A method and system for remote power monitoring for a power meter

Abstract

The present application relates to the technical field of data processing, and particularly relates to a power monitoring method and system for remote electric energy meter, the method comprising: constructing a time series factor graph model with voltage and current phasor as hidden variables and original electric parameter data as observation nodes; performing synchronous compression wavelet transform on current phasors in the electric energy state sequence to generate a time-frequency energy distribution graph and obtain a monitoring feature vector sequence; inputting the monitoring feature vector sequence into a deep auto-encoder pre-trained on normal power consumption working condition data to calculate a reconstruction error, simultaneously calculating a Lyapunov index of the sequence within a preset time window, and obtaining a negative log-likelihood probability according to a pre-established Gaussian mixture model describing the distribution of the index under normal working condition; and weighting and fusing the reconstruction error and the negative log-likelihood probability to generate a comprehensive abnormality index for judging power consumption events. The present application can realize high-precision and low-false-alarm-rate detection of power consumption events.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and more particularly to a method and system for monitoring electric energy of a remote electric energy meter. Background Art

[0002] With the rapid development of smart grid technology, remote energy meters, as key terminal devices, collect energy data that is fundamental to refined load management, electricity theft diagnosis, and equipment status assessment. Existing energy monitoring systems typically periodically upload meter parameters such as voltage, current, and power to a master server via wireless communication networks (such as GPRS and NB-IoT). However, data transmission is susceptible to network fluctuations and signal interference, leading to packet loss or corruption, resulting in incomplete time series data, which severely impacts the accuracy and reliability of subsequent analysis algorithms. Traditional energy monitoring methods often rely on setting fixed thresholds for a single electrical parameter (such as current or power), triggering alarms when the measured value exceeds the threshold. While simple, this approach has limited ability to recognize complex, nonlinear patterns of electricity usage, making it difficult to distinguish between normal load on / off fluctuations and true abnormal events. This leads to a high incidence of false positives and missed positives, and fails to meet the high monitoring accuracy and intelligence requirements of modern power grids.

[0003] To improve monitoring accuracy, existing technologies are beginning to incorporate signal processing and machine learning techniques. For example, some methods use Fourier transforms or traditional wavelet transforms to analyze current signals to extract harmonic features and identify nonlinear loads or faults. However, Fourier transforms are not suitable for analyzing time-varying, non-stationary signals, while traditional wavelet transforms suffer from energy leakage and limited time-frequency resolution when processing multi-component coupled signals, making it difficult to accurately capture the transient characteristics of events. Regarding anomaly detection models, some studies have employed supervised learning-based classifiers. However, this requires a large number of labeled samples of normal and various abnormal types. In practical applications, data acquisition is expensive and coverage of all unknown anomalies is difficult. Unsupervised methods, such as principal component analysis (PCA) and isolation forest algorithms, while not requiring anomaly labels, often use a single feature extraction dimension, are insensitive to early, weak anomalies, and require improved model robustness. Therefore, effectively addressing the issue of missing data, extracting more discerning dynamic features from high-dimensional, non-stationary power data, and combining the strengths of multiple models to achieve high-precision, low-false-alarm detection of power events remain pressing technical challenges in this field. Summary of the Invention

[0004] In order to solve the above-mentioned technical problem of how to achieve high-precision and low-false-alarm-rate detection of power consumption events, the present invention provides the following solution.

[0005] A method for monitoring electric energy of a remote electric energy meter comprises the following steps: obtaining original electric parameter data of the remote electric energy meter, constructing a time series factor graph model with voltage and current phasors as latent variables and the original electric parameter data as observation nodes, and obtaining an electric energy state sequence composed of voltage and current phasors at each sampling point; performing synchronous compression wavelet transform on the current phasors in the electric energy state sequence to generate a time-frequency energy distribution diagram, and dividing the time-frequency energy distribution diagram into a fundamental band, a low-order harmonic band, and a high-order harmonic band; calculating the energy centroid trajectory, energy kurtosis, and mutual correlation coefficient between sub-bands of each frequency band to form a monitoring characteristic. The monitoring feature vector sequence is input into a deep autoencoder pre-trained on normal power consumption data to calculate the reconstruction error, and the Lyapunov exponent of the monitoring feature vector sequence in a preset time window is calculated. The negative log-likelihood probability is obtained based on a pre-established Gaussian mixture model that describes the exponential distribution under normal working conditions. The reconstruction error and the negative log-likelihood probability are weightedly fused to generate a comprehensive anomaly index. When the comprehensive anomaly index continuously exceeds the preset decision boundary for a predetermined period of time, it is determined to be an electricity consumption event, and its start time and duration are recorded.

