An Error Prediction Method for Electronic Voltage Transformers Based on Prophet, Self-Attention Mechanism, and Time-Series Convolutional Networks

Abstract

This paper proposes an error prediction method for electronic voltage transformers based on a combined model of Prophet, self-attention mechanism, and time-series convolutional network. The method collects the ratio difference data of electronic voltage transformers and performs standardization. The standardized data is then input into the Prophet model for decomposition, and each trend component is fitted to output the prediction result. The training set data is input into the time-series convolutional network model for feature extraction and prediction. The output of the time-series convolutional network model is then weighted and combined with the periodic terms obtained from the Prophet model decomposition and output. The predicted values ​​are then input into the self-attention mechanism layer for feature extraction. The predicted values ​​output from the self-attention mechanism layer are further passed to a fully connected layer for dimensionality reduction, outputting the predicted ratio difference data of the electronic voltage transformer. This method can effectively predict the future error change trend of electronic voltage transformers, providing a strong basis for equipment verification in power systems.

Description

Technical Field

[0001] This invention relates to the field of error prediction technology for electronic voltage transformers, and specifically to an error prediction method for electronic voltage transformers based on a combined model of Prophet, self-attention mechanism, and time-series convolutional network. Background Technology

[0002] As an essential power metering device in substations, electronic voltage transformers are a crucial medium for information transmission, connecting the primary power grid and secondary control systems of the power system. They provide accurate voltage information for control, energy metering, and relay protection units. Electronic voltage transformers are also the main power equipment used by power companies in energy transaction accounting with users. Their error status assessment is particularly important for protection equipment and ensuring high measurement accuracy. However, the field operating environment of transformers is generally complex. Factors such as surrounding magnetic fields, temperature, humidity, and load inevitably affect the measurement error of transformers, resulting in poor long-term stability and metering errors. When the accuracy falls below the required precision, it directly affects the accuracy of energy metering, causing losses to power supply and consumption companies and posing potential risks to the safe and stable operation of the power grid.

[0003] With the accelerated construction of smart grids, daily electricity consumption has increased rapidly, and the level of equipment intelligence has generally improved, making instrument transformers a major component in power grid upgrades and renovations. Instrument transformers have been included in the national mandatory verification catalog, and numerous studies on methods for evaluating their accuracy are currently emerging. However, most are limited to traditional statistical methods, machine learning methods, or deep learning methods. Machine learning methods are generally only suitable for processing small amounts of data, and studying time series data by considering only a limited amount of historical data may overlook potentially useful information in the historical data, leading to insufficient prediction accuracy.

[0004] The widespread use of measuring equipment such as electronic voltage transformers has significantly increased the amount of data available. Machine learning has limitations in processing and analyzing this massive amount of data. Deep learning, with its powerful ability to analyze time series features, has become a hot research topic in the field and has certain advantages in improving the ability to predict large-sample data series. In recent years, the Prophet deep learning algorithm developed by Facebook has been widely applied in the field of time series analysis, achieving good results in areas such as power load forecasting, sales volume forecasting, and air quality index forecasting. Combining the advantages of this model, it is of great significance to predict the errors of electronic voltage transformers. Summary of the Invention

[0005] To address the problems of poor long-term stability and unstable accuracy of current electronic voltage transformers, this invention provides an error prediction method for electronic voltage transformers based on a combined model of Prophet, self-attention mechanism, and time-series convolutional network. This method can effectively predict the future error change trend of electronic voltage transformers, providing a strong basis for equipment verification in power systems. It also helps maintenance personnel select appropriate times for transformer inspection and maintenance, preventing economic losses caused by increased transformer measurement errors and improving the stability of the power system.

