A method and system for operational state analysis of a transformer

Abstract

The present application relates to the technical field of electric data processing, and more particularly, to a method and system for analyzing the operating state of a transformer, comprising: obtaining operating parameters of the transformer; dividing the power consumption period of each day according to the size of the load rate of the transformer; for any power consumption period, calculating the health sample index of the data in the period, screening out healthy sample data, and constructing a healthy data set. The present application establishes a health assessment model for each period by Mahalanobis distance, and performs weighted processing in combination with an abnormality index, so that the model has self-adaptive adjustment capability at different health levels, can avoid false positives, and can improve the sensitivity to real abnormalities. Then, the chi-square distribution is used to determine the abnormality threshold, which improves the accuracy of transformer operating state monitoring and effectively reduces the false positive and false negative rates.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric data processing. More particularly, the present application relates to a method and system for analyzing the operating state of a transformer. BACKGROUND

[0002] As the core hub equipment of the power system, the safe and stable operation of the transformer is directly related to the reliability of the entire power grid. Therefore, by collecting multi-dimensional operating parameters such as current, voltage, and oil temperature of the transformer, and using intelligent algorithms to analyze and warn the operating state in real time and accurately, it has become a key technology to ensure the safety of the power grid.

[0003] At present, the transformer state monitoring method usually relies on arranging current, voltage, temperature and other sensors at the key parts of the transformer to collect its operating parameters, and determining whether the operating state is abnormal by setting threshold or constructing statistical model. For example, some methods identify abnormalities based on the fluctuation amplitude of a single parameter, and other methods use multi-dimensional data to construct Mahalanobis distance model or other statistical learning model to detect abnormalities of newly collected data. These methods have improved the ability to identify abnormal states of the transformer to some extent, but still have some limitations.

[0004] However, most state evaluation models need to be trained based on pure healthy data, but in actual scenarios, historical operating data is often unlabeled, mixed with normal, sub-healthy, and even early fault states, making it difficult to directly use to construct accurate evaluation benchmarks. Secondly, the operating conditions of the transformer are closely related to its load level, and there are significant differences in the normal parameter fluctuation range and mode during different power consumption periods such as peak, stable, and valley. It is difficult for a single, static evaluation model to adapt to such dynamic changes, and it is easy to produce high false positives or false negatives. SUMMARY

[0005] The present application provides a method and system for analyzing the operating state of a transformer, which aims to solve the problem that in related technologies, the operating conditions of the transformer are closely related to its load level, and there are significant differences in the normal parameter fluctuation range and mode during different power consumption periods such as peak, stable, and valley. A single, static evaluation model is difficult to adapt to such dynamic changes, and is prone to high false positives or false negatives.

[0006] In a first aspect, the present application provides a method for analyzing the operating state of a transformer, comprising: obtaining operating parameters of the transformer; dividing the power consumption period of each day according to the size of the load rate of the transformer, for any power consumption period, calculating the health sample index of the data in the power consumption period, screening out the health sample data, and constructing a health data set, wherein the health sample index reflects the volatility of the data in the power consumption period and the similarity of the power consumption period to its historical period; based on the health data set, constructing a Mahalanobis distance health assessment model corresponding to the power consumption period; obtaining new operating parameters, and using the Mahalanobis distance health assessment model corresponding to the power consumption period of the new operating parameters to calculate the Mahalanobis distance value of the new operating parameters, and according to the size of the Mahalanobis distance value, analyzing the state of the transformer. By dynamically dividing the operation time according to the load rate of the transformer, and constructing a dedicated evaluation model for each period, the problem of the single static model in the prior art being difficult to adapt to the large difference in normal fluctuation range of the transformer under different working conditions such as peak, stable, and low peak is solved, and the adaptability and accuracy of state analysis are significantly improved. In addition, according to the health sample index of the data volatility and the similarity of the historical period, the sample most representing the health state of the period can be screened out from the historical data mixed with various states to construct the evaluation model, so that the model benchmark is more reliable, the accuracy and automation level of the transformer operating state analysis are improved, and the false alarm and missed alarm rates are effectively reduced.

