Power transformer fault diagnosis method based on stacked time series network

Through the stacked time series network combined with Xgboost and bidirectional gating neural network, power transformer fault diagnosis is used using gas data in oil, solving the problem of failure to effectively utilize time series data in the existing technology, and achieving high-precision fault prediction and diagnosis .

CN115146739BActive Publication Date: 2024-11-05WUHAN UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210884861.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-07-26
Publication Date
2024-11-05
Estimated Expiration
2042-07-26

AI Technical Summary

Technical Problem

The prior art cannot effectively use time series data for future time diagnosis of power transformer failures, and traditional methods lead to low diagnostic accuracy in data redundancy and modal similarity.

Method used

Using a stacked time series network method, combined with Xgboost and a bidirectional gating neural network, a fault diagnosis model is constructed by extracting and predicting gas data in oil, and a fault state output of the Softmax layer is used to use a meta-learner.

Benefits of technology

It realizes high-precision diagnosis of power transformer faults, improves the accuracy and stability of fault prediction at future moments, and solves the problem of not considering time series and low accuracy in traditional methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115146739B_ABST
    Figure CN115146739B_ABST
Patent Text Reader

Abstract

The invention discloses a method for diagnosing faults of power transformers based on a stacked time series network, comprising: collecting gas in oil of transformers in each substation; performing z-score normalization on the collected data to obtain a normalized matrix; dividing the normalized matrix into a training set and a test set in proportion; constructing a stacked time series network based on Xgboost and a bidirectional gated recurrent neural network, inputting the training set and the test set for network training; using the real-time collected data to obtain trainable data after normalization, and performing fault prediction and updating of network parameters. The invention uses Xgboost and a gated neural network to predict gas in oil data, obtains power transformer prediction data from two time series networks by a meta-learner, and obtains the fault diagnosis result of the transformer through a Softmax layer. The neural network has accurate fault diagnosis performance and stable robustness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of power transformer fault diagnosis, and more specifically, relates to a power transformer fault diagnosis method based on a stacked time series network. Background Art

[0002] High-voltage and large-capacity transformers are one of the most important electrical equipment in the power system. They have a complex structure and are expensive. They play an important role in the safe and stable operation of the power system. Once a transformer fails, it will cause significant economic losses. Power transformers are very important voltage conversion nodes and are the core equipment of substations. In addition, power transformers are expensive, and sudden failures and shutdowns will seriously endanger the safety of the power system. However, long-term preventive tests and regular maintenance cannot meet the operation and maintenance requirements of high-voltage and large-capacity transformers. How to safely supply power without affecting people's quality of life and work efficiency is a huge challenge facing my country's power workers.

[0003] Power transformer fault diagnosis is an effective means to determine the fault status of the transformer. The data accumulated during the operation of the transformer contains rich status information and is an effective basis for status assessment and fault diagnosis. With the comprehensive construction and promotion of smart grids, multi-source and heterogeneous power big data has shown explosive growth, and the traditional oil-gas fault diagnosis method cannot meet the development needs of power companies. At present, the fault diagnosis method performs fault diagnosis on the current temporal data, but there is a trend of change when the transformer fails, and new fault diagnosis technology based on time series data still needs to be studied. However, in the actual task of time series prediction, due to the similarity between modal data, there is a certain redundancy in the time series data. If a single time series model is used, the performance of the model on the target task will often be reduced. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention proposes a power transformer fault diagnosis method based on a stacked time series network, which solves the problem that the prior art relies too much on the selection of various variable parameters and cannot diagnose the fault status at future times. The method has simple operation, high diagnostic accuracy, and can easily realize the diagnosis of transformer faults at future times.

