An on-line calibration method and system for power transformer error

Abstract

The present application relates to the technical field of error calibration, in particular to an online calibration method and system for power transformer error, comprising the following steps: triggering data acquisition by setting voltage and current change rate threshold, synchronously acquiring high-frequency waveform data of primary side and secondary side of the transformer, and taking the trigger point as the center to intercept the waveform segment of transient period, the present application can accurately lock the abnormal response moment of current transformer in operation by setting threshold condition of voltage and current change rate to trigger data acquisition process, avoid the acquisition of invalid or interference data, improve the effectiveness of data, and perform multi-level wavelet packet decomposition on transient waveform, so that the energy ratio and phase difference between primary side and secondary side signals are clearly described in each frequency band, and by constructing frequency band response difference sequence and reconstructing into dynamic error feature vector, high-resolution identification of local response anomaly is effectively realized.

Description

Technical Field

[0001] This invention relates to the field of error calibration technology, and in particular to an online calibration method and system for the error of power transformers. Background Technology

[0002] With the deepening of power system reform, the separation of power plants and grids, as a key aspect of the reform, can effectively promote the diversified development of the power market. The accuracy, reliability, fairness, and impartiality of electricity metering devices involve economic interests and social benefits, becoming a common concern for power plants, grids, and other stakeholders. Current transformers are devices used to assist in measuring current data in power systems. They are connected to measuring instruments, relay protection devices, or automatic devices as needed to ensure the safe collection of data. As an important auxiliary current measuring device in substations, the error of current transformers has a significant impact on the metering of electricity.

[0003] Current technologies for error control in current transformers primarily rely on periodic manual verification or static comparison analysis, which struggles to capture dynamic error changes under complex operating conditions. Furthermore, they fail to separately model transient disturbances and steady-state processes, leading to the mixing of abrupt changes and steady-state responses, thus obscuring the dimensions of error assessment. In frequency domain analysis, they neglect the ability of cyclic frequencies to represent periodic disturbances and the nonlinear characteristics of the core, making it impossible for error identification methods to effectively distinguish between composite distortions caused by both inherent equipment characteristics and external interference. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and to propose an online calibration method and system for power transformer errors.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an online calibration method for power transformer errors, comprising the following steps:

[0006] By setting thresholds for voltage and current change rates to trigger data acquisition, high-frequency waveform data from the primary and secondary sides of the transformer are acquired synchronously. Waveform segments during transient periods are extracted with the trigger point as the center to form a transient event waveform set. Then, continuous steady-state waveform segments are selected from the region outside the transient events to establish multi-dimensional feature source data.

[0007] The transient waveform segments in the multidimensional feature source data are called, and multi-level wavelet packet decomposition is performed on the primary and secondary waveforms. The energy ratio and phase difference of the primary and secondary coefficients in each frequency band are calculated to form a frequency band response difference sequence. The frequency band response difference sequences of all frequency bands are reconstructed and arranged to generate a dynamic error feature vector.

[0008] Call the steady-state waveform segment in the multi-dimensional feature source data, extract the energy distribution information on the frequency axis and the cyclic frequency axis, construct the cyclic spectrum energy distribution map, and separate the main cyclic frequency spectrum energy and the non-main cyclic frequency spectrum energy introduced by the core saturation according to the cyclic spectrum energy distribution map, and calculate the cyclic distortion index;

[0009] By fusing the dynamic error feature vector and the cyclic distortion index, and according to the preset error mapping rules, the fusion result is converted into amplitude and phase compensation amounts to establish a comprehensive error correction matrix. Correction parameters are generated based on the comprehensive error correction matrix and written into the metering chip register of the energy meter to obtain the online calibration coefficient set.

[0010] Preferably, the steps for obtaining the multidimensional feature source data are as follows:

[0011] Based on the voltage change rate threshold and the current change rate threshold, the primary side voltage change rate, primary side current change rate, secondary side voltage change rate, secondary side current change rate and the threshold are compared, and the moment when the threshold is met is determined as the trigger point. The primary side high-frequency waveform segment and the secondary side high-frequency waveform segment are extracted with the trigger point as the center to form a transient event waveform set.

[0012] Based on the transient event waveform set, the boundary of the trigger point time range is determined, all positions within the time range are excluded, and in the remaining time range, intervals are selected continuously. High-frequency waveform segments on the primary side and high-frequency waveform segments on the secondary side are extracted from the same position to form steady-state waveform segments.

[0013] Based on the steady-state waveform segment and the transient event waveform set, align the timestamps of the primary side high-frequency waveform segment and the secondary side high-frequency waveform segment, splice the primary side high-frequency waveform segment, the secondary side high-frequency waveform segment, and the steady-state waveform segment, add trigger point identifiers and interval labels, unify the structure and field naming, and generate multi-dimensional feature source data.

[0014] Preferably, the step of obtaining the frequency band response difference sequence is as follows:

[0015] The transient waveform segments in the multidimensional feature source data are called, and the multi-level frequency band index is traversed level by level. Bandpass filtering and downsampling are performed on the primary and secondary waveforms at each level. The time axes of the primary and secondary waveforms are kept aligned and the coefficient value sequence of each frequency band is recorded to form a wavelet packet coefficient sequence.

[0016] Based on the wavelet packet coefficient sequence, the energy measure and phase measure of the primary and secondary wavelet packet coefficient sequences are statistically analyzed for each frequency band. The ratio of the energy measure and the difference of the phase measure are calculated within the same frequency band. At the same time, the corresponding time stamp and frequency band index are added to form a frequency band response difference sequence.

[0017] Preferably, the step of obtaining the dynamic error feature vector is as follows:

[0018] Based on the frequency band response difference sequence, the frequency bands are rearranged from low frequency to high frequency according to the multi-level frequency band index while maintaining the consistent order of the time series. The energy ratios and phase differences of all frequency bands are concatenated in a fixed splicing order, and the placeholders of the missing frequency band positions are filled in to generate a dynamic error feature vector.

