CN112100923A - State evaluation method for frequency converter IGBT of full-power generation system - Google Patents

Abstract

A method for evaluating the state of an inverter IGBT of a full-power power generation system provided by the present invention includes the following steps: step 1, obtaining 13 types of waveform signals; step 2, decomposing the 13 types of waveform signals obtained in step 1 to obtain a judgment fault step 3, use the wavelet decomposition coefficient to construct the feature information of the fault judgment obtained in step 2, and obtain the feature vector; step 4, extract the feature value from the feature vector obtained in step 3; step 5, construct the BP Neural network; step 6, optimize the obtained BP neural network; step 7, train the BP neural network optimized in step 6 to obtain a trained optimized BP neural network; step 8, use the optimized BP neural network trained in step 7 The BP neural network is brought into the actual operation data verification, and real-time evaluation of whether the inverter IGBT is open; the present invention solves the optimization configuration problem existing in the current offshore wind power maintenance work, and improves economy, stability and maintainability.

Description

A state evaluation method of inverter IGBT in full power generation system

technical field

The invention relates to the technical field of state evaluation of frequency converters of offshore wind turbines, in particular to a state evaluation method of frequency converter IGBTs of full power generation systems.

Background technique

The frequency converter is an important electrical equipment in large-capacity offshore wind turbines, and one of the core energy conversion equipment in wind turbines. One of the tallest devices. At present, the mainstream offshore wind turbines are full power variable speed constant frequency (VSCF) power generation systems, among which the squirrel cage asynchronous power generation system with high-speed gearbox + squirrel cage asynchronous generator + full power inverter as the technical route is more common. On the one hand, with the rapid development of my country's offshore wind power industry and the continuous increase of the single-unit capacity of offshore wind turbines, the large-capacity converter has become a development trend, which puts forward higher requirements on the operational reliability, economy and maintainability of the equipment. On the other hand, the 12 IGBTs that constitute the main circuit of the two-level inverter are controlled on and off, and its switching loss and conduction loss are very large, which is the weak link of the system reliability. As offshore wind turbines face more complex environmental conditions such as high temperature, high humidity, salt spray corrosion, lightning, typhoon, etc., the wind power fluctuates violently, and the electrical and thermal stress of the wind power converter changes drastically, giving the power semiconductor (IGBT) of the main circuit of the converter security and reliability threats. According to statistics, most of the faults in the main circuit of the inverter are caused by IGBT damage. Therefore, a new technology is needed to realize the real-time status evaluation of the IGBT components of the inverter, to detect potential faults as early as possible, and to provide guidance for the maintenance work.

At present, no one has been involved in the field of IGBT state evaluation of inverters in full-power power generation systems.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a state evaluation method of the inverter IGBT of the full power generation system, which solves the defect that the state of the inverter IGBT of the inverter of the full power generation system has not been evaluated in the prior art, resulting in poor reliability of the overall system.

In order to achieve the above object, the technical scheme adopted in the present invention is:

A method for evaluating the state of an inverter IGBT of a full-power power generation system provided by the present invention includes the following steps:

Step 1: Obtain the waveforms of the two-level full-power inverter in the normal working state, and the three-phase voltage waveform signals output by the corresponding inverters when the IGBTs of the rectifier circuit and the inverter circuit are open-circuit faults, and obtain 13 types of waveform signals;

Step 2, decompose the 13 types of waveform signals obtained in step 1 to obtain characteristic information for determining the fault;

Step 3, use the wavelet decomposition coefficient to construct the feature information of the judgment fault obtained in step 2, and obtain a feature vector;

Step 4, extract eigenvalues from the eigenvectors obtained in step 3;

Step 5, construct a BP neural network, and use the feature value obtained in step 4 as the input layer of the BP neural network obtained by constructing;

Step 6, optimize the neural network obtained in step 5 to obtain an optimized BP neural network;

Step 7, train the optimized BP neural network in step 6 to obtain the trained optimized BP neural network;

Step 8, use the optimized BP neural network trained in step 7, bring in the actual operation data for verification, and evaluate in real time whether the inverter IGBT is open.

Preferably, in step 2, the wavelet decomposition method is used to decompose the 13 types of waveform signals obtained in step 1 to obtain characteristic information for determining the fault.

Preferably, in step 5, a BP neural network is constructed, and the specific method is:

The excitation function g(x) of the BP neural network is a Sigmoid function;

The output of the hidden layer is:

The output of the output layer is:

The mean squared error is:

Among them, Y _k is the expectation;

The input layer of the neural network is set with 18 nodes; the number of neurons in the output layer is determined to be 13; the number of hidden layer nodes is 20.

