A three-phase multifunctional electric energy meter anomaly detection method and system based on deep learning
By using deep learning-based voltage and current prediction models combined with multi-dimensional feature fusion technology, the problem of insufficient detection accuracy of three-phase multi-function energy meters in complex power grid environments has been solved, achieving efficient anomaly detection and automated early warning, and improving the stability and reliability of the power grid system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WUXI HENGTONG ELECTRIC CO LTD
- Filing Date
- 2025-08-25
- Publication Date
- 2026-07-03
AI Technical Summary
Existing anomaly detection methods for three-phase multi-function energy meters cannot effectively detect latent anomalies and anomaly propagation effects, and have high false alarm and false negative rates, making them unsuitable for complex power grid environments.
A deep learning-based approach is adopted to train voltage and current prediction models through long short-term memory networks. By combining three-phase voltage/current imbalance, frequency domain harmonic characteristics, and grid structure correlation, a multi-dimensional anomaly index system is constructed to dynamically assess the comprehensive anomaly degree and detect it through an automated early warning mechanism.
It significantly improves the ability to identify complex anomalies such as harmonic distortion, load imbalance and electricity theft, reduces false alarm and missed alarm rates, improves the real-time performance and accuracy of detection, reduces the cost of manual inspection, and enhances the stability and reliability of the power grid system.
Smart Images
Figure CN121027965B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing. More specifically, this invention relates to a method and system for detecting anomalies in a three-phase multi-function energy meter based on deep learning. Background Technology
[0002] A three-phase multifunction energy meter is a powerful energy metering device capable of measuring and monitoring various electrical parameters, including three-phase voltage, current, power, power factor, frequency, and harmonics. In practical applications, the three-phase four-wire energy meter is a common form of three-phase multifunction energy meter, suitable for industrial, commercial, and some large residential building scenarios.
[0003] Because three-phase four-wire electricity meters are connected to complex power networks, multiple meters may share the same distribution line, and their electricity consumption behaviors are correlated. Therefore, latent anomalies not only affect individual meters but may also propagate through the network topology, causing neighboring meters to exhibit abnormal patterns. Most existing anomaly detection methods rely on the voltage or current thresholds of individual meters, neglecting changes in harmonic energy and three-phase imbalance, thus failing to effectively detect latent anomalies and their propagation effects. Summary of the Invention
[0004] To address the limitations of existing anomaly detection methods and their difficulty in effectively detecting anomalies in electricity meters, this invention provides solutions in the following aspects.
[0005] In the first aspect, a deep learning-based method for anomaly detection in a three-phase multi-function energy meter includes:
[0006] Select the energy meter to be tested, and input its historical voltage and current time series into the trained voltage prediction model and current prediction model respectively. Output the predicted voltage and current time series. The voltage time series includes the phase voltage time series on each phase and the voltage time series on the neutral line, and the current time series includes the phase current time series on each phase and the current time series on the neutral line.
[0007] The three-phase voltage imbalance is calculated at each moment of the historical and predicted voltage time series. Fourier transforms are performed on the historical and predicted voltage time series to calculate the voltage frequency domain anomaly of the energy meter under test. Based on the voltage frequency domain anomaly and all three-phase voltage imbalances, the voltage anomaly of the energy meter under test is calculated. The current anomaly of the energy meter under test is calculated according to the voltage anomaly calculation method. Finally, the comprehensive anomaly of the energy meter under test is obtained based on the voltage and current anomalies.
[0008] When the overall abnormality level of the electricity meter under test exceeds the preset abnormality threshold, an abnormality warning for the electricity meter is triggered.
[0009] Preferably, the training process of the voltage prediction model includes:
[0010] The voltage time series of all electricity meters in the same area are collected during historical periods. The collected voltage time series are used as the training set to train the voltage prediction model. The model type is selected as Long Short-Term Memory Network, and the mean square error is selected as the loss function.
[0011] The training set is divided into input formats of different lengths. For each input format, the model generates a predicted voltage time series of the same length based on the input of that training set.
[0012] Preferably, the training process of the current prediction model is the same as that of the voltage prediction model.
[0013] Preferably, after calculating the three-phase voltage imbalance at each moment of the historical voltage time series and the predicted voltage time series, the method further includes:
[0014] Obtain the current imbalance at each moment of the predicted current time series under any input format; determine the nearest and farthest energy meters in the power distribution network structure relative to the energy meter under test.
