CN112818601A - Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis - Google Patents

Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis Download PDF

Info

Publication number
CN112818601A
CN112818601A CN202110159154.5A CN202110159154A CN112818601A CN 112818601 A CN112818601 A CN 112818601A CN 202110159154 A CN202110159154 A CN 202110159154A CN 112818601 A CN112818601 A CN 112818601A
Authority
CN
China
Prior art keywords
vibration
working condition
generating set
neural network
unit
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.)
Withdrawn
Application number
CN202110159154.5A
Other languages
Chinese (zh)
Inventor
傅质馨
曹延
朱俊澎
袁越
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.)
Hohai University HHU
Original Assignee
Hohai University HHU
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 Hohai University HHU filed Critical Hohai University HHU
Priority to CN202110159154.5A priority Critical patent/CN112818601A/en
Publication of CN112818601A publication Critical patent/CN112818601A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • 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
    • 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
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/12—Computing arrangements based on biological models using genetic models
    • G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/08—Probabilistic or stochastic CAD

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Physiology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Genetics & Genomics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis, which comprises the steps of extracting a plurality of working condition parameters with large influence on the vibration of a generating set by adopting a correlation analysis method, constructing a vibration prediction model of the generating set by adopting the BP neural network optimized by a genetic algorithm to obtain a reference value of a health model, respectively calculating upper and lower limit values of the vibration of the generating set by adopting an error statistical analysis method based on an nonparametric kernel density estimation method and a normal distribution estimation method, and combining the limit values obtained by the two methods by adopting an entropy weight method to serve as the upper and lower limit values of the health model. According to the method, the influence of a plurality of working condition parameters related to the vibration of the hydroelectric generating set is considered when the reference value of the unit health model is determined, the upper limit value and the lower limit value of the unit health model are determined, the healthy running state of the unit can be effectively reflected, and the method is used for evaluating the health state of the unit.

