Method and system for predicting service life of all-fiber current transformer

Abstract

The invention belongs to the technical field of power distribution monitoring, and particularly relates to an all-fiber current transformer service life prediction method and system. The method comprises the following steps: acquiring historical data of all-fiber current transformer light source driving current and modulator driving voltage, and forming time sequence data; inputting the time sequence data into a pre-trained neural network model to obtain a prediction result of each state parameter within a set time; the state parameters comprise light source driving current and modulator driving voltage; respectively calculating a periodic average change value of each state parameter in the set time period according to the obtained prediction result of each state parameter in the set time period and the current value of each state parameter; according to the threshold upper / lower limit, the periodic average change value and the current value of each state parameter, the residual service life corresponding to each state parameter is calculated, the minimum value of the two is selected as the residual service life of the all-fiber current transformer, and the prediction efficiency and the prediction accuracy are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power distribution monitoring, and particularly relates to a full-optical-fiber current transformer life prediction method and system. BACKGROUND

[0002] Remaining Useful Life (RUL) prediction of equipment is one of important functions of industrial Internet of Things to realize industrial intelligence, which can predict future operation state of equipment based on health monitoring data of the equipment, so as to obtain remaining use time of the equipment before losing operation capability, and thus to formulate a corresponding predictive maintenance strategy, and improve reliability, availability and safety of industrial Internet of Things equipment. As important equipment in a direct current converter station, reliable operation of a full-optical-fiber current transformer is related to overall safety of the system, but its health state will continuously degrade with increasing use time, and when various performance degradations reach a certain threshold, it is considered that the life of the current transformer reaches an end, and if the current transformer is not timely repaired or replaced, it will bring a major hidden danger to safe operation of the power system. In order to timely take preventive equipment maintenance measures and ensure accuracy and reliability of equipment operation, it is very necessary to predict the remaining life of the current transformer. In recent years, equipment manufacturers and converter substation sites are carrying out modification in state quantity monitoring, and through monitoring of some state parameters of the current transformer and setting of a simple threshold alarm, a preliminary state maintenance mode of the equipment is realized, but the intelligent degree is insufficient, and evaluation of operation state of the current transformer and analysis of the remaining life mostly rely on manual analysis, which greatly increases workload of operation and maintenance personnel, and prediction efficiency is low and prediction accuracy is also low. SUMMARY

[0003] The application aims to provide a full-optical-fiber current transformer life prediction method and system, to solve the problem of low prediction accuracy of the existing full-optical-fiber current transformer life prediction method.

[0004] The application provides a full-optical-fiber current transformer life prediction method, which comprises the following steps:

[0005] obtaining historical data of light source driving current and modulator driving voltage of the full-optical-fiber current transformer, and constituting time sequence data, the time sequence data comprising N time sequence vectors obtained by concatenating the light source driving current and the modulator driving voltage at the same time, N≥7;

[0006] inputting the time sequence data into a pre-trained neural network model to obtain prediction results of each state parameter within a set time; the state parameters comprise the light source driving current and the modulator driving voltage;

[0007] According to the obtained prediction results of each state parameter in the set time and the current value of the state parameter, the periodic average change value of each state parameter in the set time period is calculated respectively;

[0008] According to the upper / lower threshold value of each state parameter, the periodic average change value and the current value, the residual service life corresponding to each state parameter is calculated, and the minimum value between the two is selected as the residual service life of the all-fiber current transformer.

[0009] Further, the method for calculating the residual service life corresponding to each state parameter according to the upper / lower threshold value of each state parameter, the periodic average change value and the current value is:

[0010] When the periodic average change value is positive, the upper threshold value of each state parameter is subtracted from the current value, and then divided by the periodic average change value to calculate the residual service life corresponding to each state parameter;

[0011] When the periodic average change value is negative, the lower threshold value of each state parameter is subtracted from the current value, and then divided by the periodic average change value to calculate the residual service life corresponding to each state parameter.