[0006] The present invention can realize high-precision and low-false-alarm-rate detection of power consumption events.

[0007] Furthermore, a time series factor graph model is constructed with voltage and current phasors as latent variables and the original electrical parameter data as observation nodes, including: setting the voltage and current phasors as latent variables containing amplitude and phase angle, and setting the three-phase voltage effective value, three-phase current effective value, three-phase active power and three-phase reactive power in the original electrical parameter data as observation nodes.

[0008] Among them, the functional relationship between the measured value and the hidden variable at the corresponding moment is established through the electrical physics formula; the transfer relationship between the hidden variables at adjacent moments is established through the Gaussian process, thereby establishing a reliable time series factor graph model.

[0009] Furthermore, the method further includes reconstructing the electric energy state sequence composed of the voltage and current phasors at each sampling point using a sum-product algorithm.

[0010] The present invention utilizes the sum-product algorithm to accurately infer and reconstruct the electric energy data that is lost due to unstable network transmission.

[0011] Furthermore, the time-frequency energy distribution diagram is divided into a fundamental band, a low-order harmonic band and a high-order harmonic band, including: dividing the area with a frequency of 45Hz to 55Hz in the time-frequency energy distribution diagram into a fundamental band; dividing the area with a frequency of 55Hz to 655Hz into a low-order harmonic band; and dividing the area with a frequency of 655Hz to 2500Hz into a high-order harmonic band.

[0012] Further, the deep autoencoder is a multi-layer neural network, including an encoder, a decoder and a bottleneck layer connected between the two; the encoder is used to reduce the dimension of the monitoring feature vector sequence to generate a low-dimensional representation, and the decoder is used to reconstruct the monitoring feature vector sequence from the low-dimensional representation.

[0013] By training on a large amount of normal operating data, the autoencoder can learn the internal mode of the normal data, and when abnormal data is input, the decoder will not be able to effectively reconstruct due to the difference in mode from the normal data, thereby generating a huge reconstruction error.

[0014] Further, the deep autoencoder is an autoencoder using a long short-term memory network.

[0015] Further, the Lyapunov exponent of the sequence in the preset time window is calculated, including: reconstructing the phase space based on the preset embedding dimension and time delay parameter for the data in the time window, and estimating the maximum Lyapunov exponent based on the reconstructed phase space trajectory.

[0016] Further, the maximum Lyapunov exponent is estimated, including: applying the Rosenstein algorithm to estimate the maximum Lyapunov exponent.

[0017] The present application combines the deep autoencoder and the Lyapunov exponent two kinds of unsupervised models, and can respectively detect the anomaly from the data reconstruction fidelity and the system state stability two complementary perspectives.

[0018] Further, the comprehensive anomaly index is obtained by weighted sum of the reconstruction error and the negative log-likelihood probability, and the weight coefficient is a preset normal number.

[0019] An electric energy monitoring system for a remote electric energy meter, comprising a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the above-mentioned electric energy monitoring method for a remote electric energy meter.

