A Deep Learning-Based Online Error Calibration Method and System for Mutual Inductors

Abstract

This application relates to the field of instrument transformer error calibration technology, and discloses an online instrument transformer error calibration method and system based on deep learning. The method includes: deploying a sensor array at key nodes of a high-voltage power grid to collect real-time operating parameters of the power grid, inputting these parameters into a pre-trained support vector machine model to calculate the dynamic influence coefficient of the nonlinear coupling strength between voltage and current; using this coefficient to adjust an adaptive filter to filter the real-time measurement signal; and using a neural network algorithm to extract the nonlinear error component from the filtered signal to generate a nonlinear error distribution feature vector; inputting this feature vector into a long short-term memory neural network to obtain the error compensation amount; comparing the output of the current instrument transformer with that of a standard instrument transformer to obtain the real-time measurement error; performing feedback correction through a proportional-integral controller; and finally outputting error calibration parameters to correct the measurement output of the instrument transformer in real time. This application improves the accuracy of instrument transformer measurements.

Description

Technical Field

[0001] This application relates to the field of current transformer error calibration technology, and in particular to a method and system for online calibration of current transformer errors based on deep learning. Background Technology

[0002] Instrument transformers are crucial devices in power systems for measuring voltage and current. They sense current and voltage signals in the power grid and convert these signals into low-level signals suitable for instrumentation. However, due to fluctuations in grid load, changes in voltage levels, and environmental interference, the measurement signals of instrument transformers often contain errors. These errors can lead to power system mis-dispatch, protection failures, and even system malfunctions. Traditional instrument transformer error calibration methods mainly rely on manual intervention and laboratory calibration, typically performed during power system outages or long-term maintenance. This not only wastes significant time and resources but also makes real-time and online calibration difficult to achieve.

[0003] As power systems become increasingly automated, more and more equipment requires real-time monitoring and calibration during operation to ensure system accuracy and reliability. Traditional calibration methods suffer from poor timeliness and insufficient flexibility, especially under the complex operating conditions of high-voltage power grids, where traditional technologies struggle to handle nonlinear, time-varying, and high-frequency errors. To address these issues, deep learning and adaptive control technologies have been gradually introduced into the field of power equipment calibration in recent years. These technologies, by acquiring real-time power grid operation data and combining them with machine learning algorithms, can effectively identify and compensate for transformer errors, achieving the goal of online automatic calibration.

[0004] However, achieving efficient online calibration of instrument transformer errors by combining deep learning, classical control algorithms, and reinforcement learning, especially in the face of the complex and ever-changing operating environment of the power grid, remains a challenging technical problem. This invention proposes a deep learning-based online calibration method for instrument transformer errors. By introducing technologies such as LSTM, SVM, and PI controllers, it provides a novel solution that can adjust the measurement output of the instrument transformer in real time, thereby improving the overall performance and stability of the power system. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a deep learning-based online calibration method and system for instrument transformer errors, which can efficiently and accurately correct instrument transformer measurement errors in a dynamically changing power grid environment, thereby improving the accuracy and stability of power grid measurements.

[0006] In a first aspect, this application provides an online calibration method for mutual inductor errors based on deep learning, the method comprising: Step S1: By deploying sensor arrays at key nodes of the high-voltage power grid, real-time operating condition parameters of the high-voltage power grid are collected synchronously. The real-time operating condition parameters are input into a pre-trained support vector machine model to calculate the dynamic influence coefficient representing the nonlinear coupling strength between voltage and current. Step S2: Use the dynamic influence coefficient as the adjustment parameter of the adaptive filter to filter the real-time measurement signal of the current transformer, and use a neural network algorithm to separate the residual signal representing the nonlinear error from the filtered signal, and generate a nonlinear error distribution feature vector. Step S3: Input the time change sequence of the nonlinear error distribution feature vector within the historical time period into the long short-term memory neural network to obtain the error compensation amount; Step S4: Compare the output of the current current transformer with that of the high-precision standard current transformer to obtain the real-time measurement error of the current current transformer. Use the error compensation amount as feedforward compensation and use a proportional-integral controller to perform feedback correction on the real-time measurement error. Output the final error calibration parameter and use the error calibration parameter to correct the measurement output of the current current transformer in real time.

[0007] In conjunction with the first aspect, in the first implementation of the first aspect of this application, the synchronous acquisition of real-time operating condition parameters of the high-voltage power grid in step S1 includes: The real-time operating condition parameters include voltage level data and load current data in the high-voltage power grid. The voltage level data and the load current data are sampled synchronously to obtain real-time voltage and real-time current values. The time-domain signal of the operating condition parameters is acquired, and the time-domain signal is filtered through a preset low-pass filter to obtain a preliminary denoised signal. The dominant frequency of the preliminary denoised signal is extracted, and the preliminary denoised signal is filtered a second time based on the dominant frequency to obtain the denoised operating condition parameters. The denoised operating condition parameters are normalized to obtain the final operating condition parameters.