[0019] Preferably, the step of obtaining the cyclic spectrum energy distribution map is as follows:

[0020] Call the steady-state waveform segment in the multidimensional feature source data, extract the secondary side signal time series and remove DC bias, segment it according to fixed length and fixed overlap rate and record the timestamp, calculate the cyclic autocorrelation sequence segment by segment and accumulate along the time delay axis, perform complex exponential kernel integration along the time delay axis to obtain the complex value distribution of the frequency axis and establish a joint index of the frequency axis and the cyclic frequency axis to obtain the cyclic spectral density function;

[0021] Based on the cyclic spectral density function, the maximum peak position of the energy distribution is searched on the cyclic frequency axis. The index corresponding to the maximum peak position is recorded as the main cyclic frequency index, and the bandwidth of the main peak is calculated. The maximum peak position of the energy distribution is searched on the frequency axis, and the maximum peak position is recorded as the main frequency index. The half-width of the main peak is calculated. Then, the energy set of the main cyclic frequency spectral line is labeled according to the condition that the cyclic frequency index is an integer multiple of the main cyclic frequency and the frequency index is located in the neighborhood of the main frequency's half-width. At the same time, regions where the cyclic frequency index is not an integer multiple of the main cyclic frequency and the frequency index is far from the neighborhood of the main frequency's half-width are selected. The upper quantile of the energy distribution in the region is calculated as the noise power threshold, and the region is labeled as the energy set of non-main cyclic frequency spectral lines, forming a cyclic spectral energy distribution map.

[0022] Preferably, the step of obtaining the cyclic distortion index is as follows: calculating the cyclic distortion index based on the cyclic spectrum energy distribution map.

[0023] Preferably, the steps for obtaining the comprehensive error correction matrix are as follows:

[0024] The dynamic error feature vector and the cyclic distortion index are called, the unit range is unified, and a one-to-one correspondence is established with timestamp and frequency band index. The numerical bits are concatenated in a fixed field order, and abnormal jumps are thresholded and missing positions are linearly interpolated to obtain the error fusion characterization sequence.

[0025] Based on the error fusion characterization sequence, the condition thresholds and transformation entries of the error mapping rules are read. The mapping entries are matched one by one according to the frequency band index and distortion level, and the numerical bits are split into amplitude channels and phase channels. Transformation and saturation clipping are performed respectively to generate channel identifiers and time identifiers. The compensation amounts of amplitude and phase are obtained and summarized into a comprehensive error correction matrix.

[0026] Preferably, the step of obtaining the online calibration coefficient set is as follows:

[0027] Based on the magnitude and phase compensation amounts in the comprehensive error correction matrix, the comprehensive error correction matrix is ​​filled in the order of primary side, secondary side, and channel. It is then expanded into correction parameters according to the register address mapping table, written item by item into the metering chip register, and the checksum and status bit are read back and compared. The version number and timestamp are recorded to obtain the online calibration coefficient set.

[0028] This invention also provides an online calibration system, comprising:

[0029] The data acquisition module triggers data acquisition by setting voltage and current change rate thresholds, synchronously acquires high-frequency waveform data from the primary and secondary sides of the transformer, and extracts waveform segments during transient periods centered on the trigger point to form a transient event waveform set. Then, it selects continuous steady-state waveform segments from areas outside the transient events to establish multi-dimensional feature source data.

[0030] The dynamic feature extraction module is used to call the transient waveform segments in the multidimensional feature source data, perform multi-level wavelet packet decomposition on the primary and secondary waveforms, calculate the energy ratio and phase difference of the primary and secondary coefficients in each frequency band, form a frequency band response difference sequence, reconstruct and arrange the frequency band response difference sequences of all frequency bands, and generate a dynamic error feature vector.

[0031] The steady-state feature extraction module is used to call the steady-state waveform segment in the multi-dimensional feature source data, extract the energy distribution information on the frequency axis and the cyclic frequency axis, construct the cyclic spectrum energy distribution map, and separate the main cyclic frequency spectrum energy and the non-main cyclic frequency spectrum energy introduced by the core saturation based on the cyclic spectrum energy distribution map, and calculate the cyclic distortion index.

[0032] The error correction module is used to fuse the dynamic error feature vector and the cyclic distortion index, convert the fusion result into amplitude and phase compensation amounts according to the preset error mapping rules, establish a comprehensive error correction matrix, generate correction parameters based on the comprehensive error correction matrix, and write the correction parameters into the metering chip register of the energy meter to obtain the online calibration coefficient set.

[0033] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0034] This invention triggers the data acquisition process by setting threshold conditions for voltage and current change rates, accurately pinpointing abnormal response moments in current transformer operation, avoiding the acquisition of invalid or interfering data, and improving data effectiveness. Multi-level wavelet packet decomposition of transient waveforms clearly characterizes the energy ratio and phase difference between primary and secondary signals across various frequency bands. By constructing a frequency band response difference sequence and reconstructing it into a dynamic error feature vector, high-resolution identification of local response anomalies is effectively achieved. Energy distribution is extracted through joint analysis of the frequency axis and cyclic frequency axis in the steady-state waveform, distinguishing between main cyclic frequency energy and non-main cyclic frequency energy, enhancing the accuracy of characterizing the nonlinear distortion state of the transformer. The dynamic error feature vector is fused with the cyclic distortion index, combined with error mapping rules, to transform it into targeted amplitude and phase compensation quantities. Furthermore, a comprehensive error correction matrix is ​​constructed to achieve refined dynamic correction of transformer errors. By collaboratively modeling signal characteristics under transient and steady-state conditions and extracting nonlinear errors using dual-dimensional indicators of frequency and cyclic frequency, the error correction chain is closed through the construction and writing of structured compensation quantities, enhancing the targeting of error identification. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0037] Please see Figure 1 This invention provides a technical solution: an online calibration method for the error of a power transformer, comprising the following steps:

[0038] By setting thresholds for voltage and current change rates to trigger data acquisition, high-frequency waveform data from the primary and secondary sides of the transformer are acquired synchronously. Waveform segments during transient periods are extracted with the trigger point as the center to form a transient event waveform set. Then, continuous steady-state waveform segments are selected from the region outside the transient events to establish multi-dimensional feature source data.

[0039] The transient waveform segments in the multidimensional feature source data are called, and multi-level wavelet packet decomposition is performed on the primary and secondary waveforms. The energy ratio and phase difference of the primary and secondary coefficients in each frequency band are calculated to form a frequency band response difference sequence. The frequency band response difference sequences of all frequency bands are reconstructed and arranged to generate a dynamic error feature vector.