[0006] The technical solution adopted in this invention is as follows:

[0007] An error prediction method for electronic voltage transformers based on Prophet, self-attention mechanism, and time-series convolutional network is characterized by the following steps:

[0008] Step 1: Collect the differential data of the electronic voltage transformer and perform standardization processing to eliminate non-standard data;

[0009] Step 2: Input the standardized data from Step 1 into the Prophet model for decomposition, fit each trend component, and output the prediction results;

[0010] Step 3: Input the training set data into the time series convolutional network model for feature extraction and prediction;

[0011] Step 4: Sum the output of the time series convolutional network model with the periodic terms obtained from the Prophet model decomposition, and output the predicted value;

[0012] Step 5: Input the predicted values ​​output from Step 4 into the self-attention mechanism layer for feature extraction;

[0013] Step 6: The predicted value output by the self-attention mechanism layer is passed to the fully connected layer for data dimensionality reduction processing, and the predicted data of the electronic voltage transformer ratio difference is output.

[0014] In step one, the ratio difference data of the electronic voltage transformer is collected, and z-score standardization is performed to remove some abnormal data points and missing values ​​in the historical data, and the data is concentrated around 0 with a variance of 1, so as to eliminate the problem of non-standard data affecting the model accuracy.

[0015] Step two includes the following steps:

[0016] S2.1 Input the standardized data into the Prophet model. The input is divided into two columns: the first column is the date, and the second column is the electronic voltage transformer ratio difference data corresponding to the date.

[0017] S2.2. Use the Prophet model to train the processed training set data, decompose the dataset, reflect the periodicity in days, and the random error corresponds to the unpredictable random noise in the time series.

[0018] S2.3. Use the Prophet model to fit each trend component, and observe and compare the fitting effect based on the output image.

[0019] S2.4 Continuously adjust the parameters of the Prophet model to achieve the optimal prediction effect of the Prophet model. The accumulated result is the prediction value of Prophet.

[0020] Step three includes the following steps:

[0021] S3.1 Input the standardized training set data from step one into the input layer of the time series convolutional network model. Each input layer takes the mutual inductor error data corresponding to the time series as the input of the time series convolutional network model.

[0022] S3.2 The input layer data is passed to the residual blocks of the time series convolutional network model for high-dimensional feature extraction; each residual module unit is activated by the ReLU activation function according to the stacking order of the residual blocks and then propagated backward.

[0023] S3.3. An identity mapping is performed every two convolutional layers. Each convolution is a dilated causal convolution. After the convolution, batch normalization is performed to improve classification accuracy. Then, the ReLU activation function is used to activate the resulting values, and the Dropout method is introduced to randomly discard a portion of neurons during neural network execution to reduce overfitting and improve the model's generalization ability. TCN uses expanded causal convolution to increase the receptive field and performs a step-by-step operation on the input sequence for convolution operations. Its expression is:

[0024]

[0025] In equation (1), F(i) is the convolution result of the i-th element in the sequence; h(j) is the convolution kernel; and d is the scaling factor.

[0026] S3.4 After the above calculations are performed twice, a residual connection is made, and the result of the convolution calculation and the value of the identity mapping are added together to obtain the output of one unit. In order to achieve a smooth summation, the dimensions of the convolution calculation result and the value of the identity mapping need to be the same. Therefore, a 1×1 convolution calculation is used on the path of the identity mapping to make the dimensions of the two consistent.

[0027] S3.4 Output the residual block output of the time series convolutional network model to obtain the final prediction result.

[0028] Let x be the input of the residual block, and the output o be as follows:

[0029]

[0030] In equation (2), Represents the residual. Activation() represents the activation function. This step is a process of linear transformation, mapping through the activation function, and outputting the result.

[0031] After the input data undergoes multiple convolution operations, the abstract features contained in the mutual inductor error data are effectively extracted, while still maintaining the inlineness of various data types. After multiple convolution operations, the output matrix with the same dimension as the input is obtained through the fully convolutional layer in the time series convolutional network.

