A method and system for quickly locating abnormal fluctuations in electric meter data

Abstract

The present application relates to the technical field of power quality monitoring, and particularly relates to a method and system for rapidly positioning abnormal fluctuations of electric meter data, which comprises the following steps: collecting three-phase voltage data and synchronously measuring the power grid fundamental frequency, generating a one-dimensional three-phase instantaneous projection sum, using a half-cycle sliding window that only contains historical sampling time and whose length is adaptively updated according to the power grid fundamental frequency to construct a historical mean reference, calculating the instantaneous relative offset, further solving the window disturbance energy and the correlation energy of the short-time window for the instantaneous relative offset, and fusing the stability features that can distinguish abnormal pulses and background noise to complete abnormal interval identification, interval merging and noise filtering processing and output abnormal alarm information relying on the noise threshold adaptively set by the normal operation data of the equipment. The present application combines adaptive window and double-feature fusion discrimination technology to accurately identify noise and arc fault, and greatly improves the power grid abnormal detection performance.

Description

Technical Field

[0001] This invention relates to the field of power quality monitoring technology. In particular, it relates to a method and system for rapidly locating abnormal fluctuations in electricity meter data. Background Technology

[0002] With the widespread adoption of smart grids and electricity information collection systems, smart meters, as core terminal equipment of the power grid, directly determine the effectiveness of power quality monitoring, fault early warning, and electricity safety management through their data collection and analysis performance. During operation, smart meters are prone to various abnormal data fluctuations, such as arc faults, transient overvoltages, harmonic surges, and rapid load switching. Therefore, rapid and accurate identification and location of these faults are crucial to ensure timely handling of power grid failures by maintenance personnel.

[0003] Currently, meter anomaly detection mainly falls into two categories: one is the fixed threshold method and the effective value change rate detection method; the other is the time-frequency domain feature extraction algorithm that incorporates sliding window mean filtering, wavelet transform, and Fourier transform. The former has weak noise resistance and cannot adapt to dynamic changes in the power grid frequency; the latter, although it can extract frequency domain anomaly features, generally relies on a centrally symmetric sliding window, and the window parameters need to be manually adjusted, resulting in a large overall computational cost.

[0004] Existing detection schemes have several inherent defects: the calculation of the centrally symmetric window will call future sampling data, causing normal sampling time before the anomaly occurs to be misjudged, resulting in negative deviation false positives; simple differential and high-pass filtering cannot distinguish between random abnormal pulses and strongly correlated background noise such as fundamental residual and derivative noise, and weak fault signals are easily masked by noise, resulting in frequent false alarms and missed alarms; at the same time, the algorithm has high complexity, manual parameter tuning, and poor frequency adaptation, which greatly reduces the real-time performance and detection accuracy of anomaly location. Summary of the Invention

[0005] To address the problems of existing detection schemes, such as false positives, difficulty in separating noise, complex algorithms, poor adaptability, and insufficient detection accuracy and real-time performance, this invention provides solutions in the following aspects.

[0006] In a first aspect, a method for rapidly locating abnormal fluctuations in electricity meter data includes: acquiring three-phase voltage data and simultaneously calculating the fundamental frequency of the power grid; generating a three-phase instantaneous projection sum of a one-dimensional signal based on the preprocessed three-phase instantaneous values; adaptively constructing a half-cycle sliding window containing only historical sampling times based on the fundamental frequency of the power grid, and establishing a historical average benchmark based on the half-cycle sliding window, and calculating the instantaneous relative offset in combination with the three-phase instantaneous projection sum; calculating the window disturbance energy and correlation energy within the short-time window for the instantaneous relative offset, fusing the window disturbance energy and correlation energy to obtain stability characteristics used to distinguish abnormal pulses from background noise; adaptively setting a noise threshold based on normal equipment operation data, identifying abnormal intervals based on the stability characteristics, merging and filtering noise in the abnormal intervals, and outputting abnormal alarm information.

[0007] Preferably, the step of establishing a historical mean benchmark based on a half-cycle sliding window includes: Taking any sampling time as the target time, calculate the total number of window samples corresponding to half a fundamental frequency cycle based on the meter sampling rate and the real-time fundamental frequency of the power grid, and round it down. Update the total number of window samples based on the real-time fundamental frequency of the power grid. In all three-phase instantaneous projections and data, the continuous historical three-phase instantaneous projections before the target time, which are consistent with the total number of window samples, are extracted. All extracted data are summed to obtain the cumulative sum. The cumulative sum is divided by the total number of window samples to obtain the historical average benchmark for the target time.

