TRANSFORMER STATE EVALUATION METHOD BASED ON ECHO STATE NETWORK AND DEEP RESIDUAL NEURAL NETWORK

Abstract

A transformer health state evaluation method based on a leaky-integrator echo state network includes the following steps: collecting monitoring information in each substation; performing data filtering, data cleaning and data normalization on the collected monitoring information to obtain an input matrix; inputting the input matrix into a leaky-integrator echo state network to generate trainable artificial data, and dividing the artificial data into a training set and a test set in proportion; constructing a deep residual neural network based on a squeeze-and-excitation network, and inputting the training set and the test set for network training; and performing health state evaluation and network weight update based on actual test data. Considering that a deep learning-based neural network needs a large amount of data, the present disclosure uses the leaky-integrator echo state network to generate the artificial training data.

Claims

1. A transformer health state evaluation method based on a leaky-integrator echo state network and a deep residual neural network, comprising the following steps: step 1: collecting monitoring information in each substation, comprising monitoring information of an oil test and contents of dissolved gas and furan in oil in each substation; step 2: performing data filtering, data cleaning and data normalization on the collected monitoring information to obtain an input matrix; step 3: inputting the input matrix into a leaky-integrator echo state network to generate trainable artificial data, and dividing the artificial data into a training set and a test set in proportion; step 4: constructing a deep residual neural network based on a squeeze-and-excitation network, and inputting the training set and the test set for network training; and step 5: performing health state evaluation and network weight update based on actual test data.

2. The transformer health state evaluation method according to claim 1, wherein the monitoring information collected in step 1 is specifically records of test ledgers of a transformer and an electric power company during operation, wherein each group of data comprises contents of nine key states: a breakdown voltage (BDV), water, acidity, hydrogen, methane, ethane, ethylene, acetylene, and furan, and a health state of the corresponding transformer.

3. The transformer health state evaluation method according to claim 1, wherein step 2 specifically comprises: performing moving average filtering on the data collected in step 1 to eliminate noise in the data, wherein an expression of a moving average filter is as follows: $Y\left(t\right)=\frac{1}{T_w}{\int }_{t-T_w}^{t}y\left(t\right)\mathrm{dt}$ wherein Y(t) represents an output of the filter, y(t) represents an input of the filter, t represents a length of input data, and T.sub.w represents a length of a moving window, and a value of T.sub.w determines filtering performance; then performing data cleaning, comprising error correction, duplicate deletion, specification unification, logic correction, structure conversion, data compression, incomplete/null value supplementation, and data/variable discarding to ensure data consistency; and finally normalizing processed data by a range transformation method according to the following formula: $y_i=\frac{x_i-\min \left(x\right)}{\max \left(x\right)-\min \left(x\right)}$ wherein max(x) represents a maximum value of sample data, min(x) represents a minimum value of the sample data, x.sub.i represents an i.sup.th piece of data in a sample, and y.sub.i represents an i.sup.th piece of normalized data.

4. The transformer health state evaluation method according to claim 1, wherein step 3 specifically comprises: performing model establishment and algorithm training to obtain a model of the leaky-integrator echo state network, specifically comprising: establishing the leaky-integrator echo state network, wherein a state equation is as follows:
x(t+1)=(1−γ)x(t)+γƒ(W.sup.inu(t+1)+W.sup.resx(t)+W.sub.backy(t)) wherein W.sup.in represents an input weight matrix, W.sup.res represents a weight matrix of a reservoir state, W.sup.back represents a weight matrix of an output to the reservoir state, γ represents a leakage rate, t represents time, x(t) represents a previous state of a storage pool, ƒ(⋅) represents an activation function of a neuron, u(t+1) represents an input layer, and x(t+1) represents a next state of the storage pool; and an output equation of the network is as follows:
y(t)=g(W.sup.out[x(n); u(n)]) wherein W.sup.out represents an output weight matrix of the network, and g(⋅) represents an activation function of an output layer; training the established leaky-integrator echo state network, wherein in a training process, a least-square method is used to dynamically adjust a weight of the leaky-integrator echo state network, and an L1 norm constraint is added to an objective function of the least-square method according to the following formula: $G\left(n\right)=\frac{1}{2}\underset{m=0}{\overset{n}{.Math.}}{\lambda }^{n-m}{.Math.e\left(m\right).Math.}^2+\gamma {.Math.W^{\mathrm{out}}\left(n\right).Math.}_1$ wherein λ represents a forgetting factor at time n, e(m) represents an error signal, and γ represents a regularization parameter; and finally, obtaining the leaky-integrator echo state network based on a network weight obtained through training; and inputting original data to the leaky-integrator echo state network to obtain the generated data.