[0008] In conjunction with the first aspect, in the second implementation of the first aspect of this application, the dynamic influence coefficient representing the nonlinear coupling strength between voltage and current calculated in step S1 includes: Sample pairs are formed based on the voltage level data and load current data in the operating condition parameters. These pairs are input into a support vector machine regression model, which maps them to a high-dimensional feature space. In this high-dimensional feature space, the voltage level data and the load current data are separated. The voltage level fluctuation pattern is identified as the voltage level feature, and the load current variation pattern is identified as the load current feature. A vector inner product operation is performed on the voltage level feature and the load current feature to obtain a preliminary interaction value. A weighted average method is applied to adjust the preliminary interaction value, and the interaction influence weight is calculated. Based on the interaction influence weight, the dynamic influence coefficients of the voltage level and load current are determined. The dynamic influence coefficients are then normalized to obtain standardized dynamic influence coefficients.

[0009] In conjunction with the first aspect, in the third implementation of the first aspect of this application, step S2 includes: The kernel weight distribution of the adaptive filter is dynamically configured based on the dynamic influence coefficient, wherein the dynamic influence coefficient is negatively correlated with the bandwidth of the filter kernel. The configured adaptive filter is used to convolve the real-time measurement signal to output a preliminary filtered signal. An autoencoder neural network is used to separate the fundamental component and the residual signal from the preliminary filtered signal. The residual signal is used as a nonlinear transformation error component. Statistical analysis is performed on the nonlinear transformation error component to obtain the mean and variance of the error distribution. The result is combined to generate a nonlinear error distribution feature vector.

[0010] In conjunction with the first aspect, in the fourth implementation of the first aspect of this application, step S3 includes: A long short-term memory network is used to perform time-domain analysis on the time-varying sequence of the nonlinear error distribution feature vector to extract time-domain distribution features. A time series model is constructed based on the time-domain distribution features to analyze the time-varying pattern of the nonlinear error. Key parameters are extracted from the time-varying pattern and a compensation function is constructed. The compensation function is applied to generate an error compensation amount. The error compensation amount is smoothed to obtain a stable error compensation amount.

[0011] In conjunction with the first aspect, in the fifth implementation of the first aspect of this application, extracting key parameters from the time-varying mode and constructing a compensation function includes: Based on the time-varying mode, a linear compensation function or a nonlinear neural network is dynamically selected as the compensation function. The extracted key parameters are input into the selected compensation function to calculate the preliminary compensation amount. Within a preset cycle of power grid operation, the preliminary compensation amount at multiple measurement time points is calculated by moving average to obtain the error compensation amount.

[0012] In conjunction with the first aspect, in the sixth implementation of the first aspect of this application, the final error calibration parameters output in step S4 include: The real-time measurement error is input into the proportional-integral controller. After proportional and integral operations, the real-time compensation increment is output. The final error calibration parameters are obtained based on the error compensation amount and the real-time compensation increment.

[0013] In conjunction with the first aspect, in the seventh implementation of the first aspect of this application, step S4, which uses the error calibration parameter to correct the current transformer's measurement output in real time, includes: The real-time measurement signal of the current transformer under energized conditions is acquired. The error calibration parameter is used as a correction term and applied to the original transmission equation of the current transformer in real time through an adder or multiplier. The preliminary calibration signal is directly output. The preliminary calibration signal is input into an adaptive moving average filter. The window length of the filter is dynamically adjusted according to the rate of change of the current load current. The measured signal after amplitude stabilization is output.

[0014] In conjunction with the first aspect, in the eighth implementation of the first aspect of this application, after adjusting the measurement output of the current transformer in step S4, the following is also included: The differential rate of change of the measured signal after amplitude stabilization is monitored in real time. If the differential rate of change continuously exceeds a preset threshold, it is determined to be a power grid transient process. The measured signal after amplitude stabilization is switched to the original signal calculated by the original transmission equation. Monitoring continues until the differential rate of change recovers to within the preset threshold. Then, the final output is switched to the measured signal after amplitude stabilization.

[0015] Secondly, this application provides an online calibration system for mutual inductor errors based on deep learning, the system comprising: The acquisition module is used to synchronously acquire real-time operating condition parameters of the high-voltage power grid through a sensor array deployed at key nodes of the high-voltage power grid, input the real-time operating condition parameters into a pre-trained support vector machine model, and calculate the dynamic influence coefficient representing the nonlinear coupling strength between voltage and current. The filtering module is used to filter the real-time measurement signal of the current transformer by using the dynamic influence coefficient as the adjustment parameter of the adaptive filter, and to separate the residual signal representing the nonlinear error from the filtered signal using a neural network algorithm, and to generate a nonlinear error distribution feature vector. The compensation module is used to input the time change sequence of the nonlinear error distribution feature vector within a historical time period into the long short-term memory neural network to obtain the error compensation amount; The correction module is used to compare the output of the current current transformer with that of a high-precision standard current transformer, obtain the real-time measurement error of the current current transformer, use the error compensation amount as feedforward compensation, and use a proportional-integral controller to perform feedback correction on the real-time measurement error, output the final error calibration parameter, and use the error calibration parameter to correct the measurement output of the current current transformer in real time.