[0005] To achieve the above object, the present invention provides a method for diagnosing power transformer faults based on a stacked time series network, comprising:

[0006] (1) Collecting information on gas in oil from each substation, where the information on gas in oil includes: oil testing, dissolved gas in oil, and furan content monitoring information from each substation;

[0007] (2) performing data normalization processing on the collected gas-in-oil information to obtain a normalized matrix;

[0008] (3) Divide the normalized matrix into training set and test set in proportion to train the parameters of the network;

[0009] (4) Construct a stacked time series network based on Xgboost and bidirectional gated neural network, input the training set and test set to train the stacked time series network, and learn the characteristics of gas data in transformer oil;

[0010] (5) Perform fault diagnosis on real-time oil-gas data and fine-tune the weights of the stacked time series network so that the stacked time series network can continuously learn new features.

[0011] In some optional embodiments, the gas in oil information includes: data recorded by operating transformers and power companies, wherein each set of data includes gas in oil data and its corresponding transformer fault status, and the gas in oil data contains the contents of nine key states: BDV, water content, acidity, hydrogen, methane, ethane, ethylene, acetylene, and furan content.

[0012] In some optional implementation schemes, step (2) includes: performing z-score normalization processing on the gas in oil information to obtain a normalized matrix.

[0013] In some optional implementation schemes, in step (3), the data in the oil gas information is divided into two parts, wherein a certain proportion of the data is used as a training set to train the stacked time series network, and the remaining proportion of the data is used as a test set to test the stacked time series network's transformer fault diagnosis effect.

[0014] In some optional embodiments, step (4) comprises:

[0015] (4.1) Constructing a stacked time series network based on Xgboost and a bidirectional gated neural network to extract and predict the features of gas in oil, where the construction of Xgboost includes the establishment of an integrated model, the selection of an objective function, and the solution of a loss function; the bidirectional gated neural network includes forward calculation, reverse calculation, update gate, and reset gate;

[0016] (4.2) Use Xgboost and bidirectional gated neural network to predict gas in oil and output the prediction results of gas in oil information;

[0017] (4.3) The prediction results of Xgboost and bidirectional gated neural network are trained using the meta-learner to output the prediction results of gas in oil information. The stacked time series network is used for fault diagnosis through the Softmax layer to output the fault status of the transformer.

[0018] In some optional implementation schemes, the construction of Xgboost includes the establishment of an integrated model, the selection of an objective function, and the solution of a loss function, wherein the establishment of the integrated model is to recursively construct a binary decision tree, and in the input space of the training data set, each region is recursively divided into two sub-regions using the square error minimization criterion, and the output value of each sub-region is determined; the objective function is selected to measure the error between the predicted value and the target true value, and the objective function is approximated by a second-order Taylor expansion method; the loss function is solved by using a greedy algorithm to divide the subtrees and enumerate feasible segmentation points, that is, each time a new segmentation is added to the existing leaves, and the maximum gain obtained thereby is calculated.

[0019] In some optional implementations, the bidirectional gated neural network includes forward calculation, reverse calculation, update gate and reset gate, wherein the reset gate helps to capture short-term dependencies in the time series; the update gate helps to capture long-term dependencies in the time series; the forward calculation and the reverse calculation process the input sequence in turn.

[0020] In some optional implementation schemes, the meta-learner trains and predicts the results of Xgboost and the bidirectional gated neural network, wherein the meta-learner is constructed as a linear regression model to learn and predict the results of Xgbosst and the bidirectional gated neural network.

[0021] In some optional embodiments, step (5) comprises:

[0022] For the real-time collected oil-gas data, z-score data normalization is performed, and then the normalized data is divided into training set and test set. The stacked time series network is trained for fault diagnosis. If new data categories or related influencing factors need to be added, the original stacked time series network is used as a pre-training model and all layers are activated for training.

