Transformer oil gas concentration prediction method based on online ensemble learning

Abstract

The present application relates to the technical field of transformer fault diagnosis, and is especially a transformer oil gas concentration prediction method based on online integrated learning, comprising the following steps: S1: collecting gas data in transformer oil through a gas concentration sensor and generating a data training set through data processing; S2: using a nonlinear Kalman filtering algorithm to optimize parameters of two machine learning algorithms, namely a least squares support vector machine and a countermeasure neural network; S3: training new training data obtained through optimization and outputting a final prediction result. The present application processes collected gas data based on a scenario memory gradient method, uses a nonlinear Kalman filtering algorithm to take parameters of two machine learning algorithms, namely a least squares support vector machine and a countermeasure neural network, as state variables, outputs optimal parameters, and outputs an optimal prediction result according to optimal weights, thereby improving the accuracy of the prediction result and obtaining better learning effect than a single learning device.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer fault diagnosis, in particular to a transformer oil gas concentration prediction method based on online ensemble learning. BACKGROUND

[0002] The power transformer is a key equipment of the power system, and its operation state is closely related to the stability of the power grid. The dissolved gas analysis in transformer oil is an important method to judge its operation state. Predicting the dissolved gas content in transformer oil at future time can help operation and maintenance personnel to judge the future operation trend of the transformer and master the operation state in advance to ensure stable operation.

[0003] The wavelet analysis method is a widely used decomposition method, and its combination with neural network is the main method for prediction. However, due to the complex generation mechanism of dissolved gas in oil and the influence of special operation conditions of transformer, harsh operation environment, complex electromagnetic environment and other factors, the time series of dissolved gas in oil will show nonlinear and non-stationary characteristics. The traditional offline prediction method is difficult to mine these characteristics of time series, thus leading to low prediction accuracy.

[0004] Therefore, an online learning algorithm is proposed to update the model parameters in real time according to the real-time monitored gas concentration in transformer oil, aiming to achieve higher prediction accuracy. SUMMARY

[0005] The purpose of the present application is to solve the defects proposed in the background art by proposing a transformer oil gas concentration prediction method based on online ensemble learning.

[0006] The technical scheme adopted by the present application is as follows:

[0007] The present application provides a transformer oil gas concentration prediction method based on online ensemble learning, comprising the following steps:

[0008] S1: collecting gas data in transformer oil by a gas concentration sensor and generating a data training set by data processing;

[0009] S2: using a nonlinear Kalman filter algorithm to optimize the parameters of least squares support vector machine and adversarial neural network two machine learning algorithms;

[0010] S3: training with the new training data obtained by optimization and outputting the final prediction result.

[0011] As a preferred technical solution of the present application: in S1, the data processing includes processing by a episodic memory gradient method, in which the episodic memory gradient method, the collected gas data is stored by a storage unit, and a minimum loss function is defined, when new gas data is collected, the loss gradient vector is updated, it is judged whether the loss value is reduced, if not, the loss value is calculated again until it is reduced, if the loss value is reduced, the gas data stored in the storage unit is updated, and the storage unit generates a data training set according to the updated gas data.

[0012] As a preferred technical solution of the present application: in S1, after the generation of the data training set, the gas data in the training set is interpolated to supplement the missing values.

[0013] As a preferred technical solution of the present application: in S2, the nonlinear Kalman filtering algorithm is used, the parameters of the least squares support vector machine and the adversarial neural network two machine learning algorithms are iterated as state variables, and the optimal parameters are output.

[0014] As a preferred technical solution of the present application: in the nonlinear Kalman filtering algorithm, the parameters of the least squares support vector machine and the adversarial neural network two machine learning algorithms are iterated as state variables X k+1 , as follows:

[0015] X k+1 =f(X k )+V k

[0016] Z k+1 =H k+1 X k+1 +W k+1

[0017] Wherein, f(X k ) is a state function, Z k+1 is a measurement vector, H k+1 is a measurement matrix, V k and W k+1 are system noise and measurement noise respectively;

[0018] The state one-step prediction is:

[0019]

[0020] The state one-step prediction is:

[0021]

[0022] The state one-step prediction mean square error is:

[0023]

[0024] The filter gain is:

[0025]

[0026] The state estimate is

[0027]

[0028] The state estimation mean square error is:

[0029] P k+1,k+1 = (1 - K k+1 H k+1 ) P k+1,k

[0030] wherein, represents the state prediction value of the next position; represents the state matrix of the optimal estimation value of the current position; P k+1,k is the prediction mean square error matrix of the next position; P k,k represents the estimation mean square error matrix of the current position; φ k+1,k is the state transition matrix; J k+1,k is the Jacobian matrix; is the transpose of the Jacobian matrix; Q k represents the process noise covariance; K k+1 represents the Kalman filter gain; is the transpose of the observation matrix; R k+1 is the observation noise covariance; P k+1,k+1 represents the estimation error matrix; represents the optimal estimation value of the state of the next position.

[0031] As a preferred technical solution of the present application: in the least squares support vector machine and in the anti-type neural network, iteration is performed until the iteration termination condition is met, and finally the optimal parameter value is output.

[0032] As a preferred technical solution of the present application: in S2, the evaluation index values of the prediction results of the two kinds of machine learning algorithms are calculated based on the optimal parameters obtained.

[0033] As a preferred technical solution of the present application: the evaluation index values include the accuracy of the gas data, the forward transition value and the backward transition value.

[0034] As a preferred technical solution of the present application: in S3, the evaluation index values of the two kinds of machine learning algorithms are normalized and calculated, the evaluation index values are weighted by the analytic hierarchy process, the prediction results are finally weighted by the entropy weight method, and the prediction value of this time is output based on the final weighting.

[0035] As a preferred technical solution of the present application: the analytic hierarchy process is as follows:

[0036] x=w1U1+w2U2+w3U3

[0037] Wherein, x is the score value of the gas data, U1, U2 and U3 are accuracy, forward transition value and backward transition value respectively;

[0038] The judgment matrix is constructed, and the accuracy, the forward transition value and the backward transition value are the judgment matrix criteria. The judgment matrix is composed of numerical values 1-9 and their inverse numbers, A=(u ij ) u×v , wherein A is the target, u i , u j is the evaluation index, u ij is u i The relative importance value of u j , wherein:

[0039]

[0040] Wherein, u ij Indicates the ratio of the importance of index i and index j, Indicates the ratio of the importance of index j and index i;

[0041] The judgment matrix R is composed of u ij :

[0042]

[0043] The characteristic vector w of the maximum eigenvalue v max of the judgment matrix is solved, and the specific process is as follows:

[0044] PW=μ max W

[0045] After solving the characteristic vector W, it is normalized to obtain w=(w1, w2, w3), w1, w2, w3 are the corresponding weights of the accuracy, the forward transition value and the backward transition value;

[0046] The entropy weight method is as follows:

[0047] According to the results of the analytic hierarchy process, the entropy weight evaluation matrix x=(x ab ) m×n (a∈[1,m])(b∈[1,n]) indicates the index value of the bth gas data in the ath measurement. The data of each index is processed as follows:

[0048]

[0049] Wherein a is the a-th measured gas data, b is the score data of the b-th gas data, x ab is the index value of the b-th gas data measured in the a-th time, y ab is the standardized matrix, minx b , maxx b is the minimum value and the maximum value of the same gas data;

[0050] The proportion p of the index value of the b-th gas data measured in the a-th time is calculated ab :

[0051]

[0052] The entropy value r of each gas data is calculated b :

[0053]

[0054] The entropy weight of the gas data is calculated:

[0055]

[0056] Wherein, w b represents the entropy weight of the gas data.