[0016] Compared with the prior art, the beneficial effects of the present invention are at least as follows: The deep learning-based online calibration method for instrument transformer errors provided in this application acquires voltage and load current data of the power grid in real time by deploying a sensor array, and performs noise reduction and filtering to ensure accurate acquisition of operating parameters. A support vector machine model is used to calculate the nonlinear coupling strength between voltage and current, obtaining dynamic influence coefficients to provide a basis for error correction. A neural network algorithm is used to extract nonlinear error components, generating an error distribution feature vector. A long short-term memory neural network is then used to analyze time-varying patterns and construct a compensation function to generate the error compensation amount. A proportional-integral controller is used to fine-tune the real-time measurement error, further optimizing the error compensation amount and ensuring the accuracy of error correction. This method can promptly feedback and correct errors based on dynamic changes in the power grid, eliminating the influence of factors such as power grid load fluctuations and voltage fluctuations, and ensuring that the instrument transformer output always conforms to high-precision standards.

[0017] Meanwhile, by introducing an adaptive moving average filter, the filtering window can be dynamically adjusted according to changes in load current, thereby ensuring the stability of the measurement signal. Especially during power grid transients, it can promptly restore the system to normal measurement status, ensuring the reliability of the system under various operating conditions. In summary, the method of this application not only improves the accuracy of transformer measurement error correction, but also significantly improves the automation and intelligence level of power grid monitoring, ensuring the accuracy and operational stability of the power grid. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of an embodiment of an online calibration method for mutual inductor errors based on deep learning, as described in this application. Figure 2 This is a schematic diagram illustrating the generation of error compensation amount in an embodiment of this application; Figure 3 This is a timing diagram comparing the current transformer error before and after calibration in the embodiments of this application; Figure 4 This is a schematic diagram of one embodiment of an online calibration system for mutual inductor errors based on deep learning, as described in this application. Detailed Implementation

[0020] This application provides a method and system for online calibration of transformer errors based on deep learning. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0021] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the online calibration method for mutual inductor errors based on deep learning in this application includes: Step S1: By deploying sensor arrays at key nodes of the high-voltage power grid, real-time operating condition parameters of the high-voltage power grid are collected synchronously. The real-time operating condition parameters are input into a pre-trained support vector machine model to calculate the dynamic influence coefficient representing the nonlinear coupling strength between voltage and current.

[0022] The step S1, which involves synchronously acquiring real-time operating parameters of the high-voltage power grid, includes: acquiring voltage level data and load current data from the high-voltage power grid; synchronously sampling the voltage level data and load current data to obtain real-time voltage and current values; acquiring the time-domain signal of the operating parameters; filtering the time-domain signal using a preset low-pass filter to obtain a preliminary denoised signal; extracting the dominant frequency of the preliminary denoised signal and performing a secondary filter based on the dominant frequency to obtain the denoised operating parameters; and normalizing the denoised operating parameters to obtain the final operating parameters.

[0023] Specifically, firstly, sensor arrays deployed at key nodes of the high-voltage power grid, such as substations, transmission lines, or other important monitoring locations, collect real-time voltage level data and load current data as operating parameters. To ensure data synchronization and accuracy, a synchronous clock mechanism is employed, ensuring consistent sampling frequencies for voltage and current data, typically at a sampling rate of several thousand times per second. This synchronous sampling method effectively reduces errors caused by time differences, thereby obtaining accurate real-time voltage and current values. To remove noise from the signal, a preset low-pass filter is used to initially filter the time-domain signal of the real-time operating parameters. The low-pass filter effectively removes high-frequency noise components, ensuring signal smoothness. Subsequently, the filtered signal is further processed using a wavelet transform algorithm. Wavelet transform decomposes the signal into different frequency components, helping to identify and remove high-frequency components such as electromagnetic interference or environmental noise. Threshold filtering is applied to remove unnecessary noise, and a cleaner filtered signal is obtained through wavelet reconstruction. The key dominant frequencies, i.e., the frequency components of the signal, are extracted from the filtered operating parameters. Based on these dominant frequencies, the signal is then subjected to secondary filtering to further remove high-frequency noise. Finally, a normalization method is used to standardize the denoised operating parameters, giving them uniform dimensions and scale, ensuring data stability and consistency. Ultimately, after a series of signal processing steps, standardized and accurate real-time power grid operating parameters are obtained. This processed data will serve as input data for subsequent analysis, support vector machine model training, and dynamic influence coefficient calculation, providing precise foundational data for power grid monitoring and transformer error calibration.

[0024] The calculation of the dynamic influence coefficient representing the nonlinear coupling strength between voltage and current in step S1 includes: forming sample pairs based on voltage level data and load current data in the operating condition parameters, inputting them into the support vector machine regression model, mapping the sample pairs to a high-dimensional feature space, separating the voltage level data and load current data in the high-dimensional feature space, identifying the voltage level fluctuation pattern as the voltage level feature, and the load current variation pattern as the load current feature, performing a vector inner product operation on the voltage level feature and the load current feature to obtain a preliminary interaction value, applying a weighted average method to adjust the preliminary interaction value, calculating the interaction influence weight, determining the dynamic influence coefficient of voltage level and load current based on the interaction influence weight, and normalizing the dynamic influence coefficient to obtain a standardized dynamic influence coefficient.