[0008] Preferably, the step of calculating the window perturbation energy within the short-time window includes: The local instantaneous variance, which characterizes the total disturbance energy within the window, is obtained by calculating the squared mean of the instantaneous relative offset sequence using a short-time statistical window. The number of sampling times in the short-time statistical window is obtained by scaling and rounding the number of sampling times in the adaptive historical window by a fixed ratio. The short-time statistical window is dynamically updated synchronously with the adaptive historical window. When calculating the local instantaneous variance, all instantaneous relative offsets within the short-time window, including the current target sampling time, are truncated. The squares of each instantaneous relative offset are calculated and summed. The summation result is divided by the number of sampling times in the short-time statistical window to obtain the local instantaneous variance at the corresponding time, which is used as the window disturbance energy.

[0009] Preferably, the correlation energy is calculated as follows: Retrieve all instantaneous relative offset data sequences under the complete time series to determine the total number of sampling times corresponding to the short-time statistical window; take any sampling time as the target time, and extract continuous instantaneous relative offsets from the instantaneous relative offset sequence that include the current target sampling time and whose total number is equal to the number of sampling times in the short-time window; Iterate through each group of adjacent instantaneous relative offsets within the window, multiply the two adjacent instantaneous relative offsets in each group to obtain the product value of each group of adjacent sampling times; sum the product values ​​of all adjacent points within the window to obtain the cumulative sum, and divide the cumulative sum by the total number of sampling times in the short-term statistical window to calculate the correlation energy of the current target time.

[0010] Preferably, the instantaneous relative offset is calculated as follows: Taking any sampling time as the target time, the instantaneous relative offset is obtained by comparing the three-phase instantaneous projection corresponding to the target time with the historical average benchmark corresponding to the target time; the instantaneous relative offset is used to separate the power frequency steady-state base from the low-frequency drift.

[0011] Preferably, the calculation method for the sum of the three-phase instantaneous projections is as follows: The instantaneous values ​​of the three-phase voltages A, B, and C, which are preprocessed at the same time, are directly added together to obtain the instantaneous projection sum of the three phases.

[0012] Preferably, the preprocessing step specifically includes: Isolated pulse noise introduced by the analog-to-digital converter is filtered out by median filtering or moving average algorithm; the three-phase voltage signal is normalized to eliminate gain differences between multiple acquisition channels; and the fundamental frequency of the power grid is calculated in real time by phase-locked loop or zero-crossing detection algorithm.

[0013] Secondly, a system for quickly locating abnormal fluctuations in electricity meter data includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned method for quickly locating abnormal fluctuations in electricity meter data is implemented.

[0014] The present invention has the following effects: 1. This invention constructs a historical mean benchmark corresponding to a half-cycle sliding window using historical sampling data, without calling future time-series sampling data throughout the process. This avoids the causal logic defects of traditional centrally symmetric windows and, in principle, prevents false positives where normal steady-state conditions are misjudged as abnormal. It eliminates the need for complex frequency domain iterative operations such as Fourier transform and wavelet decomposition, resulting in low computational load and small processing delay. It enables synchronous online real-time detection of sampling data, solving the problem of insufficient real-time performance in traditional solutions.

[0015] 2. This invention extracts and fuses two types of differentiated features—short-term window disturbance energy and correlation energy—to obtain stability features. It utilizes the inherent physical differences between noise and arc pulses in terms of energy intensity and temporal correlation to achieve noise suppression. Furthermore, it employs a multi-layer post-processing mechanism based on the device's own normal operating data, which adaptively generates a noise threshold, merges abnormal intervals, and filters out short-term noise intervals. This effectively separates derivative noise, periodic residual components, instantaneous glitch noise, and real arc fault pulses, overcoming the problem that traditional single-amplitude features are difficult to distinguish between noise and disturbances and have low recognition accuracy.

[0016] 3. This invention dynamically updates the length of the half-cycle sliding window to follow the real-time fundamental frequency of the power grid, and synchronously and adaptively adjusts the short-time statistical window. The noise threshold for discrimination is automatically generated based on the steady-state data of the equipment, eliminating the need for manual configuration and repeated debugging of fixed hyperparameters for different loads and frequency offsets. The entire solution has strong adaptability and is compatible with complex operating conditions such as power grid frequency fluctuations, channel gain deviations, and changes in field noise intensity, solving the technical defects of traditional algorithms with poor adaptability of fixed parameters and cumbersome field deployment and debugging. Attached Figure Description

[0017] Figure 1 This is a flowchart of steps S1-S4 in a method for quickly locating abnormal fluctuations in electricity meter data according to an embodiment of the present invention.

[0018] Figure 2 This is a structural block diagram of a system for quickly locating abnormal fluctuations in electricity meter data according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0020] Reference Figure 1 A method for quickly locating abnormal fluctuations in electricity meter data includes steps S1-S4, as detailed below: S1: Acquire three-phase voltage data and simultaneously calculate the fundamental frequency of the power grid, and generate a one-dimensional signal three-phase instantaneous projection sum based on the preprocessed three-phase instantaneous values.