Description

Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis
Technical Field
The invention belongs to the field of equipment health assessment, relates to a health assessment method for a hydroelectric generating set, and particularly relates to a health assessment method for the hydroelectric generating set based on a GA-BP neural network and error statistical analysis.
Background
The hydroelectric generating set is a core device in a hydropower station, and the healthy and stable operation of the hydroelectric generating set has great significance for the safe operation of the hydropower station. The running state of the unit can be judged by carrying out fault diagnosis and health assessment on the unit, so that a corresponding maintenance scheme is formulated. At present, most of the hydroelectric generating sets in the field of fault diagnosis are diagnostic methods based on fault symptoms, but the hydroelectric generating sets are influenced by coupling of hydraulic, mechanical, electromagnetic and other factors in the operation process, the expression forms of faults are complex and various, and the mechanisms of partial faults are still under study. Meanwhile, in the actual operation process, fewer fault samples are needed, and an accurate and comprehensive fault diagnosis model is difficult to construct. Therefore, the diagnosis method cannot fully exert the diagnosis effect in the field of hydroelectric generating set fault diagnosis.
In order to solve the problem, research proposes that a health model is constructed through a health sample of the unit, a real-time operation result is compared with the health model, and when the real-time operation result is greatly deviated from a theoretical normal result, an alarm can be sent out, so that the operation state of the unit can be monitored in real time. However, in the research, when a health model is constructed, the input variables of the model usually only consider two working condition parameters of working head and active power, and influence of other working condition parameters is ignored, and when the health model is constructed, the current method for determining the upper and lower limit ranges considers influence of the active power and the working head of the unit, the two working conditions are respectively processed in a partition mode, and the limit value in each interval is obtained according to the 3 sigma criterion. However, if the influence of more than two working condition parameters is considered, it is difficult to divide specific intervals under different working conditions due to more parameters, and it is difficult to obtain the health threshold value.
Disclosure of Invention
The purpose of the invention is as follows: in order to overcome the defects in the prior art, the method for evaluating the health of the hydroelectric generating set based on the GA-BP neural network and the error statistical analysis is provided, in order to monitor the running state of the hydroelectric generating set in real time, a health model is constructed by using sample data during the healthy running period of the hydroelectric generating set, and the reference value and the upper and lower limit values of the health model are determined and used for evaluating the health state of the hydroelectric generating set.
The technical scheme is as follows: in order to achieve the purpose, the invention provides a hydroelectric generating set health assessment method based on a GA-BP neural network and error statistical analysis, which comprises the following steps:
s1: extracting a plurality of working condition parameters which have great influence on the vibration of the unit by adopting a correlation analysis method; aiming at a plurality of working condition parameters related to the unit vibration, extracting a plurality of working condition parameters most related to the unit vibration by calculating mutual information values, contribution rates and accumulated contribution rates between each working condition parameter and the unit vibration;
s2: establishing a unit vibration prediction model by adopting a BP neural network optimized by a genetic algorithm, training and testing the unit vibration prediction model by using a data sample during the normal operation of the hydroelectric generating set by adopting an initial weight and a threshold value of the BP neural network, and obtaining a reference value of the established health model, wherein the input of the health model is extracted multiple working condition parameters, and the output of the health model is a unit vibration value;
s3: solving upper and lower limit values of a unit health model by adopting an error statistical analysis method based on a nonparametric kernel density estimation method and a normal distribution estimation method; carrying out statistical analysis on relative errors between the vibration predicted values and the vibration actual values, respectively fitting probability density function curves of the relative errors by utilizing a nuclear density estimation method and a normal distribution estimation method, and respectively solving upper and lower limit values of unit vibration under a certain confidence level;
s4: and distributing the weights of the kernel density estimation method and the normal distribution estimation method by adopting an entropy weight method, and reasonably combining the upper and lower limit values of the vibration obtained by the two methods to obtain the final upper and lower limit values of the vibration, which are used as the upper and lower limit values of the unit health model.
Further, the step S1 is specifically: and extracting a plurality of working condition parameters most relevant to the vibration of the unit by calculating mutual information values between each working condition parameter and the vibration of the unit and the contribution rate and the accumulated contribution rate.
Furthermore, the vibration of the hydroelectric generating set is related to the operation condition, and the set vibration alarm value cannot accurately reflect the actual operation condition of the hydroelectric generating set under the condition that the working condition of the hydroelectric generating set is not considered. Many working condition parameters related to the vibration of the unit, such as active power, reactive power, guide vane opening, machine flow, working head, exciting current, exciting voltage, power factor and the like, firstly, vibration correlation analysis needs to be performed on the many working condition parameters according to a mutual information theory, and specifically, the method comprises the following steps:
the working condition parameters related to the vibration of the unit are numerous, vibration correlation analysis is carried out on the numerous working condition parameters according to a mutual information theory, and a mutual information theory calculation formula is shown as (1):
Figure BDA0002935617440000021
in the formula: p (X, Y) is a joint probability density function of the variables X and Y; p (X) and p (Y) are edge probability density functions of variable X and variable Y, respectively.
Further, the step S1 extracts the operating condition parameters by using the contribution rate and the cumulative contribution rate, specifically: calculating the contribution rate of each working condition parameter, sequencing the contribution rates in a descending order, calculating the cumulative contribution rate of the first m working condition parameters, and extracting the first m working condition parameters when the cumulative contribution rate reaches a set threshold value;
