CN112084709A - Insulation Condition Evaluation Method of Large Generator Based on Genetic Algorithm and Radial Basis Neural Network - Google Patents

Abstract

The invention discloses a large-scale generator insulation state evaluation method based on genetic algorithm and radial basis neural network, comprising the following steps: 1) measuring parameters related to the insulation aging state of large-scale steam turbine generator stator bars; Step 1) The measured parameters are classified and screened, and then the parameters after classification and screening are used to construct a data set; 3) RBF neural network is established; Optimize the center and width; 5) Use the data set to train the optimized RBF neural network, and then use the negative gradient descent method to update the iterative weights; 6) Use the RBF neural network obtained in step 5) to evaluate the insulation of large generators state, this method can more accurately evaluate the insulation state of large generators.

Description

Insulation Condition Evaluation of Large Generator Based on Genetic Algorithm and Radial Basis Neural Network method

technical field

The invention relates to a large-scale generator insulation state evaluation method, in particular to a large-scale generator insulation state evaluation method based on a genetic algorithm and a radial basis neural network.

Background technique

Large-scale steam turbine generator is one of the key equipments in the power system. Its operation reliability is related to the operation stability of the power grid. It has always been highly valued by people. One of the threats to its safe operation mainly comes from the insulation system. During the operation of the motor, the stator winding is subject to the combined action of electrical, thermal, mechanical, chemical and other factors, and the insulation performance is gradually deteriorated. In the case of serious insulation aging, it will cause insulation failure of the generator. Because the residual breakdown voltage is affected by many factors, traditionally, the accuracy of predicting the residual breakdown voltage with a small number of parameters is low. It can not meet the needs of state prediction very well.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the above shortcomings of the prior art, and to provide a large-scale generator insulation state evaluation method based on genetic algorithm and radial basis neural network, which can more accurately evaluate the insulation state of large-scale generators.

In order to achieve the above object, the method for evaluating the insulation state of a large generator based on a genetic algorithm and a radial basis neural network according to the present invention includes the following steps:

1) Measure the parameters related to the insulation aging state of the large-scale turbine generator stator bar;

2) classifying and screening the parameters obtained in step 1), and then using the parameters after classification and screening to construct a data set;

3) Establish RBF neural network;

4) Using genetic algorithm to optimize the number of hidden layers, the center and width of radial basis function in RBF neural network;

5) Use the data set to train the optimized RBF neural network, and then use the negative gradient descent method to update the iterative weights;

6) Use the RBF neural network obtained in step 5) to evaluate the insulation state of the large generator.

The parameters related to the insulation aging state of large-scale steam turbine generator stator bars include insulation resistance R, polarization index PI, absorption ratio DAR, dielectric loss tanδ, dielectric loss increment Δtanδ, capacitance C and capacitance increase ΔC.

Updating iterative weights using negative gradient descent includes the following steps:

Negative gradient descent method is used to update the weights of input layer, hidden layer and output layer in iterative RBF neural network.

In step 3), an RBF neural network is constructed using the parameters classified and screened in step 2).

The specific operation process of step 4) is:

4a) Normalize the parameters;

4b) carrying out a correlation analysis on the normalized parameters to eliminate the parameters whose correlation is lower than the preset value;

4c) Generating an initial population by genetic algorithm;

计算，其中，c_i和c_j分别为第i个和第j个节点的中心向量；4d) Using the population parameters, set k=N in the Kmeans algorithm, classify the parameters, and obtain the classification center according to the classification result, that is, the center c of the hidden layer radial basis function, and the width σ of the hidden layer radial basis function according to

Calculate, where c _i and c _j are the center vectors of the i-th and j-th nodes, respectively;

4e) Use the number of neurons in the hidden layer N, the center c and the width σ of the radial basis function to build an RBF training model, train the data set, and update the weights using the negative gradient descent method until the accuracy meets the requirements or the number of iterations reaches Up to the maximum value, the trained RBF neural network is obtained;

4f) Use the trained RBF neural network to predict the remaining breakdown voltage, and then calculate the fitness of each population;

4g) Judging whether the preset termination conditions are met, when the preset termination conditions are met, then go to step 4h); otherwise, the population is selected and cross-mutated according to each degree of adaptation, and then go to step 4e);

4h) According to the training result, select the individual with the best degree of self-adaptation in the population as the number of hidden layers N of the RBF neural network, and then train the data set to obtain the parameters of the RBF neural network improved by the genetic algorithm.