[0021] This invention relies on the high-speed signal sampling module built into the smart meter to synchronously and continuously acquire analog voltage signals of phases A, B, and C, with a fixed hardware sampling frequency of 20kHz. Multi-level preprocessing is performed on the acquired raw three-phase voltage sampling data. The preprocessing process is implemented in three steps: First, a median filtering algorithm or a moving average filtering algorithm is used to filter out isolated spike pulse interference noise generated during the operation of the analog-to-digital converter; second, amplitude normalization is performed on the filtered three-phase voltage signals to eliminate amplitude deviations caused by inconsistent hardware gains across multiple acquisition channels; third, a phase-locked loop frequency discrimination algorithm or a voltage zero-crossing detection algorithm is run synchronously to calculate the fundamental frequency of the current power grid in real time. The calculated fundamental frequency is used for subsequent adaptive adjustment of the sliding window length.

[0022] After completing all preprocessing steps, the obtained instantaneous values ​​of the three-phase voltages A, B, and C are free from hardware noise and channel amplitude errors. Based on the preprocessed instantaneous values ​​of the three phases at the same sampling time, the instantaneous projection sum of the three phases is calculated, and the corresponding calculation formula is as follows: ; In the formula, Indicates the first The instantaneous algebraic sum of the three-phase voltages at the sampling moment; Indicates the first The instantaneous value of phase A voltage after filtering and normalization preprocessing at the sampling time; Indicates the first The instantaneous value of phase B voltage after filtering and normalization preprocessing at the sampling time; Indicates the first The instantaneous value of phase C voltage after filtering and normalization preprocessing at the sampling time.

[0023] Under steady-state normal operation of the power grid, the three-phase voltages are symmetrical and balanced, the positive sequence components of the three-phase fundamental waves cancel each other out, and the instantaneous projection sum of the three phases fluctuates only slightly near zero. When a transient anomaly such as an arc fault occurs in the line, the balance of the three-phase signals is broken, and the three-phase voltages undergo asymmetrical sudden changes. This algebraic sum will generate a positive or negative pulse waveform with a significantly increased amplitude.

[0024] Through the above summation operation, the original three-dimensional independent three-phase voltage time sequence signal is compressed into a one-dimensional scalar time sequence signal. Utilizing the physical property that the positive sequence components of the three-phase fundamental wave cancel each other out, the steady-state power frequency component that dominates the signal is filtered out, and only the line common-mode interference noise and arc fault abnormal pulses are retained, providing a unified and single analysis target for subsequent hierarchical feature extraction and anomaly quantification identification.

[0025] However, relying solely on one-dimensional time-series signals cannot isolate signal disturbances caused by low-frequency drift and steady-state common-mode baseline in the power grid. Conventional processing methods generally suffer from technical defects: traditional signal baseline calibration often employs centrally symmetric sliding windows or fixed-parameter filtering algorithms. The centrally symmetric window operation requires access to sampling data from future moments, introducing causality defects and easily triggering false positive alarms. Furthermore, processing methods with fixed window parameters and fixed filtering coefficients cannot adapt to the real-time dynamic fluctuations of the power grid's fundamental frequency, exhibiting poor adaptability to operating conditions and making it difficult to accurately separate the steady-state baseline from instantaneous abnormal fluctuations. To overcome these defects and accurately extract the instantaneous fluctuation characteristics of the signal, this invention sets up the following steps: S2: Based on the fundamental frequency of the power grid, a half-cycle sliding window containing only historical sampling times is adaptively constructed. Based on the half-cycle sliding window, a historical average benchmark is built, and the instantaneous relative offset is calculated by combining the three-phase instantaneous projection.

[0026] Any sampling time is selected as the target sampling time. The fixed hardware sampling rate of the smart meter and the real-time measured output of the grid fundamental frequency are obtained. The total number of sampling times corresponding to half a grid fundamental cycle is calculated using the smart meter hardware sampling rate and the real-time grid fundamental frequency. The calculated total number of sampling times is rounded down to obtain the total number of historical window samples adapted to the current grid operating conditions. Theoretically, the total number of historical window samples can be updated frame by frame in real time following the grid fundamental frequency. In this embodiment, to reduce the computational overhead of the equipment, a control logic is adopted to recalculate and refresh the total number of historical window samples every 1 second. If a large change in the grid fundamental frequency is detected, the total number of historical window samples can be recalculated and updated in real time.

[0027] The system internally stores the three-phase instantaneous projection and time-series datasets corresponding to all sampling times. From all the three-phase instantaneous projection and time-series datasets, a set of continuous three-phase instantaneous projections and data is selected, whose time-series position is earlier than the target sampling time, whose time-series position does not include the target sampling time, and whose number of data entries is equal to the total number of samples in the historical window. Numerical accumulation operations are performed on all the selected continuous three-phase instantaneous projections and data to generate the cumulative sum corresponding to the continuous three-phase instantaneous projections and data. The historical mean benchmark corresponding to the target sampling time is obtained by dividing the cumulative sum by the total number of samples in the historical window.

