Disclosure of Invention
In view of this, the present invention aims to provide a monopole ground fault location method based on wavelet decomposition and DBN, which can reliably and accurately locate a monopole ground fault point at a low sampling frequency.
In order to achieve the purpose, the invention adopts the following technical scheme:
a monopole earth fault distance measurement method based on wavelet decomposition and DBN comprises the following steps:
step S1, collecting fault voltage waveform when the circuit has monopole earth fault;
step S2, sending the fault voltage waveform into a DBN classification model, and determining the range of the transition resistance value and the category of each transition resistance value;
step S3, extracting the fault voltage waveform characteristics of each transition resistance value type by adopting wavelet decomposition, and performing single branch reconstruction processing;
step S4, after normalization processing is carried out on each wavelet component, a sub DBN regression model is respectively constructed and trained;
and step S5, superposing the results of the trained sub DBN regression models to realize fault location.
Furthermore, the DBN classification model adopts a Wavelet-DBN model.
Further, the step S3 is specifically;
step S31, adopting db4 wavelet as wavelet basis function to carry out one-layer wavelet decomposition on the fault voltage waveform, and obtaining an approximate component A and a detail component D by wavelet transformation of a one-dimensional fault signal according to the theory of multiresolution analysis, wherein A and D respectively represent low-frequency components and high-frequency components of an original fault signal; ,
and step S32, performing wavelet single branch reconstruction processing on each wavelet component to enable the length of each component to be consistent with the length of the original fault waveform.
Further, the wavelet transformation process is specifically expressed as:
Dn=HAn-1 (1)
An=GAn-1 (2)
where H and G denote a high-pass filter and a low-pass filter in a quadrature mirror filter bank, n is the number of decomposition levels, A0Is the original fault signal.
Further, the single branch reconstruction processing is carried out on the fault wavelet component according to the following formula:
wherein the content of the first and second substances,
and
the dual operators of the high pass filter H and the low pass filter G are indicated.
Further, the normalization processing specifically includes:
normalizing the wavelet components obtained by the three kinds of extraction to be between [0 and 1] according to the following formula:
wherein x isiIs the ith data, x, of wavelet coefficientsminAnd xmaxRespectively the minimum and maximum values of the wavelet coefficients.
Further, the step S5 is to reconstruct the final fault location value according to the following formula:
wherein S denotes a final position value, n denotes the number of high frequency components, and a and D denote output values of the submodels corresponding to the low frequency components and the high frequency components, respectively;
further, the overall positioning accuracy of the method is evaluated by calculating the mean absolute error MAE between the positioning result and the actual fault position:
the mean absolute error of fault location is shown as follows:
wherein, WiIs the actual fault location, preiIs the model position result and N is the number of position samples to be tested.
Further, the DBN training includes a pre-training stage and a fine-tuning stage, specifically:
step S41, in the pre-training phase, training data is sent to the model through the visible layer of the bottom RBM and each weight and offset is trained using an unsupervised, layer-by-layer greedy learning algorithm
And step S42, performing a supervised fine tuning stage, feeding back errors from the top to the bottom of each RBM layer by the top layer BP network by using a backward error propagation mechanism, and optimizing training parameters to achieve the purpose of fine tuning.
Further, the pre-training specifically comprises:
training a hidden layer in the lowest-layer RBM, wherein after the training is finished, the layer is used as a visible layer in the second layer to continue training the hidden layer in the second-layer RBM; this process continues until all hidden layers have been trained;
after the fault samples are fed into the DBN model, the energy function of each RBM layer can be represented as
Wherein v isiIs the ith neuron of the visible layer and can be denoted as v ═ v1,v2,...,vn]N is the number of visible layer neurons; h isjIs the jth neuron of the hidden layer, which can be denoted as h ═ h1,h2,...,hm]And m is the number of hidden layer neurons. θ is a parameter set containing w, a and B. w ═ wij}∈Rn×mRepresenting a weight matrix between the visible layer and the hidden layer, where wijIs the weight that connects the ith visible layer neuron and the jth hidden layer neuron. A ═ ai}∈RnAnd B ═ Bi}∈RmRespectively representing a visible layer deviation vector and a hidden layer deviation vector;
based on equation (7), when θ is fixed, the joint distribution of (v, h) is:
in the formula (I), the compound is shown in the specification,
representing the normalization factor, i.e. the sum of the energies in all possible cases.
Compared with the prior art, the invention has the following beneficial effects:
according to the invention, only single-ended fault signals need to be acquired, and the reliable and accurate positioning of the single-pole grounding fault point under the low sampling frequency can be realized.