5. The transformer health state evaluation method according to claim 1, wherein in step 3, the artificial data is divided into two parts, wherein 80% of the artificial data is used as the training set to train the deep residual neural network, and 20% of the artificial data is used as the test set to verify an effect of transformer health state evaluation performed by the network.

6. The transformer health state evaluation method according to claim 1, wherein step 4 specifically comprises: constructing a residual module, comprising a normalization layer, a fully connected layer, a squeeze-and-excitation layer, and a threshold layer, wherein the residual module is composed of eight layers of networks, wherein a first layer is the normalization layer configured to normalize the data using a regularization method, wherein a data dimension of this layer is m×n×l; and a regularization formula is as follows: $\stackrel{.fwdarw.}{x_i}=\frac{x_i-{\mu }_B}{\sqrt{{\sigma }_B^2+\epsilon }}$ wherein μ.sub.B represents an average value of the data, σ.sub.B.sup.2 represents a variance of the data, x.sub.i represents input data, {right arrow over (x)}.sub.i represents output data, and ε represents a parameter; in the regularization formula, the average value of the data is calculated according to the following formula: ${\mu }_B=\frac{1}{m}\underset{i=1}{\overset{m}{.Math.}}{x}_i$ in the regularization formula, the variance of the data is calculated according to the following formula: ${\sigma }_B^2=\frac{1}{m}\underset{i=1}{\overset{m}{.Math.}}{\left(x_i-{\mu }_B\right)}^2$ in the normalization layer, after all the data is regularized, the data is changed such that the network can learn a rule of regularization, and a corresponding formula is as follows:
y.sub.i=γx.sub.i+β wherein γ and β represent parameters learned by the neural network; a second layer is a global average pooling layer configured to reduce a dimension of the data in transmission and reduce network training parameters, wherein a data dimension of this layer is m×n×l, and a calculation formula is as follows: $z_l=\frac{1}{M\times N}\underset{i=1}{\overset{M}{.Math.}}\underset{j=1}{\overset{N}{.Math.}}u\left(i,j\right)$ wherein u(i, j) represents input data, M and N represent data dimensions, and Z.sub.l represents output data; a third layer is an excitation layer configured to capture a dependency between channels, wherein a data dimension of this layer is m×n×(l/n), and a calculation formula is as follows:
s.sub.l=σ(W.sub.2δ(W.sub.1z)) wherein δ represents a ReLU activation function, $W_1\in R^{\frac{l}{b}\times l},\mathrm{and}{W}_2\in R^{l\times \frac{l}{b}};$ a fourth layer is an activation layer, wherein a data dimension of this layer is m×n×(l/n), and a corresponding formula is as follows:
x.sub.l=F.sub.scale(u.sub.l, s.sub.l)=s.sub.lu.sub.l wherein x=[x.sub.1, x.sub.2, . . . , x.sub.l], and F.sub.scale(u.sub.l, s.sub.l) represents calculation of a product of s.sub.l and input data; a fifth layer is the excitation layer, wherein a data dimension of this layer is m×n×l; a sixth layer is an activation function layer, wherein ‘sigmoid’ is used as an activation function, and a data dimension of this layer is m×n×l; a seventh layer is a soft threshold layer, wherein α is calculated according to the following formula, and then multiplied by data output from the first layer; $\alpha =\frac{1}{1+e^{-z}}$ wherein z represents data output from the sixth layer; and an eighth layer is an additive layer configured to add data output from the seventh layer and the data input into the first layer; and forming the deep residual neural network by an input layer, a residual layer, a softmax layer, and a classification output layer, wherein a specific training process of the neural network is as follows: using a cross entropy as an error function, using data generated by the leaky-integrator echo state network as training data of the neural network, and using an Adam optimization algorithm to update a weight of the deep residual neural network, wherein in a process of using the Adam optimization algorithm to update the weight of the deep residual neural network, if an average value within a range of consecutive specified steps is less than a specified threshold, training of the neural network is stopped and a weight that is of the deep residual neural network and calculated in the step 5 is saved; a formula of a cross entropy function used in the training process is as follows: $E=-\underset{j=1}{\overset{N_{\mathrm{class}}}{.Math.}}{t}_j\log \left(y_i\right)$ wherein t.sub.j represents an actual probability that an observation result belongs to category j, and N.sub.class represents a total quantity of classification results; and a formula of the Adam optimization algorithm in the training process is as follows: $x^{n+1}=x^n-\frac{\alpha }{\sqrt{\widehat{v_n}+\epsilon }}\widehat{m_n}\widehat{v_n}=\frac{v_n}{1-{\beta }_1^n}\widehat{m_n}=\frac{m_n}{1-{\beta }_2^n}{v}_n={\beta }_1{v}_{n-1}+\left(1-{\beta }_1\right)g_n^2{m}_n={\beta }_2{m}_{n-1}+\left(1-{\beta }_2\right)g_n$ wherein α represents a step size of movement, β.sub.1 and β.sub.2 represent two hyper-parameters, {circumflex over (v)}.sub.n represents correction of an estimated deviation from a first-order moment, {circumflex over (m)}.sub.n represents correction of an estimated deviation from a second-order moment, ε represents an auxiliary constant to prevent an update collapse; g.sub.n represents a gradient of the error function, and g.sub.n.sup.2 represents a second momentum of g.sub.n.