[0032] In step four, the output of the time series convolutional network model is weighted and output with the periodic term obtained from the Prophet model decomposition, as shown in the following formula:

[0033]

[0034] In equation (3), V Prophet-TCN W represents the predicted ratio difference of the mutual inductors obtained by weighted combination of the two models; Prophet and W TCN V represents the weight coefficients of the Prophet model and the time-series convolutional network model, respectively; Prophet and V TCN These represent the predicted cross-sectional area difference values ​​of the mutual inductors obtained by the Prophet model and the time-series convolutional network model, respectively.

[0035] The model weight coefficients are determined using the standard deviation method. The calculation methods for the weights of the time series convolutional network model and the Prophet model are shown in equations (4)-(5):

[0036]

[0037] W Prophet =2-W TCN (5);

[0038] In equation (5), Prophet diff TCN represents the difference between the sample standard deviation and the standard deviation of the Prophet model prediction. diff This represents the difference between the sample standard deviation and the standard deviation of the prediction results from the time series convolutional network model, i.e.:

[0039] TCN diff =Sample std -TCNstd (6);

[0040] In equation (6), Sample std The standard deviation of the sample, TCN std The standard deviation represents the prediction results of a time-series convolutional network model.

[0041] Prophet diff =Sample std -Prophet std (7);

[0042] In equation (7), Prophet std This represents the standard deviation of the prediction results from the Prophet model.

[0043] In step five, the predicted value output from step four is input into the self-attention mechanism layer for feature extraction, specifically:

[0044] The computation process of the attention mechanism is an addressing process. Given a task-related query vector (query), and each element in the sequence stored in memory as key-value pairs, when a query equals the key, the value of the element needs to be retrieved. Unlike traditional addressing, it does not retrieve the value by address, but rather by calculating the similarity between the key and the query. The attention mechanism computation process consists of the following three steps:

[0045] S5.1 The information input is a vector, assuming there are N input information, i.e., input data X = [x1, x2, ..., x...]. N ];

[0046] S5.2 Calculate the attention distribution. Let key = value = X, then the attention distribution is:

[0047] α i =softmax(s(k) i ,q))=softmax(s(x i ,q)) (8);

[0048] In the above formula, s(k i ,q) and s(x i (q) represents the similarity calculation between the i-th input information and the query, and the attention distribution vector α is obtained by calculating through the softmax function. i .

[0049] Commonly used similarity calculation methods, such as the dot product model, are as follows:

[0050] s(x i ,q)=xi ·q (9);

[0051] In equation (9), x i Let represent the i-th element in the time series, and q represent the corresponding query vector.

[0052] S5.3, Information weighted summation, the calculation method is as follows:

[0053]

[0054] In equation (10), the attention distribution vector α i It can be calculated by equation (9), where N is the number of elements in the series, and att(q,X) represents the attention value of the input information X.

[0055] Based on this series of calculations, the attention mechanism determines key information and increases the weight of key information while decreasing the weight of non-key information through repeated iterative calculations of the above steps, thus extracting important features.

[0056] In step six, the predicted values ​​output from the self-attention mechanism layer are received. The attention predicted values ​​are a series of vectors. Each vector is input into the fully connected network for dimensionality reduction. The length of the output data is reduced to a 20×1 vector, and the predicted result image of the combined model is output sequentially.

[0057] This invention discloses an error prediction method for electronic voltage transformers based on a combined model of Prophet, self-attention mechanism, and time-series convolutional network. The technical advantages are as follows:

[0058] 1) This invention, based on a combined model of the Prophet model, self-attention mechanism, and time-series convolutional network, can fully capture time-series information and accurately predict the fluctuation of electronic voltage transformer errors, thus gaining time for early on-site decision-making and maintenance. Simultaneously, by transforming the time series into different time dimensions and combining them into an overall trend, the periodic trends and potential relationships contained in the data are extracted, thereby effectively predicting the errors of electronic voltage transformers and reflecting the future trend of their errors.

[0059] 2) The method of this invention collects historical operating data of electronic voltage transformers. The Prophet model is input with the date and the corresponding electronic voltage transformer ratio difference data. The time series data is decomposed into different time dimensions and combined into an overall trend. The optimal prediction effect is achieved by continuously adjusting the parameters of the model.