[0028] The generation calculation of the historical average benchmark only uses the three-phase instantaneous projection and data whose time sequence position is earlier than the target sampling time, and does not introduce the sampling data whose time sequence position is later than the target sampling time. The historical average benchmark can quantitatively characterize the average level of the three-phase common mode voltage signal in the half-wave period interval of the power grid before the target sampling time. Based on this historical average benchmark, a dynamic baseline without causal logic defects is built, eliminating the problem of misidentification of abnormal states caused by the traditional central symmetric sliding window to retrieve future sampling data.

[0029] Specifically, the historical mean benchmark satisfies the following relationship: ; In the formula, Indicates the first The historical mean benchmark at time t is adopted using the t. Before the moment Calculated from historical sampling moments, representing the average common-mode level within half a fundamental cycle prior to the current moment; This represents the historical backtracking offset index, indicating the index starting from the first position. Looking back at time Each sampling interval, with a value range of [value range missing]. ,in The total number of samples in the dynamic historical window is the number of sampling times contained in half a fundamental frequency period, such as the hardware sampling frequency of the electricity meter. Grid fundamental frequency hour, Each sampling time corresponds to 10ms. To reduce computational overhead, this embodiment recalculates every second based on the real-time power grid fundamental frequency. ,in Perform a rounding function; Indicates the first Time shifts forward The three-phase instantaneous projection of the position.

[0030] After obtaining the historical mean benchmark corresponding to any target sampling time, the steady-state baseline component is removed through interpolation, and the instantaneous relative offset at the target sampling time is calculated. The specific calculation formula is as follows: In the formula, Indicates the first The instantaneous relative offset at time 1. Indicates the first The instantaneous algebraic sum of the three-phase voltages at the sampling time. Indicates the first Historical average benchmark at any given time.

[0031] By subtracting the real-time acquired three-phase instantaneous projection from the adaptively updated historical steady-state baseline, the absolute amplitude measurement of the original signal is converted into a dynamic offset relative to the recent steady-state average level. This effectively eliminates the steady-state interference caused by the low-frequency drift component of the power grid and the fixed common-mode base, and separates the pure instantaneous fluctuation component in the signal.

[0032] From the perspective of signal characteristics, when the power grid equipment is operating in a normal steady state, the signal fluctuation is stable and the baseline is stable. The instantaneous relative offset generally exhibits the characteristics of small fluctuations with zero mean. When abnormal conditions such as arc disturbances occur in the line, the three-phase voltage balance is momentarily lost, and the signal generates sudden asymmetric fluctuations. The instantaneous relative offset will generate positive or negative pulses with significant amplitudes, which can intuitively and sensitively characterize instantaneous abnormal disturbances.

[0033] This embodiment uses only historical sampling data to construct a steady-state baseline and calculate the relative offset, completely avoiding the causal logic defects caused by the introduction of future sampling data in traditional centrally symmetric sliding windows. This fundamentally eliminates the false positive alarm problem of normal operating conditions being misjudged as abnormal, ensuring the real-time performance and accuracy of anomaly warnings. Simultaneously, the baseline calculation window length can be dynamically and adaptively adjusted according to the grid fundamental frequency, eliminating the need for manual configuration of fixed parameters and adapting to the complex actual operating conditions of grid frequency offset. Furthermore, by converting absolute signals to relative fluctuation characteristics, the steady-state power frequency component and low-frequency drift interference are completely eliminated, providing a clean and effective preliminary data foundation for the accurate extraction and quantitative analysis of subsequent short-term disturbance energy, time-series correlation, and other multi-dimensional features.

[0034] However, relying solely on the single-dimensional feature of instantaneous offset amplitude for anomaly detection still has limitations: random background noise in the field can also cause instantaneous amplitude fluctuations in the signal, which can easily be confused with weak arc anomaly pulses, leading to misjudgments and missed detections, and failing to achieve high-precision anomaly fluctuation identification. To address the technical shortcomings of insufficient accuracy in identifying this single feature, this invention further quantifies and analyzes the instantaneous relative offset from two differentiated dimensions: signal energy intensity and temporal correlation characteristics. These two types of features are then fused, utilizing the essential differences between noise and real fault pulses in energy and temporal correlation to achieve accurate differentiation. The specific steps are as follows: S3: Calculate the window perturbation energy and correlation energy within the short time window for the instantaneous relative offset, and fuse the window perturbation energy and correlation energy to obtain the stability characteristics used to distinguish abnormal pulses from background noise.