[0040] Call the steady-state waveform segment in the multi-dimensional feature source data, extract the energy distribution information on the frequency axis and the cyclic frequency axis, construct the cyclic spectrum energy distribution map, and separate the energy of the main cyclic frequency spectrum line and the energy of the non-main cyclic frequency spectrum line introduced by the core saturation based on the cyclic spectrum energy distribution map, and calculate the cyclic distortion index;

[0041] By fusing the dynamic error feature vector and the cyclic distortion index, and based on the preset error mapping rules, the fusion result is converted into amplitude and phase compensation quantities to establish a comprehensive error correction matrix. Correction parameters are generated based on the comprehensive error correction matrix and written into the metering chip register of the energy meter to obtain the online calibration coefficient set.

[0042] The steps for obtaining multidimensional feature source data are as follows:

[0043] Based on the voltage change rate threshold and the current change rate threshold, the primary side voltage change rate, primary side current change rate, secondary side voltage change rate, secondary side current change rate and the threshold are compared, and the moment when the threshold is met is determined as the trigger point. The primary side high-frequency waveform segment and the secondary side high-frequency waveform segment are extracted with the trigger point as the center to form a transient event waveform set.

[0044] Based on the transient event waveform set, determine the boundary of the trigger point time range, exclude all positions within the time range, select intervals in the remaining time range according to continuity, and extract high-frequency waveform segments from the same position on the primary side and the secondary side to form steady-state waveform segments.

[0045] Based on the steady-state waveform segment and the transient event waveform set, the timestamps of the primary side high-frequency waveform segment and the secondary side high-frequency waveform segment are aligned, the primary side high-frequency waveform segment, the secondary side high-frequency waveform segment and the steady-state waveform segment are spliced ​​together, trigger point identifiers and interval labels are added, the structure and field naming are unified, and multi-dimensional feature source data is generated.

[0046] Specifically, based on preset voltage and current rate of change thresholds, the real-time rate of change is first calculated for the continuously acquired digitized high-frequency sampling sequences of primary voltage, primary current, secondary voltage, and secondary current. This calculation is achieved by subtracting the values ​​of two adjacent sampling points and dividing by the sampling time interval. For example, under a sampling frequency of 1MHz, the primary voltage rate of change is calculated by subtracting the voltage value of the previous sampling point from the current sampling point's voltage value and then dividing by a 1-microsecond time interval. The same calculation is performed on all four signal streams to obtain four synchronized rate of change time sequences. Next, the settings are... The trigger threshold is set based on statistical analysis of data collected during a reference period (e.g., 10 minutes) under normal grid operation. Specifically, the average and standard deviation of the absolute values ​​of four rate-of-change sequences within the reference period are calculated. The voltage rate-of-change threshold is set as the sum of 1.8 times the average value and 2.5 times the standard deviation of the corresponding voltage rate-of-change sequence, and the current rate-of-change threshold is set as the sum of 2.0 times the average value and 3.0 times the standard deviation of the corresponding current rate-of-change sequence. Taking the primary voltage as an example, if its average rate of change is 200V / s and its standard deviation is 50V / s, then its trigger threshold is set to 1.8. * 200 + 2.5 * 50 = 360 + 125 = 485V / s. Then, the absolute values ​​of the four rate-of-change sequences calculated in real time are continuously compared with their respective thresholds. When the value of any rate of change first exceeds its set threshold, that moment is determined as the trigger point. Once the trigger point is determined, data is immediately truncated 100 milliseconds forward and 300 milliseconds backward, centered on the timestamp of the trigger point, forming a waveform segment with a total duration of 400 milliseconds. This operation is simultaneously applied to the waveform data of the primary side voltage, primary side current, secondary side voltage, and secondary side current, thereby obtaining four high-frequency waveform segments that are perfectly aligned in time. These four segments together constitute a complete record of a transient event. The set of all transient event records generated by different trigger points ultimately forms a transient event waveform set.

[0047] Based on the transient event waveform set obtained from the preceding steps, firstly, each transient event record in the set is traversed, and the start and end timestamps of each record are extracted. Since each transient event captures a 400-millisecond waveform centered on the trigger point, its time range is a clearly defined closed interval. For example, if a trigger point occurs at time T, its corresponding time range is [T - 100ms, T + 300ms]. The time ranges of all transient events are merged. If two or more time ranges overlap, they are merged into a larger continuous time range, ultimately resulting in one or more non-overlapping time periods marked as "transient". Then, within the total time span of the entire data acquisition (e.g., 24 hours), all time periods marked as "transient" are excluded, and the remaining time area is the steady-state candidate region. Next, within these steady-state candidate regions, sufficiently long continuous intervals are selected for capturing steady-state waveforms. A minimum continuous duration standard is set, such as 800 milliseconds, to ensure that the selected segment is indeed within the steady-state range. In a stable operating state, all steady-state candidate regions are scanned to identify all continuous sub-intervals with a length greater than or equal to 800 milliseconds. From these sub-intervals that meet the conditions, a fixed number of steady-state samples are selected using an equally spaced sampling method, for example, 50 samples. The length of each sample is fixed at 500 milliseconds. After selecting a 500-millisecond steady-state interval, high-frequency waveform data of primary side voltage, primary side current, secondary side voltage, and secondary side current are synchronously extracted from the same time position corresponding to the interval to ensure that these four waveform segments are precisely aligned in time. After this operation, all selected steady-state samples are collected to form a steady-state waveform segment.

[0048] Based on the acquired steady-state waveform segments and transient event waveform sets, a timestamp alignment check is first performed. For each waveform segment in the set (whether transient or steady-state), it is checked whether the start timestamps of the four waveform data sequences on the primary and secondary sides are completely consistent and whether the number of data points is equal, ensuring that there is no misalignment caused by sampling asynchrony or data loss. After confirming that the alignment is correct, all waveform segments from the transient event waveform set and all waveform segments from the steady-state waveform segment are sorted according to their respective start timestamps and concatenated into a single, chronologically ordered complete data stream. During the concatenation process, a new descriptive field is added to each data point. Specifically, a field named "Interval Label" is created. For all data points originating from the transient event waveform set, the value of this field is set to 1; for all data points originating from the steady-state waveform segment, the value of this field is set to 1. The value of the field is set to 0. At the same time, another field named "Trigger Point Identifier" is created. For data in transient events, this field records the precise timestamp of the trigger point of the corresponding event. For data in steady-state segments, the value of this field is empty or set to a predefined invalid value, such as 0. Subsequently, the entire spliced ​​data stream is subjected to structural unification processing, and the naming of all data fields is standardized. For example, they are uniformly named "Timestamp", "Primary Voltage", "Primary Current", "Secondary Voltage", "Secondary Current", "Interval Label" and "Trigger Point Identifier". It is also ensured that the data type and unit of all numerical fields are consistent. For example, the voltage unit is volt (V), the current unit is ampere (A), and the timestamp unit is microsecond (μs). After the above alignment, splicing, labeling and unification operations, a structured and information-rich multidimensional feature source data is finally generated.