[0012] Further, the set time is less than or equal to 7 days.

[0013] Further, the time series data input into the pre-trained neural network model further includes an environmental temperature, and the time series data includes N time series vectors obtained by concatenating the environmental temperature, the light source driving current and the modulator driving voltage at the same time.

[0014] Further, the neural network model is a CNN-GRU model, which includes a CNN network layer and a GRU network layer; when predicting, the time series data input into the CNN-GRU model is first processed through the CNN network layer to extract the feature quantity of the time series data, then the feature quantity is flattened and converted into a time series sequence, and then input into the GRU network layer for prediction, and the prediction results of each state parameter in the set time are output.

[0015] Further, the CNN network layer includes three convolution layers and three pooling layers.

[0016] Further, the GRU network layer includes three GRUs and one fully connected layer.

[0017] Further, the time series data is preprocessed before being input into the pre-trained neural network model; the preprocessing includes, when the time series data of the light source driving current, the modulator driving voltage and the environmental temperature has a null value, the mean of the time series data before and after the null value is filled into the null value; if a continuous missing sequence occurs, the mean of the time series data at the moment before the beginning of the missing sequence and the moment after the end of the missing sequence is filled into the entire missing sequence.

[0018] Further, the preprocessing further includes normalizing the time series data of the light source driving current, the modulator driving voltage and the environmental temperature, and when the time series data is normalized, the prediction results of the state parameters within a set time obtained in the subsequent neural network model need to be de-normalized.

[0019] The beneficial effects of the above technical solutions are that: the present application is an improved invention, aiming at the prediction of the remaining useful life of the all-fiber current transformer, the two state parameters of the light source driving current and the modulator driving voltage with high correlation degree with the life of the current transformer are selected for the RUL prediction of the current transformer, and the values of the two state parameters within a set time are predicted by using the pre-trained neural network model, then the periodic average change value of the state parameter values is analyzed according to the prediction results of the state parameters within the set time, and then the remaining life of the equipment is further analyzed according to the periodic average change value, the upper / lower threshold of each state parameter and the current value of the state parameter, without manual analysis, the prediction efficiency and accuracy are improved.

[0020] To solve the above technical problems, the present application also provides an all-fiber current transformer life prediction system, comprising a processor, the processor is used to realize the all-fiber current transformer life prediction method introduced above.

[0021] The beneficial effects of the above technical solutions are that: the present application is an improved invention, aiming at the prediction of the remaining useful life of the all-fiber current transformer, the two state parameters of the light source driving current and the modulator driving voltage with high correlation degree with the life of the current transformer are selected for the RUL prediction of the current transformer, and the values of the two state parameters within a set time are predicted by using the pre-trained neural network model, then the periodic average change value of the state parameter values is analyzed according to the prediction results of the state parameters within the set time, and then the remaining life of the equipment is further analyzed according to the periodic average change value, the upper / lower threshold of each state parameter and the current value of the state parameter, without manual analysis, the prediction efficiency and accuracy are improved. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is the all-fiber current transformer life prediction flowchart of the method embodiment of the present application;

[0023] Figure 2 is a neural network model input data structure diagram of an embodiment of the method of the present application;

[0024] Figure 3 is a CNN-GRU model network structure diagram of an embodiment of the method of the present application;

[0025] Figure 4 is a comparison diagram of the prediction results of the driving current of the light source of different models in the verification example of the present application;

[0026] Figure 5 is a comparison diagram of the prediction results of the driving voltage of the modulator of different models in the verification example of the present application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical scheme and advantages of the present application more clear and obvious, the specific embodiments of the present application will be further described below with reference to the drawings.