the contribution rate and the cumulative contribution rate are shown in equations (2) and (3):
Figure BDA0002935617440000022
m
(3)
in the formula: MIiIn order to be a mutual information value,
Figure BDA0002935617440000032
the contribution rate of the ith working condition parameter; n is a working condition parameterThe number of the numbers; phi is amThe cumulative contribution rate of the first m (m is 1,2, …, n) parameters.
Further, the process of optimizing the BP neural network by using the genetic algorithm in step S2 is as follows:
a1: defining the relevant parameters: the method mainly comprises the steps of defining the population size, the genetic algebra and the cross mutation probability, wherein the parameter coding adopts real number coding, a fitness function is the reciprocal of the error square sum of a BP neural network, and the fitness function is shown as a formula (4):
Figure BDA0002935617440000033
in the formula, n is the total number of training samples; y ishkThe real vibration value is obtained; y iskAnd the predicted value is the vibration value.
A2: initialization: encoding the initial weight and the threshold of the BP neural network, and randomly generating an initial population, namely an initial solution;
a3: and (3) fitness evaluation: calculating fitness value of the individual;
a4: genetic manipulation: adopting roulette selection, carrying out cross variation operation by a random method, and promoting population evolution and optimal solution search;
a5: and (4) cyclic judgment: and (5) terminating the operation when the iteration ending condition is reached, and outputting an optimal solution, namely an optimal initial weight and a threshold value.
The overall idea of the process is that because the initial weight and the threshold of the BP neural network are randomly selected, different convergence results of the algorithm occur due to different initial weights, and the network may fall into a local optimal solution, therefore, in the process of training the BP neural network, the optimal solution is globally searched by adopting a genetic algorithm, and the situation of obtaining the local optimal solution is effectively avoided.
And (4) training a hydroelectric generating set vibration prediction model by adopting a BP neural network optimized by a genetic algorithm, wherein the input quantity is the multiple working condition parameters determined in the step S1, and the output quantity is a set vibration value. And determining the number of the neurons of the hidden layer according to an empirical formula.
Further, the step of determining the upper and lower limit values of the health model based on the error statistical analysis method in step S3 is as follows:
b1: obtaining a deterministic vibration predicted value from the health model, and solving a relative error between the vibration predicted value and the vibration actual value, as shown in formula (5):
Figure BDA0002935617440000034
b2: fitting a probability density function curve of the vibration relative error by adopting a nonparametric kernel density estimation method, wherein the formula of the kernel density estimation is shown as the formula (6):
Figure BDA0002935617440000041
in the formula (I), the compound is shown in the specification,
Figure BDA0002935617440000042
kernel density estimation for random variables; n is the total number of samples; h is the window width; k (-) is a kernel function; deltaiAre independent and equally distributed random samples;
b3: assuming that the vibration relative error distribution obeys normal distribution, the mean value is mu and the standard deviation is sigma, and the method is expressed as formula (7), and a probability density function curve of the vibration relative error is fitted by adopting a normal distribution estimation method:
ε～N(μ,σ2) (7)
b4: respectively calculating confidence intervals, wherein the confidence intervals calculated by a nonparametric nuclear density estimation method and a normal distribution estimation method are shown as a formula (8) and a formula (9):
[yi+F-1(α1),yi+F-1(α2)] (8)
[yi-z1-α/2σ,yi+z1-α/2σ] (9)
in the formula, yiThe predicted value of the vibration is taken as the predicted value; f-1(. is) an inverse function of the probability distribution function, where α1＝α/2；α21- α/2; 1-alpha is confidence level, 1-alpha ═ alpha2-α1(0＜α＜1)；z1-α/2For coefficients, the values were obtained by looking up a standard normal distribution table, z at 95% confidence level1-α/21.96; σ is the standard deviation of the prediction error.
Further, the method for determining the upper and lower limit values of the unit health model by using the entropy weight method in step S4 includes:
because the nuclear density estimation method and the normal distribution estimation method have advantages and disadvantages during interval estimation, the two methods are combined when the upper interval and the lower interval of the vibration of the hydroelectric generating set are solved, complementation can be formed, and the reliability and the accuracy of a result are improved. Distribution of weights omega of two methods by entropy weight method1And ω2Obtaining the optimal confidence interval of [ Lcomb,Ucomb]As shown in equation (10):
Figure BDA0002935617440000043
in the formula, LcombObtaining a confidence interval lower limit for the entropy weight method; u shapecombSolving the upper limit of the confidence interval for the entropy weight method; omega1The weight is a nonparametric weight; l is1Obtaining a confidence interval lower limit for a nonparametric method; omega2Is a parametric method weight; l is2Obtaining a confidence interval lower limit for a parameter method; u shape1Solving the upper limit of the confidence interval for a nonparametric method; u shape2Obtaining the upper limit of the confidence interval for the parameter method;
will UcombAnd LcombAnd the parameters are used as the upper and lower limit values of the unit health model.
According to the method, a hydroelectric generating set health evaluation model is constructed based on a GA-BP neural network and error statistical analysis according to data samples during the healthy operation of the hydroelectric generating set, the influence of a plurality of working condition parameters on the vibration of the hydroelectric generating set is considered, the upper limit value and the lower limit value of the health model are determined, the real-time working condition parameters are input into the model in the subsequent overhaul process of the hydroelectric generating set, and the degree of the deviation of the output vibration value of the model from the normal range is judged, so that whether the hydroelectric generating set is in an abnormal operation state or not can be judged.