[0049] The steps for obtaining the frequency band response difference sequence are as follows:

[0050] The transient waveform segments in the multidimensional feature source data are called, and the multi-level frequency band index is traversed level by level. Bandpass filtering and downsampling are performed on the primary and secondary waveforms at each level. The time axes of the primary and secondary waveforms are kept aligned and the coefficient value sequence of each frequency band is recorded to form the wavelet packet coefficient sequence.

[0051] Based on the wavelet packet coefficient sequence, the energy measure and phase measure of the primary and secondary wavelet packet coefficient sequences are statistically analyzed for each frequency band. The ratio of the energy measure and the difference of the phase measure are calculated within the same frequency band. At the same time, the corresponding time stamp and frequency band index are added to form a frequency band response difference sequence.

[0052] Specifically, waveform segments marked as transient events in the multidimensional feature source data are called. For each transient event, the time series data of its corresponding primary and secondary waveforms are extracted simultaneously. The "db4" (Daubechies 4) wavelet is selected as the basis function, and the decomposition level is set to 5 levels. This choice is because the "db4" wavelet achieves a good balance between time-frequency localization characteristics and computational complexity. A 5-level decomposition can divide the signal frequency into 32 uniform sub-bands, sufficient to cover various high-frequency transient components that may occur in the transformer. The decomposition process adopts a step-by-step traversal approach, starting from level 0 (the original signal). In the first level, the primary and secondary waveforms are convolved with low-pass and high-pass filters corresponding to the "db4" wavelet, respectively, to obtain approximate and detail components. Then, these two components are downsampled by a factor of 2, i.e., every other point is sampled to generate wavelet packet coefficients for the two frequency bands of the first level. In the second level, the above filtering and downsampling operations are repeated on the coefficient sequences of the two frequency bands generated in the first level. Each frequency band is further divided into two sub-bands, generating coefficients for a total of four frequency bands. This process is iterated until the fifth layer, ultimately generating wavelet packet coefficient sequences for 32 frequency bands for both the primary and secondary waveforms. Throughout the decomposition process, the same decomposition tree structure and filter are applied to the primary and secondary waveforms, and the starting point of downsampling remains consistent, thus ensuring that the wavelet packet coefficient sequences for the primary and secondary sides are strictly aligned in time within each frequency band. Finally, the coefficients of all frequency bands obtained after decomposing each transient event are recorded in a structured manner. The record for each frequency band includes a frequency band index (composed of the layer number and the position index of that layer, for example (5, k), where k ranges from 0 to 31), as well as the corresponding coefficient value sequences for the primary and secondary sides. After summarizing the records for all frequency bands, a wavelet packet coefficient sequence is formed.

[0053] Based on the wavelet packet coefficient sequence generated in the previous step, each frequency band of each transient event is analyzed independently. First, the energy metric is calculated. For the frequency band with frequency band index k, its primary wavelet packet coefficient sequence is denoted as... The second-order wavelet packet coefficient sequence is denoted as Where n is the coefficient point index, then the primary side energy metric With secondary side energy measurement The calculation method is the sum of squares of their respective coefficient sequences, that is... and Next, the phase metric is extracted, and the time delay is determined by calculating the normalized cross-correlation function between the primary and secondary coefficient sequences. The formula for calculating the cross-correlation function is as follows: ,in For the time delay points, find the one that makes Delay point for obtaining the maximum value This delay point represents the time difference between two signals within this frequency band. This time delay is then converted into a phase difference, calculated using the following formula: ,in It is the center frequency of the k-th frequency band. This is the time sampling period corresponding to the wavelet packet coefficients of this layer. Its value is the original signal sampling period multiplied by 2 raised to the power of the decomposition layer number. After completing the calculation of energy and phase metrics, the ratio of energy metrics is calculated within the same frequency band k. And directly use the calculated phase difference value Finally, the calculation results for each frequency band, including the energy ratio and phase difference, are associated with the start time identifier of the transient event and the index of the frequency band (e.g., (5, k)) to form a record. The records generated by a transient event in all 32 frequency bands are collected to form a frequency band response difference sequence.

[0054] The steps for obtaining the dynamic error feature vector are as follows:

[0055] Based on the frequency band response difference sequence, the frequency bands are rearranged from low frequency to high frequency according to the multi-level frequency band index while maintaining the consistent order of the time series. The energy ratios and phase differences of all frequency bands are concatenated in a fixed splicing order, and the placeholders of the missing frequency band positions are filled in to generate a dynamic error feature vector.

[0056] Specifically, based on the obtained frequency band response difference sequence, the frequency band indexes are first reordered. The original frequency band indexes generated by the wavelet packet decomposition algorithm (usually in Gray code order) do not have a linear correspondence with the frequency levels. Therefore, the 32 frequency bands need to be renumbered from the lowest to the highest frequency according to the actual center frequency corresponding to each frequency band index. For example, the frequency band with the lowest center frequency is marked as number 1, the second lowest as number 2, and so on, up to number 32. During the reordering process, the energy ratio and phase difference value corresponding to each frequency band also shift accordingly, but must maintain the association with the original time stamp. That is, all operations are completed within the data of the same transient event to ensure the consistency of the time series. Next, following the new order from frequency band 1 to 32, the energy ratio and phase difference values ​​of each frequency band are concatenated and spliced ​​to form a numerical sequence of length 64. The fixed splicing order is: energy ratio of frequency band 1, phase difference of frequency band 1, energy ratio of frequency band 2, phase difference of frequency band 2, ..., energy ratio of frequency band 32, phase difference of frequency band 32. Before splicing, a validity check is performed, and an energy threshold is set. This threshold is set to 0.001% of the total energy of the original primary signal of the transient event. For example, if the total energy of the primary signal of a transient event is 10,000 units, then the energy threshold is 0.1 units. For any frequency band, if its primary energy measurement... If the signal is below this threshold, the signal in that frequency band is considered too weak, and the calculated energy ratio and phase difference are unreliable. In this case, a preset placeholder (e.g., -999.0) is used to fill the splicing position corresponding to that frequency band. After completing the splicing and placeholder filling, the resulting numerical sequence with a length of 64 is the dynamic error feature vector of the transient event.