[0025] Specifically, real-time collected voltage level data and load current data are used to form sample pairs. Each sample pair contains one voltage level data point and a corresponding load current data point. These sample pairs are input into a pre-trained Support Vector Machine (SVM) regression model. This regression model can process the input voltage and current data, and after a series of nonlinear mappings, it maps these data to a high-dimensional feature space. In this high-dimensional feature space, the SVM model can effectively separate the voltage level data and the load current data. Specifically, the fluctuation pattern of the voltage level data is extracted as voltage level features, while the variation pattern of the load current is extracted as load current features. These features reflect the complex relationship between voltage and current, especially how they interact over time. Next, a vector inner product operation is performed to match the voltage level features and the load current features. Through the inner product operation, a preliminary interaction value is obtained, which represents the initial interaction effect between voltage and load current. Because variations in grid operating conditions can lead to different strengths of these interaction values, a weighted averaging method is needed to adjust the initial interaction values. Specifically, firstly, based on the actual operating conditions of the grid, the weight of each sample pair is determined. For example, under high load conditions, load current changes are given a higher weight, while under low load or stable voltage conditions, voltage changes have a greater impact and can be given a higher weight. The weight allocation can be achieved through statistical analysis of historical data or machine learning algorithms. Then, the weighted interaction value for each sample pair is calculated by multiplying the initial interaction value by the corresponding weight to obtain the weighted interaction value. Next, the weighted interaction values ​​of all samples are summed and divided by the sum of all weights to calculate the weighted average interaction value. Finally, the weighted average interaction value is normalized, mapping it to a certain range, such as 0 to 1, to ensure its stability and applicability. Through this weighted averaging adjustment, the resulting dynamic influence coefficient can effectively reflect the nonlinear coupling strength between voltage and load current, providing accurate data support for subsequent error calibration and compensation.

[0026] Step S2: Use the dynamic influence coefficient as the adjustment parameter of the adaptive filter to filter the real-time measurement signal of the current transformer, and use a neural network algorithm to separate the residual signal representing the nonlinear error from the filtered signal, and generate the nonlinear error distribution feature vector.

[0027] Step S2 includes: dynamically configuring the kernel weight distribution of the adaptive filter based on the dynamic influence coefficient, wherein the dynamic influence coefficient is negatively correlated with the bandwidth of the filter kernel; using the configured adaptive filter to perform convolution processing on the real-time measurement signal to output a preliminary filtered signal; using an autoencoder neural network to separate the fundamental component and the residual signal from the preliminary filtered signal; using the residual signal as the nonlinear transformation error component; performing statistical analysis on the nonlinear transformation error component to obtain the mean and variance of the error distribution; and combining them to generate a nonlinear error distribution feature vector.

[0028] Specifically, the dynamic influence coefficient is used as an adjustment parameter for the adaptive filter to optimize its processing of real-time measurement signals. Specifically, the dynamic influence coefficient is negatively correlated with the bandwidth of the filter kernel. This means that when the dynamic influence coefficient is high, the filter bandwidth is smaller to capture high-frequency noise more precisely; conversely, when the dynamic influence coefficient is low, the filter bandwidth increases to accommodate a wider range of signal variations. Based on the dynamic influence coefficient, the weight distribution of the filter kernel is dynamically configured, enabling the filter to effectively filter signals according to the real-time operating conditions of the power grid and eliminate unnecessary noise. After configuring the filter, the real-time measurement signal from the transformer is convolved using the adaptive filter. The convolution operation performs calculations between the real-time signal and the filter kernel, outputting a preliminary filtered signal. This preliminary filtered signal removes most of the noise and unnecessary interference compared to the original signal, exhibiting higher smoothness and stability, which is beneficial for subsequent accurate signal analysis. A neural network algorithm is used to analyze the time-varying patterns of the preliminary filtered signal. The neural network receives the preliminary filtered signal through the input layer and extracts the error components in the signal through nonlinear transformation in the hidden layer. The hidden layer processes the signal using nonlinear activation functions such as ReLU or Sigmoid and automatically learns the complex patterns in the signal, identifying the residual signal representing nonlinear error. These nonlinear error components mainly originate from the dynamic changes of the power grid, equipment characteristics, or other external interference. Through the learning of the neural network, these error components can be effectively separated from the signal, providing accurate data support for subsequent analysis.

[0029] Once the nonlinear error components are extracted, the statistical distribution of this error component sequence is calculated using statistical analysis methods. Specifically, by calculating the mean and variance of the error components, the basic characteristics of the error distribution can be obtained, revealing its fluctuation range and variation patterns. The mean reflects the central trend of the error, while the variance shows the breadth of the error distribution. The combination of these two provides an important basis for subsequent error calibration and dynamic adjustment. Based on the calculated mean and variance, they are combined into a nonlinear error distribution feature vector. This feature vector contains the distribution characteristics of the error and can accurately represent the error characteristics of the power grid under different operating conditions. To ensure the stability and consistency of the feature vector, the mean and variance are normalized, ensuring that the feature vector remains consistent under different power grid operating conditions. The generated nonlinear error distribution feature vector will serve as input for subsequent error compensation analysis, helping the system accurately correct the measurement errors of the instrument transformers under dynamic power grid operating conditions, thereby improving the accuracy and reliability of power grid monitoring.