The present invention has the following beneficial effects:

The method for evaluating the insulation state of a large generator based on the genetic algorithm and the radial basis neural network of the present invention uses the genetic algorithm to evaluate the number of hidden layers, the center and the width of the radial basis function in the RBF neural network. Optimization, to solve the problem that it is difficult to determine the optimal number of hidden layers in the prediction of RBF neural network to improve the accuracy of prediction, then train the optimized RBF neural network, and then use the trained RBF neural network to evaluate the insulation of large generators The accuracy of the evaluation is relatively high, which effectively solves the inaccurate and incomplete problem of traditionally relying on the single-variable aging factor to predict the motor insulation state.

Description of drawings

Fig. 1 is the flow chart of the present invention;

FIG. 2 is a system diagram of the present invention.

Detailed ways

Below in conjunction with accompanying drawing, the present invention is described in further detail:

The invention aims at the evaluation and prediction of the insulation aging state of the stator of a large generator, and solves the inaccurate and incomplete problems of traditionally relying on the single-variable aging factor to predict the insulation state of the motor. At the same time, genetic algorithm is used to solve the problem that it is difficult to determine the optimal number of hidden layers in the prediction of RBF neural network, so as to improve the accuracy of prediction. At the same time, by establishing an experimental platform, a measuring device is selected to measure the required parameters. Then use the measured data to train and predict the RBF neural network. The specific process is as follows:

Radial Basis Neural Network

The radial basis neural network RBF is a three-layer forward network with a single hidden layer. The first layer is the input layer, which is composed of signal source nodes; the second layer is the hidden layer, and the transformation function of the neurons in the hidden layer is the radial basis. The function is a non-negative linear function that is radially symmetric and attenuated to the center point, and the radial basis function is a local response function; the third layer is the output layer, which is the response to the input mode; the input layer only plays the role of transmitting signals, The input layer and the hidden layer can be regarded as a connection layer with a connection weight of 1. The tasks performed by the output layer and the hidden layer are different, so their learning strategies are also different; the output layer adjusts the linear weights , adopts a linear optimization strategy, so the learning speed is faster; while the hidden layer adjusts the parameters of the activation function (Green function, Gaussian function, generally the latter), and adopts a nonlinear optimization strategy, so the learning speed is faster. slow.

The purpose of training is to find the final weights Cj, Dj and Wj of the two layers.

The training process is divided into two steps: the first step is unsupervised learning, the training determines the weights Cj and Dj between the input layer and the hidden layer; the second step is supervised learning, and the training determines the difference between the hidden layer and the output layer. The weight Wj of .

The input vector X, the corresponding target output vector Y and the width vector Dj of the radial basis function are provided before training.

When inputting the training set data for training, the expression and calculation method of each parameter are as follows:

Input vector X: X=[x ₁ , x ₂ ,..., x _n ] ^T , where n is the number of neurons in the input layer, that is, the number of types of parameters in the input training set.

Output vector Y: Y=[y ₁ , y _{2 ,} . . . , y _q ] ^T , where q is the number of neurons in the output layer, that is, the expected value.

Initialize the weights from the hidden layer to the output layer: W _K = [w _k1 , w _k2 ,..., w _kp ] ^T , where k=(1,2,...q)

Among them, p is the number of neurons in the hidden layer, and q is the number of neurons in the output layer.

The weight initialization method from the hidden layer to the output layer used by the reference center initialization method is:

其中，mink为训练集中第k个神经元中所有期望输出的最小值；maxk为训练集中第k个输出神经元中所有期望输出的最大值，其中，j＝(1,2，…，p)。

where mink is the minimum value of all expected outputs in the kth neuron in the training set; maxk is the maximum value of all expected outputs in the kth output neuron in the training set, where j=(1,2,...,p) .

Initialize the central parameter C _j =[c _j1, , c _j2 ,...c _jn ] ^T of each hidden layer neuron, and the initialization method is:

Among them, p is the total number of neurons in the hidden layer, j=(1,2,...,p).

mini is the minimum value of all input information of the ith feature in the training set, and maxi is the maximum value of all input information of the ith feature in the training set.

Initialize the width vector D _j =[d _j1 , d _j2 ,...d _jn ], the width vector affects the scope of the neuron's action on the input information: the smaller the width, the narrower the shape of the corresponding hidden layer neuron action function, Then the information located near the center of other neurons has a smaller response at the neuron, and the calculation method is:

df is the width adjustment coefficient, and its value is less than 1. Its function is to make each hidden layer neuron easier to realize the ability to perceive local information, which is beneficial to improve the local response ability of the RBF neural network.