[0057] The steps to obtain the cyclic spectrum energy distribution map are as follows:

[0058] Call the steady-state waveform segment in the multidimensional feature source data, extract the secondary side signal time series and remove DC bias, segment it according to fixed length and fixed overlap rate and record the timestamp, calculate the cyclic autocorrelation sequence segment by segment and accumulate along the time delay axis, perform complex exponential kernel integration along the time delay axis to obtain the complex value distribution of the frequency axis and establish a joint index of the frequency axis and the cyclic frequency axis to obtain the cyclic spectral density function.

[0059] Based on the cyclic spectral density function, the maximum peak position of the energy distribution is searched on the cyclic frequency axis. The index corresponding to the maximum peak position is recorded as the main cyclic frequency index, and the bandwidth of the main peak is calculated. The maximum peak position of the energy distribution is searched on the frequency axis, and the maximum peak position is recorded as the main frequency index. The half-width at half-maximum (WHM) of the main peak is calculated. Then, the energy set of the main cyclic frequency spectral lines is labeled according to the condition that the cyclic frequency index is an integer multiple of the main cyclic frequency and the frequency index is located in the neighborhood of the main frequency's WHM. At the same time, regions where the cyclic frequency index is not an integer multiple of the main cyclic frequency and the frequency index is far from the neighborhood of the main frequency's WHM are selected. The upper quantile of the energy distribution in these regions is calculated as the noise power threshold, and these regions are labeled as non-main cyclic frequency spectral line energy sets, forming a cyclic spectral energy distribution map.

[0060] Specifically, all waveform segments marked as steady-state in the multi-dimensional feature source data are called, and time series data of the secondary side signal are extracted from each segment. For example, a steady-state waveform of the secondary side current with a length of 2 seconds and a sampling frequency of 20kHz is selected to obtain a time series containing 40,000 sampling points. First, the average value of the series is calculated, and this average value is subtracted from each data point in the series to remove the DC bias. Next, the fixed length of the segment is set to 2048 sampling points, and the fixed overlap rate is 50%, that is, 1024 overlapping sampling points. The entire 40,000-point time series is divided into sliding segments according to this rule. The first segment is from point 1 to point 2048, the second segment is from point 1025 to point 3072, and so on, resulting in a total of 38 data segments. The starting point of each segment is recorded in the original sequence. Next, for each of the 38 data segments, calculate its cyclic autocorrelation sequence, which is a two-dimensional function dependent on the time delay and the cyclic frequency. After calculation, average the 38 cyclic autocorrelation sequences point-by-point at the corresponding time delay and cyclic frequency points to obtain a smoother and more stable estimate. Then, perform a discrete Fourier transform along the time delay axis on this averaged cyclic autocorrelation sequence, i.e., apply a complex exponential kernel integral operation. This transform converts the function in the time delay domain into a complex-valued function in the frequency domain, resulting in a two-dimensional complex matrix with the frequency axis and the cyclic frequency axis as its two dimensions. Finally, establish a joint index for this two-dimensional matrix, where one index p corresponds to the regular frequency and the other index q corresponds to the cyclic frequency. This two-dimensional complex matrix with the joint index is the desired cyclic spectral density function.

[0061] Based on the calculated cyclic spectral density function, which is a two-dimensional complex energy matrix with frequency index p and cyclic frequency index q as coordinates, we first project the matrix onto the cyclic frequency axis. This involves summing the energy values ​​(i.e., the squares of the complex numerical modulus) at all frequency indices p corresponding to each cyclic frequency index q, resulting in a one-dimensional energy distribution curve with respect to the cyclic frequency q. We then search for the peak point with the maximum energy value on this curve and record the cyclic frequency index corresponding to this peak point as the main cyclic frequency index. For a 50Hz power system, this frequency theoretically corresponds to 100Hz. Then, the bandwidth of this main peak is determined by finding the position where the peak energy is halfway down and recording the two cyclic frequency indices on either side of the main peak. The range between these two indices is the bandwidth of the main peak. Next, projection is performed on the frequency axis in the same way, that is, the energy values ​​at all cyclic frequency indices q corresponding to each frequency index p are added together to obtain the energy distribution curve with respect to frequency p. The position of its maximum peak is searched, and the corresponding frequency index is recorded as the main frequency index. Theoretically, this corresponds to 50Hz. The half-width at half-maximum (WHM) of this main peak is calculated, which is the width at half the peak energy. Then, the energy distribution map of the cyclic spectrum is labeled with regions, the first condition being that the cyclic frequency index q is approximately equal to the main cyclic frequency index. Integer multiples of (e.g., in) Within 0.8 times the bandwidth), and the frequency index p is located at the major frequency index. Within the half-height width neighborhood (e.g., in Within an area extending 1.5 times half the height on both the left and right sides, all elements satisfying these two conditions... The index is used to label the set of spectral energy lines of the main cyclic frequency. Meanwhile, regions that do not meet the above conditions are selected as noise and distortion regions. Specifically, the selection criterion is that the cyclic frequency index q is not the main cyclic frequency index. The frequency index p is an integer multiple of the primary frequency index p, and the frequency index p is far from the primary frequency index p. neighborhood (e.g., distance) Within all such selected areas (more than 5 times the half-width), the energy values ​​of all points are statistically analyzed, and the 99th percentile of these energy values ​​is calculated. This value is then set as the noise power threshold. At the same time, all of these areas The index pair labeled as the energy set B of non-main cyclic frequency spectral lines, after the above labeling process, forms the cyclic spectral density function, which forms the cyclic spectral energy distribution map.

[0062] The steps to obtain the cyclic distortion index are as follows:

[0063] Based on the cyclic spectrum energy distribution diagram, the cyclic distortion index is calculated using the following formula:

[0064] ;

[0065] in, The cyclic distortion index, Let be the complex values ​​of the cyclic spectral density function at frequency indices p and cyclic frequency indices q. The set of index pairs corresponding to the energy set of the main cyclic frequency spectral lines. This is the set of index pairs corresponding to the energy sets of non-main-cycle frequency spectral lines. , The frequency index corresponding to the main frequency. The loop frequency index corresponding to the main loop frequency. It is the frequency attenuation factor. Noise power threshold The non-negative truncation operator is defined as max p is the frequency index, and q is the cyclic frequency index.

