An electronic voltage transformer error prediction method and system

Abstract

This invention relates to the field of power technology and provides a method and system for predicting errors in electronic voltage transformers. The method includes: collecting operational data from operational electronic voltage transformers; constructing a physical information feature set for the electronic voltage transformers based on their physical characteristics and maintenance experience; constructing a physical loss function based on physical constraints, replacing the original loss function of the XGBoost model, and obtaining a new learning model, phy-XGB; inputting historical errors and the physical information feature set into the phy-XGB learning model to train a ratio error prediction model for the electronic voltage transformers; and calculating the average absolute value (SHAP) of the physical information features over a time period to determine the corresponding physical information features requiring focused attention. This method improves the accuracy of measurement error prediction for electronic voltage transformers and uses the SHAP values ​​of each physical information feature over a time period to determine the key areas of focus, guiding actual operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of power technology, specifically to an electronic voltage transformer error prediction method and system. Background Technology

[0002] Electronic voltage transformers are key devices for voltage measurement and signal conversion in power systems, and are the core components for sensing voltage status. Capacitive voltage divider type electronic voltage transformers (EVV transformers) are widely used in high-voltage power grids of 110kV and above due to their high insulation strength and low cost. However, the measurement errors of electronic voltage transformers are easily affected by various factors, making it difficult to maintain long-term stability, directly impacting the safe operation of the power grid and fair settlement. Therefore, if measurement errors of electronic voltage transformers can be detected and diagnosed in advance, metering disputes and relay protection malfunctions caused by measurement errors can be avoided in a timely manner. Furthermore, unnecessary operation and maintenance costs can be reduced, and the risk of power outages can be lowered, ultimately providing important support for the stable, economical, and reliable operation of the power system.

[0003] However, current research on error perception for measurement equipment is largely focused on real-time performance, which has largely solved the dilemma of traditional offline error perception methods. For example, Chinese patent application CN118131109A, "An Online Self-Detection Method and System for Error of Capacitive Voltage Transformer," proposes an online error monitoring method independent of a standard, which uses dual-sided thresholds to determine the direction of the transformer's ratio difference change. However, such methods can only reflect the current measurement error and can only map the current operating state of the measurement equipment. They are difficult to perceive measurement errors in advance, let alone control the future operating state of the measurement equipment. Meanwhile, methods for early error perception and system reliability analysis for similar equipment are mainly divided into physical-driven and data-driven methods. While physical-driven methods can roughly perceive the direction of measurement error change in advance, they are difficult to output precise error values, and cannot provide direct and effective decision support for actual operation and maintenance. Furthermore, a complete characterization of the physical change process of the equipment requires a large and comprehensive knowledge base, which is difficult to achieve in actual modeling. A series of assumptions must be used to determine model parameters, and the modeling errors introduced by this cannot be completely eliminated. Purely data-driven methods, such as the TCN network-based prediction method disclosed in Chinese invention patent application CN119862791A entitled "An Error Prediction Method and System for Electronic Voltage Transformers," have poor performance under actual operating conditions, and the system accuracy is easily affected by system fluctuations, environmental factors, and equipment noise. Most scholars explain this error by relying solely on conventional observation data without incorporating any understanding of the internal physical mechanisms of the system that generates the observation data.

[0004] In summary, there are few methods for early detection of measurement errors in electronic voltage transformers. Similar methods for detecting transformer measurement errors fall into two categories: physical-driven and data-driven. Purely physical modeling cannot definitively determine future error values ​​and is difficult to build a complete knowledge base, thus failing to support practical engineering applications. On the other hand, purely data-driven methods lack physical constraints, are susceptible to noise interference, and suffer from insufficient robustness and accuracy. Furthermore, traditional machine learning and deep learning often suffer from the "black box" dilemma, struggling to explain the impact of various features and providing targeted suggestions for operation and maintenance. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to improve the accuracy of measurement error prediction for electronic voltage transformers, and how to use the SHAP values ​​of various physical information characteristics within a time period to determine the focus of attention and guide actual operation and maintenance.

[0006] The present invention solves the above-mentioned technical problems through the following technical means:

[0007] This invention provides a method for predicting errors in electronic voltage transformers, comprising the following steps:

[0008] S1. Collect operating data of the electronic voltage transformers that have been put into operation;

[0009] S2. Based on the physical characteristics and operation and maintenance experience of electronic voltage transformers, construct a set of physical information features of electronic voltage transformers, including load temperature sensitivity coefficient LTSC, environmental comprehensive coefficient ECI, load temperature interaction coefficient LTQI, load relative fluctuation coefficient RLF, and ratio difference trend coefficient RHT.