[0020] In summary, compared with the prior art, the beneficial effects of the present invention are as follows: by constructing a time series factor graph model and applying the sum-product algorithm, the present invention can accurately infer and reconstruct the electric energy data that is missing due to unstable network transmission, effectively solving the problem of incomplete data and providing a reliable data basis for subsequent high-precision analysis. The present invention adopts synchronous compression wavelet transform technology to achieve high-resolution time-frequency characterization of non-stationary current signals, and constructs feature vectors from multiple dimensions such as energy centroid trajectory, kurtosis and sub-band cross-correlation, which greatly improves the ability to capture transient characteristics of complex power events. It integrates two unsupervised models, deep autoencoder and Lyapunov index, and performs anomaly detection from two complementary perspectives, namely data reconstruction fidelity and system state stability. Its weighted fusion strategy and duration judgment mechanism not only enhance the ability to detect early weak anomalies, but also significantly suppress false alarms caused by normal fluctuations, thereby achieving higher accuracy and robustness in monitoring power events. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematically illustrates a timing factor graph model according to an embodiment of the present invention;

[0022] Figure 2 is a schematic diagram schematically illustrating time-frequency energy distribution and frequency band division according to an embodiment of the present invention;

[0023] Figure 3 The figure schematically shows the relationship between comprehensive abnormal indicators and events according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] A method for monitoring electric energy of a remote electric energy meter comprises the following steps:

[0026] S1, obtaining original electrical parameter data of a remote electric energy meter, and constructing a time series factor graph model with voltage and current phasors as latent variables and the original electrical parameter data as observation nodes.

[0027] In one embodiment, the sum-product algorithm can also be used to reconstruct the complete power state sequence.

[0028] Specifically, the original electrical parameter data is the time series of active power, reactive power, voltage effective value and current effective value uploaded by the remote electric energy meter at regular intervals. In the time series factor graph model, the latent variables at each moment are the voltage phasor and the current phasor, each of which contains two components: amplitude and phase angle. The observation nodes at each moment are four measured values: active power, reactive power, voltage effective value and current effective value. These measured values ​​establish a functional relationship with the latent variables at the corresponding moment through electrical physics formulas. The transfer relationship between the latent variables at adjacent moments is established through the Gaussian process, that is, the phasor at the current moment obeys a Gaussian distribution with the phasor at the previous moment as the mean. When data is missing, the sum-product algorithm of the confidence propagation type is used to perform iterative message passing on the graph model to calculate the posterior probability distribution of the latent variables at the missing moment, and take its expected value as the reconstructed phasor, thereby filling the data gap and forming a complete electric energy state sequence. The time series factor graph model is as follows: Figure 1 As shown, t represents the current time.

[0029] In an optional embodiment, the voltage and current phasors are set as latent variables containing amplitude and phase angle, and the three-phase voltage effective value, three-phase current effective value, three-phase active power and three-phase reactive power in the original electrical parameter data are set as observation nodes; the sum-product algorithm infers to obtain an electric energy state sequence composed of the voltage and current phasors at each sampling point.

[0030] Specifically, the time-series factor graph model abstracts the physical relationships in the power grid into a probabilistic graph structure. Within this structure, voltage and current phasors—their amplitudes and phase angles—that cannot be directly measured but are desirable are defined as the system's latent variables. For example, at a certain time t, the latent variables for the phase A voltage are its amplitude of 221.5 volts and its phase angle of 0.1 radians. Data that can be directly measured by sensors, such as the three-phase RMS voltages of 220 volts, 221 volts, and 219 volts, and the three-phase RMS currents of 10 amperes, 10.1 amperes, and 9.9 amperes, along with the corresponding active and reactive powers, become observation nodes in the graph. These observations are linked to the latent variables through basic electrical formulas, such as the power calculation formula. These formulas constitute the factor nodes in the graph, describing the probabilistic dependencies between the variables.

[0031] Specifically, after the time series factor graph model is constructed, the sum-product algorithm is used for reasoning. The sum-product algorithm is an efficient method for information transmission on the time series factor graph. By iteratively transmitting probability information between variable nodes and factor nodes, the marginal probability distribution of each latent variable is finally calculated. For example, the observation data of multiple consecutive sampling points, such as 1000 sampling points, are comprehensively utilized to infer the most likely series of voltage and current phasor values ​​within this time period. This reasoning process can not only effectively filter out measurement noise, but also accurately track the dynamic changes of the power state, and finally output a complete time series composed of voltage and current phasors at each moment, that is, the power state sequence, which provides a high-precision data foundation for subsequent feature extraction.