[0035] To accurately quantify the disturbance intensity contained in the instantaneous relative offset and highlight the energy difference between arc anomaly pulses and weak noise disturbances, this invention performs energy statistical calculations on the time series sequence of instantaneous relative offsets using a short-time statistical window to obtain the local instantaneous variance characterizing the strength of disturbances in local intervals. The number of sampling times in the short-time statistical window is determined based on the number of sampling times in the adaptive historical window. This is achieved by scaling and rounding the number of sampling times in the adaptive historical window by a fixed ratio, allowing the short-time statistical window to dynamically update synchronously with the adaptive historical window. This adapts to the real-time fluctuations of the power grid fundamental frequency, ensuring that the short-time statistical window has a consistent transient capture capability under different power grid operating conditions.

[0036] For any target sampling time, continuous instantaneous relative offset data, including the current target sampling time and with the same number of sampling times as the short-time statistical window, is extracted from the complete instantaneous relative offset time series data. By squaring each extracted instantaneous relative offset, large-amplitude abrupt pulse components are effectively amplified, while small-amplitude stationary noise components are suppressed, highlighting the energy characteristics of abnormal disturbances. All squared values ​​within the window are summed to obtain the total disturbance energy within the short-time window. This total disturbance energy is then divided by the number of sampling times in the short-time statistical window to calculate the mean, ultimately yielding the local instantaneous variance corresponding to the current target sampling time. This local instantaneous variance quantitatively characterizes the window disturbance energy within the short-time window interval and effectively reflects the severity of transient anomalies in the voltage signal at the current moment.

[0037] Specifically, the window disturbance energy satisfies the following relationship: ; In the formula, Indicates the first The window perturbation energy at time t is the square mean of the instantaneous relative offset within the short-time statistical window, which represents the total perturbation energy within the window. Indicates the first Time shifts forward The instantaneous relative offset of the position; Indicates a short-term statistical window. In the formula, This indicates the total number of samples taken in the dynamic historical window. This indicates the window scaling factor. The value of needs to ensure that the short-time statistical window is short enough to capture transient abnormal pulses. For example... Grid fundamental frequency hour, , The corresponding time is fixed at approximately 0.5ms. This indicates the rounding function, and the update period of the short-term statistical window is consistent with the total number of samples in the dynamic historical window; This represents the relative offset index within the short-time window, with a value range of... , representing the first Based on the time frame, backtracking forward within a short time window Each sampling interval.

[0038] To distinguish between random background noise and genuine arc anomaly pulses from the perspective of signal timing characteristics, and to compensate for the insufficient accuracy of single energy feature identification, this invention conducts timing correlation analysis based on the instantaneous relative offset timing sequence. A correlation energy characterizing the degree of timing correlation is calculated using a short-time statistical window. The short-time statistical window used in this step is the same adaptive window used in the aforementioned local instantaneous variance calculation. The number of sampling times in the short-time statistical window is obtained by scaling and rounding the number of sampling times in the adaptive historical window by a fixed ratio. This allows for synchronous adaptive updates following the dynamic changes in the fundamental frequency of the power grid, ensuring the consistency and effectiveness of timing correlation analysis under different power grid operating conditions.

[0039] The specific calculation process is as follows: Retrieve the complete instantaneous relative offset time-series data sequence stored in the system to determine the total number of sampling times corresponding to the short-term statistical window under the current operating condition; select any sampling time as the target sampling time, and extract continuous instantaneous relative offset data from the complete instantaneous relative offset time-series data, including the current target sampling time, with the number of sampling times equal to the total number of samples in the short-term statistical window. Iterate through all adjacent instantaneous relative offset data within the short-term statistical window, performing pairwise multiplication on each group of immediately preceding and following instantaneous relative offsets to obtain the correlation product value of each group of adjacent sampling times; accumulate the correlation product values ​​of all adjacent sampling times within the window to obtain the cumulative sum of the products corresponding to the short-term window; divide the cumulative sum of the products by the total number of sampling times in the short-term statistical window to finally obtain the correlation energy corresponding to the current target sampling time.

[0040] This correlation energy can accurately quantify the temporal correlation characteristics of voltage fluctuation signals within a short time window: conventional interference signals such as baseline drift and random noise generated by steady-state operation of the power grid have strong temporal continuity, high numerical correlation between adjacent sampling times, and corresponding stable and positive correlation energy values; while abnormal pulses generated by arc faults have the characteristics of suddenness, randomness, and irregularity, with extremely weak signal correlation between adjacent sampling times, and correlation energy values ​​approaching zero or even negative values, which can form significant feature differences with conventional background noise, providing a core discrimination basis for subsequent multi-feature fusion and accurate identification of abnormal pulses and background noise.

[0041] Specifically, the correlation energy satisfies the following relationship: ; In the formula, This represents the correlation energy between adjacent sampling times. Normal physical signals (such as the fundamental residual trend or derivative noise) have strong correlations between adjacent points, with the product being positive and large in value; while true abnormal arc pulses are random, with weak or even negative correlations between adjacent points, and the product tending to be zero or negative. Indicates a short-term statistical window. Indicates the first Time shifts forward The instantaneous relative offset of the position. Indicates the first Time shifts forward The instantaneous relative offset of the two bits are adjacent samples in time and are multiplied together to participate in the correlation calculation.