[0010] S3. Combine physical constraints to construct a physical loss function, replace the original loss function of the XGBoost model, and obtain a new learning model phy-XGB;

[0011] S4. Input the historical error and physical information feature set into the learning model phy-XGB to train the ratio difference prediction model of the electronic voltage transformer.

[0012] S5. Calculate the average absolute SHAP value of the physical information features over the time period, and focus on the corresponding physical information features in descending order.

[0013] Furthermore, the operational data mentioned in step 1 includes temperature, humidity, load, magnetic field strength, and historical ratio error value, as shown in the following formula:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019] in, Represents temperature. Represents humidity. Represents load, Represents magnetic field strength. This represents the historical ratio error value; For the first n The temperature of the second sample, For the first n Humidity of the second sample For the first n The load of the next sampling For the first n The magnetic field strength of the next sample. For the first n The ratio error of the sampling This represents transposition.

[0020] Further, the load temperature sensitivity coefficient LTSC mentioned in step S2 is as follows:

[0021]

[0022] Among them, LTSC i For the first i The load temperature sensitivity coefficient at any given time. It is a sliding window. and For the load and temperature sequence within the window, To find the minimum value, avoid a denominator of 0. The average load within the window is used for normalization to eliminate differences in the magnitude of load and temperature.

[0023] Furthermore, the Environmental Integration Index (ECI) mentioned in step S2 is as follows:

[0024]

[0025] Among them, ECI i For the first i The environmental comprehensive coefficient at any given time , , The first i Standardized values ​​of temperature, humidity, and magnetic field strength at any given time. , , These are their respective weights.

[0026] Furthermore, the load temperature interaction coefficient LTQI mentioned in step S2 is as follows:

[0027]

[0028] Among them, LTQI i For the first i The load-temperature interaction coefficient at any given time. , The respective i Standard values ​​for load and temperature at all times.

[0029] Further, the load relative fluctuation coefficient RLF mentioned in step S2 is as follows:

[0030]

[0031] Among them, RLF i For the first i The relative load fluctuation coefficient at any given time. The load sequence within the window, To obtain the standard deviation of the sequence.

[0032] Furthermore, the ratio difference trend coefficient RHT mentioned in step S2 is as follows:

[0033]

[0034] Among them, RHT i For the first i The trend of the difference at different times, For the first i The difference in time, It is a sliding window.

[0035] Furthermore, the physical loss function described in step S3 is as follows:

[0036]

[0037] in, MSE is the original loss function of the XGBoost model; cov(,) is the covariance, as shown in the following equation:

[0038]

[0039] in, For the first i The true value of each sample For the first i Final predicted value for each sample n The total number of samples; Physical weights; P i For the first iThe load of each sample, This represents the average load of the batch. This is the predicted mean for the batch sample.

[0040] Further, step S5 includes the following steps:

[0041] S51. Conduct SHAP time-attribution analysis on the ratio difference prediction model of the electronic voltage transformer, and calculate the set of SHAP values ​​for each physical information feature on the predicted ratio difference of the electronic voltage transformer. As shown in the following formula:

[0042]

[0043] in, s For any physical information feature, f shap-phyXGB The SHAP analysis calculation formula for the learning model phy-XGB; any element Physical information characteristics j The i SHAP value at time;

[0044] S52. Conduct SHAP time-period attribution analysis on the ratio difference prediction model of electronic voltage transformers, and select the target time period. That is, the first m To the n For each time point, calculate the average absolute SHAP value for the target time period, as follows:

[0045]

[0046] in, Physical information characteristics for this time period j The average absolute SHAP value;

[0047] S53. Sort the average absolute SHAP values ​​of each physical information feature in descending order to determine the corresponding physical information features that need to be focused on.

[0048] This invention also provides an electronic voltage transformer error prediction system, which executes the above-described method during system operation and includes the following modules:

[0049] The data acquisition module is used to collect operating data from the electronic voltage transformers that have been put into operation.

[0050] The physical feature construction module is used to construct a set of physical information features of electronic voltage transformers by combining the physical characteristics and operation and maintenance experience of electronic voltage transformers, including load temperature sensitivity coefficient LTSC, environmental comprehensive coefficient ECI, load temperature interaction coefficient LTQI, load relative fluctuation coefficient RLF, and ratio difference trend coefficient RHT.