[0007] Further, the division of the power consumption period of each day comprises: calculating the average load rate of all time points in the window using a sliding window; if the average load rate is greater than a first threshold, the corresponding period is divided into a peak period; if the average load rate is greater than or equal to a second threshold and less than or equal to the first threshold, the corresponding period is divided into a stable period; if the average load rate is less than the second threshold, the corresponding period is divided into a low peak period.

[0008] Further, the calculation of the health sample index of the data in the power consumption period comprises: setting a window for the operating parameters of the power consumption period, and making the window slide on the operating parameters, and calculating the health sample index of the operating parameters in the window after each sliding. By setting a sliding window to continuously calculate the health sample index, the health characteristics of each small segment in the data stream can be captured more finely, and comprehensive and continuous evaluation of all data segments in the entire power consumption period is realized.

[0009] Further, the health sample data is screened, including: obtaining a health sample index of the running parameter in the window after each sliding, and selecting the running parameter in the window corresponding to the maximum health sample index as the health sample data. This method can exclude the interference of potential sub-health or minor abnormalities in the historical data to the greatest extent, ensure the highest purity of the constructed health evaluation model, and thus make the model have higher sensitivity for detecting deviation from the normal state.

[0010] Further, the transformer is analyzed according to the size of the Mahalanobis distance value, including: if the Mahalanobis distance value is greater than an abnormal threshold value, determining that the running state of the transformer is abnormal, and triggering a recording or alarm mechanism; and if the Mahalanobis distance value is less than or equal to the abnormal threshold value, determining that the running state of the transformer is normal.

[0011] Further, the Mahalanobis distance health evaluation model corresponding to the power consumption period is constructed, including: obtaining a health data set of the power consumption period, and calculating a mean vector of each dimension in the health data set and a covariance matrix of the health data set, and constructing the Mahalanobis distance health evaluation model based on the mean vector and the health data set.

[0012] Further, the abnormal index of each power consumption period is calculated, and the Mahalanobis distance value calculated by using the evaluation model corresponding to the power consumption period is weighted by using the abnormal index of the power consumption period. According to the overall health degree of a period, the strictness of judgment is dynamically adjusted, so that the model is more sensitive in a stable period and has higher tolerance in a fluctuating period, thereby realizing more intelligent and more scenario-aware abnormal judgment, and further reducing false positives.

[0013] Further, the abnormal index of each power consumption period is calculated, including: for any power consumption period, calculating the health sample index of the multi-dimensional data in the window after each sliding of the window in the power consumption period; and calculating the proportion of the number of health sample indexes greater than a division threshold value in all health sample indexes as the abnormal index of the power consumption period. This method avoids subjective setting of the abnormal index by experience, makes the calculation have a basis, guarantees the objectivity and reproducibility of the weighting correction process, and thus ensures the stability and reliability of the entire analysis model.

[0014] Further, the abnormal threshold value is determined by using a chi-square distribution.

[0015] The second aspect of the present application also provides a running state analysis system for a transformer, including a processor and a memory, the memory stores a computer program, and the processor executes the computer program to realize the running state analysis method for the transformer according to any one of the above.

[0016] Beneficial effects

[0017] (I) By introducing a dynamic power consumption period division mechanism based on load level, the transformer operation data can be modeled differently under different working conditions, avoiding the risk of failure of a single model in full-time monitoring. At the same time, by constructing a health sample index that integrates the instantaneous volatility and historical periodicity similarity, the health samples are accurately screened, thereby ensuring the stability of the benchmark data of the evaluation model.