The invention is innovative in that a plurality of working condition parameters influencing the unit vibration are extracted, the unit vibration is related to the operation working condition, and the actual operation condition of the unit cannot be accurately reflected by the vibration alarm value set without considering the working condition conditions of the unit, so a plurality of working condition parameters most related to the unit vibration are determined by a correlation analysis method. The innovation of the method is that the upper and lower limit values of the health model are determined, the traditional research only considering two working condition parameters of active power and working head can be partitioned according to the working conditions, the limit value is obtained according to the law of large numbers and the 3 sigma criterion, and the model constructed through the neural network cannot obtain the upper and lower limit values through the method because of more input parameters. And (3) fitting a probability density function curve of the vibration prediction relative error through statistical analysis on the vibration prediction relative error, so as to obtain an upper limit value and a lower limit value under a certain confidence interval, wherein the upper limit value and the lower limit value are used as the upper limit value and the lower limit value of the health model.
Has the advantages that: compared with the prior art, the method considers the influence of a plurality of working condition parameters on the vibration of the unit, extracts the most relevant working condition parameters from a plurality of working condition parameters through a correlation analysis method, reasonably combines confidence intervals obtained by a nonparametric kernel density estimation method and a normal distribution estimation method by adopting an entropy weight method to serve as the upper and lower limit values of a unit health model, and can reflect the health condition of the unit because training data used by a vibration prediction model contains all the operating working conditions of the unit, input real-time data into the health model, and judge the range deviating from the health interval to realize the real-time health state evaluation of the unit.
Drawings
FIG. 1 is a flow chart of a health assessment method of the present invention;
FIG. 2 shows the vibration prediction results of the BP model and the GA-BP model according to the simulation analysis of the embodiment of the present invention;
FIG. 3 is a diagram illustrating kernel density estimation under optimal window width for simulation analysis according to an embodiment of the present invention;
FIG. 4 is a diagram illustrating upper and lower limit values of unit vibration obtained by a kernel density estimation method, a normal distribution estimation method, and an entropy weight method combined method of simulation analysis according to an embodiment of the present invention;
fig. 5 is a distribution of the real value, the reference value and the limit value of the vibration according to the simulation analysis in the embodiment of the present invention.
Detailed Description
The present invention is further illustrated by the following figures and specific examples, which are to be understood as illustrative only and not as limiting the scope of the invention, which is to be given the full breadth of the appended claims and any and all equivalent modifications thereof which may occur to those skilled in the art upon reading the present specification.
The invention provides a hydroelectric generating set health assessment method based on a GA-BP neural network and error statistical analysis, and the overall design principle, the method flow and the simulation analysis of the scheme are respectively explained below.
1. Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis
The method comprises the steps of firstly extracting multiple working condition parameters which have large influence on unit vibration by adopting a correlation analysis method, secondly training and testing a prediction model of the unit vibration by utilizing a data sample during the normal operation of the unit by adopting a BP neural network optimized by a genetic algorithm to obtain a reference value of a constructed health model, then solving the upper and lower limit values of the unit health model by adopting an error statistical analysis method based on a nonparametric kernel density estimation method and a normal distribution estimation method, and reasonably combining the upper and lower limit values by an entropy weight method to obtain the upper and lower limit values of a final health model. Finally, simulation analysis proves that the provided health model can effectively reflect the healthy running state of the unit and is used for evaluating the health state of the unit.
1.1 determination of reference values for health models
The method mainly comprises the steps of constructing a BP neural network optimized by a genetic algorithm by taking a data sample of the hydroelectric generating set during the healthy operation as a vibration prediction model for training and testing, wherein the input quantity of the model is a plurality of working condition parameters most relevant to the vibration of the generating set, the output quantity of the model is a vibration value of the generating set, and after the model training is finished, the output vibration prediction value is taken as a reference value of the healthy model.
1.2 determination of upper and lower limit values of health model
And carrying out statistical analysis on the relative prediction error of the vibration predicted value and the vibration true value, and fitting a probability density function curve by adopting a kernel density estimation method and a normal distribution estimation method. Under a certain confidence level, respectively obtaining vibration estimation intervals under the two methods, then reasonably distributing the weights of the two methods through an entropy weight method, and reasonably combining the vibration estimation intervals obtained under the two methods to serve as upper and lower limit values of the health model.
2. Method flow
Fig. 1 shows a flow chart of the construction of a health model of a hydroelectric generating set.
Referring to fig. 1, the flow of the method for evaluating the health of the hydroelectric generating set is as follows:
s1: extracting a plurality of working condition parameters which have large influence on the vibration of the unit by adopting a correlation analysis method, and extracting a plurality of working condition parameters most relevant to the vibration of the unit by calculating mutual information values, contribution rates and accumulated contribution rates between each working condition parameter and the vibration of the unit aiming at the plurality of working condition parameters relevant to the vibration of the unit;
s2: optimizing initial weight and threshold of a BP neural network by adopting a genetic algorithm, training and testing a prediction model of unit vibration by using a data sample during normal operation of the hydroelectric generating set so as to obtain a reference value of a constructed health model, wherein the input of the model is extracted multiple working condition parameters, and the output of the model is a unit vibration value;
s3: and solving the upper and lower limit values of the unit health model by adopting an error statistical analysis method based on a nonparametric kernel density estimation method and a normal distribution estimation method. Carrying out statistical analysis on relative errors between the vibration predicted values and the vibration actual values, respectively fitting probability density function curves of the relative errors by utilizing a nuclear density estimation method and a normal distribution estimation method, and respectively solving upper and lower limit values of unit vibration under a certain confidence level;