[0028] The present application aims at the problem that the existing artificial analysis leads to low efficiency and low accuracy of RUL prediction of the all-fiber current transformer, and proposes a method and system for RUL prediction based on a neural network model. For the prediction of the remaining useful life of the all-fiber current transformer, the two state parameters of the light source driving current and the modulator driving voltage which have a high degree of correlation with the life of the current transformer are selected for RUL prediction of the current transformer. The pre-trained neural network model is used to predict the parameter values of the two state parameters within a set time, and then the periodic average change of the state parameter values is analyzed according to the predicted values of the state parameters. Then, the remaining life of the equipment is further analyzed according to the periodic average change, the upper / lower threshold of each state parameter and the current value of the state parameter. No manual analysis is required, which improves the prediction efficiency and prediction accuracy, avoids errors caused by directly using the neural network model for RUL prediction, efficiently and accurately analyzes the life of the all-fiber current transformer, avoids the major hidden dangers to the safe operation of the power system caused by not timely maintenance or replacement of the equipment, and ensures the accuracy and reliability of the operation of the equipment.

[0029] Method embodiment

[0030] The present application provides a method for predicting the life of an all-fiber current transformer. The implementation principle of the method is as shown in Figure 1 , which will be described in detail below.

[0031] 1. Determine the key parameters for predicting the life of the current transformer.

[0032] The light source driving current and the modulator driving voltage are more relevant to the service life of the all-fiber current transformer. The service life of the transformer can be estimated by predicting and analyzing the two state parameters. Therefore, the light source driving current and the modulator driving voltage values in a set time are predicted according to the existing light source driving current and modulator driving voltage. Further, considering that environmental parameters such as outdoor temperature also affect the operation state of the device, as a preferred embodiment, in order to further improve the prediction accuracy, the light source driving current and the modulator driving voltage are predicted by the environmental temperature, the light source driving current and the modulator driving voltage, and the service life is estimated according to the prediction result. In actual use, the light source driving current and the modulator driving voltage in a set time can also be predicted by combining other influencing factors on the basis of the light source driving current and the modulator driving voltage.

[0033] 2. Data access.

[0034] With the continuous development of Internet of Things technology, the real-time and accuracy of device data collection have been greatly improved. With the help of the Internet of Things platform, the operation data of the current transformer can be collected, uploaded and persisted. Then, based on a large amount of historical data, subsequent data processing and prediction analysis processes can be performed.

[0035] 3. Data preprocessing.

[0036] 1) Null data processing: Due to measurement errors, sensor abnormalities and other factors, the collected various raw data often contains some null values. According to the time sequence characteristics of the data, the null data is completed by using the mean value method of the front and rear data, and the mean value of the front and rear time sequence data of the null data is filled into the null data. If a continuous missing sequence occurs, the entire missing sequence is filled and completed by using the mean value of the time point before the missing sequence and the time point after the missing sequence.

[0037] 2) Normalization processing: The data dimensions of different influencing factors are different, and the numerical difference is large. In order to ensure the consistency of the influence of different parameters on the output value, and to improve the convergence speed of the neural network loss function and the model prediction accuracy, the raw data such as the light source driving current, the environmental temperature and the modulator driving voltage are normalized. Min-Max normalization or Z-Score normalization processing can be used. In this embodiment, the Min-Max method is used for normalization processing, and the calculation formula is:

[0038]

[0039] Wherein, x is the sample data to be normalized; x * is the normalized value; x min , x maxThe minimum value and the maximum value in the sample data before normalization, respectively.

[0040] 3) Feature information coupling: Obtain the historical data of the light source driving current and the modulator driving voltage of the all-fiber current transformer, and form time series data. When using a neural network model to predict the time series data, a time series feature map needs to be used as input. Since the environmental temperature, light source driving current and modulator driving voltage are essentially independent time series, the feature information of these parameters needs to be coupled. In this embodiment, the word vector method used in processing natural language is used for coupling. The environmental temperature, light source driving current and modulator driving voltage at the same time are concatenated into a time series vector. At this time, the time series data includes N time series vectors obtained by concatenating the environmental temperature, light source driving current and modulator driving voltage at the same time. Alternatively, the light source driving current and modulator driving voltage at the same time are concatenated into a time series vector. At this time, the time series data includes N time series vectors obtained by concatenating the light source driving current and modulator driving voltage at the same time, thereby obtaining the time series data after feature information coupling. The larger N is, the more data is needed for prediction, and the better the prediction effect is. However, the larger N is, the larger the amount of calculation is. The smaller N is, the less data is needed for prediction, and the smaller the amount of calculation is. Therefore, N≥7, which means that at least 7 days of time series vectors obtained by concatenating the environmental temperature, light source driving current and modulator driving voltage at the same time are needed. Considering the amount of calculation and the prediction effect, the value of N in this embodiment is 16. In other embodiments, N can be set according to specific prediction requirements and computer operation capacity.