[0051] The prediction model building module is used to combine physical constraints to construct a physical loss function, replace the original loss function of the XGBoost model, and obtain a new learning model phy-XGB.

[0052] The prediction model training module is used to input the historical error and physical information feature set into the learning model phy-XGB to train the ratio difference prediction model of the electronic voltage transformer.

[0053] The SHAP value calculation module is used to calculate the average absolute SHAP value of physical information features over a time period and sort them in descending order, identifying the physical information features with larger average absolute SHAP values ​​as key focus objects.

[0054] The advantages of this invention are:

[0055] (1) This invention proposes a physical information enhancement mechanism to make physical mechanisms such as load fluctuation, environmental interference, and periodic characteristics visible, transforming qualitative experience of power operation and maintenance into quantitative features, enhancing the physical interpretability of the model, laying the foundation for subsequent model construction and operation and maintenance decisions, and more importantly, accurately predicting the measurement error of electronic voltage transformers under the condition of multi-factor coupling.

[0056] (2) This invention proposes a physical loss function that combines the variation law of the ratio difference of electronic voltage transformers and incorporates it into the XGBoost machine learning framework to construct the physical information machine learning model phy-XGB, which makes the model prediction more in line with objective laws and greatly improves the prediction accuracy.

[0057] (3) This invention proposes an improved method for calculating the SHAP value of each physical information feature of an electronic voltage transformer from point to time period, explains the influence of each feature on the prediction ratio difference, and thus supports more precise operation and maintenance control, which has good practical value and economic benefits. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating an electronic voltage transformer error prediction method according to an embodiment of the present invention.

[0059] Figure 2 This is a schematic diagram comparing the prediction results of the method of this invention with those of other models.

[0060] Figure 3 This is a schematic diagram of SHAP analysis of various physical information features in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1

[0063] This embodiment provides an error prediction method for electronic voltage transformers, aiming to address the problems of limited prediction methods for measurement errors in electronic voltage transformers, large errors in existing methods for similar transformers, difficulty in integrating the advantages of data and physical driving forces, and lack of corresponding model interpretation, making it difficult to provide targeted suggestions for operation and maintenance. The specific implementation process is as follows: Figure 1 As shown, it includes the following steps:

[0064] S1. Collect operating data of the electronic voltage transformers in operation; the operating data includes temperature, humidity, load, magnetic field strength, and historical ratio error values, as shown in the following formula:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070] in, Represents temperature. Represents humidity. Represents load, Represents magnetic field strength. This represents the historical ratio error value; For the first n The temperature of the second sample, For the first n Humidity of the second sample For the first n The load of the next sampling For the first n The magnetic field strength of the next sample. For the first n The ratio error of the sampling This represents transposition.

[0071] S2. Based on the physical characteristics and operation and maintenance experience of electronic voltage transformers, construct a physical information feature set for electronic voltage transformers, including the load temperature sensitivity coefficient LTSC, environmental comprehensive coefficient ECI, load temperature interaction coefficient LTQI, load relative fluctuation coefficient RLF, and ratio difference trend coefficient RHT; specifically including:

[0072] Engineering practice shows that the measurement error of electronic voltage transformers is affected by both load temperature changes and ambient temperature changes. Load temperature changes alter the threshold voltage of the internal semiconductor devices, the resistance of resistive elements, and the dielectric constant of the capacitors. Ambient temperature changes affect the temperature distribution of the equipment casing and internal components through heat conduction and convection. However, the two contribute differently to the measurement error, necessitating the quantification of this relative sensitivity. The load temperature sensitivity coefficient (LTSC) is given by the following formula:

[0073]

[0074] Among them, LTSC i For the first i The load temperature sensitivity coefficient at any given time. This is a sliding window (12 in this embodiment, representing two hours). and For the load and temperature sequence within the window, To find the minimum value, avoid a denominator of 0. The average load within the window is used for normalization to eliminate differences in the magnitude of load and temperature.