Calculate the output z _j of the hidden layer neurons,

Among them, Cj is the center vector of the jth neuron in the hidden layer, Dj is the width vector of the jth neuron in the hidden layer, and ||·|| is the Euclidean norm;

Calculate the output Y=[y ₁ , y ₂ , ..., y _q ] ^T of the output layer neurons,

Among them, w _kj is the adjustment weight between the kth neuron in the output layer and the jth neuron in the hidden layer;

Among them, w _kj (t) is the adjustment weight between the k-th output neuron and the j-th hidden layer neuron during the t-th iteration calculation; c _ji (t) is the j-th hidden layer neuron For the center component of the i-th input neuron in the t-th iteration calculation; d _ji (t) is the width corresponding to the center c _ji (t); η is the learning factor; E is the evaluation function of the RBF neural network, where ,

Among them, _Olk is the expected output value of the kth output neuron at the lth input sample; _ylk is the actual output value of the network at the lth input sample of the kth output neuron.

When E is less than the expected error ε, the training ends.

Genetic Algorithm

Genetic algorithm (Genetic Algorithm) is a random search method evolved from the evolutionary laws of biology (survival of the fittest, survival of the fittest genetic mechanism). It was first proposed by Professor J. Holland in the United States in 1975. It operates directly on structural objects, and there is no restriction on derivation and function continuity; it has inherent implicit parallelism and better global optimization capabilities; using probabilistic optimization methods, it can automatically obtain and guide the optimization search space, adaptively adjusts the search direction, and does not require definite rules. These properties of genetic algorithm can well solve the problems of combinatorial optimization and machine learning, and it is a key technology in modern intelligent computing.

The basic operation process of the genetic algorithm is:

a) Initialization: set the evolutionary algebra counter t=0, set the maximum evolutionary algebra T, and randomly generate M individuals as the initial population P(0);

b) Individual evaluation: calculate the fitness of each individual in the group P(t);

c) Selection operation: The selection operator is applied to the group. The purpose of selection is to directly inherit the optimized individual to the next generation or generate new individuals through pairing and crossover and then inherit it to the next generation. The selection operation is based on the adaptation of individuals in the group. on the basis of degree evaluation;

d) Crossover operation: The crossover operator is applied to the population, and the core role in the genetic algorithm is the crossover operator;

e) Mutation operation: The mutation operator is applied to the population, that is, the gene values on some loci of the individual strings in the population are changed. After the population P(t) is selected, crossed, and mutated, the next generation population is obtained. P(t+1);

f) Judgment of termination condition: if t=T, the individual with the maximum fitness obtained in the evolution process is taken as the output of the optimal solution, and the calculation is terminated.

Genetic algorithm is a search algorithm used to solve optimization in computational mathematics. It is a kind of evolutionary algorithm. It is formed based on the idea of genetic variation of biological genes. Genetic algorithm is a global optimization algorithm, which can avoid falling into the local optimal solution and over-fitting, and can greatly improve the convergence speed of optimization through the method of crossover mutation.

Genetic Algorithm firstly performs coding and then further completes gene selection, crossover and mutation.

The steps of genetic algorithm optimization of RBF neural network are:

1) In order to facilitate the crossover variation of genes, the present invention adopts binary coding to generate an initial population, and generate any network parameters of N groups of center vectors c _i and width parameters δ _i ;

2) Use the gradient descent method to train the network parameters c _i and δ _i , and obtain the weight _wi of the network and the fitness f of the individual population, and the fitness f is:

Among them, E is the error function, C _max =max{E};

3) Perform crossover mutation operation on individuals with high fitness in step 2);

For example, randomly select individual a1, its corresponding fitness is f1, randomly select individual a2, its fitness is f2, select individual an, its fitness is fn; select some individuals with higher fitness in the crossover mutation. Part of the binary code is exchanged or reversed; and a new binary code is generated, namely new c _i and δ _i ;

4) Continue to use step 2) to train the new populations c _i and δ _i obtained in step 3).

5) When the error function meets the error accuracy requirements, the training stops, and the c _i and δ _i with the highest fitness at this time are taken as the optimal parameter values of the RBF neural network.