[0018] (II) The Mahalanobis distance is used to establish a health evaluation model for each period, and the abnormal index is combined for weighted processing, so that the model has self-adaptive adjustment ability under different health levels, which can avoid false positives and improve the sensitivity to real abnormalities. Then, the chi-square distribution is used to determine the abnormal threshold, which improves the accuracy of transformer operation state monitoring and effectively reduces the false positive and false negative rates. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flowchart illustrating the operation state monitoring of a transformer according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0021] As shown in Figure 1 S101: Collecting the operation parameters of the transformer and dividing the power consumption period.

[0022] Specifically, by deploying various high-precision sensors at key positions of the transformer, the multi-dimensional operation parameters of the transformer are collected in real time. Specifically, the operation parameters can include but are not limited to current, voltage, oil temperature, ambient temperature, winding temperature, etc., and form a multi-dimensional time series data set.

[0023] Considering that the operation condition of the transformer is closely related to its load level, and the load level shows obvious periodicity within a day, this embodiment introduces dynamic division of power consumption period based on load rate. Specifically, for any natural day, the load rate data of the transformer in the natural day is collected in real time, and a sliding window is set for smoothing processing to remove noise effects. As a preferred scheme, the size of the sliding window can be set to 1 hour, and in other embodiments, the size of the sliding window can also be set to 45 minutes or 30 minutes, which can be adjusted according to the specific implementation scenario. Then, the sliding window is slid on the real-time load rate sequence with a step of 1, and the average value of all load rate samples in the window is calculated. Based on the average value, the 24 hours of a day are dynamically divided into three different power consumption periods, namely peak period, stable period and low peak period. The specific division process is as follows.

[0024] If the average load rate in the sliding window is greater than the first threshold, the time period is divided into a peak period, and in the embodiment, the empirical value of the first threshold is 70%, because this setting aims to cover the time period when industrial production or residential electricity consumption is the most concentrated, at which time the transformer is under the greatest pressure, and its normal range of operating parameters is significantly different from other time periods. If the average load rate in the sliding window is greater than or equal to the second threshold and less than or equal to the first threshold, it is divided into a stable period. This is the normal operating state of the transformer, and in the embodiment, the empirical value of the second threshold is 40%. If the average load rate in the sliding window is less than the second threshold, it is divided into an off-peak period, because the load is the lowest at this time, which usually corresponds to night or production intervals, so the average load rate of this period is low. It should be noted that the first threshold is greater than the second threshold.

[0025] Through the above steps, the original continuous data stream can be divided into discrete data segments with a clear working condition background, laying a foundation for subsequent construction of differentiated health assessment models for different working conditions.

[0026] S102: Screening health samples of each time period.

[0027] For any electricity consumption period, when constructing the corresponding assessment model, it is necessary to first screen a benchmark data set that can represent a good health state from the massive historical data. Therefore, in the embodiment, a health sample index construction method is proposed that fuses instant volatility and historical period similarity. Taking the off-peak period as an example, the construction process is as follows.

[0028] First, the collected operating parameters are standardized, and then in the embodiment, a health sample index is constructed to quantify whether the multi-dimensional data in a sliding window of size N in any electricity consumption period is a health sample, where N is 20, and in other embodiments, N can be 30 or 15, etc. The construction is based on the fact that a true health data segment not only has small internal fluctuations, but also has a similar fluctuation pattern to the historical same period.

[0029] As mentioned above, a health sample index calculation method is provided, which is: . In the formula, represents the health sample index of the multi-dimensional data in the window of the i-th off-peak period, represents the variance of the i-th dimension data in the i-th off-peak period, represents the average of the variances of the i-th dimension data in all off-peak periods, represents the variance of the i-th dimension data in the i-th off-peak period, represents the variance of the i-th dimension data in the i-th off-peak period, represents the average of the variances of the i-th dimension data in all off-peak periods, represents the variance of the i-th dimension data in the i-th off-peak period, Indicates the first of all off-peak periods The maximum value of the variance of the dimensional data. Indicates the number of data dimensions. Indicates the first Correction coefficients for multidimensional data within window N during off-peak periods. This represents an exponential function with the natural base e. It is used to measure whether the single-dimensional volatility during a certain off-peak period deviates from the normal volatility benchmark of the off-peak period, thus serving as a basic indicator for whether there is a healthy sample in that window. The larger the value, the smaller the health sample index. Divide by This is for normalization purposes.