[0015] For the predicted three-phase voltage imbalance under any input format, calculate the average Euclidean distance between the energy meter under test and the neighboring energy meter and the farthest energy meter respectively, and use the ratio of the average Euclidean distance between the energy meter under test and the farthest energy meter to the average Euclidean distance between the energy meter under test and the neighboring energy meter as the first parameter.
[0016] For the predicted current imbalance, the mean Euclidean distance between the energy meter under test and the neighboring energy meter and the farthest energy meter is calculated respectively. The ratio of the mean Euclidean distance between the energy meter under test and the farthest energy meter to the mean Euclidean distance between the energy meter under test and the neighboring energy meter is used as the second parameter.
[0017] The degree of structural abnormality of the energy meter under test is obtained by adding the first parameter and the second parameter to the corresponding input format.
[0018] Preferably, the process of obtaining the voltage frequency domain anomaly degree includes:
[0019] For the historical voltage time series of the energy meter under test under any input format, calculate the ratio of the total energy of odd harmonics to the fundamental energy as the third parameter; for the predicted voltage time series of the energy meter under test under any input format, calculate the ratio of the total energy of odd harmonics to the fundamental energy as the fourth parameter, calculate the difference between the third and fourth parameters, and calculate the exponent of the difference. Sum the exponents of the differences of all voltage time series and average them to obtain the voltage frequency domain anomaly degree under the corresponding input format.
[0020] Preferably, the process of obtaining the degree of voltage anomaly of the energy meter under test includes:
[0021] Construct a historical unbalance sequence for all three-phase voltage unbalances of the energy meter under test under any input format, and construct a predicted unbalance sequence for all three-phase voltage unbalances of the predicted voltage time series of the energy meter under test under any input format.
[0022] Calculate the Euclidean distance between the historical imbalance sequence and the predicted imbalance sequence;
[0023] The product of the voltage frequency domain anomaly and the Euclidean distance is used as the voltage anomaly level under the corresponding input format.
[0024] Preferably, after obtaining the degree of voltage anomaly and the degree of current anomaly, the method further includes:
[0025] The voltage anomaly level and current anomaly level calculated based on any input format are added together to obtain the characteristic anomaly level of the energy meter under the corresponding input format.
[0026] Preferably, the process of obtaining the overall anomaly level includes:
[0027] The product of the structural anomaly degree and the feature anomaly degree under any input format is taken as the comprehensive anomaly degree under the corresponding input format;
[0028] The average of the overall anomaly levels of all input formats is taken as the final overall anomaly level of the energy meter under test.
[0029] Secondly, a deep learning-based three-phase multi-function energy meter anomaly detection system 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 deep learning-based three-phase multi-function energy meter anomaly detection method described in any one of the claims is implemented.
[0030] The beneficial effects of this invention are:
[0031] This invention introduces deep learning models and multi-dimensional feature fusion technology to construct an anomaly detection method for three-phase multi-functional energy meters based on voltage-prediction and current-prediction models. This effectively solves the problems of insufficient detection accuracy, high false alarm rate, and high missed alarm rate of traditional methods in complex power grid environments. By modeling voltage and current time series using a long short-term memory network, and combining time-domain imbalance analysis (three-phase voltage / current imbalance), frequency-domain harmonic characteristics (odd harmonic energy ratio), and power grid structural correlation (mean ratio of Euclidean distance between adjacent and distant energy meters), a multi-dimensional anomaly index system is constructed, significantly improving the ability to identify complex anomalies such as harmonic distortion, load imbalance, and electricity theft. Simultaneously, by dynamically adapting the input format to time series data of different lengths and combining the product calculation of structural anomaly degree and feature anomaly degree, dynamic evaluation of the comprehensive anomaly degree is achieved, effectively reducing false alarm and missed alarm rates. Furthermore, it supports adaptive threshold determination to adapt to the operating characteristics of power grids in different regions. Furthermore, this method, through automated early warning mechanisms and multimodal data fusion, not only improves the real-time performance and accuracy of anomaly detection, but also provides maintenance personnel with comprehensive diagnostic information, thereby reducing the cost of manual inspections and enhancing the stability and reliability of the power grid system. Attached Figure Description
[0032] Figure 1 This is a flowchart of steps S1-S3 in a method for detecting anomalies in a three-phase multi-functional energy meter based on deep learning, according to an embodiment of the present invention. Detailed Implementation
[0033] 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.
