大学 A kind of transformer fault detection method and relevant apparatus
CN111722046A (en) * 2020-07-01 2020-09-29 昆明理工大学 Transformer fault diagnosis method based on deep forest model
CN112561058A (en) * 2020-12-15 2021-03-26 广东工业大学 Short-term photovoltaic power prediction method based on Stacking-ensemble learning

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110516757A (en) * 2019-09-02 2019-11-29 广东工业大学 A kind of transformer fault detection method and relevant apparatus
CN111722046A (en) * 2020-07-01 2020-09-29 昆明理工大学 Transformer fault diagnosis method based on deep forest model
CN112561058A (en) * 2020-12-15 2021-03-26 广东工业大学 Short-term photovoltaic power prediction method based on Stacking-ensemble learning

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
刘宇俊 等: "基于纵横交叉优化BP神经网络的变压器故障诊断研究", 《陕西电力》 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114109737A (en) * 2021-07-01 2022-03-01 国电电力宁夏新能源开发有限公司 Wind turbine generator hydraulic station state diagnosis system and method
CN114109737B (en) * 2021-07-01 2024-03-08 国电电力宁夏新能源开发有限公司 Wind turbine generator hydraulic station state diagnosis system and method
CN113985207A (en) * 2021-10-28 2022-01-28 国网北京市电力公司 Method, system and device for monitoring faults of power grid operation equipment and storage medium
CN114878674A (en) * 2022-04-28 2022-08-09 国电南瑞南京控制***有限公司 Transformer winding defect diagnosis method based on comprehensive characteristics of winding stress and magnetic leakage parameter of fusion algorithm

Similar Documents

Publication Publication Date Title
CN112990593A (en) Transformer fault diagnosis and state prediction method based on CSO-ANN-EL algorithm
CN108228716B (en) SMOTE _ Bagging integrated sewage treatment fault diagnosis method based on weighted extreme learning machine
CN112070128B (en) Transformer fault diagnosis method based on deep learning
CN112116058B (en) Transformer fault diagnosis method for optimizing multi-granularity cascade forest model based on particle swarm algorithm
CN108197648A (en) A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models
CN110110887A (en) To the prediction technique of low-voltage platform area line loss per unit
CN110009030B (en) Sewage treatment fault diagnosis method based on stacking meta-learning strategy
CN109492748B (en) Method for establishing medium-and-long-term load prediction model of power system based on convolutional neural network
CN108509601B (en) Food flavor assessment method based on big data analysis
CN110689069A (en) Transformer fault type diagnosis method based on semi-supervised BP network
CN110363230B (en) Stacking integrated sewage treatment fault diagnosis method based on weighted base classifier
CN108717149A (en) Diagnosis Method of Transformer Faults based on M-RVM fusion dynamic weightings AdaBoost
CN111046961B (en) Fault classification method based on bidirectional long-time and short-time memory unit and capsule network
CN112147432A (en) BiLSTM module based on attention mechanism, transformer state diagnosis method and system
CN113177587B (en) Generalized zero sample target classification method based on active learning and variational self-encoder
CN113011997A (en) Power grid user electricity utilization abnormal behavior detection method
CN110826611A (en) Stacking sewage treatment fault diagnosis method based on weighted integration of multiple meta-classifiers
CN112215279B (en) Power grid fault diagnosis method based on immune RBF neural network
CN111275204B (en) Transformer state identification method based on hybrid sampling and ensemble learning
CN114022311A (en) Comprehensive energy system data compensation method for generating countermeasure network based on time sequence condition
CN116842337A (en) Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model
CN115293400A (en) Power system load prediction method and system
CN110766215B (en) Wind power climbing event prediction method based on feature adaptive selection and WDNN
CN114330486A (en) Power system bad data identification method based on improved Wasserstein GAN
CN112836876A (en) Power distribution network line load prediction method based on deep learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210618