Preferably, the neural network obtained in step 5 is optimized, and the specific method is:

The neural network is optimized and improved by genetic algorithm, and the optimized BP neural network is obtained.

Preferably, in step 7, the BP neural network optimized in step 6 is trained, specifically:

Use the training set to train the optimized BP neural network, optimize the weights and thresholds, and use the optimized weights and thresholds as the initial weights and thresholds of the BP neural network to obtain a trained optimized BP neural network.

Compared with the prior art, the beneficial effects of the present invention are:

The present invention provides a state evaluation method for the inverter IGBT of a full-power power generation system, which can provide real-time evaluation results of the state of the key component IGBT in the two-level inverter under the condition of existing monitoring quantities, so as to provide full-power power generation at sea. Provide reference for system inverter maintenance. It can solve the optimization configuration problem existing in the current offshore wind power maintenance work, and improve the economy, stability and maintainability;

The BP neural network has strong nonlinear mapping ability, and can establish the connection between the fault type of the inverter IGBT and the fault result; the wavelet decomposition is beneficial to extract the relevant characteristic parameters. However, the traditional BP neural network also has problems such as low precision and difficulty in obtaining the optimal weights. The invention adopts the genetic algorithm improved BP neural network, and optimizes the weights and thresholds of the neural network by integrating the model established by the genetic algorithm and the neural network, so as to solve the limitation of the neural network itself. The optimized model can diagnose the IGBT fault of the inverter very well, and the accuracy of the diagnosis result is higher, which can effectively locate and diagnose the complex inverter fault, improve the maintenance efficiency, and reduce the loss when the fault occurs.

Description of drawings

FIG. 1 is a schematic flow chart of the present invention.

Figure 2 is a schematic diagram of the structure of a two-level PWM inverter.

Figure 3 is a flowchart of the improved BP neural network.

Figure 4 is the training diagram of the improved BP neural network.

Figure 5 is the prediction diagram of the improved BP neural network.

Detailed ways

The present invention will be described in further detail below with reference to the accompanying drawings.

As shown in FIG. 1 , a method for evaluating the state of an inverter IGBT of a full-power power generation system provided by the present invention includes the following steps:

Step 1: Collect and filter the three-phase voltage waveform data on the grid side during the operation of the two-level full-power inverter, and obtain the waveforms of the two-level full-power inverter in the normal working state, as well as the IGBTs of the rectifier circuit and the inverter circuit. There are 13 types of waveform signals in the three-phase voltage waveform signal output by the corresponding inverter in the case of an open-circuit fault, and the 13 types of waveform signals correspond to the normal state or the fault state respectively;

Step 2, use wavelet decomposition to decompose the 13 types of waveform signals to obtain characteristic information for determining the fault;

Step 3, use the wavelet decomposition coefficient to construct the feature information of the judgment fault obtained in step 2, and obtain a feature vector;

Step 4, extract the normalized eigenvalues from the obtained eigenvectors, use the eigenvalues as the input layer of the neural network, and go to step 5;

Step 5, setting the network hidden layer and output layer, constructing a BP neural network, and training the constructed BP neural network to obtain a trained BP neural network;

Step 6, using multiple swarm algorithms to improve and optimize the trained BP neural network to obtain an optimized BP neural network;

Step 7, train the optimized BP neural network to obtain the trained optimized BP neural network;

Step 8: Use the trained optimized BP neural network, bring in the actual operation data for verification, and evaluate in real time whether the inverter IGBT is open.

specifically:

Use wavelet decomposition to decompose 13 types of waveform signals. The specific methods are:

The wavelet decomposition is set as a 5-layer decomposition form, and the original signal is decomposed as follows:

X=A ₅ +D ₁ +D ₂ +D ₃ +D ₄ +D ₅

Wherein, X is the original waveform signal; A5 is the high frequency part contained in the _fifth layer decomposition _; D1, D2, _D3 , _D4 and _D5 are the low frequency part contained after each layer decomposition _;

A ₅ , D ₁ , D ₂ , D ₃ , D ₄ and D ₅ all have corresponding wavelet packet decomposition coefficients, and feature vectors are constructed according to the coefficients obtained from the decomposition:

S _A = [C ₁ C ₂ C ₃ C ₄ C ₅ C ₆ ]

C ₁ , C ₂ , C ₃ , C ₄ , C ₅ and C ₆ are the decomposition coefficients of each layer of phase A, and S _A is the eigenvector of phase A;

S=[S _A S _B S _C ]

Among them, S is the three-phase eigenvector;

The elements in the eigenvector S are normalized by the following formula, and the normalized eigenvalues are obtained:

In the formula, y is the normalized value, x is the sample value, x _min is the minimum value in the sample data, and x _max is the maximum value of the data in the sample.