[0032] S2, performing synchronous compression wavelet transform on the current signal in the reconstructed electric energy state sequence to generate a time-frequency energy distribution diagram, and dividing the time-frequency energy distribution diagram into a fundamental band, a low-order harmonic band, and a high-order harmonic band, calculating the energy centroid trajectory, energy kurtosis, and mutual correlation coefficient between sub-bands of each frequency band to form a monitoring feature vector sequence;

[0033] Specifically, the amplitude signal of the reconstructed current phasor sequence is extracted, and its time-frequency coefficient matrix is ​​obtained using a continuous wavelet transform based on the Morley wavelet. The energy in this matrix is ​​then compressed along the frequency axis at the instantaneous frequency points of the signal, resulting in a time-frequency energy distribution diagram with concentrated energy. According to my country's power grid frequency standards, the frequency range from 48 Hz to 52 Hz is defined as the fundamental band, the range from 52 Hz to 500 Hz as the low-order harmonic band, and the range from 500 Hz to half the sampling frequency as the high-order harmonic band. At each time point, the weighted average frequency of the energy along the frequency axis in each of the three frequency bands is calculated as the energy centroid. The fourth-order central moment of the energy distribution, or energy kurtosis, is calculated, and the Pearson cross-correlation coefficients of the energy sequences between the fundamental band and the low-order harmonic band, and between the low-order harmonic band and the high-order harmonic band, are calculated. The eight values ​​(the three energy centroids, the three energy kurtosis, and the two cross-correlation coefficients) are combined into a feature vector, forming a time-synchronized monitoring feature vector sequence.

[0034] like Figure 2 As shown in the figure, the frequency band division is based on the physical characteristics of different electrical phenomena in the power system. The fundamental band, which ranges from 45 Hz to 55 Hz, is centered on the standard operating frequency of my country's power grid, 50 Hz. The energy concentration in this frequency band is a core indicator of power quality. During normal operation, more than 99 percent of the energy should be concentrated here. For example, a stable power grid may have a fundamental energy value as high as 10,000 units, while the energy in other frequency bands may total less than 100 units. Any significant energy distribution that deviates from this range may indicate a frequency excursion or severe disturbance.

[0035] The low-harmonic band covers the frequency range from 55 Hz to 655 Hz, corresponding to the 2nd to 13th harmonics of the 50 Hz fundamental. Harmonics in this range are typically generated by large nonlinear loads, such as rectifiers and inverters in industrial production. For example, when a six-pulse commutation device starts up, distinct energy bands can be observed at 250 Hz (the 5th harmonic) and 350 Hz (the 7th harmonic) in the time-frequency energy diagram. The high-harmonic band extends from 655 Hz to 2500 Hz, corresponding to the 14th to 50th harmonics. These higher-frequency disturbances are often associated with the rapid switching of modern power electronic equipment or transient events such as arc faults. For example, a small arc fault may produce a brief but concentrated bright spot in the 1500 Hz to 2000 Hz range. By dividing the frequency bands, the disturbance characteristics of these different sources can be effectively separated, enabling precise identification of the fault type.

[0036] S3, input the monitoring feature vector sequence into a deep autoencoder pre-trained on normal power consumption data to calculate the reconstruction error, and at the same time calculate the Lyapunov exponent of the monitoring feature vector sequence within a preset time window, and obtain its negative log-likelihood probability based on a pre-established Gaussian mixture model that describes the exponential distribution under normal operating conditions.

[0037] In an optional embodiment, the deep autoencoder is a multi-layer neural network, including an encoder, a decoder and a bottleneck layer connected therebetween; the encoder is used to reduce the dimension of the monitoring feature vector sequence to generate a low-dimensional representation, and the decoder is used to reconstruct the monitoring feature vector sequence from the low-dimensional representation.