[0075] The insulation performance and electromagnetic characteristics of electronic voltage transformers are affected by the combined influence of temperature, humidity, and magnetic field strength. A single environmental factor cannot fully reflect the equipment's operating status; therefore, an Environmental Comprehensive Index (ECI) is established, calculated using a weighted summation method. The weights are determined by the standardized temperature, humidity, and magnetic field strength and their corresponding historical ratio difference sequences Y. The ECI is expressed as follows:

[0076]

[0077] Among them, ECI i For the first i The environmental comprehensive coefficient at any given time , , The first i Standardized values ​​of temperature, humidity, and magnetic field strength at any given time. , , These are their respective weights.

[0078] At high temperatures, the heat dissipation efficiency of equipment drops sharply. Operational experience and experimental data both indicate that the temperature characteristics of many electronic components contain significant nonlinear components, which may cause nonlinear drift in the differential of electronic voltage transformers. Therefore, the load temperature interaction coefficient LTQI is used to capture nonlinear changes in the differential. During the operation of an electronic voltage transformer, the Joule heat generated by the load current and the ambient temperature jointly determine the operating temperature of the equipment, forming a nonlinear thermal balance relationship: the heat generated by the load is proportional to the square of the current, while the ambient temperature affects the heat dissipation efficiency; the actual temperature rise is determined by both. The aforementioned load temperature interaction coefficient LTQI is as follows:

[0079]

[0080] Among them, LTQI i For the first i The load-temperature interaction coefficient at any given time. , The respective i Standard values ​​for load and temperature at all times.

[0081] The ratio error of electronic voltage transformers is highly sensitive to load shocks, and load fluctuations introduce additional dynamic error components. Operational experience shows that the error characteristics of electronic voltage transformers differ across different load ranges, and relevant standards stipulate that the ratio error of electronic voltage transformers only needs to meet the standard within the rated load range. Therefore, a load relative fluctuation coefficient (RLF) is introduced, as follows:

[0082]

[0083] Among them, RLF i For the first i The relative load fluctuation coefficient at any given time. The load sequence within the window, To obtain the standard deviation of the sequence.

[0084] The voltage ratio difference exhibits significant time-periodic characteristics: it displays a regular "peak-valley-peak" pattern. Therefore, a ratio difference trend coefficient (RHT) is introduced to reflect the direction and rate of change of the ratio difference. The ratio difference trend coefficient RHT is given by the following formula:

[0085]

[0086] Among them, RHT i For the first i The trend of the difference at different times, For the first i The difference in time, The sliding window is 12 in this embodiment, representing a two-hour window.

[0087] Through the above process, combined with physical information and operation and maintenance experience, a set of five physical information characteristics was constructed.

[0088] S3. When the primary load increases, the main circuit current will rise synchronously. According to the principle of electromagnetic induction, the increase in current will significantly enhance the power frequency electric field strength around the busbar. The high-voltage arm capacitor of the electronic voltage transformer is closer to this electric field. Due to the increased electric field strength, the dielectric of the high-voltage arm capacitor will generate more intense dipole polarization motion. The energy loss caused by this motion will increase with the square of the electric field strength, which is manifested as an increase in the loss tangent (tanδ) of the dielectric. The increased loss directly leads to an increase in the equivalent resistance of the high-voltage arm capacitor. Under the action of capacitive current, the heat loss increases significantly, causing the temperature of the high-voltage arm capacitor to gradually rise from the ambient temperature. The increase in temperature will further cause a change in the dielectric constant of the dielectric. Commonly used dielectric fillers such as polypropylene have a negative temperature coefficient. Therefore, the capacitance of the high-voltage arm capacitor will eventually decrease with increasing temperature, resulting in a positive specific gravity ε that increases with increasing load. However, this increasing trend has a clear physical saturation characteristic, that is, there is a certain upper limit to the increase of specific gravity with increasing load. This upper limit depends on the saturation characteristics of the dielectric and the insulation performance of the equipment. This change is not an absolute increase in the secondary voltage, but rather a relative increase in the measured value compared to the true value. Essentially, the primary load alters the voltage division characteristics of the electronic voltage transformer through links such as electric field and temperature, ultimately manifesting as a positive shift in the ratio difference.