[0030] Step S3: Input the time change sequence of the nonlinear error distribution feature vector within the historical time period into the long short-term memory neural network to obtain the error compensation amount.

[0031] Step S3 includes: using a long short-term memory network to perform time-domain analysis on the time-varying sequence of the nonlinear error distribution feature vector, extracting time-domain distribution features, constructing a time series model based on the time-domain distribution features, analyzing the time-varying pattern of the nonlinear error, extracting key parameters from the time-varying pattern and constructing a compensation function, applying the compensation function to generate the error compensation amount, smoothing the error compensation amount, and obtaining a stable error compensation amount.

[0032] Specifically, the extracted nonlinear error distribution feature vector is input into a Long Short-Term Memory (LSTM) neural network for time-domain analysis. The LSTM model identifies the temporal variation patterns of the error by analyzing the historical time-varying sequences of the nonlinear error components. Specifically, the LSTM model processes error signals containing dynamic changes in the power grid, identifying and extracting the temporal distribution features of the signals. Based on the extracted temporal distribution features, a time-series model is constructed to analyze the time-varying patterns of the nonlinear error, capturing the variation patterns of the error in different time periods. These patterns reflect the dynamic impact of factors such as voltage fluctuations and load fluctuations in the power grid on measurement errors. Through model learning, key parameters such as the periodic fluctuations of the error, voltage fluctuations, and the amplitude of load current changes are extracted. Key parameters are extracted from the time-varying patterns, and a compensation function is constructed. This compensation function is then used to generate the error compensation amount; the detailed process will be explained later. To further improve the accuracy of error compensation, a smoothing method is used to optimize the error compensation amount, making the compensation value more stable and continuous. Through this smoothing process, the obtained stable error compensation amount can effectively cope with sudden fluctuations in the power grid and maintain consistency and high accuracy under different operating conditions. The error compensation amount generated by applying the compensation function is a preliminary compensation result calculated based on historical data and the time-varying pattern of power grid operating conditions. However, in actual operation, the operating conditions of the power grid will continue to change, especially under circumstances such as sudden events, instantaneous load fluctuations, or instantaneous voltage fluctuations. The error patterns of historical data may not be fully applicable or need to be appropriately adjusted.

[0033] The process of extracting key parameters from time-varying modes and constructing compensation functions includes: dynamically selecting a linear compensation function or a nonlinear neural network as the compensation function based on the time-varying mode; inputting the extracted key parameters into the selected compensation function; calculating the preliminary compensation amount; and performing a sliding average calculation on the preliminary compensation amount at multiple measurement time points within a preset cycle of power grid operation to obtain the error compensation amount.

[0034] Specifically, such as Figure 2 The diagram illustrates the generation of error compensation. A compensation function is constructed based on the extracted key parameters. This function generates a compensation value adapted to the current power grid state, taking into account the periodic fluctuations and amplitude characteristics extracted from the LSTM network. This corrects the measurement signal output by the instrument transformer, reducing errors caused by dynamic changes in the power grid. Specifically, the compensation function can be divided into linear and nonlinear forms to adapt to the error characteristics under different power grid operating conditions. If the time-varying pattern of the error exhibits a relatively linear trend, a linear compensation function can be used. ,in, This is the error compensation amount. , and These are the amplitudes of error, voltage fluctuation, and load current fluctuation, respectively. , and The weighting coefficients, learned by the LSTM model, reflect the degree of influence of these factors on the measurement error. Linear compensation functions are suitable for relatively simple power grid operating conditions and stable error changes, and can improve measurement accuracy through simple weighted correction. However, under complex power grid operating conditions, the nonlinear effects of voltage fluctuations, load changes, and other factors may cause the time-varying pattern of the error to exhibit a more complex relationship. In this case, a nonlinear compensation function is more appropriate. Nonlinear compensation functions are usually modeled using deep learning algorithms (such as neural network regression models), which can learn the complex patterns in power grid errors and fit the error compensation function through the time-domain features in the training data. The form of a nonlinear compensation function is usually as follows: , It is a function learned through neural networks or other nonlinear modeling methods, capable of automatically generating compensation values ​​based on input voltage, load current, and error amplitude. This nonlinear compensation function can flexibly respond to dynamic changes in power grid conditions, especially under conditions of large load fluctuations and drastic voltage changes, providing more accurate compensation. Multiple measurement time points are set for the entire power grid operating cycle. For the measurement error at each time point, the corresponding compensation amount is calculated using the compensation function, and the average of the compensation amounts from multiple measurement time points is calculated as the final error compensation amount.