[0041] Further, the time series data after feature information coupling is used as the element of neural network input by using the method of sliding window. In this embodiment, the width of the sliding window is set to 16, and the step is set to 1, both in days. Finally, the structure of the input data is obtained, as shown in the following figure: Figure 2 wherein t is the time scale of the coordinate axis; T is the time, representing a day; n is the step; and T+n represents the nth day after T.

[0042] 4) Using the pre-trained neural network model to predict the light source driving current and the modulator driving voltage of the all-fiber current transformer within a certain time.

[0043] In theory, the remaining service life of the all-fiber current transformer can be directly estimated according to the neural network prediction model, but in the actual prediction process, considering that the long-term prediction error will be transmitted, the prediction accuracy will continuously decrease with the increase of the prediction time length, and the prediction accuracy will also sharply decrease when the prediction time span is larger, so the remaining service life of the all-fiber current transformer cannot be directly predicted by the neural network prediction model. Therefore, the neural network model trained is not used to directly predict the remaining service life of the equipment, but the neural network model trained is used to first perform short-term prediction on the state parameters, then the subsequent change trend of the parameters is analyzed according to the short-term prediction result of the state parameters, and finally the remaining life of the equipment is further analyzed according to the trend, and the short term refers to a set time less than or equal to 7 days, and in the embodiment, the short term is set to 7 days, and in other embodiments, the short term can be set to 6 days, 5 days or 4 days, etc.

[0044] The neural network model can use an existing neural network model such as an RNN (recurrent neural network) model, a CNN (convolutional neural network) model, a CNN-GRU (convolutional neural network-gated recurrent unit) model, etc. In use, the time series data coupled with the feature information is input into the neural network model to directly obtain the light source driving current prediction value and the modulator driving voltage prediction value within the set time.

[0045] Compared with a single CNN or GRU model, the combination model of CNN-GRU can fully combine the advantages between the two neural networks, has stronger feature extraction and time series data processing capability, and is more suitable for the prediction scene of the remaining life of the equipment. Therefore, in the embodiment, the neural network model selects the CNN-GRU model, the CNN-GRU model includes a CNN network layer and a GRU network layer, the application designs an all-fiber current transformer RUL prediction model based on the CNN-GRU combined neural network by combining the characteristics of the CNN and the GRU model, the model first processes the preprocessed time series data through the CNN network layer to extract the feature quantity of the time series data, then performs a flattening operation on the feature quantity to convert it into a time series sequence, inputs it into the GRU network layer for prediction, and performs anti-normalization on the predicted data to obtain a prediction result with practical significance, and the model structure is as shown in Figure 3

[0046] Specifically, the CNN-GRU model parameters are designed as follows:

[0047] ​1) CNN Network Layer Structure: The CNN network layer structure used for feature extraction in the model includes three convolutional layers (Conv2D) and three pooling layers (MaxPooling2D). More convolutional layers mean more deep feature information can be obtained, but deeper layers are also more prone to overfitting. After multiple experiments, a three-layer CNN structure was finally determined. To fully utilize the data, the convolutional kernel size was set to 3×3, with a stride of 1. The number of kernels in the first layer was 32, the second layer had 64, and the third layer had 128. The pooling layer size was chosen to be 2×2, and max pooling was used to reduce the dimensionality of the convolutional layer output.