[0034] Reference Figure 1 A method for detecting anomalies in a three-phase multi-function energy meter based on deep learning includes steps S1-S3, as detailed below:
[0035] S1: Select the energy meter to be tested, input its historical voltage and current time series into the trained voltage prediction model and current prediction model respectively, and output the predicted voltage and current time series; the voltage time series includes the phase voltage time series on each phase and the voltage time series on the neutral line, and the current time series includes the phase current time series on each phase and the current time series on the neutral line.
[0036] Time series forecasting models such as LSTM (Long Short-Term Memory) can simulate the voltage or current variation patterns of electricity meters under normal conditions by learning the dependencies in historical data.
[0037] In one embodiment, a data acquisition terminal is used to collect phase voltage and phase current data for each phase of a three-phase multi-function energy meter, as well as voltage and current data on the neutral line. Data from all three-phase multi-function energy meters in the same area are collected, resulting in phase voltage and phase current sequences for each phase of each energy meter, and neutral line voltage and current sequences for each energy meter. The data collection frequency is set to once per minute, collecting data from the past 100 days.
[0038] The collected data will be further preprocessed, including denoising (frequency domain denoising or time domain smoothing) and data normalization (eliminating physical dimensions).
[0039] The collected phase voltage and neutral voltage sequences are used as training sets to train the voltage prediction model, typically choosing the mean square error as the loss function. Similarly, the phase current and neutral current sequences are used to train the current prediction model, with the selection of the current prediction model and loss function consistent with that of the voltage prediction model.
[0040] Furthermore, during model training, the input training set is divided into input formats with different time lengths. For example, the training set can be divided into several data segments with a time length of 10, or it can be divided into several data segments with a time length of 20, etc.
[0041] For each input format of the training set, the model generates a predicted output time series of the same length based on the input of that training set.
[0042] For example, when the input time length is 10, the model will generate predicted voltage and current sequences for phases A, B, and C, respectively, and predicted voltage and current sequences for the neutral line, each with a length of 10. When the input time length is 20, the model will generate predicted voltage and current sequences for phases A, B, and C, respectively, and predicted voltage and current sequences for the neutral line, each with a length of 20. This allows the model to be trained to process three-phase data and neutral line data using input formats of different time lengths, enabling it to adapt to the changing patterns of three-phase data at different time scales.
[0043] During model training, at each time point, the model comprehensively considers the data characteristics of three-phase voltage, three-phase current, and neutral line voltage and current. By adjusting the model parameters, the mean square error between the predicted voltage and current sequences output by the model and the actual three-phase voltage and current sequences is minimized. To ensure model performance, methods such as cross-validation can be used to evaluate the model's accuracy and reliability in processing three-phase data. After training is complete, the model is validated using a test set, which should contain actual voltage and current data for the three phases and the neutral line.
[0044] Thus, the trained voltage prediction model and current prediction model are obtained, that is, the voltage prediction model and current prediction model with different input time lengths are obtained.
[0045] Furthermore, select the energy meter to be tested, and collect a series of data from the energy meter in the same way as the above-mentioned data collection operations, that is, obtain the historical phase voltage sequence on each phase and the voltage timing sequence on the neutral line of the energy meter to be tested, as well as the phase current timing sequence on each phase and the current timing sequence on the neutral line.
[0046] Furthermore, for voltage prediction models and current prediction models with a single input format, all historical voltage and current time series of the above-mentioned energy meter under test are input into the corresponding voltage prediction model and current prediction model respectively, and the predicted voltage and current time series with the same input format are output.
[0047] S2: Calculate the three-phase voltage imbalance at each moment of the historical voltage time series and the predicted voltage time series respectively; perform Fourier transform on the historical voltage time series and the predicted voltage time series and calculate the voltage frequency domain anomaly degree of the energy meter under test; calculate the voltage anomaly degree of the energy meter under test based on the voltage frequency domain anomaly degree and all three-phase voltage imbalance degrees; calculate the current anomaly degree of the energy meter under test according to the calculation method of voltage anomaly degree; obtain the comprehensive anomaly degree of the energy meter under test based on the voltage anomaly degree and the current anomaly degree.