[0030] The formula not only considers whether the real-time fluctuation range of the multidimensional data within the current window is within the normal range, but also introduces the similarity of the fluctuation pattern with historical data from the same period as a correction factor, thereby achieving a dual-constraint screening of healthy samples. This approach effectively avoids misjudgments caused by relying solely on a single amplitude indicator, making the identification of healthy samples more accurate.

[0031] It should be noted that transformer operating conditions often exhibit patterns such as diurnal cycles and weekday / restday cycles. For example, off-peak periods generally correspond to reduced load at night, and operating data at the same time on different dates will show relatively stable fluctuation patterns. If only the instantaneous fluctuation amplitude is considered, while ignoring this cyclical pattern, some short-term abnormal fluctuations may be misclassified as normal healthy samples. Therefore, considering cyclical similarity is to utilize the historical regularity of transformer operation, avoid interference from occasional noise in the selection of healthy samples, make the health data more representative and stable, and thus improve the reliability and accuracy of subsequent evaluation models.

[0032] For the The correction coefficients for multidimensional data within window N during off-peak periods are obtained as follows: The calculation formula is: In the formula, This represents the correction coefficient for the multidimensional data within window N during the i-th low-peak period. The correction coefficient reflects the degree of consistency between the data pattern of this window and the overall historical data of the same period. This represents the window in the o-th dimension of the corresponding time period on day m. The similarity between the corresponding sequence and the sequence corresponding to window N in the o-th dimension of the i-th low-peak period (the similarity is the reciprocal of the sum of the DTW distance and the preset hyperparameters. The preset hyperparameters avoid a denominator of 0, for example, the preset hyperparameters are 1, and can map the range of similarity values ​​to 0-1). Indicates the number of consecutive days. represents the number of data dimensions. In this embodiment, the number of adjacent days is 10 days.

[0033] In the formula, if the window N is highly similar to the fluctuation pattern of the historical adjacent days, then is larger, and after squaring, the contribution is further amplified, so that is also increased. If the similarity is generally low, then the value of is small. The introduction of the correction coefficient makes the identification of healthy samples not only depend on whether the fluctuation within the window is normal, but also further combines the similarity of the historical periodicity, thereby improving the judgment ability of the healthy samples.

[0034] In summary, according to the above calculation method, all low-peak periods are traversed to find the window corresponding to the maximum, and the multi-dimensional data in the window is taken as the healthy data set. Since the larger the value of is, the more stable the fluctuation amplitude of the data in the window is, and the more consistent the fluctuation pattern is with the historical period, so the data corresponding to the window can most accurately represent the best healthy state, thereby providing a reliable benchmark sample for the subsequent evaluation model.

[0035] S103: Based on the health samples of each power consumption period, an evaluation model corresponding to each power consumption period is constructed.

[0036] Specifically, taking the low-peak period as an example, the health data set corresponding to the low-peak period is obtained according to the above method, and the health data set contains M sample points, and each sample point is a feature vector composed of d-dimensional data. Then the mean vector of each dimension and the covariance matrix of the health data set are calculated, and finally a Mahalanobis distance health evaluation model is constructed based on the mean vector and the health data set. For any new sample, input it into the Mahalanobis distance health evaluation model to obtain the distance value of the new sample in the evaluation model, which quantifies the distance of the new sample point from the center of the health data distribution after considering the correlation between dimensions.

[0037] S104: Real-time monitoring of the running state of the transformer based on the evaluation model.