[0048] 2) GRU Network Layer Structure: The GRU network layer structure contains three GRU layers and one fully connected (Dense) layer. Experiments show that prediction accuracy increases with the number of GRU layers. Appropriately increasing the number of GRU layers can improve the model's learning ability, but too many layers can lead to overfitting. Therefore, in this embodiment, the number of neurons in the first GRU layer is set to 128, and the number of neurons in the second and third GRU layers is set to 64. After each GRU layer, the Dropout method is used to randomly deactivate neurons to prevent overfitting. The Dropout value is set to 0.5. Finally, the prediction result is output through the fully connected (Dense) layer.

[0049] 3) Activation Function: The activation function in the hidden layer of the CNN-GRU model can be any existing activation function used in neural networks, such as the Sigmoid (non-linear activation) function, the Tanh (hyperbolic tangent) function, or the ReLU (rectified linear function). Preferably, to alleviate the gradient vanishing problem that may occur in the model, the ReLU function is selected as the activation function in this embodiment.

[0050] 4) Optimization Algorithm: Existing optimization algorithms are used to optimize the CNN-GRU model, such as the Adam optimization algorithm, SGD optimization algorithm, and PMSprop optimization algorithm. Since the Adam optimization algorithm combines the advantages of SGD and PMSprop optimization algorithms, it uses momentum updates and learning rate adjustments for adaptive optimization, resulting in a more computationally efficient and faster convergence model. Therefore, this embodiment uses the Adam optimizer for optimization, with a learning rate of 0.0008, 500 training epochs, and a batch size of 32.

[0051] 5. Lifespan calculation of all-fiber current transformers.

[0052] After the neural network model outputs the predicted values of the light source driving current value and the modulator driving voltage within the set time, the predicted values are subjected to inverse normalization processing to obtain short-term prediction results with practical significance. According to the periodic average change value of each state parameter, the upper and lower thresholds of the state parameter, and the current value of the state parameter, the RUL corresponding to the two state parameters is calculated, and the minimum value between the two is selected as the RUL of the all-fiber current transformer.

[0053] When the RUL is positive, the RUL is calculated according to the upper threshold of the state parameter, the current value of the state parameter, and the periodic average change value, and the calculation formula is:

[0054]

[0055] When the RUL is negative, the RUL is calculated according to the lower threshold of the state parameter, the current value of the state parameter, and the periodic average change value, and the calculation formula is:

[0056]

[0057] wherein RUL is the remaining useful life; T max , T min are the upper threshold value and the lower threshold value of the state parameter respectively; Y is the current value of the state parameter; is the periodic average change value of the state parameter. When the period is daily, the unit of RUL is also daily.

[0058] 6. Determine the evaluation index.

[0059] The evaluation index is a quantitative index for measuring the prediction accuracy of the model. In the present application, three measurement indexes, namely Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R 2 ) are used to evaluate the effectiveness of the prediction model.

[0060] The calculation formula of MAE is:

[0061]

[0062] The calculation formula of RMSE is:

[0063] ​​

[0064] R 2 The calculation formula is:

[0065]

[0066] Wherein, n is the total number of prediction, y i ′ is the predicted value of the i th state parameter, y i The i th state parameter is the true value, and y is the average value of the true value of the state parameter.

[0067] System embodiment

[0068] The application provides a full-fiber current transformer life prediction system, which comprises a processor, and the processor is used for realizing a full-fiber current transformer life prediction method.

[0069] Verification example

[0070] The application compares the light source driving current prediction results of the CNN-GRU model, the existing GRU (gated recurrent unit) model and the CNN-LSTM (convolutional-long short-term memory network) model. Figure 4 As shown in the table, according to the prediction results of different models, the MAE, RMSE and R 2 Three kinds of measurement indexes, and the indexes are compared as shown in Table 1.