[0042] After acquiring the local instantaneous variance and correlation energy within the short-time window, in order to further remove steady-state background interference, highlight the true arc abnormal pulse characteristics, and achieve accurate differentiation between noise and fault signals, this invention constructs highly discriminative stability features through differential fusion, completing the effective coupling of two-dimensional features. The specific fusion calculation formula is as follows: In the formula, Indicates the first The stability characteristics corresponding to each moment represent the remaining energy of the real abnormal pulse after removing background temporal correlation interference; Indicates the first The local instantaneous variance at time t is used to characterize the total disturbance energy intensity of the signal within a short time window; It represents the correlation energy between adjacent sampling times, used to characterize the temporal correlation of signals within a short time window.

[0043] It should be noted that for random abnormal pulses caused by arc faults, the signal timing is disordered, the values ​​at adjacent sampling times have no stable correlation, and the correlation energy... The values ​​are extremely small or even negative; while the local instantaneous variance Because the concentration of abnormal pulse energy will increase significantly, the stability feature obtained by subtracting the two will further amplify the abnormal amplitude, and the value will be significantly greater than zero, which can accurately highlight the real abnormal disturbance component.

[0044] This differential fusion operation has a clear physical discrimination mechanism: For conventional background interference signals such as periodic residual components and derivative noise under steady-state operation of the power grid, the signal time sequence is highly continuous and the correlation between adjacent data is high, so that the local instantaneous variance value and the correlation energy value are approximately equal, and the stability feature obtained by the difference between the two approaches zero, which can effectively suppress background interference; For random abnormal pulses caused by arc faults, the pulse signal has the characteristics of suddenness, non-periodicity and temporal disorder, and the signal time sequence correlation is extremely weak, and the corresponding correlation energy value is extremely small or even negative, while the local instantaneous variance will increase significantly due to the concentration of abnormal pulse energy, so that the calculated stability feature value is significantly greater than zero, which can accurately highlight the real abnormal disturbance components.

[0045] Through the aforementioned variance-correlation differential cancellation mechanism, there is no need to perform complex frequency domain analytical operations such as Fourier transform and wavelet transform. By simply fusing time domain features, it is possible to accurately separate strongly correlated steady-state background noise and residual periodic components, while fully preserving the unique random pulse characteristics of arc faults. This fundamentally distinguishes between conventional noise interference and real power grid anomalies, significantly improving the output signal-to-noise ratio and recognition accuracy of abnormal pulse signals. It provides highly reliable core discrimination features for subsequent high-precision anomaly interval determination, noise filtering, and anomaly alarm.

[0046] The stability features obtained through the above steps have effectively separated strongly correlated background noise from abnormal arc pulses, significantly improving the feature recognition of abnormal disturbances and accurately highlighting the true abnormal fluctuation components in the voltage signal. However, the actual operating conditions of the power grid are complex and variable, and fixed discrimination thresholds are difficult to adapt to baseline fluctuations and noise levels under different loads and environments, easily leading to poor generalization of anomaly identification. At the same time, instantaneous single-point feature mutations are prone to generating scattered pseudo-anomalies, causing fragmentation of abnormal intervals and alarm redundancy, making it impossible to directly output accurate and reliable anomaly results. To overcome the problems of poor adaptability of fixed thresholds and the accuracy of identification due to scattered noise interference, this invention adaptively matches the static noise discrimination threshold based on the historical normal operating data of the equipment, and combines interval optimization strategies to filter out pseudo-anomalies and regularize the true abnormal intervals, ultimately generating standardized anomaly alarm information. The specific steps are as follows: S4: Adaptively set the noise threshold based on the normal operating data of the equipment, identify abnormal intervals based on stability characteristics, merge and filter out noise from abnormal intervals, and output abnormal alarm information.

[0047] It should be noted that the squelch threshold is calculated by multiplying the statistical maximum value of all stability features by a fixed multiplier of 2 based on the steady-state normal operation data of the device in the first second after power-on. Abnormal intervals are identified based on stability features, and these intervals are merged and noise filtered out to output abnormal alarm information.

[0048] After solving the stability features at each sampling time step, edge transition detection is performed on the stability features under all time series based on a preset noise threshold. By distinguishing the rising and falling edge timing positions of the stability features crossing the noise threshold, the boundary of each voltage abnormal fluctuation is accurately identified, and the original abnormal interval is completely divided. The specific edge recognition and interval division execution logic is as follows: Select any sampling time within the time series as the target sampling time. Retrieve target sampling time Corresponding stability features At the same time, the time sequence position is retrieved earlier than the target sampling time. Stability characteristics corresponding to the preceding sampling time of one bit stability characteristics Stability characteristics Each with a pre-adaptively configured noise threshold By comparing the numerical values, the signal edge type is distinguished based on the combination relationship between the two sets of comparison results, and the starting and ending boundaries of the abnormal interval are located and marked.