[0048] Electricity meters typically calculate electrical energy by measuring the product of voltage and current. When the voltage is unbalanced, the amplitude and phase of the voltage may deviate, resulting in inaccurate voltage signals measured by the electricity meter, which in turn affects the accuracy of the energy calculation. Similarly, current imbalance will cause deviations in the amplitude and phase of the current signal, affecting the measurement accuracy of the electricity meter.
[0049] By calculating the three-phase voltage imbalance of the energy meter under test, the degree of asymmetry in the amplitude or phase of the three-phase voltage can be measured. The higher the value, the worse the voltage quality, which may be caused by load imbalance, fault, or electricity theft. Similarly, by calculating the current imbalance of the energy meter under test, the degree of asymmetry in the three-phase current can be measured, reflecting the user's load characteristics (such as the connection of single-phase high-power equipment).
[0050] Taking the predicted voltage and current time series of the energy meter under test output by the prediction model in any input format as an example, in one embodiment, the maximum value of the difference between the root mean square value of the three-phase phase voltage and the average value of the root mean square value of the three-phase phase voltage at each predicted moment is calculated, and then divided by the ratio of the average value of the root mean square value of the three-phase phase voltage to the voltage imbalance of the energy meter under test at a single moment. Similarly, the current imbalance of the energy meter under test at a single moment is calculated.
[0051] Under normal circumstances, the imbalance of nearby electricity meters should be similar to that of the meter under test, while the imbalance of the farthest electricity meter should be significantly different. If the outlier deviates significantly from this pattern, it may indicate a fault in the network structure or data anomalies.
[0052] Taking the predicted voltage and current timing sequences of the energy meter under test as an example, using the prediction model output of any input format, in one embodiment, the neighboring energy meters and the farthest energy meter of the energy meter under test are first determined. For example, the energy meters that are directly connected to the energy meter under test or have the fewest hops in the distribution network topology (such as a tree structure) are identified as the neighboring energy meters of the energy meter under test, and the energy meter that is topologically farthest from the energy meter under test in the network is identified as the farthest energy meter.
[0053] Then, for the voltage imbalance at all times in the predicted voltage time series, the voltage imbalance at each time moment is treated as a vector, and the mean Euclidean distance between the energy meter under test and the neighboring energy meters is calculated. The ratio of the mean Euclidean distance between the energy meter under test and the farthest energy meter to the mean Euclidean distance between the energy meter under test and the neighboring energy meters is used as the first parameter. Similarly, for the current imbalance at all times in the predicted current time series, the current imbalance at each time moment is treated as a vector, and the mean Euclidean distance between the energy meter under test and the neighboring energy meters is calculated. The ratio of the mean Euclidean distance between the energy meter under test and the farthest energy meter to the mean Euclidean distance between the energy meter under test and the neighboring energy meters is used as the second parameter.
[0054] Finally, the first parameter and the second parameter are added together to obtain the degree of structural abnormality of the energy meter under the corresponding input format.
[0055] In power systems, the voltage and current imbalance of electricity meters is closely related to the network structure. When the network structure is normal, the voltage and current imbalances of neighboring electricity meters are usually similar, and the Euclidean distance between them is relatively small. However, due to the increased topological distance, the imbalance of the furthest electricity meter will be larger, and the Euclidean distance between it and its neighbors will be larger. This difference can be quantified by calculating the first and second parameters. If these parameters deviate significantly from the normal range in a measurement, such as an abnormally large sum of the first and second parameters, it may indicate a fault in the network structure, such as incorrect line connections or improper meter installation, thus helping staff quickly locate the fault.
[0056] In addition, normal voltage / current in a power system should be a pure sine wave (50 / 60Hz fundamental frequency). Abnormal events (such as electricity theft, nonlinear load faults) can introduce harmonics, especially odd harmonics (3rd, 5th, 7th, etc.). In three-phase systems, even harmonics are usually caused by severe asymmetrical faults (very rare), while odd harmonics are more common and strongly correlated with electricity theft; therefore, the energy of odd harmonics also needs to be considered.
[0057] Since noise and transient fluctuations in time-domain waveforms (such as voltage curves) can mask harmonic characteristics, frequency domain conversion can clearly separate the fundamental and harmonic components in the energy spectrum, making anomalous characteristics quantifiable.
[0058] Taking the predicted voltage timing of the energy meter under test as an example, with any input format of the prediction model output, in one embodiment, firstly, the historical voltage timing and the predicted voltage timing of the energy meter under test are subjected to Fourier transform, and the features are extracted, including odd harmonics and fundamental energy.