[0060] 3) The Prophet model can transform time series into different time dimensions, such as hour, day, week, year, etc., and combine them into an overall trend. It can adapt well to and extract the periodic trends contained in the data, and has a strong ability to predict medium and long-term time series.

[0061] 4) The method of this invention introduces a time-series convolutional network to enhance the feature extraction capability of time-series data. It extracts features strongly correlated with transformer errors from a large number of time series data and performs predictions. The time-series convolutional network exhibits excellent performance in data feature extraction, especially in predicting time series with strong nonlinear characteristics, demonstrating higher accuracy.

[0062] 5) The method of this invention introduces a self-attention mechanism to better mine and utilize the latent relationships in the data sequence, thereby improving the prediction effect and outputting time series prediction data of the same length as the input. The introduction of the self-attention mechanism allows the latent relationships in the data sequence to be better mined and utilized, thereby improving the prediction effect.

[0063] 6) The method of the present invention can better predict the future error change trend of electronic voltage transformers, providing a strong basis for equipment verification in power systems, preventing economic losses caused by increased transformer measurement errors, and improving the stability of power systems. Attached Figure Description

[0064] Figure 1 This is a decomposition diagram of the Prophet model.

[0065] Figure 2 This is a graph showing the growth trend of the mutual inductor ratio difference data obtained from the Prophet model decomposition.

[0066] Figure 3 A daily periodic chart of the mutual inductor ratio difference data obtained from the Prophet model decomposition.

[0067] Figure 4 The figure shows the prediction results of the ratio difference of electronic voltage transformers based on the Prophet model.

[0068] Figure 5 The figure shows the prediction results of the ratio difference of an electronic voltage transformer based on a time-series convolutional network.

[0069] Figure 6 This is a structural diagram of the self-attention mechanism.

[0070] Figure 7 This is a flowchart of error prediction for electronic voltage transformers based on a combined model.

[0071] Figure 8 The figure shows the error prediction results of electronic voltage transformers based on the combined model. Detailed Implementation

[0072] The ratio difference is an important indicator of the error of an electronic voltage transformer. Let X = (x1, x2, ..., x...). t The input is a sequence of historical electronic voltage transformer differential data, and the prediction target is the electronic voltage transformer differential data at n future time points. The electronic voltage transformer differential data are not spatially correlated, therefore it is a single time series rather than a spatiotemporal series.

[0073] An error prediction method for electronic voltage transformers based on a combined model of Prophet, self-attention mechanism, and time-series convolutional network includes the following steps:

[0074] Step 1: Input data preprocessing:

[0075] The data collected from the field operation of electronic voltage transformers may contain some abnormal data due to environmental interference and measurement deviations during the data collection process. To avoid these abnormal data from adversely affecting the model training, z-score standardization is required to remove some abnormal data points and missing values ​​from the historical data, and to concentrate the data around 0 with a variance of 1, so as to eliminate the problem of non-standard data affecting the model accuracy.

[0076] The z-score standardization formula is as follows:

[0077]

[0078] Where x represents the collected EVT error data, x * The normalized value of the EVT error data. Let σ be the mean of the EVT error data, and σ be the standard deviation of the error data.

[0079] Step 2: Prophet model processing:

[0080] The Prophet model is a machine learning framework for predicting time series trends. Prophet uses an additive regression model to study the nonlinear trends of time series. This model can not only handle outliers and missing values ​​in time series but also predict their trends. Its focus is on analyzing and processing the periodicity, directional nature, and holiday effects of time series. The Prophet model decomposes time series data into four parts: a trend function g(t) that analyzes the nonlinear changes in the time series; s(t) that reflects periodic changes, which can be expressed in units of years, months, etc.; potential random fluctuations h(t), such as holidays; and an error ε representing unpredictable and unmanifested anomalous changes. By fitting these four factors together, the predicted results for the time series can be obtained.