s4: and (3) reasonably distributing the weights of the kernel density estimation method and the normal distribution estimation method by adopting an entropy weight method, and reasonably combining the upper and lower limit values of the vibration obtained by the two methods to obtain the final upper and lower limit values of the vibration, which are used as the upper and lower limit values of the unit health model.
In this embodiment, the mutual information theory is adopted as the correlation analysis method in step S1:
the vibration of the hydroelectric generating set is related to the operation condition, and the set vibration alarm value cannot accurately reflect the actual operation condition of the hydroelectric generating set under the condition that the working condition of the hydroelectric generating set is not considered. The working condition parameters related to the vibration of the unit are numerous, such as active power, reactive power, guide vane opening, machine passing flow, working water head, exciting current, exciting voltage, power factor and the like, firstly, vibration correlation analysis is needed to be carried out on the numerous working condition parameters according to a mutual information theory, and a mutual information theory calculation formula is shown as (1):
Figure BDA0002935617440000071
in the formula: p (X, Y) is a joint probability density function of the variables X and Y; p (X) and p (Y) are edge probability density functions of variable X and variable Y, respectively.
After obtaining the mutual information value between each working condition parameter and the vibration value, extracting the most relevant working condition parameter to the unit vibration through the contribution rate and the accumulated contribution rate, specifically: calculating the contribution rate of each working condition parameter, sequencing the contribution rates in a descending order, calculating the cumulative contribution rate of the first m working condition parameters, and extracting the first m working condition parameters when the cumulative contribution rate reaches a set threshold value;
the calculation formula of the contribution rate is shown as formula (2):
Figure BDA0002935617440000072
the cumulative contribution ratio calculation formula is shown in formula (3):
Figure BDA0002935617440000073
in the formula: MIiIn order to be a mutual information value,
Figure BDA0002935617440000074
the contribution rate of the ith working condition parameter; n isThe number of working condition parameters; phi is amThe cumulative contribution rate of the first m (m is 1,2, …, n) parameters.
Calculating mutual information value MI of each working condition parameter and unit vibrationiAnd then, calculating the contribution rate and the accumulated contribution rate according to a formula (2) and a formula (3), sequencing the contribution rates of the obtained working condition parameters in a descending order, calculating the accumulated contribution rate phi of the first m working condition parameters, and if the phi is greater than or equal to 0.8, considering the m working condition parameters as the most relevant working condition parameters.
In this embodiment, in step S2, a unit vibration prediction model is constructed by using a BP neural network optimized by a genetic algorithm, and the process of optimizing by using the genetic algorithm is as follows:
(1) defining the relevant parameters: the method mainly comprises the steps of defining the population size, the genetic algebra and the cross mutation probability, wherein the parameter coding adopts real number coding, a fitness function is the reciprocal of the error square sum of a BP neural network, and the fitness function is shown as a formula (4):
Figure BDA0002935617440000075
in the formula, n is the total number of training samples; y ishkThe real vibration value is obtained; y iskAnd the predicted value is the vibration value.
(2) Initialization: and coding the initial weight and the threshold of the BP neural network, and randomly generating an initial population, namely an initial solution.
(3) And (3) fitness evaluation: an individual fitness value is calculated.
(4) Genetic manipulation: and (4) adopting roulette selection, and performing cross variation operation by a random method to promote population evolution and optimal solution search.
(5) And (4) cyclic judgment: and (5) terminating the operation when the iteration ending condition is reached, and outputting an optimal solution, namely an optimal initial weight and a threshold value.
The overall idea of the process is that because the initial weight and the threshold of the BP neural network are randomly selected, different convergence results of the algorithm occur due to different initial weights, and the network may fall into a local optimal solution, therefore, in the process of training the BP neural network, the optimal solution is globally searched by adopting a genetic algorithm, and the situation of obtaining the local optimal solution is effectively avoided.
And (4) training a hydroelectric generating set vibration prediction model by adopting a BP neural network optimized by a genetic algorithm, wherein the input quantity is the multiple working condition parameters determined in the step S1, and the output quantity is a set vibration value. And determining the number of the neurons of the hidden layer according to an empirical formula.
In step S3 of this embodiment, the steps of determining the upper and lower limit values of the health model based on the error statistical analysis method are as follows:
(1) obtaining a deterministic vibration predicted value from the health model, and solving a relative error between the vibration predicted value and the vibration actual value, as shown in formula (5):
Figure BDA0002935617440000081
in the formula, y is a vibration predicted value;
Figure BDA0002935617440000082
the vibration is the real value.
(2) Fitting a probability density function curve of the vibration relative error by adopting a nonparametric kernel density estimation method, wherein the formula of the kernel density estimation is shown as the formula (6):
Figure BDA0002935617440000083
in the formula (I), the compound is shown in the specification,
Figure BDA0002935617440000084
kernel density estimation for random variables; n is the total number of samples; h is the window width; k (-) is a kernel function; deltaiAre independent and equally distributed random samples.
(3) Assuming that the vibration relative error distribution obeys normal distribution, the mean value is mu and the standard deviation is sigma, and the method is expressed as formula (7), and a probability density function curve of the vibration relative error is fitted by adopting a normal distribution estimation method:
ε～N(μ,σ2) (7)
(4) confidence intervals are respectively obtained, and the confidence intervals obtained by a nonparametric nuclear density estimation method and a normal distribution estimation method are shown as a formula (8) and a formula (9).
[yi+F-1(α1),yi+F-1(α2)] (8)
[yi-z1-α/2σ,yi+z1-α/2σ] (9)
In the formula, yiThe predicted value of the vibration is taken as the predicted value; f-1(. is) an inverse function of the probability distribution function, where α1＝α/2；α21- α/2; 1-alpha is confidence level, 1-alpha ═ alpha2-α1(0＜α＜1)；z1-α/2For coefficients, the values were obtained by looking up a standard normal distribution table, z at 95% confidence level1-α/21.96; σ is the standard deviation of the prediction error.
In step S4 of this embodiment, the method for determining the upper and lower limit values of the unit health model by using the entropy weight method includes:
because the nuclear density estimation method and the normal distribution estimation method have advantages and disadvantages during interval estimation, the two methods are combined when the upper interval and the lower interval of the vibration of the hydroelectric generating set are solved, complementation can be formed, and the reliability and the accuracy of a result are improved. The weights of the two methods are reasonably distributed by using an entropy weight method to obtain the optimal confidence interval [ L [ ]comb,Ucomb]As shown in equation (10):
Figure BDA0002935617440000091
in the formula, LcombObtaining a confidence interval lower limit for the entropy weight method; u shapecombSolving the upper limit of the confidence interval for the entropy weight method; omega1The weight is a nonparametric weight; l is1Obtaining a confidence interval lower limit for a nonparametric method; omega2Is a parametric method weight; l is2Obtaining a confidence interval lower limit for a parameter method; u shape1Solving the upper limit of the confidence interval for a nonparametric method; u shape2The upper limit of the confidence interval is determined for the parametric method.
In the invention, indexes such as Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) are adopted to evaluate the performance of the BP neural network after genetic algorithm optimization, and the calculation formulas of MAE, MAPE and RMSE are shown as (11), (12) and (13):
Figure BDA0002935617440000092
Figure BDA0002935617440000093
Figure BDA0002935617440000094
in the formula, yiA vibration prediction value output for the ith model;
Figure BDA0002935617440000095
the ith vibration real value is obtained.
And evaluating the accuracy of the confidence interval by using indexes such as Prediction Interval Coverage (PICP), prediction interval width (PINAW) and the like. The calculation formulas of PICP and PINAW are shown in (11) and (12):
Figure BDA0002935617440000096
Figure BDA0002935617440000101
wherein n is the number of prediction samples,
Figure BDA0002935617440000104
when the predicted target value is within the confidence interval, n is 1, otherwise, n is 0; u shapeiUpper limit of confidence interval, LiIs the lower confidence interval limit.
3. Simulation analysis
For explaining the effect of the invention, the following takes a certain hydropower station 1# unit in China as a research object, and takes the following guide bearing x-direction swing as an example, and monitoring data of the unit in a healthy operation state are obtained.
3.1 extraction of Multi-Condition parameters of hydroelectric generating set
The data sample adopted in the invention contains eight working condition parameters of unit active power, reactive power, guide vane opening, machine flow, working water head, exciting current, exciting voltage, power factor and the like and the x-direction swing value of the lower guide bearing. Table 1 shows the mutual information between the operating condition parameters and the x-direction throw of the lower guide bearing.
Figure BDA0002935617440000102
The contribution rate and the cumulative contribution rate of each condition parameter are calculated and arranged in descending order of the contribution rate, as shown in table 2.
Figure BDA0002935617440000103
As can be seen from table 2, after the accumulated contribution rates of the operating condition parameters are sorted in a descending order, the accumulated contribution rate of the exciting current exceeds 0.8, so that six main operating condition parameters affecting the x-direction throw of the lower guide bearing are respectively a working water head, active power, guide vane opening, reactive power, machine passing flow and exciting current.
3.2 establishment of vibration prediction model of hydroelectric generating set
3000 pieces of data are selected from the unit health operation data samples to serve as health samples for training and testing of a unit vibration prediction model. The 3000 selected data comprise different stable working conditions of the unit, different operation working conditions of the unit can be reflected, and before training, the data of the sample set are uniformly normalized to be between 0 and 1. And (3) adopting the test set sample for verifying the model, respectively inputting the test set sample into the BP neural network and the GA-BP neural network, and performing inverse normalization on the output result to obtain a vibration prediction value. The comparison between the output value and the actual value of the x-direction throw of the lower guide bearing of the hydro-electric machine set in the two methods is shown in figure 2. The MAE, MAPE and RMSE results for the BP network and GA-BP neural network predictions are shown in Table 3.
Figure BDA0002935617440000111
3.3 statistical analysis of vibration prediction errors of hydroelectric generating set
The 1000 groups of data are divided into two groups, the first 900 groups of data are used for statistical analysis of vibration prediction errors, and the last 100 groups of data are used for verifying the result of interval estimation. And obtaining the vibration predicted value of 900 groups of data through a GA-BP model, and calculating the relative error between the vibration predicted value and the true value. And respectively fitting a probability density function curve of the relative error by adopting a nonparametric kernel density estimation method and a normal distribution estimation method. In the nonparametric method, the kernel function is a Gaussian kernel, and the optimal window width is h-0.947752. The probability density function image obtained by fitting with the nonparametric method and the parametric method is shown in fig. 3.
Confidence intervals were determined for each of the two methods, where α was 0.05 at a given significance level and 95% confidence, at which time α was obtained1And alpha2Respectively 0.025 and 0.975, calculating the confidence interval of the test set sample under the two methods, and respectively connecting the upper limit and the lower limit of each point of the test data sample to obtain an interval estimation result under the 95% confidence level. The weight of the two methods is calculated by an entropy weight method, and the optimal interval estimation can be obtained. The weight of the nonparametric method and the weight of the parametric method are respectively omega obtained by the entropy weight method1＝0.4686，ω20.5314, fig. 4 shows the estimation interval after the combination is obtained according to the weight. PICP and PINAW values for the nonparametric, parametric, and combinatorial methods are shown in table 4.
Figure BDA0002935617440000112
3.4 hydroelectric generating set health assessment verification
And taking the health prediction value obtained by the GA-BP vibration prediction model as a reference value of the hydroelectric generating set health model, and taking an estimation interval obtained by combining a non-parametric method and a parametric method as a limit value of the health model. Fig. 5 shows the distribution of the real vibration value, the reference value and the limit value of 100 groups of data, and it can be seen from the graph that the real vibration value fluctuates around the reference value and most of the data are within the upper and lower limits, which illustrates that during the healthy operation of the unit, the constructed health model can reflect the healthy operation condition of the unit.