[0035] Step S4: Compare the output of the current current transformer with that of the high-precision standard current transformer to obtain the real-time measurement error of the current current transformer. Use the error compensation amount as feedforward compensation and use a proportional-integral controller to perform feedback correction on the real-time measurement error. Output the final error calibration parameters and use the error calibration parameters to correct the measurement output of the current current transformer in real time.

[0036] In step S4, the final error calibration parameters are output by: inputting the real-time measurement error into the proportional-integral controller, performing proportional and integral operations, outputting the real-time compensation increment, and obtaining the final error calibration parameters based on the error compensation amount and the real-time compensation increment.

[0037] Specifically, by comparing the output of the current instrument transformer with that of a high-precision standard instrument transformer, the real-time measurement error of the current instrument transformer is obtained. The real-time measurement error is then processed by a proportional-integral controller to generate a real-time compensation increment. The error compensation amount is then fine-tuned using the real-time compensation increment to further optimize the compensation effect. The process of fine-tuning the error compensation amount based on the real-time measurement error using a proportional-integral controller is as follows: Since the real-time measurement error may be affected by various factors, such as instantaneous fluctuations and long-term deviations, instantaneous fluctuations are usually caused by factors such as grid load, equipment switching, faults, and environmental changes, leading to rapid changes in error. Long-term deviations are caused by factors such as equipment aging, long-term grid instability, and equipment performance drift, leading to the gradual accumulation of error. These instantaneous fluctuations and long-term deviations make the real-time measurement error more complex, so it is necessary to appropriately correct the real-time measurement error.

[0038] This application introduces a proportional-integral (PI) controller to calculate the real-time compensation increment to correct the real-time measurement error of the current transformer. Specifically, the real-time measurement error is first input into the PPI controller, which calculates the real-time compensation increment through both proportional and integral operations. The proportional component is based on the real-time measurement error at the current moment. The formula for calculating the compensation increment is as follows: , It is the proportional gain coefficient, which determines the magnitude of the proportional response error. It is a moment The real-time measurement error is compensated by the following formulas: the proportional part is calculated by directly multiplying the error by a proportionality coefficient to generate a compensation amount proportional to the current error; the integral part is calculated by accumulating the real-time measurement error over a certain time range, that is, correcting the accumulation of the real-time error over time. The formula is as follows: , From 0 to the current time The cumulative value of error, It is a moment The real-time measurement error is eliminated by the integral part, which accumulates the errors to eliminate long-term, persistent small errors, especially systematic deviations caused by equipment aging, long-term changes in grid load, etc. The combined result of the proportional and integral parts is... + As the final real-time compensation increment, the real-time compensation increment is weighted and added to the error compensation amount to obtain a new compensation amount, which is the final error calibration parameter. This error calibration parameter will be applied to the measurement signal of the instrument transformer to correct its output in real time, ensuring that the measured value of the instrument transformer is consistent with the output of the standard instrument transformer, thereby eliminating errors caused by power grid fluctuations, load changes, etc., and ensuring measurement accuracy and system stability.

[0039] In step S4, the real-time correction of the current transformer's measurement output using error calibration parameters includes: acquiring the real-time measurement signal of the transformer under energized conditions, using the error calibration parameters as correction terms, applying them in real-time to the transformer's original transmission equation through an adder or multiplier, directly outputting a preliminary calibration signal, inputting the preliminary calibration signal into an adaptive moving average filter, dynamically adjusting the filter's window length according to the rate of change of the current load current, and outputting a stabilized measurement signal.

[0040] Specifically, firstly, the measurement signals of the current transformer under energized conditions are acquired in real time, reflecting the current operating state of the power grid. After acquiring the real-time measurement signals, the previously calculated error calibration parameters are used as correction terms and applied in real time to the original transfer equation of the current transformer. The original transfer equation is a mathematical expression describing the relationship between the current transformer measurement signals and actual physical quantities of the power grid (such as current and voltage). For example, for a current transformer (CT), its transfer equation can be expressed as: ,in, It is the output current of the current transformer. The actual current of the power grid, The conversion ratio represents the proportional relationship between the grid current and the transformer's output current. For current transformers, if the measured current value has an error, the error calibration parameter can adjust this current value to ensure a more accurate output from the transformer, i.e., a corrected current value. It can be represented as: ,in, The initial calibration signal, after correction to the error calibration parameters, represents a preliminary correction of the original measurement signal, but may still contain minor fluctuations or noise. To further smooth the signal and improve measurement accuracy, the initial calibration signal is input into an adaptive moving average filter. The filter's window length is dynamically adjusted based on the rate of change of the current load current. Specifically, when the load current fluctuates significantly, the filter window length adaptively increases to better smooth the volatile signal; when the load current changes less, the window length decreases to ensure a rapid response to minor fluctuations in the power grid. Finally, this filtering process outputs a stabilized measurement signal—a stable measurement value after removing fluctuations and noise—ensuring the accuracy and stability of the transformer output.

[0041] In step S4, after adjusting the measurement output of the current transformer, the following steps are also included: real-time monitoring of the differential change rate of the measured signal after amplitude stabilization; if the differential change rate continuously exceeds a preset threshold, it is determined to be a power grid transient process, and the measured signal after amplitude stabilization is switched to the original signal calculated by the original transmission equation. Monitoring continues until the differential change rate recovers to within the preset threshold, and then the final output is switched to the measured signal after amplitude stabilization.