[0081] The Prophet model studies a univariate time series prediction and regression problem that considers only the time dimension. For example... Figure 1 As shown, the Prophet model decomposes the time series into a trend term, a periodic term, a holiday term, and an error term. The trend term g(t) is the main term in Prophet, and it is divided into two types: a piecewise linear function and a logistic regression function. Here, the piecewise linear function is used. The model is as follows:

[0082] g(t)=(k+a(t)Tδ)·t+(m+a(t)Tγ) (12);

[0083] Where k is the growth rate; δ is the change in the growth rate; and m is the offset.

[0084] s(t) represents the periodic term. Since there may be various periodic trends such as weekly, monthly, and yearly cycles in time series, Prophet uses Fourier series to approximate these periodic characteristics. The periodic function is represented by a sine or cosine function, and the Fourier series form is as follows:

[0085]

[0086] Where T represents the period; N represents the number of periods in the model; a n b n This indicates an adjustable parameter. The more periods there are, the more complex periodic functions can be fitted.

[0087] In real life, holidays and major events often lack a cyclical pattern, but they can still significantly impact time series. Prophet collects special holidays from various countries and allows users to add or set special time points based on actual circumstances and needs.

[0088] The processed data is input into the Prophet model. The input consists of two columns: the first column represents the date, and the second column represents the corresponding electronic voltage transformer ratio difference data. Prophet is used to decompose the dataset, with the periodic term primarily reflecting the number of days, and the random error mainly corresponding to unpredictable random noise in the sequence. Subsequently, the Prophet model is trained on the processed training set, fitting each trend component. The fitting effect is observed and compared based on the output image, and the model parameters are continuously adjusted to achieve the optimal model prediction effect. The accumulated result is the Prophet prediction value.

[0089] The Prophet model is used for time series decomposition to reflect the changing trends of the time series using growth trend and periodic trend (reflected by daily periodicity) images. Figure 2 The figure shown is a graph illustrating the growth trend of the mutual inductor ratio difference data obtained from the Prophet model decomposition. Figure 3 The image shows a daily periodic chart of the transformer ratio difference data obtained from the Prophet model decomposition. It can be observed that the ratio difference data exhibits certain fluctuations over time, displaying both periodicity and randomness, with some peaks. Between 13:00 and 17:00, the transformer ratio difference is at a relatively high level, while from 3:00 AM to 6:00 AM it is at a relatively low level. This may be related to rising temperatures or increased load, causing daily fluctuations in the ratio difference.

[0090] After the above steps, the prediction results of the ratio difference of electronic voltage transformers using the Prophet model are obtained, such as... Figure 4 As shown, the black dots represent the original actual values ​​of the time series, the blue lines represent the predicted values ​​obtained by fitting the time series, and the light blue shaded area represents the uncertainty range, that is, a confidence interval of the time series, which is also known as the reasonable upper and lower bounds.

[0091] Step 3: Time-series convolutional network processing:

[0092] Time-series convolutional networks (CNNs) are a novel type of neural network model based on traditional one-dimensional convolutional neural networks, incorporating architectures such as dilated causal convolution and residual connections, for time-series modeling or time-series data processing. They can extract temporal features from large amounts of time-series samples and predict their trends, while effectively overcoming the performance degradation problem of deep networks during training.

[0093] The training data is input into a time-series convolutional network model for feature extraction and prediction. Input layer data is passed to the residual blocks of the time-series convolutional network for high-dimensional feature extraction. Each residual block unit, following the stacking order of the residual blocks, activates the result using the ReLU activation function before propagating it backward. An identity mapping is performed every two convolutional layers. After each convolution, ReLU activation is applied, and Dropout is introduced to prevent overfitting. After two such calculations, a residual connection is performed, adding the convolution result to the identity mapping value, resulting in the output of a single unit. A 1×1 convolution is used along the identity mapping path to maintain dimensionality consistency. Finally, the output of the time-series convolutional network residual blocks is output to obtain the final prediction result.