[0089] Therefore, this positive correlation is necessary for predicting electronic voltage transformer performance, but its strength must be limited by the dielectric characteristics and engineering standards of the electronic voltage transformer. An excessively strong positive correlation violates physical laws. In summary, to balance statistical prediction accuracy with the constraints of equipment physical characteristics, a physical loss function is constructed to replace the original loss function of the XGBoost model, resulting in a new learning model, phy-XGB. The physical loss function is as follows:

[0090]

[0091] in, MSE is the original loss function of the XGBoost model; cov(,) is the covariance, as shown in the following equation:

[0092]

[0093] in, For the first i The true value of each sample For the first i Final predicted value for each sample n The total number of samples; The physical weights are optimized using cross-validation. In this embodiment, 0.1 is used as the initial value, and the final value is optimized to 0.05. P i For the first i The load of each sample, This is the average load of the batch. This is the predicted mean for the batch sample.

[0094] When cov(,) is positive but small, the penalty is weak and does not affect the model's learning of normal positive correlation. If the value is large, the predicted value changes beyond the range allowed by the medium's saturation characteristics, and the penalty is strong, forcing the model to reduce the strength of the positive correlation and return to the physically reasonable range. When cov(,) is negative, no penalty is given, and this situation will be indirectly corrected by the prediction loss MSE.

[0095] S4. Input the historical error and physical information feature set into the learning model phy-XGB to train the ratio difference prediction model of the electronic voltage transformer.

[0096] S5. Calculate the average absolute SHAP value of the physical information features over the time period, and focus on the corresponding physical information features in descending order. The specific implementation includes the following steps:

[0097] S51. Conduct SHAP time-attribution analysis on the ratio difference prediction model of the electronic voltage transformer, and calculate the set of SHAP values ​​for each physical information feature on the predicted ratio difference of the electronic voltage transformer. As shown in the following formula:

[0098]

[0099] in, s For any physical information feature, f shap-phyXGB The SHAP analysis calculation formula for the learning model phy-XGB; any element Physical information characteristics j The i SHAP value at time;

[0100] S52. Conduct SHAP time-period attribution analysis on the ratio difference prediction model of electronic voltage transformers, and select the target time period. That is, the first m To the n For each time point, calculate the average absolute SHAP value for the target time period, as follows:

[0101]

[0102] in, Physical information characteristics for this time period jThe average absolute SHAP value;

[0103] S53. Arrange the average absolute SHAP values ​​of each physical information feature in descending order to determine the corresponding physical information features that require key attention. This provides practical interpretation and guidance for operation and maintenance, and offers more effective decision support for the refined scheduling, resource optimization, and safe and stable operation of the power grid, thus powerfully promoting the robust construction and high-quality development of the new power system.

[0104] In this embodiment, RMSE, MAE, and R are selected for performance evaluation of the model. 2 The evaluation index is as follows:

[0105]

[0106]

[0107]

[0108] in, N The total number of samples, y i This is the actual value. For predicted values, The mean is used. RMSE, as a comprehensive index of error analysis, reflects the accuracy of prediction, while MAE assesses the degree of fluctuation in model prediction error, reflecting the robustness and stability of the model. The better the model fits the regression curve, the closer it is to 1.

[0109] The method provided in this embodiment can predict the ratio difference of electronic voltage transformers in real-world scenarios. It predicts future ratio differences by using historical ratio difference data and physical information characteristics, fully utilizing collected shallow observable data, and forcing the model to adapt to the actual process through a physical loss function. This method helps to optimize the management of power grid equipment, improve the utilization efficiency of power equipment, reduce system operating costs, and support more precise operation and maintenance control, demonstrating significant practical value and economic benefits.

[0110] This embodiment also provides a simulation of the above method. Taking actual phase electronic voltage transformer data from a 110kV substation in a certain area in 2016 as an example, actual measurements were performed. The experiment was conducted according to the model proposed in this embodiment, and compared with current advanced models. The results are shown in Table 1 and... Figure 2 As shown.

[0111] Table 1 Comparison results with various advanced models

[0112]

[0113] In terms of prediction accuracy, the model proposed in this embodiment has the highest accuracy and shows a significant improvement over current advanced models, where R0 2 The performance is improved by at least 1.75%, RMSE is reduced by 91.9%, and MAE is reduced by 93.4%. The reasons why the model proposed in this embodiment performs much better than existing mature methods are as follows: First, other models deviate from the concept of fitting data to the changes in complex systems, resulting in overfitting; Second, other models lack physical rule constraints, and although the predicted sequences conform to the corresponding change trends, they have a low grasp of details; Third, deep learning may mislearn discontinuous features in the dataset, resulting in systematic bias, but in actual engineering, such low-quality datasets are often encountered; Fourth, the splitting strategy of other machine learning methods, represented by tree structures, is not suitable for early perception of rapidly changing differences. In addition, the enhancement of physical information features makes XGBoost, which is based on "classification-regression", have stronger generalization ability.