The experiment of each group is actually to extract a feature vector composed of 18 feature elements from the waveform of each fault.

In the described step 5, the BP neural network algorithm is adopted to construct the model in the step 6, and the specific method is as follows:

algorithm design

Select the eigenvalues obtained in step 4, first set the excitation function g(x) as the Sigmoid function

Set the input layer to the hidden layer weight (ω _ij ), the threshold (a _j ), the hidden layer to the output layer weight (ω _jk ), the threshold (b _k ), the number of nodes in the input layer, hidden layer and output layer For n, l, m, the output of the hidden layer (H _j ) and the output of the output layer (O _k ) can be expressed as:

The output layer mean square error (E _k ) can be expressed as:

Y _k is the expectation.

According to the output, observe whether the error is within the allowable range. If it is not within the allowable range, continue the operation. If it is within the allowable range, end the operation.

Input layer design, from the fault feature vector obtained by wavelet analysis, the input layer of the neural network is designed with 18 nodes, and the sample fault data is 18-dimensional;

In the design of the output layer, for the single-tube open-circuit fault of the inverter, there are 13 fault states, that is, the state of the inverter during normal operation and the state of 12 IGBT open-circuit faults, respectively. Therefore, we determine the number of neurons in the output layer as 13. This paper defines failure modes as 1, 2, 3, …, 13. Among them, 1 represents the normal state of the inverter, and 2 to 13 represent the fault states corresponding to IGBTs with different numbers.

13-bit binary fault codes are used for the output of the neural network and correspond to different fault modes respectively;

In the hidden layer design, the number of hidden layer nodes is obtained by using the empirical formula, and the optimal solution is selected by comparing them. After repeated testing, the number of hidden layer nodes is finally determined to be 20.

Training function settings, the training function of the neural network selects trainlm, the transfer function from the input layer to the hidden layer selects the tansing function, and the function from the hidden layer to the output layer selects the purelin function.

Preferably, the genetic algorithm is used to improve the BP neural network algorithm in step 6:

Genetic algorithm (Genetic Algorithm) can improve the global search ability of neural network, and can solve the problem that neural network is easy to fall into local minimum. Here, the BP neural network is improved by using multi-swarm algorithm.

Use genetic algorithm to optimize and improve, specifically:

First, encode the weights and thresholds of the BP neural network;

Choose a moderate function, which is a standard used to distinguish good and bad individuals, and it is a function used to calculate individual fitness according to evolutionary goals. The function formula is expressed as:

n is the number of output nodes of the neural network, ri is the actual value of the _ith node in the neural network, pi is the predicted value of the _ith node, and k is the coefficient.

The selection operation is performed using the roulette algorithm, and the population evolution is performed using the real number crossover method:

Set up the variogram:

a _max and a _min are the maximum and minimum values of a _ij , f(g) is the evolution function, and g is the current evolution times.

The optimized weight threshold is used as the initial weight and threshold of the neural network.

Train the network using the training set.

In the step 8, the specific method for evaluating the inverter IGBT according to the established improved BP neural network is:

Collect the three-phase voltage data on the grid side of the inverter, and use the 5-layer wavelet decomposition to extract the eigenvectors;

The BP neural network is optimized by genetic algorithm, and the parameters are set as follows:

Table 1 Genetic algorithm parameter settings

Use the optimized weight threshold as the initial weight and threshold of the neural network to run the neural network.

In step 8, the actual state of the 12 IGBTs in the inverter is evaluated according to the 13-dimensional vector output in step 7.

Example

The full-power power generation system adopts a two-level mode, and the core rectifier and inverter circuit is composed of 12 high-power IGBTs. The overall structure is shown in Figure 2; specifically:

Collect the three-phase voltage operation data on the grid side of the inverter;

The extracted data is extracted by wavelet decomposition, and the number of decomposition layers is 5.

X=A ₅ +D ₁ +D ₂ +D ₃ +D ₄ +D ₅

Wherein, X is the original waveform signal; A ₅ is the high frequency part included in the fifth layer decomposition; D ₁ , D ₂ , D ₃ , D ₄ and D ₅ are the low frequency part included in the decomposition of each layer.