Detailed Description
The invention is further explained by the following embodiments in conjunction with the drawings.
Referring to fig. 1, the present invention provides a monopole ground fault location method based on wavelet decomposition and DBN, comprising the following steps:
step S1, collecting fault voltage waveform when the circuit has monopole earth fault;
step S2, sending the fault voltage waveform into a DBN classification model, and determining the range of the transition resistance value and the category of each transition resistance value;
step S3, extracting the fault voltage waveform characteristics of each transition resistance value type by adopting wavelet decomposition, and performing single branch reconstruction processing;
step S4, after normalization processing is carried out on each wavelet component, a sub DBN regression model is respectively constructed and trained;
and step S5, superposing the results of the trained sub DBN regression models to realize fault location.
In the present embodiment, referring to fig. 2, the MMC-HVDC system model rectifies the fault voltage waveform around the time of the sudden change of line voltage at a sampling frequency of 20kHz for the lateral recording device. The neutral point grounding method used in the system is a grounding mode of clamping a large resistor on the direct current side.
The rectifier side adopts constant direct current voltage and constant reactive power control, and the inverter side adopts constant active power and constant reactive power control. The transmission line adopts a frequency conversion overhead line with the length of 200 km. The system model structure is shown in fig. 2, the detailed parameters are shown in table 1, and the overhead line tower structure is shown in fig. 3. The system simulation sets the occurrence time of the anode ground fault to be 1s, and collects the voltage waveform of the anode ground at the rectifying side at the sampling frequency of 20kHz when the fault occurs.
TABLE 1 System parameters
System parameter
|
Numerical value
|
Rated DC voltage/kV
|
±250
|
Rated power/MW
|
400
|
DC line length/km
|
200
|
Number of submodules per bridge arm
|
200
|
capacitance/mF of each submodule
|
10
|
Bridge arm inductance/mH
|
50
|
Neutral point grounding resistance/k omega
|
100
|
Winding line voltage/kV
|
380/220
|
Winding connection mode
|
Yn/D |
In an embodiment, in the step S2, the DBN classification model is used to classify the fault voltage waveform according to a range of transition resistances, so as to reduce the interference of the transition resistance factors on the fault location.
Simulating the transition resistance of 0.1 omega to 0.HVDC system model by using the MMC-HVDC system model of the step S110000 omega unipolar ground faults and collect the corresponding single ended fault voltage waveforms. Then, the fault voltage waveforms are classified into four categories of 0.1 Ω -1000 Ω, 1200 Ω -3000 Ω, 3300 Ω -6000 Ω and 6400 Ω -10000 Ω according to different spans of the transition resistance. Each class contains 10 transition resistance values to train the DBN classification model, and the corresponding transition resistance ranges are shown as C in fig. 4set1-4As shown. And in the subsequent steps, a Wavelet-DBN model is built for each transition resistance range.
In an embodiment, the step S3 is specifically;
step S31, adopting db4 wavelet as wavelet basis function to carry out one-layer wavelet decomposition on the fault voltage waveform, and obtaining an approximate component A and a detail component D by wavelet transformation of one-dimensional fault signals according to the theory of multi-resolution analysis, wherein A and D respectively represent the low-frequency component and the high-frequency component of the original fault signals;
the wavelet transformation process is specifically expressed as:
Dn=HAn-1 (1)
An=GAn-1 (2)
where H and G denote a high-pass filter and a low-pass filter in a quadrature mirror filter bank, n is the number of decomposition levels, A0Is the original fault signal. The discrete wavelet decomposition tree is shown in figure 5.
In step S32, since the wavelet component signal length obtained by decomposition decreases with the increase of the number of wavelet decomposition layers, it is not favorable for the construction of the subsequent regression model. Therefore, it is necessary to perform wavelet single-branch reconstruction processing on each wavelet component so that the length of each component is consistent with the length of the original fault waveform. And (4) performing single-branch reconstruction processing on the fault wavelet component according to a formula (3).
Wherein the content of the first and second substances,
and
the dual operators of the high pass filter H and the low pass filter G are indicated.
When a single-pole ground fault occurs, the fault recording equipment extracts 200 sampling points as an original fault waveform at the frequency of 20kHz according to the time when a signal suddenly changes. Wavelet components of each frequency band of the fault waveform are extracted through wavelet decomposition, and single-branch reconstruction processing is carried out on each wavelet component.