[0042] Specifically, after applying error calibration parameters to the transformer's measurement signal and outputting a stabilized measurement signal through a filter, the stabilized signal undergoes differential processing to calculate the differential rate of change. This rate of change represents the rate at which the stabilized measurement signal changes over time, reflecting the dynamic changes in the power grid and potential transient processes. If the differential rate of change continuously exceeds a preset threshold, it indicates that the power grid may have experienced instantaneous load changes or faults, such as sudden power grid fluctuations caused by short-circuit faults or lightning strikes. To avoid misinterpreting such transient processes as normal fluctuations, the system switches the current stabilized measurement signal to the original signal calculated from the original transmission equation. This ensures that the transformer's measurement is not incorrectly corrected during power grid transients, preventing interference with error compensation results from transient processes. When the transient process ends, the system continuously monitors the differential rate of change until it returns to within the preset threshold, indicating that the power grid has returned to a stable state. At this point, it switches back to the stabilized measurement signal to ensure that the transformer's measurement signal accurately reflects the stable operating state of the power grid.

[0043] like Figure 3 The figure shows a time-series diagram comparing the current transformer error before and after calibration. The original error curve (solid black line) represents the change of the uncalibrated error over time during the current transformer measurement process. The calibrated error curve (dashed black line) shows the change of the current transformer's measurement error over time after error calibration; typically, the calibrated error is small and close to zero. The error reduction area (gray shaded area) indicates the range of measurement error reduction after error calibration, demonstrating the improvement effect of the error before and after calibration. This figure clearly shows that the measurement error of the current transformer is significantly corrected through error calibration, especially in the middle and later stages of the time series, where the error reduction is obvious, reflecting the effectiveness of calibration technology in improving measurement accuracy.

[0044] The above describes a deep learning-based online calibration method for current transformer errors in embodiments of this application. The following describes a deep learning-based online calibration system for current transformer errors in embodiments of this application. Please refer to [link to relevant documentation]. Figure 4 One embodiment of the deep learning-based online calibration system for mutual inductor errors in this application includes: The acquisition module is used to synchronously acquire real-time operating parameters of the high-voltage power grid through sensor arrays deployed at key nodes of the high-voltage power grid. The real-time operating parameters are then input into a pre-trained support vector machine model to calculate the dynamic influence coefficient representing the nonlinear coupling strength between voltage and current.

[0045] The filtering module is used to filter the real-time measurement signal of the current transformer by using the dynamic influence coefficient as the adjustment parameter of the adaptive filter, and uses a neural network algorithm to separate the residual signal representing the nonlinear error from the filtered signal and generate the nonlinear error distribution feature vector.

[0046] The compensation module is used to input the time change sequence of the nonlinear error distribution feature vector within a historical time period into the long short-term memory neural network to obtain the error compensation amount.

[0047] The correction module is used to compare the output of the current current transformer with that of the high-precision standard current transformer, obtain the real-time measurement error of the current current transformer, use the error compensation amount as feedforward compensation, and use a proportional-integral controller to perform feedback correction on the real-time measurement error, output the final error calibration parameters, and use the error calibration parameters to correct the measurement output of the current current transformer in real time.

[0048] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0049] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0050] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A deep learning-based online calibration method for mutual inductor errors, characterized in that, The method includes: Step S1: By deploying sensor arrays at key nodes of the high-voltage power grid, real-time operating condition parameters of the high-voltage power grid are collected synchronously. The real-time operating condition parameters are input into a pre-trained support vector machine model to calculate the dynamic influence coefficient representing the nonlinear coupling strength between voltage and current. Step S2: Use the dynamic influence coefficient as the adjustment parameter of the adaptive filter to filter the real-time measurement signal of the current transformer, and use the neural network algorithm to separate the residual signal representing the nonlinear error from the filtered signal and generate the nonlinear error distribution feature vector. Step S3: Input the time change sequence of the nonlinear error distribution feature vector within the historical time period into the long short-term memory neural network to obtain the error compensation amount; Step S4: Compare the output of the current current transformer with that of the high-precision standard current transformer to obtain the real-time measurement error of the current current transformer. Use the error compensation amount as feedforward compensation and use a proportional-integral controller to perform feedback correction on the real-time measurement error. Output the final error calibration parameters and use the error calibration parameters to correct the measurement output of the current current transformer in real time.

2. The method according to claim 1, characterized in that, The real-time operating parameters of the high-voltage power grid that are synchronously collected in step S1 include: The real-time operating condition parameters include voltage level data and load current data in the high-voltage power grid. The voltage level data and the load current data are sampled synchronously to obtain real-time voltage and real-time current values. The time-domain signal of the operating condition parameters is acquired, and the time-domain signal is filtered through a preset low-pass filter to obtain a preliminary denoised signal. The dominant frequency of the preliminary denoised signal is extracted, and the preliminary denoised signal is filtered a second time based on the dominant frequency to obtain the denoised operating condition parameters. The denoised operating condition parameters are normalized to obtain the final operating condition parameters.