[0094] Temporal series convolutional networks (TSNs) possess the ability to extract features from time series with strong nonlinear characteristics. Therefore, they can extract features strongly correlated with transformer errors from large amounts of time series data and make predictions. Applying causal convolution to TSNs ensures that the output at a given time step is convolved only with the input corresponding to that time step in the previous layer and inputs from even earlier time steps. This prevents the "leakage" of historical or future data, ensuring information integrity. Applying dilated convolution to TSNs, by increasing the kernel size and dilation coefficients, expands the data receptive range, giving the model a longer convolutional "memory" function and maintaining high computational efficiency even with parallel input of multidimensional data.

[0095] The prediction results of the mutual inductor ratio difference obtained by using a time series convolutional network are as follows: Figure 5 As shown. Figure 5 The black line represents the actual value of the ratio difference data for electronic voltage transformers, while the red line represents the predicted value. Because the time-series convolutional network model has strong feature extraction capabilities, the predicted values ​​are generally close to the actual values.

[0096] Step 4: Weight the output of the time series convolutional network model with the periodic terms obtained from the Prophet model decomposition:

[0097] To obtain the prediction results of the combined model, a weighted combination method is generally required. This involves first calculating the product of the predicted value of each model and its corresponding weight, then summing the values ​​obtained from each model to obtain the overall value, and finally dividing the sum by the number of models to obtain the predicted value of the combined model. The setting of model weights is crucial to the accuracy of model prediction. The standard deviation method is used to determine the model weight coefficients. The weighted combination calculation method is shown in formulas (3)-(7). This method integrates the prediction results of the time-series convolutional network model with the prediction results of the Prophet model, using them as input to the self-attention mechanism for subsequent processing.

[0098] Step 5: Feature extraction from the self-attention mechanism layer:

[0099] Self-attention mechanisms can effectively enhance the feature correlation of time series data. Self-attention mechanisms employ a parallel processing mode, integrating features of the input sequence, such as… Figure 6 The a1, b1, c1, and d1 values ​​are mapped to different subspaces, thus obtaining attention vectors for each different subspace. These attention vectors are then combined using a fully connected layer.

[0100] The predicted values ​​from the Prophet model and the time-series convolutional network module are input together into the self-attention mechanism layer for feature extraction. This prevents the disappearance of features containing important information and highlights important information, making it easier for the model to learn long-distance interdependencies in a sequential manner. The input information of the self-attention mechanism is a row of vectors. The attention distribution is calculated using formulas (8)-(10), and the information is weighted and summed to obtain the corresponding attention value. Through the above method, the model is trained to determine key information, which is strengthened while the weight of non-key information is weakened. At this time, the length of the output sequence is the same as the length of the input sequence.

[0101] Feature extraction in step five is only one component of the model. The final data, after processing, is output as the final image. To illustrate the role of feature extraction in the model, the prediction results of a model without self-attention are presented for comparison. Figure 8 This will be introduced together after step six.

[0102] Step Six: Data Dimensionality Reduction Processing

[0103] The predicted values ​​output by the self-attention mechanism layer are then passed to the fully connected network for data dimensionality reduction. Specifically, the predicted values ​​output by the self-attention mechanism are a series of vectors that are passed to the fully connected layer. After processing by the fully connected layer, the length of the output data is reduced to a 20×1 vector, and the predicted result image of the combined model is output sequentially.

[0104] Based on the above steps, the error process of an electronic voltage transformer is as follows: Figure 7 As shown, the final prediction result of the electronic voltage transformer error is as follows: Figure 8 The PtcnAttn curve is shown in the figure. To demonstrate the effectiveness of the model's feature extraction, the prediction results of the model without a self-attention mechanism are used for comparison. Figure 8 The curve represented by Ptcn reflects this. It can be seen that the electronic voltage transformer error prediction model proposed in this invention can accurately predict the fluctuation of the EVT ratio, and the prediction of the overall trend is in good agreement with the actual value, thus verifying the effectiveness of the self-attention mechanism in extracting the characteristics of the transformer's historical operating data.

