Electronic current transformer fault early warning model training method and fault early warning method

Abstract

The invention discloses an electronic current transformer fault early warning model training method, and also discloses a current transformer fault early warning method for performing early warning by using a model obtained by the training method. According to the fault model training method provided by the invention, the Transform network is combined with the BiLSTM network for prediction, and the problem of precision reduction during long-time sequence data prediction is solved. Meanwhile, the NRBO algorithm is adopted to iterate the learning rate of the BiLSTM network, the number of neurons in the hidden layer and the regularization coefficient in the training process, the problems that the training time is prolonged, the fitting effect is reduced and the prediction precision is reduced due to the fact that relevant parameters are set through experience of research and development personnel are solved, and prediction is more accurate.

Description

Technical Field

[0001] This invention belongs to the field of active DC electronic current transformer fault early warning technology, and relates to a fault early warning model training method for electronic current transformers, as well as a fault early warning method using the trained model for early warning. Background Technology

[0002] As the core equipment of the control and protection system for DC transmission projects, the safe operation and normal functioning of the ECT (Electronic Transmission Control System) are essential for ensuring the power generation benefits of the project. However, due to its long-term operation in complex environments such as strong electromagnetic interference and high and low outdoor temperatures, the ECT's operational stability is increasingly evident. Therefore, how to use time-series monitoring data during ECT operation to provide early warning of potential faults is of great significance and has promising application prospects. Currently, time-series prediction methods can be mainly divided into statistical methods and machine learning methods. Statistical methods include the Autoregressive Integrated Moving Average (ARIMA) model and the Exponential Smoothing (ES) model. The accuracy of these methods depends heavily on the distribution of the dataset. They are not only sensitive to outliers and difficult to capture nonlinear relationships, but their accuracy may also decline over the long term. Machine learning methods mainly include Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Random Forest (RF). While these methods have improved accuracy, they still cannot solve the problem of declining accuracy in long-term series data prediction. Summary of the Invention

[0003] The purpose of this invention is to provide a fault early warning method for electronic current transformers, which solves the problem of decreased accuracy in long-term data prediction in the prior art.

[0004] Another objective of this invention is to provide a method for early warning of current transformer faults.

[0005] The technical solution adopted in this invention is a training method for an electronic current transformer fault early warning model, which is carried out according to the following steps: Step 1: Analyze the historical fault data of ECT, select the time series data of the merged unit data level and the laser drive current as the monitoring data, and export the monitoring data into training set and test set; Step 2: Initialize the Transformer-BiLSTM model and set the initial parameters. Train the model using the training set and iteratively set the parameters using the NRBO algorithm. Step 3: After the iteration is complete, select the optimal value of the set parameters to establish a fault early warning model; Step 4: Test the fault warning model using a test set. When the test results meet the standards, the fault warning model can be used for early warning.

[0006] The invention is further characterized by: In step 1, after selecting the time series data of the merged unit data level and the laser drive current, the obtained data is sliced ​​according to a certain time interval and divided into training set and test set in a 7:3 ratio.

[0007] In step 2, when initializing the Transformer-BiLSTM model, the maximum number of encodings in the Transformer is set to 512, the number of heads in the multi-head self-attention mechanism is 4, and the number of key channels is set to 32. At the same time, the internal parameters of BiLSTM are trained using the Adam algorithm, the ReLU function is selected as the activation function in the hidden layer, the learning rate decrease factor is set to 0.1, and the number of iterations is 20.

[0008] The parameters set in step 2 include the learning rate, the number of hidden layer neurons, and the regularization coefficient.

[0009] The iterative process of the NRBO algorithm in step 2 is as follows: Step 2.1: Set upper and lower limits for the parameters to be optimized, and then generate random initial populations within the corresponding boundaries. Each population consists of fuzzy decision variables and vectors. The population generation rule is as follows:

[0010] in, and These are the upper and lower bounds of the learning rate, respectively. Indicates the first in the population The individual in the first Position in each dimension Representing a random number between 0 and 1, the initial population is obtained by grouping each individual. ; Step 2.2: Calculate the fitness of each individual in the population, and determine the position corresponding to the best and worst fitness of each individual. , ; Step 2.3: Apply the Newton-Raphson optimization algorithm to calculate the... Population matrix in the next iteration The Location update of each individual The calculation formula is:

[0011] in, For the current number Population matrix in the next iteration The The location of each individual This represents a normally distributed random number with a mean of 0 and a variance of 1. Indicates the updated baseline; Step 2.4: Update the amount based on location Calculate candidate direction vectors , and The calculation method is as follows:

[0012]

[0013]

[0014] Then based on the candidate direction vector , and The population matrix after this iteration is obtained. The The position of an individual The calculation method is as follows:

[0015] in, and All numbers are random numbers between 0 and 1; Step 2.5: Introduce random perturbations for optimization, specifically in conjunction with the optimal location. and Constructing new solutions with enhanced quality The expression is:

[0016] in This is the optimal solution for the individual. All random numbers are introduced perturbations. This represents the set of position vectors of all individuals in the entire population at the t-th iteration. Step 2.6: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] As the individual optimal solution, it replaces the global worst position from the previous iteration. This yields a set containing the optimal configuration of learning rate, number of hidden layer neurons, and regularization coefficients.