[0057] The transformer oil gas concentration prediction method based on online integrated learning provided by the present application has the following beneficial effects compared with the prior art:

[0058] The present application processes the collected gas data based on the scenario memory gradient method, takes the parameters of the least square support vector machine and the adversarial neural network two machine learning algorithms as state variables through the nonlinear Kalman filtering algorithm, outputs the optimal parameters, evaluates the evaluation index values of the two machine learning algorithms, finally outputs the optimal weight, and outputs the optimal prediction result, thereby improving the accuracy of the prediction result and obtaining better learning effect than a single learner. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is the method flowchart of the preferred embodiment of the present application;

[0060] Figure 2 is the scenario memory gradient method flowchart in the preferred embodiment of the present application;

[0061] Figure 3 is the algorithm parameter optimization flowchart in the preferred embodiment of the present application;

[0062] Figure 4 is the weight distribution method flowchart in the preferred embodiment of the present application. DETAILED DESCRIPTION

[0063] It should be noted that the embodiments in the embodiments and the features in the embodiments can be combined with each other without conflict, and the technical solutions in the embodiments will be described clearly and completely in combination with the drawings of the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0064] With reference to Figure 1 The preferred embodiment of the application provides an online integrated learning-based transformer oil gas concentration prediction method, comprising the following steps:

[0065] S1: collecting gas data in transformer oil by a gas concentration sensor and generating a data training set by data processing;

[0066] S2: optimizing parameters of two machine learning algorithms of least squares support vector machine and adversarial neural network by using a nonlinear Kalman filter algorithm;

[0067] S3: training by new training data obtained by optimization and outputting final prediction results.

[0068] In S1, the data processing includes processing by a scenario memory gradient method, in which the collected gas data is stored by a storage unit, and a minimum loss function is defined. When new gas data is collected, the loss gradient vector is updated, and it is determined whether the loss value is reduced. If not, the loss value is calculated again until it is reduced. If the loss value is reduced, the gas data stored in the storage unit is updated, and the storage unit generates a data training set according to the updated gas data.

[0069] In S1, after the data training set is generated, the gas data in the training set is interpolated to supplement the missing values.

[0070] In S2, the nonlinear Kalman filter algorithm is used to take the parameters of the two machine learning algorithms of least squares support vector machine and adversarial neural network as state variables for iteration, and output the optimal parameters.

[0071] In the nonlinear Kalman filter algorithm, the parameters of the two machine learning algorithms of least squares support vector machine and adversarial neural network are taken as state variables X k+1 , and the specific process is as follows:

[0072] X k+1 =f(X k )+V k

[0073] Z k+1 =Hk+1 X k+1 +W k+1

[0074] where f(X k ) is a state function, Z k+1 is a measurement vector, H k+1 is a measurement matrix, V k and W k+1 are system noise and measurement noise, respectively;

[0075] The state one-step prediction is:

[0076]

[0077] The state one-step prediction is:

[0078]

[0079] The state one-step prediction mean square error is:

[0080]

[0081] The filter gain is:

[0082]

[0083] The state estimate is

[0084]

[0085] The state estimate mean square error is:

[0086] P k+1,k+1 = (1 - K k+1 H k+1 ) P k+1,k

[0087] where denotes the state prediction value at the next position; denotes the state matrix of the optimal estimate value at the current position; P k+1,k is the prediction mean square error matrix at the next position; P k,k denotes the estimation mean square error matrix at the current position; φ k+1,k is the state transition matrix; J k+1,k is the Jacobian matrix; is the transpose of the Jacobian matrix; Q k denotes the process noise covariance; K k+1 denotes the Kalman filter gain; is the transpose of the observation matrix; R k+1 is the observation noise covariance; P k+1,k+1 denotes the estimation error matrix; The state optimal estimation value of the next position is represented.

[0088] In the least squares support vector machine and the anti-type neural network, iterations are performed until an iteration termination condition is met, and finally an optimal parameter value is output.

[0089] In the S2, evaluation index values of prediction results of the two kinds of machine learning algorithms are calculated based on the obtained optimal parameters respectively.

[0090] The evaluation index values include accuracy, forward transition value and backward transition value of the gas data.

[0091] In the S3, the evaluation index values of the two kinds of machine learning algorithms are normalized and calculated, the evaluation index values are weighted by the analytic hierarchy process, the prediction results are finally weighted by the entropy weight method, and the prediction value of this time is output based on the final weighting.