[0114] The target period for SHAP analysis is from 9:00 AM on December 6th to 9:00 PM on December 10th. (See Table 2 and...) Figure 3 As shown, the Environmental Comprehensive Index (ECI) is the core dominant characteristic of the ratio difference change, with the highest and positive average absolute SHAP value. Considering the physical meaning of ECI incorporating the synergistic effects of temperature, humidity, and magnetic field, this indicates that the overall environmental conditions suppressed ratio difference drift during this period, preventing risks such as equipment insulation aging, metal corrosion, or electromagnetic imbalance. The Load Temperature Interaction Index (LTQI) is a secondary influencing characteristic during this period, with an average absolute SHAP value at a moderate level, only slightly increasing the ratio difference; its effect is far weaker than the suppressive effect of ECI. The average absolute SHAP values ​​of RTH, LTSC, and RLF are all below 0.001, and their contribution to the ratio difference change during this period is negligible. Overall, the equipment ratio difference status was relatively stable during this period. During operation and maintenance, the ratio difference drift warning level can be appropriately reduced for this period, but continuous monitoring of the environmental conditions corresponding to ECI is necessary to avoid subsequent fluctuations in environmental factors causing a reversal in the direction of ECI's effect and increasing the risk of ratio difference drift.

[0115] Table 2 SHAP Analysis Results

[0116]

[0117] In summary, the method provided in this embodiment solves the problem of the lack of existing electronic voltage transformer ratio difference prediction technology. By integrating physical information characteristics and actual operation and maintenance experience, it realizes a practical and highly accurate electronic voltage transformer ratio difference prediction method. At the same time, the system integrates an improved SHAP analysis mechanism, which can realize time-period analysis of physical information characteristics, providing decision support for actual operation and maintenance.

[0118] Example 2

[0119] It should be further explained that, based on the same inventive concept, this embodiment provides an electronic voltage transformer error prediction system. When the system is running, it executes the method described in Embodiment 1, including the following modules:

[0120] The data acquisition module is used to collect operating data from the electronic voltage transformers that have been put into operation.

[0121] The physical feature construction module is used to construct a set of physical information features of electronic voltage transformers by combining the physical characteristics and operation and maintenance experience of electronic voltage transformers, including load temperature sensitivity coefficient LTSC, environmental comprehensive coefficient ECI, load temperature interaction coefficient LTQI, load relative fluctuation coefficient RLF, and ratio difference trend coefficient RHT.

[0122] The prediction model building module is used to combine physical constraints to construct a physical loss function, replace the original loss function of the XGBoost model, and obtain a new learning model phy-XGB.

[0123] The prediction model training module is used to input the historical error and physical information feature set into the learning model phy-XGB to train the ratio difference prediction model of the electronic voltage transformer.

[0124] The SHAP value calculation module is used to calculate the average absolute SHAP value of physical information features over a time period and sort them in descending order, identifying the physical information features with larger average absolute SHAP values ​​as key focus objects.

[0125] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention.

Claims

1. An electronic voltage transformer error prediction method, characterized by, Includes the following steps: S1. Collect operating data of the electronic voltage transformers that have been put into operation; S2. Based on the physical characteristics and operation and maintenance experience of electronic voltage transformers, construct a set of physical information features of electronic voltage transformers, including load temperature sensitivity coefficient LTSC, environmental comprehensive coefficient ECI, load temperature interaction coefficient LTQI, load relative fluctuation coefficient RLF, and ratio difference trend coefficient RHT. S3. Combine physical constraints to construct a physical loss function, replace the original loss function of the XGBoost model, and obtain a new learning model phy-XGB; The physical loss function is as follows: where, is the MSE, i.e., the original loss function of the XGBoost model; cov(, ) is the covariance, as follows: wherein, is the true value of the i th sample, is the final prediction value of the i th sample, n is the total number of samples; is the physical weight; P i is the load of the i th sample, is the batch load mean, is the batch sample prediction mean; S4. Input the historical error and physical information feature set into the learning model phy-XGB to train the ratio difference prediction model of the electronic voltage transformer. S5. Calculate the average absolute SHAP value of the physical information features over the time period, and focus on the corresponding physical information features in descending order.