[0017] In step 2, training terminates when the fitness tends to stabilize or the number of iterations reaches its maximum value.

[0018] In step 4, the fault warning model is tested using a test set, and the coefficient of determination is used. As an evaluation indicator, the calculation method is as follows:

[0019] in, For the observed values, Model predictions The average of the predicted values, when If the model is found to be suitable for fault diagnosis, then the iteration count is increased until the model meets the requirements.

[0020] Another technical solution adopted in this invention is a fault early warning method for current transformers, which is carried out according to the following steps: Step 1: Extract the data level received by the selected merging unit and the time series data of the laser drive current during the ECT operation; Step 2: Input the data level received by the parallel unit and the time series data of the laser driving current into the current transformer fault early warning model to obtain the prediction data; Step 3: A fault occurs when the laser drive current is greater than 1100mA or the data level is less than 500mV.

[0021] The beneficial effects of this invention are: The model training method proposed in this application uses a BiLSTM network to learn and predict time-series data of electronic current transformers, achieving better results than traditional statistical methods such as ARIMA and ES. Furthermore, the special gating unit structure introduced in BiLSTM effectively solves the accuracy degradation problem of models such as SVM, RNN, and RF when predicting long-term series data. Simultaneously, this application uses the NRBO algorithm to further iteratively obtain the optimal values ​​for the learning rate, number of hidden layer neurons, and regularization coefficient of BiLSTM, overcoming a series of problems caused by setting empirical values, such as increased training time, decreased fitting effect, and reduced prediction accuracy. In addition, the encoder structure and self-attention mechanism introduced in the Transformer neural network solve the feature vanishing and efficiency degradation problems that occur in serial data processing, more effectively extracting key information from the time-series signal. The results of the numerical example analysis demonstrate that the prediction accuracy of the model proposed in this paper is significantly improved compared to other models. Attached Figure Description

[0022] Figure 1 This is a flowchart of the fault early warning model training method for sub-type current transformers of the present invention; Figure 2 This is a comparison chart of the relative errors in predicting drive current using the model trained by this invention and the conventional model; Figure 3 This is a comparison chart of the relative errors in predicting received data levels using the model trained by this invention and a conventional model. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0024] Example 1: like Figure 1 As shown, the training method for the fault early warning model of electronic current transformers is carried out according to the following steps: Step 1: Analyze the historical fault data of ECT, select the time series data of the merged unit data level and the laser drive current as the monitoring data, and export the monitoring data into training set and test set; Step 2: Initialize the Transformer-BiLSTM model and set the initial parameters. Train the model using the training set and iteratively set the parameters using the NRBO algorithm. Step 3: After the iteration is complete, select the optimal value of the set parameters to establish a fault early warning model; Step 4: Test the fault warning model using a test set. When the test results meet the standards, the fault warning model can be used for early warning.

[0025] The Transformer-BiLSTM model combines BiLSTM and Transformer networks. It encodes time-series data through the Transformer network and adds the encoded information to the time-series signal as the input to the BiLSTM network. This effectively avoids the feature vanishing and efficiency degradation problems that occur in the BiLSTM model when processing data serially.

[0026] Example 2: like Figure 1 As shown, the training method for the fault early warning model of electronic current transformers is carried out according to the following steps: Step 1: Analyze the historical fault data of ECT, select the time series data of the merged unit data level and the laser drive current, slice the obtained data according to a certain time interval, and divide it into training set and test set according to a 7:3 ratio; Step 2: Initialize the Transformer-BiLSTM model and set the initial parameters. Train the model using the training set and iteratively set the parameters using the NRBO algorithm. Step 3: After the iteration is complete, select the optimal value of the set parameters to establish a fault early warning model; Step 4: Test the fault warning model using a test set. When the test results meet the standards, the fault warning model can be used for early warning.