[0066] Specifically, the formula: The above formula calculates the cyclic distortion index, which reflects the degree of signal distortion caused by nonlinear effects such as transformer core saturation, by quantifying the ratio of the weighted energy of non-main cyclic frequency components to the total energy of the main cyclic frequency components. The method subtracts the background noise power before calculating the distortion energy and removes energy contributions below the noise level using a non-negative truncation operator, thereby improving the exponent's sensitivity to detecting weak distortions and its noise resistance. A specially designed weighting function is also introduced. The function consists of two parts, a logarithmic term. This allows higher-order cyclic frequency components, which are far from the fundamental cyclic frequency, to receive greater weight, since these higher-order components are typically associated with more severe nonlinear distortion. The exponential term... This results in higher weighting for distortion components closer to the dominant frequency, thus increasing the impact of in-band distortion on measurement accuracy. Through this weighting method, the resulting cyclic distortion index not only reflects the overall energy of the distortion but also the relative importance of distortion components at different frequency positions.

[0067] The parameter is obtained by performing a series of signal processing steps on a steady-state waveform segment from a multidimensional feature source data. This includes extracting the secondary side signal, removing the DC bias, segmenting the data, calculating and averaging the cyclic autocorrelation sequence, and finally performing a Fourier transform. The resulting parameter is a two-dimensional complex matrix and serves as the fundamental data source for this formula. Its value directly reflects the energy and phase information of the signal at specific frequencies and cyclic frequencies. For example, in one calculation, the index pairs... Located in set A, its corresponding complex value is And the index pairs in set B At this location, the corresponding complex value is .

[0068] The steps to obtain the parameter are as follows: this parameter represents the index pair corresponding to the energy set of the main circulating frequency spectral lines. The set is first determined by the energy projection method to determine the dominant frequency index. and main loop frequency index And calculate its spectral peak half-width and bandwidth, then according to the preset conditions, i.e., the cyclic frequency index is... Integer multiples of and frequency index located at Within the neighborhood of [value], filter out all index pairs that meet the criteria. This constitutes set A, which represents the main periodic components of the signal and serves as the benchmark for calculating the total energy of the signal. For example, it can be calculated as follows: , Set A is determined to contain index pairs. A set of points.

[0069] The steps to obtain the parameter are as follows: this parameter represents the index pair corresponding to the set of spectral energy lines of the non-main cycle frequency spectrum. The identification process of the set A is carried out simultaneously with the determination of set A, by selecting a cyclic frequency index that is not the primary cyclic frequency index. Integer multiples of the primary frequency index, and the frequency index is far from the primary frequency index. The neighborhood is used to define the scope of this set. This set represents the non-periodic or weakly periodic components caused by factors such as nonlinear distortion and noise, and is the main object for measuring the degree of signal distortion. For example, in an analysis, set B is determined to contain index pairs. and Areas with equal points.

[0070] The steps to obtain the parameter are as follows: the parameter is a weighting function, and its calculation formula is... Its calculation depends on the index of the current point. and a pre-determined main frequency index Main loop frequency index and frequency attenuation factor The value of this function is dynamically calculated based on the characteristics of the cyclic spectrum itself, without the need to collect data from external sources. It is used to adjust the contribution of distortion components at different locations in the final exponent calculation.

[0071] The steps for obtaining the parameter are as follows: This parameter is the frequency index corresponding to the main frequency. It is determined by summing and projecting the energy of the cyclic spectral density function along the cyclic frequency axis, and searching for the location of the maximum peak point on the resulting frequency energy distribution curve. It identifies the fundamental frequency position where the energy is most concentrated in the signal. For China's 50Hz power system, if the frequency resolution is 1Hz / index, then... The value should be 50.

[0072] The steps to obtain the parameter are as follows: this parameter is the loop frequency index corresponding to the main loop frequency, and its determination method is the same as... Similarly, the location of the maximum peak point is determined by summing and projecting the energy of the cyclic spectral density function along the frequency axis and searching on the resulting cyclic frequency energy distribution curve. This peak typically corresponds to twice the fundamental frequency, i.e., 100Hz. If the cyclic frequency resolution is 2Hz / index, then... The value should be 50.

[0073] The steps for obtaining the parameter are as follows: This parameter is the frequency attenuation factor, used to control the attenuation rate of the Gaussian part in the weighting function. Its value is set based on the spectral width of the main frequency peak. A reasonable setting method is to correlate its value with the standard deviation or half-width at half maximum (FWHM) of the energy distribution of the main frequency peak. It can be set to a fixed multiple of FWHM, for example, set If the full width at half maximum (FWHM) of the dominant frequency peak is measured to be 10 index units, then It can be set to 8.49; for ease of calculation, a fixed value based on empirical adjustment is set here. This value ensures sensitivity to in-band distortion while also taking into account a certain range of out-of-band distortion.

[0074] The steps to obtain the parameter are as follows: This parameter is the noise power threshold, which is calculated by statistically analyzing the energy values ​​of all points within the selected non-main cyclic frequency spectral line energy set B region in the cyclic spectrum energy distribution map. These energy values ​​are then compiled into a dataset, and the 99th percentile of this dataset is calculated as... For example, if 10,000 energy points are collected within region B, and they are sorted from smallest to largest, then the value of the 9900th point is taken. If this value is 0.08, then... .

[0075] Calculations based on parameters:

[0076] The parameters for the simulation are set as follows:

[0077] Set A contains a point Its energy .

[0078] Set B contains two points and Their energies are as follows:

[0079] .

[0080] .

[0081] Other parameters: , , , .

[0082] Calculate the denominator:

[0083] ;

[0084] Calculate the weight of each point in the molecule :

[0085] For point :

[0086] ;

[0087] ;

[0088] ;

[0089] For point :

[0090] ;

[0091] ;

[0092] ;

[0093] Calculate the numerator:

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] Calculate the cyclic distortion index :

[0100] ;

[0101] The results show that the calculated cyclic distortion index in this steady-state waveform analysis is 0.001236. The evaluation criterion is set as follows: If... This indicates that the current transformer is operating in the linear region with a very low distortion level. This indicates the presence of slight nonlinear distortion. If the value is less than 0.005, it indicates that the distortion is relatively severe, possibly caused by core saturation, etc. The current calculation result of 0.001236 is much less than 0.005, indicating that the current transformer corresponding to this steady-state segment is in good operating condition and has extremely low nonlinear distortion.