[0092] The analytic hierarchy process is specifically as follows:

[0093] x = w1U1 + w2U2 + w3U3

[0094] Wherein, x is the score value of the gas data, U1, U2 and U3 are accuracy, forward transition value and backward transition value respectively;

[0095] A judgment matrix is constructed, and accuracy, forward transition value and backward transition value are used as the judgment matrix criteria. The judgment matrix is composed of numerical values 1-9 and their inverse numbers, A = (u ij ) u×v , wherein A is the target, u i , u j is the evaluation index, and u ij is the relative importance value of u i to u j , wherein:

[0096]

[0097] Wherein, u ij represents the ratio of the importance of index i to index j, represents the ratio of the importance of index j to index i;

[0098] The judgment matrix R is composed of u ij :

[0099]

[0100] The characteristic vector w of the maximum eigenvalue μ max of the judgment matrix is solved, and the specific process is as follows:

[0101] PW = μ max W

[0102] After solving for the eigenvector W, normalize it to obtain w = (w1, w2, w3), where w1, w2, and w3 are the corresponding weights of accuracy, forward transition value, and backward transition value.

[0103] The entropy weight method is described in detail below.

[0104] Based on the results of the analytic hierarchy process, construct the entropy weight evaluation matrix x = (x ab ) m×n (a∈[1,m])(b∈[1,n]) represents the index value of the b-th gas data in the a-th measurement. The data of each index are processed as follows:

[0105]

[0106] Where a represents the gas data from the a-th measurement, b represents the score of the b-th gas data, and x... ab Let y be the index value of the b-th gas data from the a-th measurement. ab For the normalized matrix, minx b maxx b These are the minimum and maximum values ​​for the same gas data.

[0107] Calculate the weight p of the index value of the b-th gas data from the a-th measurement. ab :

[0108]

[0109] Calculate the entropy value r of each gas data. b :

[0110]

[0111] Calculate the entropy weights of gas data:

[0112]

[0113] Among them, w b Entropy weights represent gas data.

[0114] In this embodiment, various gas data in the transformer oil are collected using a gas concentration sensor, with reference to... Figure 2 The system stores the collected gas data in storage units. Whenever a new dataset arrives, a loss function is calculated. If, after multiple projection offsets, the loss value is less than a set value, the storage content is updated, retaining the most effective historical information. The data in the storage units, along with new data of a finite length, serves as the training set for the next prediction. This approach minimizes the forgetting of historical information while maintaining limited memory and fast computation.

[0115] Referring to Figure 3 , after the data training set is generated, the gas data therein is interpolated to supplement missing values. Then, through a nonlinear Kalman filtering algorithm, parameters of two machine learning algorithms, least squares support vector machine and adversarial neural network, are taken as state variables, and parameters of the least squares support vector machine are taken as state variables X k+1 For example, known data is taken as input, and the specific process is as follows:

[0116]

[0117] Z k+1 =H k+1 X k+1 +W k+1

[0118] Wherein, f1(X k ) and f2(X k ) are state functions of the value and the rate of change, respectively, Z k+1 is a measurement vector, H k+1 is a measurement matrix, V k and W k+1 are system noise and measurement noise, respectively;

[0119] It is assumed that one-step prediction of the state according to parameters of the least squares support vector machine is as follows:

[0120]

[0121] Wherein, α k+1,k and β k+1,k represent the value optimal estimate and the rate optimal estimate of the parameters of the least squares support vector machine at the kth position;

[0122] The mean square error of one-step prediction of the state is as follows:

[0123]

[0124]

[0125] The filtering gain is as follows:

[0126]

[0127] The state estimate is as follows:

[0128]

[0129] The mean square error of the state estimate is as follows:

[0130] P k+1,k+1 =(1-K k+1 H k+1 )P k+1,k

[0131] wherein, represents the state prediction value of the next position; P k+1,k is the predicted mean square error matrix of the next position; P k,k represents the estimated mean square error matrix of the current position; φ k+1,k is the state transition matrix; J k+1,k is the Jacobian matrix, i.e. the partial derivative of the state equation with respect to the state variable; is the transpose of the Jacobian matrix; Q k represents the process noise covariance; K k+1 represents the Kalman filter gain; is the transpose of the observation matrix; R k+1 is the observation noise covariance; P k+1,k+1 represents the estimation error matrix; represents the optimal state estimation value of the next position.