7. The transformer health state evaluation method according to claim 1, wherein step 5 specifically comprises: performing data filtering and data cleaning on the data according to step 2, then generating an artificial data set according to step 3, and finally performing health state evaluation based on the test set in step 4, wherein if a new data category or a relevant influencing factor needs to be added, the original network is used as a pre-training model to activate all layers for training.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present disclosure is described in further detail with reference to the accompanying drawings and embodiments.

[0045] FIG. 1 is a flowchart of implementing a transformer health state evaluation method based on a leaky-integrator echo state network and a deep residual neural network according to an embodiment of the present disclosure;

[0046] FIG. 2 shows a structure of leaky-integrator echo state network according to an embodiment of the present disclosure;

[0047] FIG. 3 shows a structure of a residual module according to an embodiment of the present disclosure; and

[0048] FIG. 4 shows a structure of a model based on a leaky-integrator echo state network and a deep residual neural network according to an embodiment of the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] To make the objectives, technical solutions, and advantages of the present disclosure clearer, the present disclosure is further described below in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present disclosure, rather than to limit the present disclosure.

[0050] The present disclosure is intended to provide a new evaluation method for a health state of a transformer, to achieve higher evaluation efficiency and accuracy and resolve a problem that a traditional method relies too much on selection of variable parameters and cannot evaluate the health state in the case of a plurality of variables.

[0051] The present disclosure is implemented by the following technical solutions.

[0052] As shown in FIG. 1, a transformer health state evaluation method based on a leaky-integrator echo state network and a deep residual neural network according to an embodiment of the present disclosure includes the following steps.

[0053] Step 1: Collect monitoring information of an oil test and contents of dissolved gas and furan in oil in each substation.

[0054] The data in step 1 is collected from records of test ledgers of a transformer and an electric power company during operation, where each group of data includes contents of nine key states: a BDV, water, acidity, hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), and furan, and a health state of the corresponding transformer.

[0055] Step 2: Perform data filtering, data cleaning and data normalization on the collected monitoring information to obtain an input matrix.

[0056] With reference to DL/T 1685-2017 Guide for Condition Evaluation of Oil Immersed Power Transformers, this embodiment classifies the transformer state into the following four levels: normal, cautionary, abnormal, and severe, as shown in Table 1.

TABLE-US-00001 TABLE 1 Corresponding relationship between an evaluation level and a relative deterioration degree and corresponding state description Relative deterioration Corresponding Sample degree state State description quantity 0.0-0.2 Normal Each index is stable and within a standard 217 limit. The transformer can operate normally. 0.2-0.5 Cautionary The index is developing towards the limit. 139 The transformer can continue to operate, but needs to be monitored and overhauled normally as planned. 0.5-0.8 Abnormal The index varies greatly, approaching or 103 slightly exceeding the standard limit. Power-outage overhaul shall be properly arranged. 0.8-1.0 Severe One or more indexes seriously exceed the 72 standard limit. Power-outage overhaul must be arranged as soon as possible.

[0057] Different transformers have unequal health state data samples. Sample data needs to be generated.

[0058] Step 3: Input the input matrix into a leaky-integrator echo state network to generate trainable artificial data, and divide the artificial data into a training set and a test set in proportion.

[0059] A model of the leaky-integrator echo state network is shown in FIG. 2. The leaky-integrator echo state network generates the artificial data according to the following formula:

y(t)=g(W.sup.out[x(n); u(n)])

[0060] Then, artificial data generated according to a formula x(t+1) is input into a deep residual neural network for training. A data set is divided into two parts, where 80% of the data set is used as the training set to train the deep residual neural network, and 20% of the data set is used as the test set to verify an effect of classifying the heath state by the network.

[0061] Step 4: Construct a deep residual neural network based on a squeeze-and-excitation network, and input the training set and the test set for network training.