[0049] When the first set of numerical judgment conditions is met: target sampling time Corresponding stability characteristics The value is greater than the noise threshold. And the stability features corresponding to the preceding sampling time. The value is less than or equal to the squelch threshold. When the signal stability characteristic jumps from the steady-state reference interval to the anomaly detection interval, this jump is considered an anomaly rising edge; the current target sampling time is... The starting index of an independent voltage fluctuation range is determined, and the sampling time is... Add an exception start marker.

[0050] Within the continuous time series range where the abnormal starting index has been marked, the comparison and determination of the above dual features and the noise threshold are continuously performed at each sampling time; when the second set of numerical determination conditions appears: target sampling time Corresponding stability characteristics The value is less than or equal to the squelch threshold. And the stability features corresponding to the preceding sampling time. The value is greater than the noise threshold. When the signal stability characteristic returns from the anomaly detection interval to the steady-state reference interval, this transition is considered an abnormal falling edge; the current target sampling time... The end index of the currently activated abnormal voltage fluctuation range is determined, and the sampling time is... Add an exception end marker.

[0051] Further supplementary boundary condition handling rules: If the first sampling time at the start of the timing sequence satisfies... The first sampling time is directly marked as the abnormal start index; if the end position of the time series is still in In the event of an abnormal state or a missing falling edge, the last sampling time of the time series will be forcibly marked as the end index of the abnormal interval to ensure that all abnormal intervals are completely closed.

[0052] For each pair of abnormal start and end indices that appear one-to-one in the time series, all continuous sampling moments between the abnormal start and end indices of that group are extracted. All sampling moments in this interval are uniformly marked as voltage abnormal fluctuation sampling moments. Through the above-mentioned rising edge and falling edge pairing method, multiple sets of original abnormal intervals that are independent of each other in the time series and do not overlap are completely divided. Each set of original abnormal intervals completely covers all sampling data from the appearance to the disappearance of a single voltage disturbance.

[0053] In real-world power grid operation, conditions such as motor switching and sudden load changes generate multiple short-duration disturbance pulses. Direct judgment of these pulses would result in a large number of fragmented abnormal intervals that are separate and have very short intervals, causing redundant and chaotic alarm information. To address this, this invention adds a logic for merging adjacent abnormal intervals: a pre-configured interval merging duration threshold is set, with 5 milliseconds selected as the merging criterion in this embodiment; the time interval between any two independent original abnormal intervals is defined as the duration between the end index of the first original abnormal interval and the start index of the second original abnormal interval; if this duration is less than the preset merging duration threshold, the two original abnormal intervals are merged into a single complete and continuous abnormal event, eliminating fragmented alarms.

[0054] After completing the interval merging operation, isolated noise interval filtering is then performed: the interval merging duration threshold is pre-configured, and in this embodiment, 5 milliseconds is selected as the merging judgment standard. This duration is less than 10 milliseconds, which is half the power frequency cycle, and can merge continuous adjacent short-time pulses generated by load switching; the minimum effective anomaly duration threshold is set to half the fundamental frequency cycle of the power grid, i.e., 10 milliseconds, based on the fact that the duration of real arc and short-circuit fault disturbances will exceed half the power frequency cycle, and signals shorter than this duration are only instantaneous noise spikes; real power anomalies such as arc faults and line short circuits have the physical characteristic of continuous existence, and the disturbance duration generally exceeds half the fundamental frequency cycle of the power grid, so they can be completely retained.

[0055] After completing the optimization and screening of all intervals, a standardized abnormal alarm record is generated for each valid abnormal event that is retained. The alarm record fully stores three types of core information: the sampling timestamp corresponding to the first occurrence of the abnormal event, the overall duration of the abnormal event, and the original three-phase voltage waveform segment corresponding to the abnormal interval. Retaining the original waveform segment facilitates maintenance personnel to conduct post-event fault tracing and waveform reproduction analysis.

[0056] The above operation steps use threshold edge transition detection instead of single-point threshold judgment to accurately lock the complete boundary of the anomaly and avoid false triggering by steady-state small-amplitude noise; through the short-distance anomaly interval merging strategy, the continuous short-time pulses brought by load switching are integrated to reduce repeated redundant alarms; instantaneous noise spikes are eliminated based on the minimum duration threshold, and random interference is distinguished from real arc and short-circuit faults from the time dimension; the entire set of discrimination criteria is adaptively generated based on the normal operation data of the equipment itself, without the need for manual adjustment of parameters for different sites, and the robustness and adaptability of the algorithm are significantly improved.