[0023] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0024] The power transformer fault diagnosis method based on stacked time series network of the present invention takes into account the problems of not considering time series and low accuracy in the typical transformer fault diagnosis process, and uses stacked time series network for fault diagnosis. Taking into account the different prediction capabilities and feature extraction capabilities of stacked time series network, a fault diagnosis model of stacked time series network based on Xgboost and bidirectional gated neural network is innovatively constructed. On this basis, considering the problem of inconsistent units in actual engineering data, the model input data is z-score normalized. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a method for diagnosing power transformer faults based on a stacked time series network provided by an embodiment of the present invention;

[0026] Figure 2 It is an Xgboost structure provided by an embodiment of the present invention;

[0027] Figure 3 It is a bidirectional gated neural network structure provided by an embodiment of the present invention;

[0028] Figure 4 It is a structure based on a stacked time series network model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0030] The purpose of the present invention is to provide a new fault diagnosis method for transformer status, which has higher fault efficiency and accuracy, and solves the problem that traditional methods do not consider fault diagnosis in time series situations.

[0031] The present invention is achieved by adopting the following technical solutions:

[0032] like Figure 1 As shown, the power transformer fault diagnosis method based on the stacked time series network of the embodiment of the present invention comprises the following steps:

[0033] Step 1: First, collect the oil test of each substation, the monitoring information of dissolved gas and furan content in the oil;

[0034] The data in step 1 are collected from the operating transformer and the test data of the power company, and each set of data includes the contents of nine key states, namely BDV, water content, acidity, hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), and furan content, and their corresponding transformer fault states.

[0035] Step 2: Perform data normalization on the collected gas in oil information to obtain a normalized matrix;

[0036] Specifically, step 2 can be implemented in the following ways:

[0037] The data of gas in oil are normalized by using the z-score normalization method, and the formula is:

[0038]

[0039] Among them, μ and σ are the mean and variance of the original data set.

[0040] This embodiment divides the transformer fault status into the following four levels: low temperature overheating fault (T1), medium temperature overheating fault (T2), high temperature overheating fault (T3), partial discharge fault (PD), low energy discharge fault (D1), high energy discharge fault (D2), as shown in Table 1:

[0041] Table 1 Transformer fault status

[0042] Fault Status Number of samples Low temperature overheating fault (T1) 256 Medium temperature overheating fault (T2) 179 High temperature overheating fault (T3) 141 Partial Discharge Fault (PD) 98 Low energy discharge fault (D1) 221 High energy discharge fault (D2) 194

[0043] Step 3: Divide the normalized matrix into training set and test set in proportion;

[0044] The normalized matrix dataset is divided into two parts, 80% of which is used as a training set to train the stacked time series network, and 20% is used as a test set to test the classification effect of the stacked time series network on the fault state.

[0045] Step 4: Build a stacked time series network based on Xgboost and bidirectional gated neural network, and input the training set and test set for network training;

[0046] The construction method of the stacked time series network in step 4 is: first build the Xgboost model, such as Figure 2 As shown, ; then construct a bidirectional gated neural network, such as Figure 3 As shown; Finally, the meta-learner is used to learn the results of Xgboost and bidirectional gated neural network, as shown Figure 4 shown.

[0047] Among them, step 4 can be specifically implemented in the following ways:

[0048] Step 4.1: Building the Xgboost model is the process of recursively building a binary decision tree. The tree ensemble model is as follows:

[0049]

[0050] in, Measures the error between the predicted value and the target true value. is the predicted value of the model, y i is the target true value, l represents the number of decision trees combined with the number of trees that need to be adjusted, and f l is the lth tree, x i represents the i-th input sample, and F represents the set of all tree models; is a regular term to control the complexity of the model to prevent overfitting of the model, Ω(f) = λT + 1 / 2·λ||w|| 2 In the formula, T is the number of leaf nodes, λ and γ represent the penalty coefficients for the model, and w represents the weight of the leaf node in a tree. The model adds an incremental function f in each round. t (x i ) minimizes the loss function to the greatest extent. The objective function of the tth round can be written as:

[0051]

[0052] To find the objective function that minimizes t (x i ), the second-order Taylor expansion method is used to approximate the objective function. Define the sample set of each leaf of the jth tree as I j ={i|q(x i =j)}. Among them, and are the first-order derivative and second-order derivative of the loss function respectively. From this we can get:

[0053]

[0054] definition And taking partial derivative of w, we get:

[0055]

[0056] Substituting the weights into the objective function, we get:

[0057]

[0058] The smaller the loss function, the better the model. A greedy algorithm is used to divide the subtree and enumerate feasible split points, that is, new splits are added to the existing leaves each time and the maximum gain obtained is calculated.