The new RBF neural network designed after the fusion of the genetic algorithm is used to evaluate the state of the stator bars as follows:

1) Select the electrical parameters and non-electrical parameters related to the aging state of the stator bar insulation of large-scale steam turbine generators as the parameter types to be measured. Insulation resistance R, polarization index PI and absorption ratio DAR can well reflect the moisture and contamination of the insulation, and are sensitive parameters for the investigation of the moisture absorption and contamination characteristics of stator insulation, dielectric loss tanδ, dielectric loss increment Δtanδ, capacitance C and capacitance increase ΔC can reflect the size of air gaps, delaminations and overall insulation defects in the insulation. Generally, the aging degree of insulation is often evaluated according to the change trend of tanδ and C with voltage. If Δtanδ or ΔC is close to 0, it indicates that the insulation flawless. When the insulation contains a large number of air gaps, with the increase of the AC voltage applied to the insulation, the leakage current flowing through the insulation will change abruptly with the enhancement of the air gap discharge, and the first and second current surge points will appear. Therefore, the AC current increase rate ΔI is related to the distribution and quantity of the internal air gaps, and can also reflect the overall aging state of the insulation. Partial discharge damages the insulation the most seriously. The discharge mainly comes from the breakdown of the air gap inside the insulation. The maximum partial discharge _Qmax reflects the largest local defect inside the insulation, which can effectively judge the local aging of the insulation. Studies have shown that the above characteristic quantities are all related to the residual breakdown voltage. According to the IEC standard, the residual breakdown voltage U _BD of the stator bar insulation can be used to reflect the degree of insulation aging. When U _BD drops to half of the initial breakdown voltage, the insulation is considered to have reached the end of its life.

2) Measure the accelerated aging wire rod of the large turbo-generator to obtain the required parameters;

The thermal aging test is carried out on the stator bar in an oven or by means of a heating plate, and a vibration exciter is used to apply vibration excitation to the stator bar to simulate the vibration of the stator bar, and a voltage is applied to both ends of the bar to simulate its vibration. The effects of electrical aging. Insulation resistance measurement was performed using an IR tester in terms of measurement parameters. The IR tester includes an ohmmeter, a built-in generator, and a generator. The generator is used to generate a high DC voltage, apply the voltage to the surface of the stator insulation, and cause current to flow around the surface of the insulation, giving an IR reading in ohms , and then calculate PI with the insulation resistance values at 1min and 10min, and calculate DAR with the insulation resistance values at 30s and 60s. Use a dielectric loss tester to measure the dielectric loss tanδ and the dielectric loss increment Δtanδ. Calculate the capacitance C and the capacitance increase ΔC by measuring the capacitance voltage and capacitance current. The value of the leakage current is measured by the current transformer. A partial discharge measuring instrument is used to measure the maximum partial discharge quantity Q _max . Use a voltmeter to measure the residual breakdown voltage. In order to better control the temperature during the aging test, the present invention adopts the infrared lattice temperature measurement method to detect the temperature. The measurement of the deformation amount can be done by a displacement sensor.

3) Use Pearson correlation analysis to classify and filter the measured parameters, and remove the factors that are not very correlated with the residual breakdown voltage. The Pearson algorithm is used to measure the correlation between two variables X and Y ( Linear correlation), its value is between -1 and 1, and its calculation formula is:

Among them, ρ _{x, y} is the correlation coefficient, cov(X, Y) is the covariance of the two variables, δ _x and δ _y are the standard deviations of the two parameters, respectively. When the correlation coefficient is greater than 0.7, the two variables are considered to be The correlation is very high, otherwise, it is considered as redundant parameters and eliminated. In order to improve the convergence speed and accuracy of the model, the parameters selected above are normalized, namely

Among them, x, y∈R ⁿ ; x _min =min(x); x _max =max(x);

4) RBF neural network modeling improved by genetic algorithm

4a) Initialize the dataset

Normalize the experimentally measured dielectric parameter data to reduce the influence of data of different magnitudes on prediction and reduce the amount of calculation;

4b) performing a correlation analysis on the dielectric parameter data obtained in step 4a) to eliminate parameters with low correlation;

4c) Generating an initial population by genetic algorithm;

计算，其中，c_i和c_j分别是第i个和第j个节点的中心向量；4d) Using the population parameters, let k=N in the Kmeans algorithm, classify the dielectric parameter data, and obtain the classification center according to the classification result, that is, the center c of the hidden layer radial basis function, and the width σ can be determined according to

Calculate, where c _i and c _j are the center vectors of the i-th and j-th nodes, respectively;

4e) Use the number of neurons in the hidden layer N, the center c and the width σ of the radial basis function to build an RBF training model, train the data set, and update the weights using the negative gradient descent method until the accuracy meets the requirements or the number of iterations reaches The maximum value is the trained RBF neural network;

4f) Use the trained RBF neural network to predict the remaining breakdown voltage, and then calculate the fitness of each population, that is, the mean absolute percentage error MAPE;

4g) Judging whether the condition is stopped, if not, then select crossover mutation for the population according to each adaptive degree, and then turn to step 4e), if yes, then turn to step 4h);

4h) according to the training result, select the individual with the best fitness in the population as the number N of hidden layers in the RBF neural network, and then train the entire training set data to obtain the parameters of the RBF neural network improved by the genetic algorithm;

4i) The RBF neural network improved by genetic algorithm is used to predict the remaining breakdown voltage to be predicted, and the predicted value of the remaining breakdown voltage is obtained.