[0102] The steps to obtain the comprehensive error correction matrix are as follows:

[0103] The dynamic error feature vector and the cyclic distortion index are called, the unit range is unified, and a one-to-one correspondence is established with timestamp and frequency band index. The numerical bits are concatenated according to a fixed field order, and abnormal jumps are thresholded and missing positions are linearly interpolated to obtain the error fusion characterization sequence.

[0104] Based on the error fusion characterization sequence, the condition thresholds and transformation entries of the error mapping rules are read. The mapping entries are matched one by one according to the frequency band index and distortion level, and the numerical bits are split into amplitude channels and phase channels. Transformation and saturation clipping are performed respectively to generate channel identifiers and time identifiers. The compensation amounts of amplitude and phase are obtained and summarized into a comprehensive error correction matrix.

[0105] Specifically, the dynamic error feature vector and cyclic distortion index generated in the aforementioned steps are used. First, the dimensions of these two features from different sources are unified. The energy ratio in the dynamic error feature vector is dimensionless, the phase difference is in radians, and the cyclic distortion index is also dimensionless. To facilitate fusion, the phase difference is divided by... Normalization to The energy ratio and cyclic distortion index are already within a suitable range and require no additional processing. Next, using the timestamp as a reference, the dynamic error feature vector corresponding to each transient event is associated with the cyclic distortion index calculated during the nearest steady-state period before and after the transient event. Since a dynamic error feature vector contains information from 32 frequency bands, and the cyclic distortion index is a scalar, the value of the cyclic distortion index is copied 32 times, establishing a one-to-one correspondence with the 32 frequency band data in the dynamic error feature vector, forming a data table containing timestamps, frequency band indices, normalized energy ratios, normalized phase differences, and cyclic distortion indices. Then, according to the frequency... Using a fixed field order for "indexed", "normalized energy ratio", "normalized phase difference", and "cyclic distortion index", data from each timestamp is concatenated into a longer feature sequence. This sequence is then processed for outliers, with an outlier rejection threshold set. This threshold is determined through statistical analysis of the feature sequence calculated on a reference dataset (e.g., the initial 24-hour data collection). Specifically, the average and standard deviation of the differences between adjacent timestamps for each feature field are calculated, and the threshold is set to the average plus 3.5 times the standard deviation. For example, if the average difference in the energy ratio is 0.01 and the standard deviation is 0.005, then the outlier threshold is 0.01. + 3.5 * 0.005 = 0.0275. Traverse the concatenated feature sequence. If the absolute value of the difference between a data point and the value of its predecessor exceeds this threshold, the point is marked as an anomaly and removed. For the missing positions that originally existed in the dynamic error feature vector (marked by the placeholder -999.0) and the empty positions generated after removing the anomaly jump, linear interpolation is used to fill them. That is, the average value of the two nearest valid data points before and after the position is used to fill them. After completing all the above processing, a complete and uniform error fusion representation sequence is obtained.

[0106] Based on the error fusion representation sequence generated in the previous step, the preset error mapping rules are first loaded. These rules are multidimensional lookup tables, established through offline experiments and simulations. The construction process is as follows: In a laboratory environment, different types and intensities of transient disturbances and varying degrees of saturation excitation are applied to a standard transformer. Simultaneously, the primary and secondary side signals are precisely measured, and the corresponding dynamic error feature vectors and cyclic distortion indices are calculated. These are then compared with the actual amplitude and phase errors measured by standard instruments. Through machine learning methods such as regression analysis and decision trees, a mapping relationship is established from "band index" and "distortion level" (divided by the cyclic distortion index) to "amplitude compensation coefficient" and "phase compensation angle." This relationship is then solidified into a rule table. Each row of the table is a conversion entry, containing conditional thresholds (such as band index range and distortion level) and corresponding conversion parameters. For example, a rule might be: when the band index is between 5 and 8, and the distortion level is "medium" (cyclic distortion index), the error is considered to be "moderate" (associated with the cyclic distortion index). If the true exponent is between 0.02 and 0.05, then the amplitude compensation coefficient is 1.0015 and the phase compensation angle is -0.002 radians. Next, the data in the error fusion representation sequence are processed line by line. For each data point, its frequency band index is first read, and its distortion level is determined according to the preset distortion level threshold (e.g., <0.02 is "low", 0.02-0.05 is "medium", >0.05 is "high"). Then, using the frequency band index and distortion level as keys, a matching conversion entry is searched in the error mapping rule table. After finding a matching entry, the normalized energy ratio and normalized phase difference values ​​in the data point are extracted and sent to the amplitude channel and phase channel respectively for conversion. In the amplitude channel, the energy ratio is multiplied by the found amplitude compensation coefficient; in the phase channel, the phase difference is added to the found phase compensation angle. After the conversion, saturation clipping is performed, and a reasonable range for the amplitude compensation amount is set, for example... The reasonable range of phase compensation amount, for example If the calculated compensation amount exceeds this range in radians, it is forcibly set to the boundary value of the range. Finally, a channel identifier ("amplitude" or "phase") and the original time identifier are added to each calculated amplitude and phase compensation amount. The compensation amounts of all time points and all frequency bands are summarized to form a comprehensive error correction matrix.

[0107] The steps for obtaining the online calibration coefficient set are as follows:

[0108] Based on the magnitude and phase compensation amounts in the comprehensive error correction matrix, the comprehensive error correction matrix is ​​filled in the order of primary side, secondary side, and channel. It is then expanded into correction parameters according to the register address mapping table, written item by item into the metering chip register, and the checksum and status bit are read back and compared. The version number and timestamp are recorded to obtain the online calibration coefficient set.