[0059] Step 4.2: The composition of the bidirectional gated neural network is as follows:

[0060] The bidirectional GRU structure can transform the input data x i =[x1,...,x n ] T The processing is performed in both the forward and reverse directions, and then the two obtained feature vectors are concatenated together as another expression of the input vector.

[0061] 1) The forward calculation formula is:

[0062]

[0063] 2) The reverse calculation formula is:

[0064]

[0065] Among them, r t 、z t is the update gate and reset gate at the current time step; g t is the hidden state at the current moment; h t-1 and h t are the states of the previous moment and the current moment respectively; W xz , W xr and W xg is connected to the input vector x t The weight matrix W hz , W hr and W hg is connected to the previous unit state vector h t-1 The weight matrix of z 、b r and b g is the bias vector; σ is the activation function, which is the sigmoid function and the gating signal of the bidirectional gated neural network; tanh is the activation function that scales the data to between [-1,1]. Finally, the forward calculation is and the reverse calculation The results are added together as the final output of the model.

[0066]

[0067] Step 4.3: The meta-learner is constructed as a linear regression model, which is composed as follows:

[0068] The prediction results of Xgboost and bidirectional gated neural network form a new training set, and the univariate linear regression model is trained to build a stacked time series network. The univariate linear regression model is:

[0069] y=mx+b (10)

[0070] The results of stacked time series network predictions are input into the Softmax layer to obtain the fault status of the transformer.

[0071] For simplicity, the accuracy is calculated using the following formula. In practical applications, the output results of the softmax layer can also be considered comprehensively. The probability of each set of data corresponding to each transformer state label is not only selected as the diagnosis result, but also when the second largest probability value in the softmax is not significantly different from the maximum probability value, the two diagnosis results can be considered comprehensively.

[0072]

[0073] Among them, TP refers to the number of positive predictions as positive, FP refers to the number of negative predictions as positive, and FN refers to the number of positive predictions as negative.

[0074] Step 5: Combine actual test data for fault diagnosis and fine-tuning of the neural network. Input the transformer data set obtained by monitoring for fault diagnosis. The final diagnosis accuracy is 98.87%.

[0075] The method for performing fault diagnosis and network parameter fine-tuning in combination with actual test data in step 5 is as follows: for the real-time collected oil-gas data, the data is normalized according to step 2, and then the data is divided into a training set and a test set according to step 3, the stacked time series network is trained, and finally fault diagnosis is performed according to the test set of step 5. If new data categories or related influencing factors need to be added, the original stacked time series network is used as a pre-training model, and all layers are activated for training.

[0076] The present invention collects gas in oil from transformers in each substation; performs data normalization processing on the collected monitoring information to obtain a normalized matrix; divides the normalized matrix into a training set and a test set in proportion; constructs a stacked time series network based on Xgboost and a bidirectional gated neural network, inputs the training set and the test set for network training; and uses the trainable data obtained after normalization of the real-time collected data to perform fault diagnosis and update network parameters. The present invention considers the problem of gas in oil data in power transformer data, and the fault diagnosis accuracy of gas in oil data is poor. Xgboost and a bidirectional gated neural network are used to predict gas in oil data, and finally the prediction result of gas in oil is obtained by a meta-learner, and the fault state of the transformer is obtained by a Softmax layer. The neural network has accurate evaluation performance and stable robustness.