The delamination phenomenon and air gap distribution after aging of mica were observed by scanning electron microscopy and thermogravimetric analysis, and the aging conditions of delamination and air gap were classified into 4 grades, as shown in Table 1:

Table 1

4g) Analytic Hierarchy Process: On the basis of the problem situation and strategic objectives, through subjective judgment and analysis of the interaction and mutual inclusion between the problem factors, the various factors are summarized according to different levels, and the hierarchical structure diagram of the problem is obtained. . Generally speaking, the multi-level analysis structure is the target layer, the standard layer and the decision plan layer in order from high to low. In this way, more complex problems can be transformed into multi-level problems with clear order and strong logic. After establishing the hierarchical structure diagram of the problem, the AHP should rank the relative importance of each factor at the same level, and determine the weight of the factors according to the sequence results, so as to provide a quantitative standard for the final decision, because the motor environment runs The temperature has a certain influence on the aging speed of the stator bar, so the residual breakdown voltage predicted by the neural network, the ambient temperature and the aging level of the generator during operation are comprehensively considered according to the principle of AHP. , giving the final running status evaluation result.

Claims (5)

1. a large-scale generator insulation state assessment method based on genetic algorithm and radial basis neural network, is characterized in that, comprises the following steps: 1) Measure the parameters related to the insulation aging state of the large-scale turbine generator stator bar; 2) classifying and screening the parameters obtained in step 1), and then using the parameters after classification and screening to construct a data set; 3) Establish RBF neural network; 4) Using genetic algorithm to optimize the number of hidden layers, the center and width of radial basis function in RBF neural network; 5) Use the data set to train the optimized RBF neural network, and then use the negative gradient descent method to update the iterative weights; 6) Use the RBF neural network obtained in step 5) to evaluate the insulation state of the large generator. 2. the large-scale generator insulation state assessment method based on genetic algorithm and radial basis neural network according to claim 1, is characterized in that, the parameter relevant with large-scale steam turbine generator stator bar insulation aging state comprises insulation resistance R , polarization index PI, absorption ratio DAR, dielectric loss tanδ, dielectric loss increment Δtanδ, capacitance C and capacitance increase ΔC. 3. the large-scale generator insulation state assessment method based on genetic algorithm and radial basis neural network according to claim 1, is characterized in that, utilizes negative gradient descent method to update iteration weight and comprises the following steps: Negative gradient descent method is used to update the weights of input layer, hidden layer and output layer in iterative RBF neural network. 4. the large-scale generator insulation state assessment method based on genetic algorithm and radial basis neural network according to claim 1, is characterized in that, utilizes the parameter after step 2) classification screening to construct RBF neural network in step 3). 5. the large-scale generator insulation state assessment method based on genetic algorithm and radial basis neural network according to claim 1, is characterized in that, the concrete operation process of step 4) is: 4a) Normalize the parameters; 4b) carrying out a correlation analysis on the normalized parameters to eliminate the parameters whose correlation is lower than the preset value; 4c) Generating an initial population by genetic algorithm; 4d) Using the population parameters, set k=N in the Kmeans algorithm, classify the parameters, and obtain the classification center according to the classification result, that is, the center c of the hidden layer radial basis function, and the width σ of the hidden layer radial basis function according to

Calculate, where c _i and c _j are the center vectors of the i-th and j-th nodes, respectively; 4e) Use the number of neurons in the hidden layer N, the center c and the width σ of the radial basis function to build an RBF training model, train the data set, and update the weights using the negative gradient descent method until the accuracy meets the requirements or the number of iterations reaches Up to the maximum value, the trained RBF neural network is obtained; 4f) Use the trained RBF neural network to predict the remaining breakdown voltage, and then calculate the fitness of each population; 4g) Judging whether the preset termination conditions are met, when the preset termination conditions are met, then go to step 4h); otherwise, the population is selected and cross-mutated according to each degree of adaptation, and then go to step 4e); 4h) According to the training result, select the individual with the best degree of self-adaptation in the population as the number of hidden layers N of the RBF neural network, and then train the data set to obtain the parameters of the RBF neural network improved by the genetic algorithm.

Images