In this embodiment, the step S4 specifically includes: and after normalization processing is carried out on each wavelet component, a sub DBN regression model is respectively constructed. Normalizing the wavelet components of each wavelet component obtained by the three kinds of extraction in the step to be between [0 and 1] according to a formula (4).
Wherein x isiIs the ith data, x, of wavelet coefficientsminAnd xmaxRespectively, the minimum and maximum values of each wavelet coefficient. After normalization, the model is returned to the DBN for training.
In this embodiment, in order to ensure the generalization performance of the model, the overlapping portion between the transition resistances, i.e. the characteristic waveforms corresponding to 0.1-1200 Ω, 1000-. Taking a unipolar ground fault with a transition resistance value of 1100 Ω as an example, the fault voltage waveform is sent to the DBN classification model, since the resistance value is in a range not covered by each category, and thus the output category may be Cset1 or Cset2, but the DBN regression model corresponding to both categories has learned the characteristic information of including it in Rset1 or Rset2 of the resistance value, and thus the fault can still be accurately located.
In this embodiment, the step S5 is specifically to reconstruct the final fault location value according to the following formula:
wherein S denotes a final position value, n denotes the number of high frequency components, and a and D denote output values of the submodels corresponding to the low frequency components and the high frequency components, respectively;
the overall positioning accuracy of the method is evaluated by calculating the mean absolute error MAE between the positioning result and the actual fault location:
the mean absolute error of fault location is shown as follows:
wherein, WiIs the actual fault location, preiIs the model position result and N is the number of position samples to be tested.
Example 1:
in this embodiment, a fault voltage waveform corresponding to a ground fault is collected according to the MMC-HVDC system, a transition resistance range of the fault voltage waveform is 0.1 Ω to 10000 Ω, a fault of 0 Ω to 1000 Ω is regarded as a low-resistance fault, and a fault of 1000 Ω to 10000 Ω is regarded as a high-resistance fault. When the single-pole grounding fault occurs, the fault occurrence time is set to be 1s, a fault point is arranged at every 1km, and 40 fault voltage waveform data are collected at each fault point according to different transition resistance values. A total of 8,000 fault voltage data were collected as an overall sample of the DBN classification model and the Wavelet-DBN fault location model.
In this embodiment, the training of the preferred DBN model: RBMs are random neural network models with a two-layer structure that are fully connected between layers, but unconnected within a layer. The main idea of the DBN model is to train the weights of the deep neural network by using a layer-by-layer greedy learning algorithm, so as to mine deeper data characteristics and realize complex mapping from input to output of training data. The DBN model of the three-layer structure is shown in fig. 5.
In this embodiment, the normalized wavelet coefficients are sent to a DBN for training, where the DBN training includes a pre-training stage and a fine-tuning stage, and specifically includes:
in the pre-training phase, training data is sent to the model through the visible layer of the bottom RBM, and each weight and offset is trained using an unsupervised, layer-by-layer greedy learning algorithm. First, the hidden layers in the lowest layer RBM are trained. After training is complete, this layer will be used as the visible layer in the second layer to continue training the hidden layer in the second layer RBM. This process continues until all hidden layers have been trained. After the fault samples are sent into the DBN model, the energy function of each RBM layer can be expressed as
Wherein v isiIs the ith neuron of the visible layer and can be denoted as v ═ v1,v2,...,vn]N is the number of visible layer neurons; h isjIs the jth neuron of the hidden layer, which can be denoted as h ═ h1,h2,...,hm]And m is the number of hidden layer neurons. θ is a parameter set containing w, a and B. w ═ wij}∈Rn×mRepresenting a weight matrix between the visible layer and the hidden layer, where wijIs the weight connecting the ith visible layer neuron and the jth hidden layer neuron. A ═ ai}∈RnAnd B ═ Bi}∈RmRepresenting the visible layer disparity vector and the hidden layer disparity vector, respectively. Based on equation (7), when θ is fixed, the joint distribution of (v, h) is:
in the formula (I), the compound is shown in the specification,
representing the normalization factor, i.e. the sum of the energies in all possible cases.
After the unsupervised training of the RBM is over, a supervised fine tuning phase will be performed. The top layer BP network feeds back errors from the top to the bottom of each RBM layer by using a backward error propagation mechanism, and optimizes training parameters to achieve the aim of fine adjustment.