[0077] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0078] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for fault diagnosis of power transformer based on stacked time series network, characterized in that: include: (1) Collecting information on gas in oil from each substation, where the information on gas in oil includes: oil testing, dissolved gas in oil, and furan content monitoring information from each substation; (2) performing data normalization processing on the collected gas-in-oil information to obtain a normalized matrix; (3) Divide the normalized matrix into training set and test set in proportion to train the parameters of the network; (4) Construct a stacked time series network based on Xgboost and bidirectional gated neural network, input the training set and test set to train the stacked time series network, and learn the characteristics of gas data in transformer oil; (5) Perform fault diagnosis on real-time oil-gas data and fine-tune the weights of the stacked time series network so that the stacked time series network can continuously learn new features; Step (4) comprises: (4.1) A stacked time series network based on Xgboost and a bidirectional gated neural network is constructed to extract and predict the features of gas in oil. The construction of Xgboost includes the establishment of an integrated model, the selection of an objective function, and the solution of a loss function. The establishment of an integrated model is to recursively construct a binary decision tree. In the input space of the training data set, each region is recursively divided into two sub-regions using the square error minimization criterion, and the output value of each sub-region is determined. The objective function is selected to measure the error between the predicted value and the target true value, and the second-order Taylor expansion method is used to approximate the objective function. The solution of the loss function uses a greedy algorithm to divide the subtrees and enumerate feasible segmentation points, that is, each time a new segmentation is added to the existing leaves, and the maximum gain obtained is calculated. Among them, the tree ensemble model is It is a measure of the error between the predicted value and the target true value. is the predicted value of the model, y i is the target true value, l represents the number of decision trees combined with the number of trees that need to be adjusted, and f l is the lth tree, x i represents the i-th input sample, F represents the set of all tree models, is the regularization term that controls the complexity of the model, Ω(f)=γT+1 / 2·λ||w|| 2 , T is the number of leaf nodes, λ and γ represent the penalty coefficients for the model, and w represents the weight of the leaf node in a tree; The bidirectional gated neural network includes forward calculation, reverse calculation, update gate and reset gate. The reset gate helps to capture the short-term dependency in the time series; the update gate helps to capture the long-term dependency in the time series; the forward calculation and reverse calculation process the input sequence in turn; The forward calculation formula is: The reverse calculation formula is: r t 、z t is the update gate and reset gate at the current time step; g t is the hidden state at the current moment; h t-1 and h t are the states of the previous moment and the current moment respectively; W xz , W xr and W xg is connected to the input vector x t The weight matrix W hz , W hr and W hg is connected to the previous unit state vector h t-1 The weight matrix of z 、b r and b g is the bias vector; σ is the activation function, which is the sigmoid function and the gating signal of the bidirectional gated neural network; tanh is the activation function that scales the data to between [-1,1]; (4.2) Use Xgboost and bidirectional gated neural network to predict gas in oil and output the prediction results of gas in oil information; (4.3) The prediction results of Xgboost and bidirectional gated neural network are trained by meta-learner, and the prediction results of gas in oil are output. The stacked time series network is fault diagnosed through the Softmax layer, and the fault status of the transformer is output. The prediction results of Xgboost and bidirectional gated neural network form a new training set, and the univariate linear regression model is trained to construct the stacked time series network.

2. The method according to claim 1, characterized in that The gas in oil information includes: data recorded by operating transformers and power companies, wherein each set of data includes gas in oil data and its corresponding transformer fault status, and the gas in oil data contains the contents of nine key states: BDV, water content, acidity, hydrogen, methane, ethane, ethylene, acetylene, and furan content.

3. The method according to claim 1, characterized in that Step (2) includes: performing z-score normalization processing on the gas in oil information to obtain a normalized matrix.

4. The method according to any one of claims 1 to 3, characterized in that: In step (3), the data in the oil gas information is divided into two parts, wherein a certain proportion of the data is used as a training set to train the stacked time series network, and the remaining proportion of the data is used as a test set to test the stacked time series network's effect on transformer fault diagnosis.

5. The method according to claim 4, characterized in that Step (5) comprises: For the real-time collected oil-gas data, z-score data normalization is performed, and then the normalized data is divided into training set and test set. The stacked time series network is trained for fault diagnosis. If new data categories or related influencing factors need to be added, the original stacked time series network is used as a pre-training model and all layers are activated for training.