[0062] The improved deep residual neural network in step 4 is constructed as follows: first, building a residual module, where a residual layer is composed of 50 residual modules, as shown in FIG. 3, and the residual module includes a normalization layer, a regularization layer, a fully connected layer, a squeeze-and-excitation layer, and a threshold layer; and then, forming a model by an input layer, the residual layer, a softmax layer, and a classification output layer. FIG. 4 shows a structure of the deep residual neural network, and a gate activation function of the deep residual neural network is σ. In the present disclosure, an activation function is ‘sigmoid’, and a state activation function is ‘tanh’.

[0063] In step 4, the constructed deep residual neural network is trained by using the input training set and test set.

[0064] Hyper-parameter settings of the network are shown in Table 2. After that, the deep residual neural network is trained and verified according to step 4 in the implementation to obtain a health state evaluation model.

TABLE-US-00002 TABLE 2 Hyper-parameter settings of the deep residual neural network Item Value Optimizer Adam Initial learning rate 0.01 Residual module 50 Minibatchsize 28 Maxepoch 1000

[0065] For simplicity, the following formula is used to calculate an accuracy rate. In practical application, an output result of the softmax layer can also be comprehensively considered. Each group of data corresponds to a probability of each transformer state label. A type corresponding to a maximum probability can be selected as a diagnosis result. In addition, when there is no significant difference between the second largest probability and the maximum probability in the softmax layer, the two diagnosis results can be comprehensively considered.

[00012]$\mathrm{Acc}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}$

[0066] In the above formula, TP represents a quantity of positive categories that are predicted as positive categories, TN represents a quantity of negative categories that are predicted as negative categories, FP represents a quantity of negative categories that are predicted as positive categories, and FN represents a quantity of positive categories that are predicted as negative categories.

[0067] Step 5: Perform health state evaluation and network weight update based on actual test data. The data set, obtained through monitoring, of the transformer is input for health evaluation. A final diagnosis accuracy rate is 93.4511%. Some evaluation results are selected, as shown in Table 3.

TABLE-US-00003 TABLE 3 Some fault diagnosis results based on the deep residual neural network Evaluation Actual result health state Normal Cautionary Abnormal Severe Normal Normal 99.47% 0.08% 0.08% 0.37% Cautionary Cautionary 0.05% 99.46% 0.00% 0.50% Abnormal Abnormal 0.06% 0.05% 98.99% 0.90% Normal Normal 99.62% 0.06% 0.03% 0.29% Severe Severe 0.03% 0.00% 0.22% 99.75% Normal Normal 98.86% 0.06% 0.09% 1.00% Abnormal Abnormal 0.05% 0.08% 98.99% 0.88% Abnormal Abnormal 0.10% 0.05% 98.81% 1.04% Normal Normal 99.62% 0.08% 0.09% 0.21% Severe Severe 0.07% 0.07% 0.44% 99.42% Abnormal Abnormal 0.10% 0.03% 99.00% 0.87% Severe Severe 0.01% 0.01% 0.98% 99.01% Abnormal Abnormal 0.00% 0.01% 98.85% 1.14% Severe Severe 0.00% 0.00% 0.00% 100.00% Normal Normal 99.35% 0.05% 0.04% 0.56% Abnormal Abnormal 0.05% 0.91% 98.95% 0.08% Normal Normal 99.29% 0.09% 0.04% 0.58% Cautionary Cautionary 0.70% 99.20% 0.05% 0.06% Severe Cautionary 0.06% 0.07% 0.64% 99.23% Abnormal Abnormal 0.02% 0.26% 99.66% 0.06% Cautionary Cautionary 0.22% 99.64% 0.05% 0.09% Cautionary Normal 0.68% 99.18% 0.09% 0.05% Severe Severe 0.10% 0.00% 0.52% 99.38% Cautionary Cautionary 0.00% 100.00% 0.00% 0.00% Severe Severe 0.09% 0.01% 0.54% 99.37% Cautionary Cautionary 0.20% 99.70% 0.01% 0.09% Severe Severe 0.02% 0.01% 1.00% 98.97%

[0068] The performing health state evaluation and network weight update based on actual test data in step 5 includes: for real-time monitoring data, performing data filtering and data cleaning according to step 2, then generating an artificial data set according to step 3, and finally performing health state evaluation based on the test set in step 4. If a new variable or a relevant influencing factor needs to be added, the original neural network is used as a pre-training model to activate all the layers for training.

[0069] It should be understood that those of ordinary skill in the art can make improvements or transformations based on the above description, and all these improvements and transformations should fall within the protection scope of the appended claims of the present disclosure.