[0132] In the least squares support vector machine, the covariance matrix and the state vector are updated, and iteration is performed until the iteration termination condition is met, and the optimal parameters are output. In the adversarial neural network, the optimal parameter selection is output by the nonlinear Kalman filter algorithm, the measurement error and the prediction error are balanced, and the optimal parameter selection is given. After new data arrives, the accuracy, backward transfer value and forward transfer value of the last prediction result of the two algorithms are given. Among them, the accuracy measures the gap between the predicted value and the actual value, the backward transfer ability measures the influence of the model learning a new task on the current task, and the forward transfer ability measures the influence of learning the past task on the current task.

[0133] With reference to Figure 4 , the accuracy, backward transfer value and forward transfer value of the two machine learning algorithms are normalized, and the evaluation index values are weighted by the analytic hierarchy process:

[0134] x = w1U1 + w2U2 + w3U3

[0135] wherein, x is the score value of the gas data, U1, U2 and U3 are the accuracy, forward transfer value and backward transfer value respectively;

[0136] A judgment matrix is constructed, with the accuracy 1, the forward transfer value 2 and the backward transfer value 3 as the judgment matrix criteria. The judgment matrix is composed of numerical values 1-9 and their inverses, A = (u ij ) u×v , wherein A is the target, u i , u j is the evaluation index, and u ij is u i u jThe relative importance values of the two indicators can be set as follows: 1—same importance; 3—slightly more important of the former compared to the latter; 5—more important of the former compared to the latter; 7—obviously more important of the former compared to the latter; 9—strongly more important of the former compared to the latter; 2, 4, 6, 8—intermediate values between the adjacent judgments; wherein:

[0137]

[0138] wherein, u 12 represents the ratio of the importance of the accuracy rate 1 to the forward transition value 2, represents the ratio of the importance of the forward transition value 2 to the accuracy rate 1;

[0139] The judgment matrix R is composed of u ij :

[0140]

[0141] The characteristic vector w of the maximum eigenvalue μ max of the judgment matrix is solved, and the specific process is as follows:

[0142] PW = μ max W

[0143] After the characteristic vector W is solved, it is normalized to obtain w = (w1, w2, w3), wherein w1, w2 and w3 are the corresponding weights of the accuracy rate, the forward transition value and the backward transition value;

[0144] The final weighting of the prediction result is performed by the entropy weight method:

[0145] According to the analytic hierarchy process result, an entropy weight evaluation matrix x = (x ab ) m×n (a∈[1,10])(b∈[1,25]) represents the index value of the bth gas data in the ath measurement, and the data of each index is processed as follows:

[0146]

[0147] wherein, a is the gas data in the ath measurement, b is the score data of the bth gas data, x ab is the index value of the bth gas data in the ath measurement, y ab is a standardization matrix, minx b and maxx b are the minimum value and the maximum value of the same gas data;

[0148] The proportion p ab of the index value of the bth gas data in the ath measurement is calculated:

[0149]

[0150] Calculate the entropy value r of each gas data b :

[0151]

[0152] Calculate the entropy weight of gas data:

[0153]

[0154] Wherein, w b represents the entropy weight of gas data.

[0155] Based on the final weighting, output the prediction value of this time combined with the output of the two algorithms.

[0156] It is apparent to those skilled in the art that the present application is not limited to the details of the foregoing exemplary embodiments, and that the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all respects as illustrative and not restrictive, and the scope of the present application should be defined by the appended claims rather than the above description, and it is intended to encompass all changes falling within the meaning and scope of the equivalent elements of the claims. Any reference signs in the claims should not be considered as limiting the claims to which they belong.

[0157] Furthermore, it should be understood that, although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be combined appropriately to form other embodiments that those skilled in the art can understand.