[0038] After the model is built, it can be used for real-time condition monitoring of the transformer. Specifically, when a new data sample is collected, first determine the power consumption period (peak period, stable period or low peak period) to which the new data sample belongs, and call the Mahalanobis distance evaluation model of the corresponding period to calculate the Mahalanobis distance value. In the actual monitoring process, when the new data sample is input into the Mahalanobis distance health evaluation model, the Mahalanobis distance value is calculated and compared with the preset abnormal threshold value. If the Mahalanobis distance value is less than or equal to the threshold value, it means that the difference between the new data sample and the health data set is within the normal fluctuation range, and the transformer operating state is normal; otherwise, if the Mahalanobis distance value is greater than the threshold value, it means that the data sample deviates significantly from the healthy state, and should be judged as abnormal, and the recording or alarm mechanism is triggered.

[0039] In another embodiment, the abnormality index of each power consumption period can also be calculated, and the Mahalanobis distance value calculated by using the evaluation model corresponding to the power consumption period is weighted by using the abnormality index of the power consumption period. Specifically, the abnormality index of any power consumption period is obtained, taking the low peak period as an example, including: determining the division threshold, the experience value of the division threshold is 0.8, then calculating the health sample index of the multi-dimensional data in the window after the window is slid in the low peak period each time, the proportion of the number of health sample indexes greater than the division threshold in all health sample indexes is calculated, the greater the proportion, the more stable and healthy the mode of the data segments in the period, that is, the healthier the period as a whole, then the weighted Mahalanobis distance value is smaller after weighting the calculated Mahalanobis distance value, otherwise, the smaller the proportion, the poorer the health of the period as a whole, and the more fluctuations or abnormalities, once the new sample deviates greatly from the health data, it is more likely to be a real anomaly, then the weighted Mahalanobis distance value is larger after weighting the calculated Mahalanobis distance value, and the model is more strict in judgment.

[0040] For the determination of the abnormal threshold value, the difference between the new data sample and the health data set in the Mahalanobis distance health evaluation model can be measured by the Mahalanobis distance. Since the health data set obeys the multivariate normal distribution, the square value of the Mahalanobis distance obeys the chi-square distribution with the degree of freedom equal to the number of data dimensions in the statistical theory. Based on this property, the chi-square distribution can be used to determine the threshold value between the health state and the abnormal state. Specifically, first, determine the degree of freedom of the chi-square distribution according to the characteristic dimension number of the health data set; then set a significance level, for example, 0.05, 0.01 or 0.001, which reflects the maximum false positive probability allowed by the system; then find the critical value corresponding to the degree of freedom and the confidence level in the chi-square distribution table, which is the abnormal threshold value of the Mahalanobis distance.

[0041] The present application also provides a system for analyzing the operating state of a transformer. The system comprises a processor and a memory storing computer program instructions which, when executed by the processor, implement a method for analyzing the operating state of a transformer according to the first aspect of the present application.

[0042] The system also comprises a communication bus and a communication interface and other components well known to those skilled in the art, the settings and functions of which are known in the art and thus will not be described here.

[0043] In the present application, the aforementioned memory can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. For example, the computer-readable storage medium can be any appropriate magnetic storage medium or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random-Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store the desired information and that can be accessed by an application, a module, or both. Any such computer storage media can be part of the device or accessible or connectable thereto. Any application or module described in the present application can be implemented using computer-readable / executable instructions stored or otherwise held by such computer-readable media.

[0044] The above-described embodiments only express several implementation manners of the present application, which are described in a more specific and detailed manner, but should not be understood as limiting the scope of the application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application.