2. The method for predicting errors in an electronic voltage transformer according to claim 1, characterized in that, The operational data mentioned in step 1 includes temperature, humidity, load, magnetic field strength, and historical ratio error value, as shown in the following formula: in, Represents temperature. Represents humidity. Represents load, Represents magnetic field strength. This represents the historical ratio error value; For the first n The temperature of the second sample, For the first n Humidity of the second sample For the first n The load of the next sampling For the first n The magnetic field strength of the next sample. For the first n The ratio error of the sampling This represents transposition.

3. The method for predicting errors in an electronic voltage transformer according to claim 2, characterized in that, The load temperature sensitivity coefficient LTSC mentioned in step S2 is as follows: Among them, LTSC i For the first i The load temperature sensitivity coefficient at any given time. It is a sliding window. and For the load and temperature sequence within the window, To find the minimum value, avoid a denominator of 0. The average load within the window is used for normalization to eliminate differences in the magnitude of load and temperature.

4. The method for predicting errors in an electronic voltage transformer according to claim 2, characterized in that, The Environmental Integration Index (ECI) mentioned in step S2 is as follows: Among them, ECI i For the first i The environmental comprehensive coefficient at any given time , , The first i Standardized values ​​of temperature, humidity, and magnetic field strength at any given time. , , These are their respective weights.

5. The method for predicting errors in an electronic voltage transformer according to claim 2, characterized in that, The load temperature interaction coefficient LTQI mentioned in step S2 is as follows: Among them, LTQI i For the first i The load-temperature interaction coefficient at any given time. , The respective i Standard values ​​for load and temperature at all times.

6. The method for predicting the error of an electronic voltage transformer according to claim 2, characterized in that, The load relative fluctuation coefficient RLF mentioned in step S2 is as follows: Among them, RLF i For the first i The relative load fluctuation coefficient at any given time. The load sequence within the window, To obtain the standard deviation of the sequence.

7. The method for predicting errors in an electronic voltage transformer according to claim 2, characterized in that, The ratio difference trend coefficient RHT mentioned in step S2 is as follows: Among them, RHT i For the first i The trend of the difference at different times, For the first i The difference in time, It is a sliding window.

8. The method for predicting the error of an electronic voltage transformer according to claim 1, characterized in that, Step S5 includes the following steps: S51. Conduct SHAP time-attribution analysis on the ratio difference prediction model of the electronic voltage transformer, and calculate the set of SHAP values ​​for each physical information feature on the predicted ratio difference of the electronic voltage transformer. As shown in the following formula: in, s For any physical information feature, f shap-phyXGB The SHAP analysis calculation formula for the learning model phy-XGB; any element Physical information characteristics j The i SHAP value at time; S52. Conduct SHAP time-period attribution analysis on the ratio difference prediction model of electronic voltage transformers, and select the target time period. That is, the first m To the n For each time point, calculate the average absolute SHAP value for the target time period, as follows: in, Physical information characteristics for this time period j The average absolute SHAP value; S53. Sort the average absolute SHAP values ​​of each physical information feature in descending order to determine the corresponding physical information features that need to be focused on.

9. An electronic voltage transformer error prediction system, characterized in that, Includes the following modules: The data acquisition module is used to collect operating data from the electronic voltage transformers that have been put into operation. The physical feature construction module is used to construct a set of physical information features of electronic voltage transformers by combining the physical characteristics and operation and maintenance experience of electronic voltage transformers, including load temperature sensitivity coefficient LTSC, environmental comprehensive coefficient ECI, load temperature interaction coefficient LTQI, load relative fluctuation coefficient RLF, and ratio difference trend coefficient RHT. The prediction model building module is used to combine physical constraints to construct a physical loss function, replace the original loss function of the XGBoost model, and obtain a new learning model phy-XGB. The physical loss function is as follows: in, MSE is the original loss function of the XGBoost model; cov(,) is the covariance, as shown in the following equation: in, For the first i The true value of each sample For the first i Final predicted value for each sample n The total number of samples; Physical weights; P i For the first i The load of each sample, This represents the average load of the batch. Predict the mean for the batch sample; The prediction model training module is used to input the historical error and physical information feature set into the learning model phy-XGB to train the ratio difference prediction model of the electronic voltage transformer. The SHAP value calculation module is used to calculate the average absolute SHAP value of physical information features over a time period and sort them in descending order, identifying the physical information features with larger average absolute SHAP values ​​as key focus objects.

Images