Claims

1. A method for predicting gas concentration in transformer oil based on online ensemble learning, characterized in that: Includes the following steps: S1: Collect gas data in transformer oil using a gas concentration sensor and process the data to generate a data training set; S2: The nonlinear Kalman filter algorithm is used to optimize the parameters of two machine learning algorithms: least squares support vector machine and adversarial neural network. S3: Train using the new training data obtained through optimization and output the final prediction result; In S1, the data processing includes processing using the context memory gradient method. In the context memory gradient method, the collected gas data is stored in a storage unit, and a minimum loss function is defined. When new gas data is collected, the loss gradient vector is updated, and it is determined whether the loss value has decreased. If it has not decreased, the loss value is recalculated until it decreases. If the loss value has decreased, the gas data stored in the storage unit is updated, and the storage unit generates a data training set based on the updated gas data. In step S2, the evaluation index values ​​of the prediction results of the two machine learning algorithms are calculated based on the optimal parameters of the two algorithms. The evaluation index values ​​include the accuracy of the gas data, the forward transfer value, and the backward transfer value. In step S3, the evaluation index values ​​of the two machine learning algorithms are normalized, the evaluation index values ​​are weighted by the analytic hierarchy process, the prediction results are finally weighted by the entropy weighting method, and the prediction value is output based on the final weighting.

2. The method for predicting gas concentration in transformer oil based on online ensemble learning according to claim 1, characterized in that: In step S1, after the data training set is generated, the gas data in the training set is interpolated to supplement the missing values.

3. The method for predicting gas concentration in transformer oil based on online ensemble learning according to claim 2, characterized in that: In S2, a nonlinear Kalman filter algorithm is used to iterate the parameters of two machine learning algorithms, least squares support vector machine and adversarial neural network, as state variables and output the optimal parameters.

4. The method for predicting gas concentration in transformer oil based on online ensemble learning according to claim 3, characterized in that: In the aforementioned nonlinear Kalman filter algorithm, the parameters of both least squares support vector machine and adversarial neural network machine learning algorithms are used as state variables. The details are as follows: ； ； in, For state functions, For measurement vectors, For the measurement matrix, and These are system noise and measurement noise, respectively. The state is predicted in one step as follows: ； The state is predicted in one step as follows: ； The mean square error of the state prediction step is: ； The filter gain is: ； The state is estimated as follows: ； The mean square error of the state estimation is: ； in, This indicates the predicted state value for the next position; The state matrix representing the optimal estimate of the current position; This is the mean square error matrix for the prediction of the next position; The mean square error matrix representing the estimated position; This is the state transition matrix; It is a Jacobian matrix; This is the transpose of the Jacobian matrix; Represents the process noise covariance; Indicates the Kalman filter gain; This is the transpose of the observation matrix; To observe the noise covariance; Represents the estimation error matrix; This represents the optimal estimate of the state at the next position.

5. The method for predicting gas concentration in transformer oil based on online ensemble learning according to claim 4, characterized in that: In the least squares support vector machine and the anti-neural network, the iteration continues until the iteration termination condition is met, and finally the optimal parameter values ​​are output.

6. The method for predicting gas concentration in transformer oil based on online ensemble learning according to claim 5, characterized in that: The Analytic Hierarchy Process (AHP) is described in detail below: ； in, The score for the gas data. , and These are accuracy, forward shift value, and backward shift value, respectively. Construct a judgment matrix, using accuracy, forward transition value, and backward transition value as the judgment criteria. The judgment matrix consists of values ​​1-9 and their reciprocals. ,in, With the goal, , As evaluation indicators, for right The relative importance values, where: ； in, Indicators With indicators The ratio of importance Indicators With indicators The ratio of importance; Depend on Composition of judgment matrix : ； Finding the largest eigenvalue from the judgment matrix eigenvectors The details are as follows: ； Solving for eigenvectors Then, normalize it to obtain , , , These are the corresponding weights of accuracy, forward shift value, and backward shift value; The entropy weight method is described in detail below. Construct an entropy weight evaluation matrix based on the results of the analytic hierarchy process. Indicates the first The first measurement The index values ​​for each gas data point are processed as follows: ； in, For the first Gas data from the previous measurement For the first The score data for each gas data point For the first The first measurement The index value of each gas data point, For a standardized matrix, , These are the minimum and maximum values ​​for the same gas data. Calculate the first The first measurement The proportion of the index value of each gas data : ； Calculate the entropy value of each gas data. : ； Calculate the entropy weights of gas data: ； in, Entropy weights represent gas data.

Images