[0109] Specifically, based on the generated comprehensive error correction matrix, which contains the amplitude and phase compensation amounts calculated for different frequency bands and distortion levels, these compensation amounts are first organized into a structured data block according to a preset channel order. The order is typically primary-side voltage amplitude compensation, primary-side voltage phase compensation, primary-side current amplitude compensation, primary-side current phase compensation, and the corresponding four secondary-side compensation amounts. Since this method primarily calibrates secondary-side measurement errors, the primary-side compensation amounts are usually set to zero, while the secondary-side compensation amounts are directly taken from the corresponding values ​​in the comprehensive error correction matrix. After filling, the register address mapping table of the metering chip is loaded. This table defines in detail the addresses and functions of each register inside the chip used to store correction coefficients. For example, address 0x1000 is used to store the amplitude gain correction coefficient of the voltage channel, address 0x1002 is used to store its phase correction coefficient, etc. Then, the filled structured data block is expanded item by item into a series of... The calibration parameters are then written to the corresponding register address of the metering chip. This may involve data format conversion, such as converting floating-point compensation values ​​to fixed-point or integer formats that the chip can recognize. Next, these calibration parameters are written one by one to the register address of the metering chip through a communication interface (such as SPI or I2C). After each parameter is written, a readback operation is immediately performed to read the data from the same address and compare it with the written value. At the same time, the status register of the chip is checked to confirm whether the write operation was successful and whether there is an error flag set. In addition, the checksum (e.g., CRC16) of all written parameters is calculated and compared with the checksum automatically calculated by the chip based on the written data. Only when the readback data is consistent, the status bit is normal, and the checksum matches is the parameter confirmed to have been written successfully. After all parameters are successfully written, the version number of the current calibration operation (an auto-incrementing sequence number) and the timestamp accurate to the second are recorded in a specific information register. The set of all calibration parameters that have been written and verified successfully ultimately constitutes the online calibration coefficient set.

[0110] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method of on-line calibration of power transformer error, characterized by, The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps:

2. The method of on-line calibration of power transformer error according to claim 1, characterized in that, The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps:

3. The method of on-line calibration of power transformer error according to claim 1, wherein, The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the 4. The method of on-line calibration of power transformer error of claim 1, wherein, The acquisition step of the dynamic error feature vector is: According to the frequency band response difference sequence, rearranging from low frequency to high frequency according to the multi-level frequency band index and keeping the consistent order of time sequence, concatenating the energy ratio and phase difference of all frequency bands in fixed splicing order and filling the placeholder of missing frequency band position, a dynamic error feature vector is generated.

5. The method of on-line calibration of power transformer error as claimed in claim 1, wherein, The acquisition step of the cyclic spectrum energy distribution diagram is: Calling the steady-state waveform segment in the multi-dimensional feature source data, extracting the secondary side signal time sequence and removing the DC bias, segmenting according to the fixed length and fixed overlap rate and recording the time stamp, calculating the cyclic autocorrelation sequence segment by segment and accumulating along the time delay axis, performing complex exponential kernel integration along the time delay axis to obtain the frequency axis complex value distribution and establish the joint index of frequency axis and cycle frequency axis, and obtaining the cyclic spectrum density function; According to the cyclic spectrum density function, searching for the maximum peak position of energy distribution on the cycle frequency axis, recording the index corresponding to the maximum peak position as the main cycle frequency index, and counting the bandwidth where the main peak is located, searching for the maximum peak position of energy distribution on the frequency axis, recording the maximum peak position as the main frequency index, and counting the half width of the main peak, then according to the conditions that the cycle frequency index is an integer multiple of the main cycle frequency and the frequency index is located in the neighborhood range of the half width of the main frequency, the main cycle frequency spectrum line energy set is marked, and the region where the cycle frequency index is not an integer multiple of the main cycle frequency and the frequency index is far away from the neighborhood range of the half width of the main frequency is selected, the upper quantile of energy distribution in the region is counted as the noise power threshold, and the region is marked as the non-main cycle frequency spectrum line energy set, forming a cyclic spectrum energy distribution diagram.

6. The method of on-line calibration of power transformer error of claim 1, wherein, The acquisition step of the cyclic distortion index is: according to the cyclic spectrum energy distribution diagram, the cyclic distortion index is calculated.

7. The method of online calibration of power transformer error as claimed in claim 1 wherein, The acquisition step of the comprehensive error correction matrix is: Calling the dynamic error feature vector and the cyclic distortion index, unifying the dimension range and establishing a one-to-one correspondence with the time stamp and the frequency band index, splicing the value bits in fixed field order and removing the abnormal jump threshold and linearly interpolating the missing position, obtaining the error fusion representation sequence; According to the error fusion representation sequence, reading the conditional threshold and conversion items of the error mapping rule, matching the mapping items one by one according to the frequency band index and the distortion level, and splitting the value bits into amplitude channel and phase channel, respectively, converting and saturating clipping, generating channel identifier and time identifier, obtaining the compensation amount of amplitude and phase, and summarizing the comprehensive error correction matrix.

8. The method of online calibration of power transformer error as claimed in claim 1 wherein, The acquisition step of the online calibration coefficient set is: According to the compensation amount of amplitude and phase in the comprehensive error correction matrix, filling the comprehensive error correction matrix according to the primary side and secondary side and channel order, and expanding the correction parameter according to the register address mapping table, writing into the meter chip register one by one and reading back to compare the checksum and status bit, recording the version number and time stamp, and obtaining the online calibration coefficient set.

9. The on-line calibration system of the on-line calibration method of errors of power transformers according to any of claims 1-8, characterized in that, It comprises: The data acquisition module triggers data acquisition by setting voltage and current change rate threshold, synchronously acquires high-frequency waveform data of the primary side and the secondary side of the mutual inductor, intercepts waveform fragments of transient period with the trigger point as the center to form a transient event waveform set, selects continuous steady-state waveform fragments from the area outside the transient event, and establishes multi-dimensional characteristic source data; The dynamic feature extraction module is configured to call the transient waveform fragments in the multi-dimensional characteristic source data, perform multi-level wavelet packet decomposition on the primary side and the secondary side waveforms, calculate the energy ratio and phase difference of the primary side and the secondary side coefficients in each frequency band to form a frequency band response difference sequence, reconstruct and arrange the frequency band response difference sequences of all frequency bands, and generate a dynamic error feature vector; The steady-state feature extraction module is configured to call the steady-state waveform fragments in the multi-dimensional characteristic source data, extract energy distribution information on the frequency axis and the cyclic frequency axis, construct a cyclic spectrum energy distribution diagram, separate main cyclic frequency spectrum line energy and non-main cyclic frequency spectrum line energy introduced by core saturation according to the cyclic spectrum energy distribution diagram, and calculate a cyclic distortion index; The error correction module is configured to fuse the dynamic error feature vector and the cyclic distortion index, convert the fusion result into compensation amounts of amplitude and phase according to a preset error mapping rule, establish a comprehensive error correction matrix, generate a correction parameter according to the comprehensive error correction matrix, and write the correction parameter into a metering chip register of the electric energy meter to obtain an online calibration coefficient set.

Images