Extract each decomposition coefficient to construct a eigenvector:

S _A = [C ₁ C ₂ C ₃ C ₄ C ₅ C ₆ ]

C ₁ , C ₂ , C ₃ , C ₄ , C ₅ and C ₆ are the decomposition coefficients of each layer, and S _A is the phase A feature vector;

S=[S _A S _B S _C ]

S is the three-phase eigenvector;

Using the genetic algorithm as the optimization strategy, a BP neural network is established, as shown in Figure 3.

Normalize the elements in the feature vector as follows:

In the formula, y is the normalized value, x is the sample value, x _min is the minimum value in the sample data, and x _max is the maximum value of the data in the sample.

Set the excitation function to the sigmoid function,

By setting the hidden layer weights, thresholds and input layer learning rate, the output of the hidden layer and output layer can be expressed as:

The output mean squared error can be expressed as:

Encode the weights and thresholds of the BP neural network;

Choose a moderate function, which is a standard used to distinguish good and bad individuals, and it is a function used to calculate individual fitness according to evolutionary goals. The function formula is expressed as:

n is the number of output nodes of the neural network, ri is the actual value of the _ith node in the neural network, pi is the predicted value of the _ith node, and k is the coefficient.

The selection operation is performed using the roulette algorithm, and the population evolution is performed using the real number crossover method:

Set up the variogram:

a _max and a _min are the maximum and minimum values of a _ij , f(g) is the evolution function, and g is the current evolution times.

The optimized weight threshold is used as the initial weight and threshold of the neural network, and the network is trained using the training set. The training results are shown in Figure 4.

Continuously adjust the evolution function, and finally optimize the weight range. For the squirrel cage full power generation system. Using the improved BP neural network to predict the state of each IGBT, the accuracy rate can reach more than 95%, as shown in Figure 5. It greatly improves the efficiency and accuracy of operation and maintenance analysis.

Claims (5)

1. a state evaluation method of full power generation system frequency converter IGBT, is characterized in that, comprises the following steps: Step 1: Obtain the waveforms of the two-level full-power inverter in the normal working state, and the three-phase voltage waveform signals output by the corresponding inverters when the IGBTs of the rectifier circuit and the inverter circuit are open-circuit faults, and obtain 13 types of waveform signals; Step 2, decompose the 13 types of waveform signals obtained in step 1 to obtain characteristic information for determining the fault; Step 3, use the wavelet decomposition coefficient to construct the feature information of the judgment fault obtained in step 2, and obtain a feature vector; Step 4, extract eigenvalues from the eigenvectors obtained in step 3; Step 5, construct a BP neural network, and use the feature value obtained in step 4 as the input layer of the BP neural network obtained by constructing; Step 6, optimize the neural network obtained in step 5 to obtain an optimized BP neural network; Step 7, train the optimized BP neural network in step 6 to obtain the trained optimized BP neural network; Step 8, use the optimized BP neural network trained in step 7, bring in the actual operation data for verification, and evaluate in real time whether the inverter IGBT is open. 2. The method for evaluating the state of an inverter IGBT in a full-power power generation system according to claim 1, wherein in step 2, the 13 types of waveform signals obtained in step 1 are decomposed by using a wavelet decomposition method to obtain a judgment Characteristic information of the fault. 3. the state evaluation method of a kind of full power generation system frequency converter IGBT according to claim 1, is characterized in that, in step 5, constructs BP neural network, and the concrete method is: The excitation function g(x) of the BP neural network is a Sigmoid function; The output of the hidden layer is:

The output of the output layer is:

The mean squared error is:

Among them, Y _k is the expectation; The input layer of the neural network is set with 18 nodes; the number of neurons in the output layer is determined to be 13; the number of hidden layer nodes is 20. 4. the state evaluation method of a kind of full power generation system inverter IGBT according to claim 1 is characterized in that, the neural network obtained in step 5 is optimized, and the concrete method is: The neural network is optimized and improved by genetic algorithm, and the optimized BP neural network is obtained. 5. the state evaluation method of a kind of full power generation system frequency converter IGBT according to claim 1, is characterized in that, in step 7, the BP neural network optimized in step 6 is trained, specifically: Use the training set to train the optimized BP neural network, optimize the weights and thresholds, and use the optimized weights and thresholds as the initial weights and thresholds of the BP neural network to obtain a trained optimized BP neural network.

Images