3. The method according to claim 2, characterized in that, The dynamic influence coefficients representing the nonlinear coupling strength between voltage and current calculated in step S1 include: Sample pairs are formed based on the voltage level data and load current data in the operating condition parameters. These pairs are input into a support vector machine regression model, which maps them to a high-dimensional feature space. In this high-dimensional feature space, the voltage level data and the load current data are separated. The voltage level fluctuation pattern is identified as the voltage level feature, and the load current variation pattern is identified as the load current feature. A vector inner product operation is performed on the voltage level feature and the load current feature to obtain a preliminary interaction value. A weighted average method is applied to adjust the preliminary interaction value, and the interaction influence weight is calculated. Based on the interaction influence weight, the dynamic influence coefficients of the voltage level and load current are determined. The dynamic influence coefficients are then normalized to obtain standardized dynamic influence coefficients.

4. The method according to claim 1, characterized in that, Step S2 includes: The kernel weight distribution of the adaptive filter is dynamically configured based on the dynamic influence coefficient, wherein the dynamic influence coefficient is negatively correlated with the bandwidth of the filter kernel. The configured adaptive filter is used to convolve the real-time measurement signal to output a preliminary filtered signal. An autoencoder neural network is used to separate the fundamental component and the residual signal from the preliminary filtered signal. The residual signal is used as a nonlinear transformation error component. Statistical analysis is performed on the nonlinear transformation error component to obtain the mean and variance of the error distribution. The result is combined to generate a nonlinear error distribution feature vector.

5. The method according to claim 1, characterized in that, Step S3 includes: A long short-term memory network is used to perform time-domain analysis on the time-varying sequence of the nonlinear error distribution feature vector to extract time-domain distribution features. A time series model is constructed based on the time-domain distribution features to analyze the time-varying pattern of the nonlinear error. Key parameters are extracted from the time-varying pattern and a compensation function is constructed. The compensation function is applied to generate an error compensation amount. The error compensation amount is smoothed to obtain a stable error compensation amount.

6. The method according to claim 5, characterized in that, Extracting key parameters from the time-varying mode and constructing a compensation function includes: Based on the time-varying mode, a linear compensation function or a nonlinear neural network is dynamically selected as the compensation function. The extracted key parameters are input into the selected compensation function to calculate the preliminary compensation amount. Within a preset cycle of power grid operation, the preliminary compensation amount at multiple measurement time points is calculated by moving average to obtain the error compensation amount.

7. The method according to claim 1, wherein the final error calibration parameters output in step S4 include: The real-time measurement error is input into the proportional-integral controller. After proportional and integral operations, the real-time compensation increment is output. The final error calibration parameters are obtained based on the error compensation amount and the real-time compensation increment.

8. The method according to claim 1, characterized in that, Step S4, which involves using the error calibration parameters to correct the current transformer's measurement output in real time, includes: The real-time measurement signal of the current transformer under energized conditions is acquired. The error calibration parameter is used as a correction term and applied to the original transmission equation of the current transformer in real time through an adder or multiplier. The preliminary calibration signal is directly output. The preliminary calibration signal is input into an adaptive moving average filter. The window length of the filter is dynamically adjusted according to the rate of change of the current load current. The measured signal after amplitude stabilization is output.

9. The method according to claim 8, characterized in that, After adjusting the measurement output of the current transformer in step S4, the following steps are also included: The differential rate of change of the measured signal after amplitude stabilization is monitored in real time. If the differential rate of change continuously exceeds a preset threshold, it is determined to be a power grid transient process. The measured signal after amplitude stabilization is switched to the original signal calculated by the original transmission equation. Monitoring continues until the differential rate of change recovers to within the preset threshold. Then, the final output is switched to the measured signal after amplitude stabilization.

10. A deep learning-based online calibration system for mutual inductor errors, used to implement the deep learning-based online calibration method for mutual inductor errors as described in any one of claims 1-9, characterized in that, The system includes: The acquisition module is used to synchronously acquire real-time operating condition parameters of the high-voltage power grid through a sensor array deployed at key nodes of the high-voltage power grid, input the real-time operating condition parameters into a pre-trained support vector machine model, and calculate the dynamic influence coefficient representing the nonlinear coupling strength between voltage and current. The filtering module is used to filter the real-time measurement signal of the current transformer by using the dynamic influence coefficient as the adjustment parameter of the adaptive filter, and to separate the residual signal representing the nonlinear error from the filtered signal using a neural network algorithm, and to generate a nonlinear error distribution feature vector. The compensation module is used to input the time change sequence of the nonlinear error distribution feature vector within a historical time period into the long short-term memory neural network to obtain the error compensation amount; The correction module is used to compare the output of the current current transformer with that of a high-precision standard current transformer, obtain the real-time measurement error of the current current transformer, use the error compensation amount as feedforward compensation, and use a proportional-integral controller to perform feedback correction on the real-time measurement error, output the final error calibration parameter, and use the error calibration parameter to correct the measurement output of the current current transformer in real time.

Images