[0059] Then, for the historical voltage time series of the energy meter under test, the ratio of its total odd harmonic energy to the fundamental energy is calculated as the third parameter; for the predicted voltage time series of the energy meter under test, the ratio of its total odd harmonic energy to the fundamental energy is calculated as the fourth parameter.
[0060] Finally, the difference between the third and fourth parameters is calculated, and the exponent of the difference is calculated. The exponents of the differences for all voltage timing sequences are summed and averaged to obtain the voltage frequency domain anomaly degree under the corresponding input format.
[0061] It should be noted that if only frequency domain analysis is used to analyze the anomalies of the energy meter under test, the dynamic changes in the time domain may be ignored.
[0062] Furthermore, by constructing a historical unbalance sequence of all three-phase voltage imbalances in the historical voltage time series of the energy meter under test, and constructing a predicted unbalance sequence of all three-phase voltage imbalances in the predicted voltage time series of the energy meter under test under a single input format;
[0063] The Euclidean distance between the historical imbalance series and the predicted imbalance series is calculated to quantify the difference between the predicted series and the series in terms of the trend of imbalance.
[0064] Further combining frequency domain and time domain characteristics, the product of the voltage frequency domain anomaly degree and the Euclidean distance between the historical imbalance degree sequence and the predicted imbalance degree sequence is used as the voltage anomaly degree of the energy meter under the corresponding input format.
[0065] The degree of current anomaly can be calculated similarly based on the degree of voltage anomaly described above.
[0066] After obtaining the voltage and current abnormality levels of the energy meter under test, the voltage and current abnormality levels are added together to obtain the characteristic abnormality level of the energy meter under test in the corresponding input format.
[0067] Finally, the product of the structural anomaly degree and the feature anomaly degree corresponding to the single input format calculated above is taken as the comprehensive anomaly degree under the corresponding input format.
[0068] To elaborate, the aforementioned structural anomaly levels reflect the rationality of the power grid topology. If the imbalance between the meter under test and its neighboring meters is small, but the difference with the furthest meter is large, the first and second parameters will approach 0, indicating a normal structure. If the first and second parameters increase significantly, it may indicate local load imbalance, line faults, or topology errors. The characteristic anomaly level reflects the meter's own state, quantifying its own anomalies based on the frequency and time domain characteristics of voltage or current. Combining the structural and characteristic anomaly levels, the overall anomaly level is only significantly high when both are large (disconnected from neighboring meters), allowing for precise identification of individual anomalies and filtering of regional events.
[0069] The average of the overall anomaly levels of all input formats is taken as the final overall anomaly level of the energy meter under test.
[0070] S3: When the overall abnormality level of the energy meter under test exceeds the preset abnormality threshold, an energy meter abnormality warning is triggered.
[0071] After calculating the overall abnormality level of the electricity meter to be tested in the above S2, historical electricity meter data (normal electricity meters and abnormal electricity meters) are collected, and the distribution of its overall abnormality index is calculated. The 95th percentile of the normal sample is taken as the abnormality threshold to ensure that 95% of the meters are not falsely reported.
[0072] Furthermore, if the overall abnormality level of the energy meter under test exceeds the above-mentioned abnormality threshold, the energy meter is determined to be an abnormal energy meter, and an early warning signal is issued.
[0073] The system includes a processor and a memory, the memory storing computer program instructions. When the computer program instructions are executed by the processor, the method for detecting anomalies in a three-phase multi-function energy meter based on deep learning according to the first aspect of the present invention is implemented.
[0074] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.
[0075] 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 anomaly detection in a three-phase multi-functional energy meter based on deep learning, characterized in that, include: Select the energy meter to be tested, and input its historical voltage and current time series into the trained voltage prediction model and current prediction model respectively. Output the predicted voltage and current time series. The voltage time series includes the phase voltage time series on each phase and the voltage time series on the neutral line, and the current time series includes the phase current time series on each phase and the current time series on the neutral line. The three-phase voltage imbalance is calculated at each moment of the historical and predicted voltage time series. Fourier transforms are performed on the historical and predicted voltage time series to calculate the voltage frequency domain anomaly of the energy meter under test. Based on the voltage frequency domain anomaly and all three-phase voltage imbalances, the voltage anomaly of the energy meter under test is calculated. The current anomaly of the energy meter under test is calculated according to the voltage anomaly calculation method. Finally, the comprehensive anomaly of the energy meter under test is obtained based on the voltage and current anomalies. When the overall abnormality level of the electricity meter under test exceeds the preset abnormality threshold, an abnormality warning for the electricity meter is triggered.