Claims

1. A method for operational state analysis of a transformer, characterized in that, The method comprises the following steps: obtaining the operating parameters of the transformer; dividing the power consumption period of each day according to the size of the load rate of the transformer, comprising: using a sliding window to calculate the average load rate of all time points in the window; if the average load rate is greater than a first threshold, the corresponding period is divided into a peak period; if the average load rate is greater than or equal to a second threshold and less than or equal to the first threshold, the corresponding period is divided into a stable period; if the average load rate is less than the second threshold, the corresponding period is divided into a low-peak period, wherein the first threshold is greater than the second threshold; for any power consumption period, calculate the health sample index of the data in the power consumption period, comprising: setting a window for the operating parameters of the power consumption period, and sliding the window, after each sliding, considering the instantaneous fluctuation amplitude of the multi-dimensional data in the window and its periodic similarity with the historical data of the same period, calculating the health sample index of the operating parameters in the window, the calculation formula is: In the formula, Indicates the first During each electricity consumption period, the window Health sample index of internal multidimensional data, Indicates the first During the electricity consumption period, the first Variance of dimensional data Indicates the first of all electricity consumption periods The mean of the variance of the dimensional data. Indicates the first of all electricity consumption periods The maximum value of the variance of the dimensional data. Indicates the number of data dimensions. Indicates the first Correction coefficients for multidimensional data within window N of a single electricity consumption period. This represents an exponential function with base e. It is used to measure whether the single-dimensional volatility of a certain electricity consumption period deviates from the normal volatility benchmark of the electricity consumption period; represents the correction coefficient of the multi-dimensional data in the window N in the i-th electricity consumption period, reflecting the consistency degree of the data pattern of the window with the overall historical same period, and the calculation formula is: ; represents the similarity between the sequence corresponding to the window of the m-th day corresponding period and the o-th dimension and the sequence corresponding to the window N of the i-th electricity consumption period corresponding period and the o-th dimension, represents the number of adjacent days, represents the number of data dimensions; screening out the health sample data and constructing a health data set; based on the health data set, constructing the Mahalanobis distance health evaluation model corresponding to the power consumption period; obtaining new operating parameters, and using the Mahalanobis distance health evaluation model corresponding to the power consumption period to which the new operating parameters belong, calculating the Mahalanobis distance value of the new operating parameters, and according to the size of the Mahalanobis distance value, analyzing the state of the transformer.

2. The operating state analysis method for a transformer according to Claim 1, characterized by, Screening out health sample data, comprising: obtaining the health sample index of the operating parameters in the window after each sliding, and selecting the operating parameters in the window corresponding to the maximum health sample index as the health sample data.

3. The operating state analysis method for a transformer according to Claim 1, characterized by, According to the size of the Mahalanobis distance value, analyzing the state of the transformer, comprising: if the Mahalanobis distance value is greater than an abnormal threshold, it is determined that the operating state of the transformer is abnormal, and a recording or alarm mechanism is triggered; if the Mahalanobis distance value is less than or equal to the abnormal threshold, it is determined that the operating state of the transformer is normal.

4. The operating state analysis method for a transformer according to Claim 1, characterized by, Constructing the Mahalanobis distance health evaluation model corresponding to the power consumption period, comprising: obtaining the health data set of the power consumption period, and calculating the mean vector of each dimension in the health data set and the covariance matrix of the health data set, and constructing the Mahalanobis distance health evaluation model based on the mean vector and the health data set.

5. The operating state analysis method for a transformer according to Claim 1, characterized by, Further comprising: calculating the abnormal index of each power consumption period, and using the abnormal index of the power consumption period to weight the Mahalanobis distance value calculated by using the evaluation model corresponding to the power consumption period.

6. The operating state analysis method for a transformer according to Claim 5, characterized by, Calculating the abnormal index of each power consumption period, comprising: for any power consumption period, calculating the health sample index of the multi-dimensional data in the window after each sliding in the window; statistically calculating the proportion of the number of health sample indexes greater than the division threshold in all health sample indexes in all windows as the abnormal index of the power consumption period.

7. The operating state analysis method for a transformer according to Claim 3, characterized by, The method for determining the abnormal threshold, comprising: determining the abnormal threshold based on the chi-square distribution.

8. An operating condition analysis system for a transformer, comprising a processor and a memory, characterized in that, The memory stores a computer program, and the processor executes the computer program to realize the method for analyzing the operating state of the transformer according to any one of claims 1-7.

Images