[0105] This invention discloses an error prediction method for electronic voltage transformers based on a combined model of Prophet, self-attention mechanism, and time-series convolutional network, which has the following characteristics:

[0106] (1) The Prophet time series model used is powerful and its strong flexibility makes the periodicity easy to adjust. For the predicted trend, flexible assumptions can be made according to one's own needs. Its good applicability means that the presence of outliers, missing values ​​and missing normal measurement intervals in the data will not have a significant impact on the prediction results. Moreover, it has a fast fitting speed and good performance. Its interpretability makes the model parameters easy to interpret and can be adjusted according to the current situation.

[0107] (2) In time-series convolutional networks, the application of dilated causal convolution enables them to have a larger field of view with fewer layers, thereby improving their ability to receive historical data. The introduction of residual networks "connects" network layers that are one or more layers apart, effectively solving the gradient vanishing problem in complex models. Time-series convolutional networks can take sequences of arbitrary length as input and output sequences of equal length at the output layer, which are then passed to the next unit. Their architecture can be adjusted to any length as needed, making them highly adaptable.

[0108] (3) The self-attention mechanism adopts a parallel processing mode, mapping the features of the input sequence to different subspaces to obtain attention vectors for each different subspace, and then using a fully connected layer to aggregate these attention vectors together. This parallel self-attention mechanism can analyze subspace information at different locations and with different representations, and has a powerful feature extraction capability.

[0109] (4) This invention presents an error prediction method for electronic voltage transformers based on a combined model of Prophet, self-attention mechanism, and time-series convolutional network. The introduction of a time-series convolutional network improves the model's performance in extracting data features, especially for predicting time series with strong nonlinear characteristics. The introduction of a self-attention mechanism further extracts features from the time series and performs data dimensionality reduction, which can fully utilize the useful information in the historical operating data of the electronic voltage transformer.

Claims

1. A method for predicting errors in electronic voltage transformers based on Prophet, self-attention mechanism, and time-series convolutional networks, characterized in that... Includes the following steps: Step 1: Collect the differential data of the electronic voltage transformer and perform standardization processing to eliminate non-standard data; Step 2: Input the standardized data from Step 1 into the Prophet model for decomposition, fit each trend component, and output the prediction results; Step 3: Input the training set data into the time series convolutional network model for feature extraction and prediction; Step 4: Sum the output of the time series convolutional network model with the periodic terms obtained from the Prophet model decomposition, and output the predicted value; Step 5: Input the predicted values ​​output from Step 4 into the self-attention mechanism layer for feature extraction; Step 6: The predicted value output from the self-attention mechanism layer is passed to the fully connected layer for data dimensionality reduction processing, and the predicted data of the electronic voltage transformer ratio difference is output. In step four, the output of the time series convolutional network model is weighted and output with the periodic term obtained from the Prophet model decomposition, as shown in the following formula: （3）； In equation (3), This represents the predicted ratio difference of the mutual inductors obtained by weighted combination of the two models. and These represent the weight coefficients of the Prophet model and the time-series convolutional network model, respectively. and These represent the predicted inductor ratio difference values ​​obtained from the Prophet model and the time-series convolutional network model, respectively. The model weight coefficients are determined using the standard deviation method. The calculation methods for the weights of the time series convolutional network model and the Prophet model are shown in equations (4)-(5): （4）； （5）； In equation (5), This represents the difference between the sample standard deviation and the standard deviation of the Prophet model prediction. This represents the difference between the sample standard deviation and the standard deviation of the prediction results from the time series convolutional network model, i.e.: （6）； In equation (6), The standard deviation of the sample The standard deviation represents the prediction results of a time-series convolutional network model. （7）； In equation (7), This represents the standard deviation of the prediction results from the Prophet model.

2. The electronic voltage transformer error prediction method based on Prophet, self-attention mechanism, and time-series convolutional network as described in claim 1, characterized in that: In step one, the ratio difference data of the electronic voltage transformer is collected, and z-score standardization is performed to remove some abnormal data points and missing values ​​in the historical data, and the data is concentrated around 0 with a variance of 1, so as to eliminate the problem of non-standard data affecting the model accuracy.