[0057] This invention also provides a system for quickly locating abnormal fluctuations in electricity meter data. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a method for rapidly locating abnormal fluctuations in electricity meter data according to the first aspect of the present invention. The system also includes other components well-known to those skilled in the art, such as a communication bus and a communication interface. Their configuration and functions are known in the art and will not be described further here.

[0058] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for rapidly locating abnormal fluctuations in electricity meter data, characterized in that, include: Acquire three-phase voltage data and simultaneously calculate the fundamental frequency of the power grid. Generate a one-dimensional signal's three-phase instantaneous projection sum based on the preprocessed three-phase instantaneous values. A half-cycle sliding window containing only historical sampling times is adaptively constructed based on the fundamental frequency of the power grid. A historical mean benchmark of the three-phase instantaneous projection sum is built based on the half-cycle sliding window, and the instantaneous relative offset is calculated by combining the three-phase instantaneous projection sum. The calculation of window perturbation energy and correlation energy within a short-term statistical window for the instantaneous relative offsets includes: obtaining the mean square of the instantaneous relative offset sequence using a short-term statistical window to obtain the local instantaneous variance characterizing the total perturbation energy within the window; the number of sampling times in the short-term statistical window is obtained by scaling and rounding the number of sampling times in the adaptive historical window by a fixed ratio; the short-term statistical window is dynamically updated synchronously with the adaptive historical window; when calculating the local instantaneous variance, all instantaneous relative offsets within the short-term statistical window, including the current target sampling time, are truncated; the squares of each instantaneous relative offset are calculated and summed; the summation result is divided by the number of sampling times in the short-term statistical window to obtain the local instantaneous variance at the corresponding time, which is used as the window perturbation energy; Retrieve all instantaneous relative offset data sequences under the complete time series to determine the total number of sampling times corresponding to the short-term statistical window; take any sampling time as the target sampling time, extract continuous instantaneous relative offsets from the instantaneous relative offset data sequence that include the current target sampling time and whose total number is equal to the number of sampling times in the short-term statistical window; traverse each group of adjacent instantaneous relative offsets within the window, multiply the two adjacent instantaneous relative offsets in each group pairwise to obtain the product value of each group of adjacent sampling times; sum the product values ​​of all adjacent points within the window to obtain the cumulative sum, and divide the cumulative sum by the total number of sampling times in the short-term statistical window to calculate the correlation energy of the current target sampling time; The window perturbation energy and correlation energy are fused to obtain stability features used to distinguish abnormal pulses from background noise; The noise threshold is adaptively set based on the normal operating data of the equipment. Abnormal intervals are identified based on stability characteristics. Abnormal intervals are merged and noise is filtered out, and abnormal alarm information is output.

2. The method for rapidly locating abnormal fluctuations in electricity meter data according to claim 1, characterized in that, The method of establishing a historical mean benchmark based on a half-cycle sliding window includes: Taking any sampling time as the target sampling time, calculate the total number of window samples corresponding to half a fundamental frequency cycle based on the meter sampling rate and the real-time fundamental frequency of the power grid, and round it down. Update the total number of window samples based on the real-time fundamental frequency of the power grid. In all three-phase instantaneous projections and data, extract the continuous historical three-phase instantaneous projections before the target sampling time that are consistent with the total number of samples in the window. Add all the extracted data to obtain the cumulative sum. Divide the cumulative sum by the total number of samples in the window to obtain the historical average benchmark at the target sampling time.

3. The method for rapidly locating abnormal fluctuations in electricity meter data according to claim 1, characterized in that, The instantaneous relative offset is calculated as follows: Taking any sampling time as the target sampling time, the difference between the three-phase instantaneous projection corresponding to the target sampling time and the historical average benchmark corresponding to the target sampling time is used as the instantaneous relative offset; the instantaneous relative offset is used to separate the power frequency steady-state base and the low-frequency drift.

4. The method for rapidly locating abnormal fluctuations in electricity meter data according to claim 1, characterized in that, The calculation method for the sum of the three-phase instantaneous projections is as follows: The instantaneous values ​​of the three-phase voltages A, B, and C, which are preprocessed at the same time, are directly added together to obtain the instantaneous projection sum of the three phases.

5. The method for rapidly locating abnormal fluctuations in electricity meter data according to claim 1, characterized in that, The preprocessing steps specifically include: Isolated pulse noise introduced by the analog-to-digital converter is filtered out by median filtering or moving average algorithm; the three-phase voltage signal is normalized to eliminate gain differences between multiple acquisition channels; and the fundamental frequency of the power grid is calculated in real time by phase-locked loop or zero-crossing detection algorithm.

6. A system for rapidly locating abnormal fluctuations in electricity meter data, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the method for rapidly locating abnormal fluctuations in electricity meter data according to any one of claims 1-5.

Images