2. The method for anomaly detection of a three-phase multi-functional energy meter based on deep learning according to claim 1, characterized in that, The training process of the voltage prediction model includes: The voltage time series of all electricity meters in the same area are collected during historical periods. The collected voltage time series are used as the training set to train the voltage prediction model. The model type is selected as Long Short-Term Memory Network, and the mean square error is selected as the loss function. The training set is divided into input formats of different lengths. For each input format, the model generates a predicted voltage time series of the same length based on the input of that training set.
3. The method for detecting anomalies in a three-phase multi-functional energy meter based on deep learning according to claim 2, characterized in that, The training process for the current prediction model is the same as that for the voltage prediction model.
4. The method for anomaly detection of a three-phase multi-functional energy meter based on deep learning according to claim 2, characterized in that, After calculating the three-phase voltage imbalance at each moment of the historical and predicted voltage time series, the following is also included: Obtain the current imbalance at each moment of the predicted current time series under any input format; determine the nearest and farthest energy meters in the power distribution network structure relative to the energy meter under test. For the predicted three-phase voltage imbalance under any input format, calculate the average Euclidean distance between the energy meter under test and the neighboring energy meter and the farthest energy meter respectively, and use the ratio of the average Euclidean distance between the energy meter under test and the farthest energy meter to the average Euclidean distance between the energy meter under test and the neighboring energy meter as the first parameter. For the predicted current imbalance, the mean Euclidean distance between the energy meter under test and the neighboring energy meter and the farthest energy meter is calculated respectively. The ratio of the mean Euclidean distance between the energy meter under test and the farthest energy meter to the mean Euclidean distance between the energy meter under test and the neighboring energy meter is used as the second parameter. The degree of structural abnormality of the energy meter under test is obtained by adding the first parameter and the second parameter to the corresponding input format.
5. The method for anomaly detection of a three-phase multi-functional energy meter based on deep learning according to claim 4, characterized in that, The process of obtaining the voltage frequency domain anomaly level includes: For the historical voltage time series of the energy meter under test under any input format, calculate the ratio of the total energy of odd harmonics to the fundamental energy as the third parameter; for the predicted voltage time series of the energy meter under test under any input format, calculate the ratio of the total energy of odd harmonics to the fundamental energy as the fourth parameter, calculate the difference between the third and fourth parameters, and calculate the exponent of the difference. Sum the exponents of the differences of all voltage time series and average them to obtain the voltage frequency domain anomaly degree under the corresponding input format.
6. The method for anomaly detection of a three-phase multi-functional energy meter based on deep learning according to claim 5, characterized in that, The process of obtaining the degree of voltage anomaly of the energy meter under test includes: Construct a historical unbalance sequence for all three-phase voltage unbalances of the energy meter under test under any input format, and construct a predicted unbalance sequence for all three-phase voltage unbalances of the predicted voltage time series of the energy meter under test under any input format. Calculate the Euclidean distance between the historical imbalance sequence and the predicted imbalance sequence; The product of the voltage frequency domain anomaly and the Euclidean distance is used as the voltage anomaly level under the corresponding input format.
7. The method for anomaly detection of a three-phase multi-functional energy meter based on deep learning according to claim 6, characterized in that, After obtaining the degree of voltage and current anomalies, the following is also included: The voltage anomaly level and current anomaly level calculated based on any input format are added together to obtain the characteristic anomaly level of the energy meter under the corresponding input format.
8. The method for anomaly detection of a three-phase multi-functional energy meter based on deep learning according to claim 6, characterized in that, The process of obtaining the overall anomaly level includes: The product of the structural anomaly degree and the feature anomaly degree under any input format is taken as the comprehensive anomaly degree under the corresponding input format; The average of the overall anomaly levels of all input formats is taken as the final overall anomaly level of the energy meter under test.
9. A three-phase multi-functional energy meter anomaly detection system based on deep learning, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the deep learning-based anomaly detection method for a three-phase multi-function energy meter according to any one of claims 1-8.