[0027] Example 3: like Figure 1 As shown, the training method for the fault early warning model of electronic current transformers is carried out according to the following steps: Step 1: Analyze the historical fault data of ECT, select the time series data of the merged unit data level and the laser drive current as the monitoring data, and export the monitoring data into training set and test set; Step 2: Initialize the Transformer-BiLSTM model and set the maximum number of encodings in the Transformer to 512, the number of heads in the multi-head self-attention mechanism to 4, and the number of key channels to 32. Set the learning rate, number of hidden layer neurons, and regularization coefficient of BiLSTM. Train the model using the training set and iterate the learning rate, number of hidden layer neurons, and regularization coefficient using the NRBO algorithm. Step 3: After the iteration is complete, select the optimal value of the set parameters to establish a fault early warning model; Step 4: Test the fault warning model using a test set. When the test results meet the standards, the fault warning model can be used for early warning.

[0028] Example 4: like Figure 1 As shown, the training method for the fault early warning model of electronic current transformers is carried out according to the following steps: Step 1: Analyze the historical fault data of ECT, select the time series data of the merged unit data level and the laser drive current as the monitoring data, and export the monitoring data into training set and test set; Step 2: Initialize the Transformer-BiLSTM model and set the initial parameters, then train the model using the training set; Step 3: Set upper and lower limits for the parameters to be optimized, and then generate random initial populations within the corresponding boundaries. Each population consists of fuzzy decision variables and vectors. The population generation rule is as follows:

[0029] in, and These are the upper and lower bounds of the learning rate, respectively. Indicates the first in the population The individual in the first Position in each dimension Representing a random number between 0 and 1, the initial population is obtained by grouping each individual. ; Step 4: Calculate the fitness of each individual in the population, and determine the position corresponding to the best and worst fitness of each individual. , ; Step 5: Apply the Newton-Raphson optimization algorithm to calculate the... Population matrix in the next iteration The Location update of each individual The calculation formula is:

[0030] in, For the current number Population matrix in the next iteration The The location of each individual This represents a normally distributed random number with a mean of 0 and a variance of 1. Indicates the updated baseline; Step 6: Update the amount based on location Calculate candidate direction vectors , and The calculation method is as follows:

[0031]

[0032]

[0033] Then based on the candidate direction vector , and The population matrix after this iteration is obtained. The The position of an individual The calculation method is as follows:

[0034] in, and All numbers are random numbers between 0 and 1; Step 7: Introduce random perturbations for optimization, specifically in conjunction with the optimal location. and Constructing new solutions with enhanced quality The expression is:

[0035] in This is the optimal solution for the individual. All random numbers are introduced perturbations. This represents the set of position vectors of all individuals in the entire population at the t-th iteration. Step 8: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] As the individual optimal solution, it replaces the global worst position from the previous iteration. This yields a set containing the optimal configuration of learning rate, number of hidden layer neurons, and regularization coefficients; Step 9: Terminate training when the fitness tends to stabilize or the number of iterations reaches the maximum value, and select the optimal value of the set parameters to establish a fault warning model; Step 10: Test the fault warning model using a test set. When the test results meet the standards, the fault warning model can be used for early warning.

[0036] Example 5: like Figure 1 As shown, the training method for the fault early warning model of electronic current transformers is carried out according to the following steps: Step 1: Analyze the historical fault data of ECT, select the time series data of the merged unit data level and the laser drive current as the monitoring data, and export the monitoring data into training set and test set; Step 2: Initialize the Transformer-BiLSTM model and set the initial parameters. Train the model using the training set and iteratively set the parameters using the NRBO algorithm. Step 3: After the iteration is complete, select the optimal value of the set parameters to establish a fault early warning model; Step 4: Test the fault warning model using a test set and use the coefficient of determination. The test results are evaluated when If the model is suitable for fault diagnosis, then the number of iterations is increased until the model meets the requirements.

[0037] Example 6: like Figure 1 As shown, the fault early warning method for current transformers is implemented according to the following steps: Step 1: Extract the data level received by the selected merging unit and the time series data of the laser drive current during the ECT operation; Step 2: Input the data level received by the parallel unit and the time series data of the laser driving current into the current transformer fault early warning model to obtain the prediction data; Step 3: A fault occurs when the laser drive current is greater than 1100mA or the data level is less than 500mV. Operation is terminated when the data level is less than 350mV.

[0038] The NRBO-Transformer-BiLSTM model obtained using the training method of this application was used for prediction, while RNN, BiLSTM, Transformer, and Transformer-BiLSTM models were used as controls. The prediction errors are as follows: Figure 2 and Figure 3 As shown in the table below, the performance of the model was evaluated using the coefficient of determination, mean absolute error, and root mean square error. The evaluation of the laser drive current prediction is shown in the table below:

[0039] The evaluation of the received data level of the merging unit is shown in the table below:

[0040] As shown in the table above, the prediction model obtained using the training method provided in this application has a higher coefficient of determination, as well as lower mean absolute error and root mean square error, compared to conventional prediction models. Therefore, the model trained in this application is more accurate in predicting the data level received by the merging unit and the laser drive current, and can effectively predict information during the operation of the electronic current transformer. Predicting future operating information using existing operating data for fault diagnosis yields more accurate results.