In the present embodiment, an appropriate hidden layer and the number of neurons in each hidden layer are selected based on the high-frequency component and the low-frequency component obtained by the wavelet decomposition. Finally, it was determined that both the high and low frequency DBN models use a two-layer hidden layer structure, with 500 and 1000 neuron numbers in each layer of the high frequency component model, respectively, and 750 and 1500 neuron numbers in each layer of the low frequency component model, respectively, in the neuron model. If more than one level of wavelet decomposition is used, the DBN models corresponding to the plurality of high frequency components obtained by the decomposition all adopt the same network structure. Run over 120 pre-training sessions with a batch size of 16 in each layer, set the pre-trained unsupervised learning rate to 0.005 and the fine-tuned supervised learning rate to 0.001. The parameters of the DBN classification model are consistent with the DBN high frequency component regression model. The difference between the two is that the number of neurons in the output layer of the former is 4 (classified into 4 classes), and the activation function of this layer is "softmax", aiming at normalization. The latter has a number of neurons in the output layer of 1 (position value), an activation function of "linear", and a mean square error used as a loss function to calculate the positioning error during training.
Referring to fig. 6, the fault samples are classified into four classes (C)set1-4) So as to reduce the influence of the transition resistance factor on the final fault positioning precision. The fault voltage waveform is directly used as an input feature of the DBN model, and the classification mechanism of the DBN is used for realizing sample classification. In the course of training the DBN classification model, a total of 8000 fault samples (excluding the undetected resistance values for each class) were used, with 80% of the samples used to train the model and the remaining 20% used to test classification performance. The confusion matrix for the test results is shown in FIG. 6, where the label number corresponds to Cset1-4A category. The classification test result shows that the test precision of the classification model is 100% for the ground fault sample covered by each resistance class, which shows that the classification model can accurately evaluate the transition resistance value range of the fault.
In the present embodiment, the method of the present invention is compared with SVM and BPNN fault location methods. And obtaining the optimal parameters in the comparison model by a grid search method. Parameters 'C' and 'g' in the SVM model are optimized, and the search range is as follows: c belongs to [0.0001, 100], g belongs to [50, 5000 ]; the number of hidden layers and the number of neurons of each layer of the BPNN are optimized, and the parameter search range is as follows: the hidden layer number is equal to [1, 4], and the number of neurons in each layer is [64, 128, 256, 512, 1024 ]. The mean absolute error was used as an evaluation index for two comparative models, and the optimum parameters are shown in Table 3-1. It is noted that 512-: the number of neurons in the input layer is 512, the number of neurons in the hidden layer 1 is 256, the number of neurons in the hidden layer 2 is 128, and the number of neurons in the output layer is 1.
TABLE 2 control model parameter settings
Dividing fault voltage waveform into four types of C through DBN classification modelset1-4Then according to the formula and Rset1-4And training the DBN regression model by the corresponding characteristic waveform. For a single resistance class, 2200 fault waveform samples were used in training the Wavelet-DBN regression model, 80% of which were used for model training and the remaining 20% were used for testing localization effects. Meanwhile, the fault voltage waveform is directly used as training data of each model and is compared with a positioning method using wavelet decomposition, so that the influence of the wavelet decomposition feature extraction method on the positioning result is tested. When training a comparison model that does not use wavelet decomposition, only 80% of the fault samples for each resistance class need to be used directly to train the model, with the remaining 20% tested, similar to training a single wavelet component localization model. Table 3 shows the fault location error for each comparative model. As can be seen from table 3, compared with other positioning methods, the average absolute error of the fault positioning result in each resistance interval by using the wavelet-DBN method is greatly reduced, and the fault positioning performance has a great advantage in precision. Meanwhile, when a high-resistance ground fault occurs, the fault positioning effect of the Wavelet-DBN modelThe result was still very excellent. From the performance of the regression model, the DBN model can locate the unipolar ground fault more accurately and reliably than the BPNN and SVM. This is because it is difficult for a machine learning method that relies on manual feature extraction to accurately identify the fault features, thereby increasing the positioning error.
TABLE 3 fault location test results for each model
In this example, in order to further test the high resistance (1000 Ω -10000 Ω) fault location performance of the proposed method, one fault point was simulated every 30km from the rectifier side, and there were a total of 7 unipolar ground fault points. High resistance fault voltage waveforms with different resistance values are collected and input to the position model for testing. And comparing the positioning result of each fault point with the actual fault position, and obtaining a position error. According to the proposed method and the comparison method, high resistance fault data of 10 different transition resistances were tested at each fault point, and the corresponding mean absolute error was obtained by calculation. The high resistance fault test results for each model are shown in table 4. By comparing the MAE of each model, it can be easily seen that the method provided by the embodiment can obtain the best positioning accuracy at the high-resistance positions of different positions.
TABLE 4 high resistance Fault location test results
The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.