Claims (7)

1. A hydroelectric generating set health assessment method based on a GA-BP neural network and error statistical analysis is characterized by comprising the following steps:
s1: extracting a plurality of working condition parameters which have great influence on the vibration of the unit by adopting a correlation analysis method; extracting a plurality of working condition parameters most relevant to the vibration of the unit by calculating mutual information values between each working condition parameter and the vibration of the unit and contribution rates and accumulated contribution rates;
s2: establishing a unit vibration prediction model by adopting a BP neural network optimized by a genetic algorithm, training and testing the unit vibration prediction model by using a data sample during the normal operation of the hydroelectric generating set by adopting an initial weight and a threshold of the BP neural network, taking a vibration prediction value output by the unit vibration prediction model as a reference value of a health model, inputting the health model into extracted multiple working condition parameters, and outputting the parameters as a unit vibration value;
s3: carrying out statistical analysis on relative errors between the vibration predicted values and the vibration actual values, respectively fitting probability density function curves of the relative errors by utilizing a nuclear density estimation method and a normal distribution estimation method, and respectively solving upper and lower limit values of unit vibration under a confidence level;
s4: and distributing the weights of the kernel density estimation method and the normal distribution estimation method by adopting an entropy weight method, and reasonably combining the upper and lower limit values of the vibration obtained by the two methods to obtain the final upper and lower limit values of the vibration, which are used as the upper and lower limit values of the unit health model.
2. The hydroelectric generating set health assessment method based on the GA-BP neural network and the error statistical analysis according to claim 1, wherein the step S1 specifically comprises: and extracting a plurality of working condition parameters most relevant to the vibration of the unit by calculating mutual information values between each working condition parameter and the vibration of the unit and the contribution rate and the accumulated contribution rate.
3. The hydroelectric generating set health assessment method based on the GA-BP neural network and the error statistical analysis as claimed in claim 2, wherein the step S1 adopts a mutual information theory as a correlation analysis method, specifically:
vibration correlation analysis is carried out on a plurality of working condition parameters according to a mutual information theory, and a mutual information theory calculation formula is shown as (1):
Figure FDA0002935617430000011
in the formula: p (X, Y) is a joint probability density function of the variables X and Y; p (X) and p (Y) are edge probability density functions of variable X and variable Y, respectively.
4. The hydroelectric generating set health assessment method based on the GA-BP neural network and the error statistical analysis according to claim 3, wherein the step S1 is to extract working condition parameters by using the contribution rate and the accumulated contribution rate, and specifically comprises the following steps: calculating the contribution rate of each working condition parameter, sequencing the contribution rates in a descending order, calculating the cumulative contribution rate of the first m working condition parameters, and extracting the first m working condition parameters when the cumulative contribution rate reaches a set threshold value;
the contribution rate and the cumulative contribution rate are shown in equations (2) and (3):
Figure FDA0002935617430000021
Figure FDA0002935617430000022
in the formula: MIiAs mutual informationThe value of the one or more of,
Figure FDA0002935617430000023
the contribution rate of the ith working condition parameter; n is the number of working condition parameters; phi is amThe cumulative contribution rate of the first m (m is 1,2, …, n) parameters.
5. The hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis according to claim 1, wherein the process of optimizing BP neural network by using genetic algorithm in step S2 is as follows:
a1: defining the relevant parameters: the method comprises the following steps of defining the population size, genetic algebra and cross mutation probability, wherein parameter coding adopts real number coding, a fitness function is the reciprocal of the sum of squares of errors of a BP neural network, and the fitness function is shown in a formula (4):
Figure FDA0002935617430000024
in the formula, n is the total number of training samples; y ishkThe real vibration value is obtained; y iskAnd the predicted value is the vibration value.
A2: initialization: encoding the initial weight and the threshold of the BP neural network, and randomly generating an initial population, namely an initial solution;
a3: and (3) fitness evaluation: calculating fitness value of the individual;
a4: genetic manipulation: adopting roulette selection, carrying out cross variation operation by a random method, and promoting population evolution and optimal solution search;
a5: and (4) cyclic judgment: and (5) terminating the operation when the iteration ending condition is reached, and outputting an optimal solution, namely an optimal initial weight and a threshold value.
6. The hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis of claim 1, wherein the step of determining the upper and lower limit values of the health model based on the error statistical analysis method in step S3 is as follows:
b1: obtaining a deterministic vibration predicted value from the health model, and solving a relative error between the vibration predicted value and the vibration actual value, as shown in formula (5):
Figure FDA0002935617430000025
b2: fitting a probability density function curve of the vibration relative error by adopting a nonparametric kernel density estimation method, wherein the formula of the kernel density estimation is shown as the formula (6):
Figure FDA0002935617430000031
in the formula (I), the compound is shown in the specification,
Figure FDA0002935617430000032
kernel density estimation for random variables; n is the total number of samples; h is the window width; k (-) is a kernel function; deltaiAre independent and equally distributed random samples;
b3: assuming that the vibration relative error distribution obeys normal distribution, the mean value is mu and the standard deviation is sigma, and the method is expressed as formula (7), and a probability density function curve of the vibration relative error is fitted by adopting a normal distribution estimation method:
ε～N(μ,σ2) (7)
b4: respectively calculating confidence intervals, wherein the confidence intervals calculated by a nonparametric nuclear density estimation method and a normal distribution estimation method are shown as a formula (8) and a formula (9):
[yi+F-1(α1),yi+F-1(α2)] (8)
[yi-z1-α/2σ,yi+z1-α/2σ] (9)
in the formula, yiThe predicted value of the vibration is taken as the predicted value; f-1(. is) an inverse function of the probability distribution function, where α1＝α/2；α21- α/2; 1-alpha is confidence level, 1-alpha ═ alpha2-α1(0＜α＜1)；z1-α/2Is a coefficient ofInquiring a standard normal distribution table to obtain; σ is the standard deviation of the prediction error.
7. The hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis according to claim 1, wherein the method for determining the upper and lower limit values of the generating set health model by using the entropy weight method in step S4 comprises:
distribution of weights omega of two methods by entropy weight method1And ω2Obtaining the optimal confidence interval of [ Lcomb,Ucomb]As shown in equation (10):
Figure FDA0002935617430000033
in the formula, LcombObtaining a confidence interval lower limit for the entropy weight method; u shapecombSolving the upper limit of the confidence interval for the entropy weight method; omega1The weight is a nonparametric weight; l is1Obtaining a confidence interval lower limit for a nonparametric method; omega2Is a parametric method weight; l is2Obtaining a confidence interval lower limit for a parameter method; u shape1Solving the upper limit of the confidence interval for a nonparametric method; u shape2Obtaining the upper limit of the confidence interval for the parameter method;
will UcombAnd LcombAnd the parameters are used as the upper and lower limit values of the unit health model.
CN202110159154.5A 2021-02-05 2021-02-05 Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis Withdrawn CN112818601A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110159154.5A CN112818601A (en) 2021-02-05 2021-02-05 Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110159154.5A CN112818601A (en) 2021-02-05 2021-02-05 Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis

Publications (1)

Publication Number Publication Date
CN112818601A true CN112818601A (en) 2021-05-18

Family

ID=75861672

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110159154.5A Withdrawn CN112818601A (en) 2021-02-05 2021-02-05 Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis

Country Status (1)

Country Link
CN (1) CN112818601A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113642251A (en) * 2021-08-31 2021-11-12 佛山众陶联供应链服务有限公司 Data analysis and prediction method and system for powder making quality of architectural ceramic spray
CN114324979A (en) * 2022-01-06 2022-04-12 北京轩宇空间科技有限公司 Accelerometer vibration detuning error evaluation method and device, storage medium and equipment
CN115034485A (en) * 2022-06-16 2022-09-09 华北电力大学 Wind power interval prediction method and device based on data space
CN118013468A (en) * 2024-04-03 2024-05-10 华电电力科学研究院有限公司 Method, device, equipment and medium for monitoring health degree of wind turbine generator component

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113642251A (en) * 2021-08-31 2021-11-12 佛山众陶联供应链服务有限公司 Data analysis and prediction method and system for powder making quality of architectural ceramic spray
CN113642251B (en) * 2021-08-31 2024-05-28 佛山众陶联供应链服务有限公司 Data analysis and prediction method and system for building ceramic spray powder preparation quality
CN114324979A (en) * 2022-01-06 2022-04-12 北京轩宇空间科技有限公司 Accelerometer vibration detuning error evaluation method and device, storage medium and equipment
CN114324979B (en) * 2022-01-06 2024-04-02 北京轩宇空间科技有限公司 Accelerometer vibration imbalance error assessment method, accelerometer vibration imbalance error assessment device, storage medium and accelerometer vibration imbalance error assessment device
CN115034485A (en) * 2022-06-16 2022-09-09 华北电力大学 Wind power interval prediction method and device based on data space
CN118013468A (en) * 2024-04-03 2024-05-10 华电电力科学研究院有限公司 Method, device, equipment and medium for monitoring health degree of wind turbine generator component
CN118013468B (en) * 2024-04-03 2024-06-11 华电电力科学研究院有限公司 Method, device, equipment and medium for monitoring health degree of wind turbine generator component

Similar Documents

Publication Publication Date Title
CN112818601A (en) Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis
Cui et al. Deep learning-based time-varying parameter identification for system-wide load modeling
CN105973594B (en) A kind of rolling bearing fault Forecasting Methodology based on continuous depth confidence network
Liao et al. Voltage sag estimation in sparsely monitored power systems based on deep learning and system area mapping
US6606580B1 (en) Fault diagnosis
CN109459671A (en) A kind of switch cabinet state monitoring method of the improvement neural network based on genetic algorithm
CN113805064B (en) Lithium ion battery pack health state prediction method based on deep learning
CN111709490A (en) Fan health state assessment method based on GRU neural network
CN110059845B (en) Metering device clock error trend prediction method based on time sequence evolution gene model
CN109992895B (en) Method for extracting and predicting equipment performance degradation trend
CN108717579B (en) Short-term wind power interval prediction method
CN113298297A (en) Wind power output power prediction method based on isolated forest and WGAN network
CN107993012A (en) A kind of adaptive electric system on-line transient stability appraisal procedure of time
CN114118789B (en) Radar transmitter state evaluation method based on fuzzy comprehensive judgment and comprehensive weighting
CN112746934B (en) Method for diagnosing fan fault through self-association neural network
CN109710661A (en) Based on Global Genetic Simulated Annealing Algorithm to the method for high-pressure heater state analysis
CN112834927A (en) Lithium battery residual life prediction method, system, device and medium
CN112288147B (en) Method for predicting insulation state of generator stator by BP-Adaboost strong predictor
CN112485622A (en) Partial discharge pattern recognition method based on GA-BPNN
CN110705859A (en) PCA-self-organizing neural network-based method for evaluating running state of medium and low voltage distribution network
CN109494716A (en) Wind power output power confidence interval prediction technique based on Bootstrap
CN114611372A (en) Industrial equipment health prediction method based on Internet of things edge calculation
CN113379116A (en) Cluster and convolutional neural network-based line loss prediction method for transformer area
Hu et al. A hybrid model for predicting the degradation trend of hydropower units based on deep learning
CN115856623A (en) Motor fault diagnosis method based on uniformity and kurtosis calculation

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
WW01 Invention patent application withdrawn after publication

Application publication date: 20210518

WW01 Invention patent application withdrawn after publication