Claims

1. A method for training a fault early warning model of an electronic current transformer, characterized in that, According to the following steps: Step 1: analyze the historical failure data of ECT, select the time series data of the combined unit data level and the laser drive current as the monitoring data, and divide the monitoring data into training set and test set; Step 2: initialize the Transformer-BiLSTM model and set the initial setting parameters, train the model using the training set, and continuously iterate the setting parameters through the NRBO algorithm; Step 3: After iteration, select the optimal value of the setting parameters as the fault early warning model; Step 4: test the fault early warning model using the test set, and when the test result meets the standard, the fault early warning model at this time can be used for early warning.

2. The electronic current transformer fault early-warning model training method according to claim 1, characterized in that, The data obtained after selecting the time series data of the combined unit data level and the laser drive current in step 1 is sliced according to a certain time interval, and is divided into training set and test set according to the ratio of 7:

3. 3.The electronic current transformer fault pre-warning model training method of claim 1, wherein, In step 2, when initializing the Transformer-BiLSTM model, the maximum encoding number in the Transformer is set to 512, the number of heads in the multi-head self-attention mechanism is set to 4, the key channel number is set to 32, the Adam algorithm is used to train the internal parameters of BiLSTM, the activation function in the hidden layer is selected as Relu function, the learning rate reduction factor is set to 0.1, and the iteration number is 20 times.

4. The electronic current transformer fault early-warning model training method according to claim 1, characterized in that, The setting parameters in step 2 include learning rate, number of hidden layer neurons and regularization coefficient.

5. The electronic current transformer fault early-warning model training method according to claim 4, characterized in that, The iteration process of the NRBO algorithm in step 2 is as follows: Step 2.1: set the upper and lower limits for the parameters to be optimized, then generate a random initial population within the corresponding boundary, each population consists of fuzzy decision variables and vectors, and the population generation rule is: wherein, and are the upper and lower bounds of the learning rate, respectively, denotes the position of the th individual in the th dimension in the population, denotes a random number between 0 and 1, and the initial population is obtained by collecting each individual; Step 2.2: Calculate the fitness of each individual in the population, derive the position corresponding to the best and worst fitness of each individual , ; Step 2.3: Apply the Newton-Raphson optimization algorithm to calculate the... Population matrix in the next iteration The Location update of each individual The calculation formula is: wherein is the current iteration number, is the population matrix of the iteration number is the position of the individual number is a random number of a normal distribution with mean 0 and variance 1, is the update reference; Step 2.4: Position update amount Computing candidate direction vectors , and are computed as follows: The position of the i-th individual of the population matrix after this iteration is then computed as follows: , and ​​​​ wherein and are random numbers between 0 and 1. Step 2.5: Introduce random perturbations for optimization, in particular in combination with best positions and Construct new solution with enhanced quality , the expression being: wherein is the optimal solution for the individual, is a random number belonging to the introduced perturbation vector, represents the set of position vectors of all individuals in the entire population at the tth iteration; Step 2.6: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] As the individual optimal solution, it replaces the global worst position from the previous iteration. This yields a set containing the optimal configuration of learning rate, number of hidden layer neurons, and regularization coefficients.

6. The electronic current transformer fault early-warning model training method according to claim 5, characterized in that, In step 2, when the fitness tends to be stable or the iteration number reaches the maximum value, the training is terminated.

7. The electronic current transformer fault early-warning model training method according to claim 1, characterized in that, The step 4 tests the failure early warning model by using the test set, and uses the coefficient of determination As an evaluation index, the calculation method is as follows: where, is the observed value, is the model predicted value, is the average of the predicted values, when the model can be used for fault diagnosis, otherwise increase the number of iterations until the model meets the requirements.

8. A method for current transformer fault early warning, characterized in that, The model obtained by the electronic current transformer fault early warning model training method of claim 6 is obtained by the following steps: Step 1: extract the selected combined unit receiving data level and laser drive current time series data during ECT operation; Step 2: input the combined unit receiving data level and laser drive current time series data into the current transformer fault early warning model to obtain the predicted data; Step 3: when the laser drive current is greater than 1100mA or the data level is less than 500mV, it represents a fault, and when the data level is less than 350mV, the operation is terminated.

Images