3. The electronic voltage transformer error prediction method based on Prophet, self-attention mechanism, and time-series convolutional network as described in claim 1, characterized in that: Step two includes the following steps: S2.1 Input the standardized data into the Prophet model. The input is divided into two columns: the first column is the date, and the second column is the electronic voltage transformer ratio difference data corresponding to the date. S2.

2. Use the Prophet model to train the processed training set data, decompose the dataset, reflect the periodicity in days, and the random error corresponds to the unpredictable random noise in the time series. S2.

3. Use the Prophet model to fit each trend component, and observe and compare the fitting effect based on the output image. S2.4 Continuously adjust the parameters of the Prophet model to achieve the optimal prediction effect of the Prophet model. The accumulated result is the prediction value of Prophet.

4. The electronic voltage transformer error prediction method based on Prophet, self-attention mechanism, and time-series convolutional network according to claim 1, characterized in that: Step three includes the following steps: S3.1 Input the standardized training set data from step one into the input layer of the time series convolutional network model. Each input layer takes the mutual inductor error data corresponding to the time series as the input of the time series convolutional network model. S3.2 The input layer data is passed to the residual blocks of the time series convolutional network model for high-dimensional feature extraction; each residual module unit is activated by the ReLU activation function according to the stacking order of the residual blocks and then propagated backward. S3.

3. An identity mapping is performed every two convolutional layers. Each convolution is a dilated causal convolution. After the convolution, the ReLU activation function is used to activate the resulting values, and the Dropout method is introduced to perform a step-by-step operation on the input sequence for convolution. Its expression is: （1）； In equation (1), F(i) For the first in the sequence i The convolution result of 1 element; h(j) For convolution kernel; d As an expansion factor; S3.4 After the above calculations are performed twice, a residual connection is performed, and the result of the convolution calculation is added to the value of the identity mapping. The result is the output of one unit. A 1×1 convolution calculation is performed on the path of the identity mapping to make the two dimensions consistent. S3.4 Output the residual block output of the time series convolutional network model to obtain the final prediction result; set up x It is the input and output of the residual block. o As shown below: （2） In equation (2), Represents residuals; Indicates the activation function; After the input data undergoes multiple convolution operations, the abstract features contained in the mutual inductor error data are effectively extracted, while still maintaining the inlineness of various data types. After multiple convolution operations, the output matrix with the same dimension as the input is obtained through the fully convolutional layer in the time series convolutional network.

5. The electronic voltage transformer error prediction method based on Prophet, self-attention mechanism, and time-series convolutional network according to claim 1, characterized in that: In step five, the predicted value output from step four is input into the self-attention mechanism layer for feature extraction, specifically: The attention mechanism computation process consists of the following three steps: S5.1 The information input is a row of vectors, assuming that... N Input information, i.e., input data ; S5.2 Calculate the attention distribution, assuming key=value= X Then the attention distribution is: （8）； In the above formula, and Indicates the first i The similarity between each input and the query is calculated, and the attention distribution vector is obtained by using the softmax function. ; Commonly used similarity calculation methods, such as the dot product model, are as follows: （9）； In equation (9), This represents the i-th element in the time series. This represents the corresponding query vector; S5.3, Information weighted summation, the calculation method is as follows: （10）； In equation (10), the attention distribution vector It can be calculated from equation (9). N The number of elements in the series. Indicates input information X The attention value.

6. The electronic voltage transformer error prediction method based on Prophet, self-attention mechanism, and time-series convolutional network according to claim 1, characterized in that: In step six, the predicted values ​​output from the self-attention mechanism layer are received. The attention predicted values ​​are a series of vectors. Each vector is input into the fully connected network for dimensionality reduction. The length of the output data is reduced to a 20×1 vector, and the predicted result image of the combined model is output sequentially.

Images