[0071] Table 1

[0072]

[0073] The application compares the light source driving current prediction results of the CNN-GRU model, the existing GRU (gated recurrent unit) model and the CNN-LSTM (convolutional-long short-term memory network) model. Figure 5 As shown in the table, according to the prediction results of different models, the MAE, RMSE and R 2 Three kinds of measurement indexes, and the indexes are compared as shown in Table 2.

[0074] Table 2

[0075]

Claims

1. An all-fiber current transformer lifetime prediction method, characterized by, The method comprises the following steps: acquiring historical data of light source driving current and modulator driving voltage of the all-fiber current transformer, and constructing time series data, wherein the time series data comprises N time series vectors obtained by concatenating the light source driving current and the modulator driving voltage at the same time, and N is greater than or equal to 7; inputting the time series data into a pre-trained neural network model to obtain prediction results of state parameters in a set time, wherein the state parameters comprise the light source driving current and the modulator driving voltage; calculating a periodic average change value of each state parameter in the set time period according to the prediction results of the state parameters in the set time and current values of the state parameters; calculating a remaining service life corresponding to each state parameter according to a threshold upper limit / threshold lower limit of each state parameter, the periodic average change value and the current value of each state parameter, and selecting a minimum value between the two as the remaining service life of the all-fiber current transformer.

2. The all-fiber current transformer lifetime prediction method according to claim 1, characterized in that, The method for calculating the remaining service life corresponding to each state parameter according to the threshold upper limit / threshold lower limit of each state parameter, the periodic average change value and the current value of each state parameter comprises: when the periodic average change value is positive, calculating the remaining service life corresponding to each state parameter by subtracting the current value from the threshold upper limit of each state parameter and dividing the result by the periodic average change value; when the periodic average change value is negative, calculating the remaining service life corresponding to each state parameter by subtracting the current value from the threshold lower limit of each state parameter and dividing the result by the periodic average change value.

3. The all-fiber current transformer lifetime prediction method according to claim 1 or 2, characterized in that, The set time is less than or equal to 7 days. 4.The all-fiber current transformer lifetime prediction method according to claim 1, characterized in that, The time series data input into the pre-trained neural network model further comprises an environmental temperature, and the time series data comprises N time series vectors obtained by concatenating the environmental temperature, the light source driving current and the modulator driving voltage at the same time.

5. The all-fiber current transformer lifetime prediction method according to claim 1 or 4, characterized in that, The neural network model is a CNN-GRU model, which comprises a CNN network layer and a GRU network layer; when performing prediction, the time series data input into the CNN-GRU model is processed by the CNN network layer to extract feature quantities of the time series data, the feature quantities are then flattened and converted into time series sequences, and the time series sequences are input into the GRU network layer to perform prediction, thereby outputting the prediction results of the state parameters in the set time.

6. The all-fiber current transformer lifetime prediction method according to claim 5, characterized in that, The CNN network layer comprises three convolution layers and three pooling layers.

7. The all-fiber current transformer lifetime prediction method according to claim 5 or 6, characterized in that, The GRU network layer comprises three GRUs and one fully connected layer. 8.The all-fiber current transformer lifetime prediction method according to claim 4, characterized in that, Before the time series data is input into the pre-trained neural network model, the time series data is preprocessed; the preprocessing comprises: when there is null data in the time series data of the light source driving current, the modulator driving voltage and the environmental temperature, the mean values of the time series data before and after the null data are filled into the null data; if there is a continuous missing sequence, the mean values of the time series data at the moment before the beginning of the missing sequence and the moment after the end of the missing sequence are filled into the entire missing sequence. 9.The all-fiber current transformer lifetime prediction method according to claim 8, characterized in that, The preprocessing further comprises normalizing the time series data of the light source driving current, the modulator driving voltage and the environmental temperature; when normalizing each time series data, the prediction results of the state parameters in the set time obtained in the subsequent neural network model need to be de-normalized.

10. An all-fiber current transformer life prediction system comprising a processor, wherein, The processor is configured to implement the all-fiber current transformer life prediction method according to any one of claims 1-9.

Images