[0038] Specifically, the deep autoencoder is a symmetrically structured neural network. For example, the encoder may consist of three hidden layers. The input layer receives a 20-dimensional monitoring feature vector composed of features such as frequency band energy. The number of neurons in the network layer is then reduced from 20 to 15, then to 10, and finally to a bottleneck layer of only 5 neurons. This process is called dimensionality reduction encoding, where the encoder learns to extract the most core and representative 5-dimensional features from the complex 20-dimensional input data, forming a low-dimensional representation.

[0039] The decoder part performs the inverse operation of the encoder. It receives the 5-dimensional low-dimensional representation from the bottleneck layer and goes through a series of expansion layers, for example, the number of neurons increases from 5 to 10, then to 15, and finally outputs a 20-dimensional vector with the same dimension as the input layer. The task of the decoder is to use this compressed core feature to reconstruct the original monitoring feature vector sequence as accurately as possible. The training goal of the entire network is to minimize the difference between the input vector and the reconstructed output vector, for example, the mean square error. Through training on a large amount of normal operating data, the autoencoder can learn the internal patterns of normal data. When abnormal data is input, the decoder will not be able to effectively reconstruct due to the difference in pattern from normal data, resulting in a large reconstruction error.

[0040] In other embodiments, the deep autoencoder is an autoencoder employing a long short-term memory network.

[0041] Meanwhile, the Rosenstein algorithm can be used to reconstruct the phase space of the monitoring feature vector sequence within a preset length of time window, and calculate the largest Lyapunov exponent. A Gaussian mixture model composed of multiple weighted Gaussian distributions has been trained on a large number of Lyapunov exponent samples under normal operating conditions. For the newly calculated Lyapunov exponent, calculate its probability density under the model, and take the negative logarithm value to get the negative log-likelihood probability.

[0042] In an optional embodiment, calculating the Lyapunov exponent of the sequence within a preset time window comprises: reconstructing the phase space of the data within the time window based on the preset embedding dimension and time delay parameter, and estimating the largest Lyapunov exponent based on the reconstructed phase space trajectory.

[0043] Specifically, phase space reconstruction is a process of expanding one-dimensional time series data into high-dimensional space to reveal its internal dynamic characteristics. Assuming there is a reconstruction error time series output by the autoencoder, select data within a time window, such as 500 consecutive data points. Then set the embedding dimension m to 5 and the time delay τ to 2 sampling points. The first reconstructed phase space vector is composed of the 1st, 3rd, 5th, 7th, and 9th data points of the sequence. The second vector is composed of the 2nd, 4th, 6th, 8th, and 10th data points, and so on. Through this process, the one-dimensional time series is converted into a series of points moving in five-dimensional space, and the trajectory formed by connecting these points is the reconstructed phase space trajectory.

[0044] After obtaining the phase space trajectory, an algorithm such as Rosenstein is applied to estimate the maximum Lyapunov exponent. The maximum Lyapunov exponent measures the degree of chaos in the system dynamics, that is, the speed at which two trajectories with extremely close initial positions separate over time. During the calculation, adjacent pairs of points are found in the phase space, and the change in their distance over time is tracked. A positive Lyapunov exponent indicates that the system has chaotic characteristics, which is a typical manifestation of the normal operation of a complex power system. For example, under normal operating conditions, the calculated maximum Lyapunov exponent may stabilize at around 0.05. If the system state changes suddenly, such as an oscillation or a fault occurs, the dynamic behavior will become more unpredictable, resulting in accelerated trajectory separation, and the maximum Lyapunov exponent may increase sharply to 0.8, thereby serving as a sensitive abnormality indication signal.

[0045] S4, performing weighted fusion of the reconstruction error and the negative log-likelihood probability to generate a comprehensive anomaly index. When the comprehensive anomaly index continuously exceeds the preset decision boundary for a predetermined period of time, it is determined to be an electricity consumption event, and its start time and duration are recorded.

[0046] In an optional embodiment, the comprehensive abnormality index is obtained by performing a weighted summation on the reconstruction error and the negative log-likelihood probability, and the weight coefficient is a preset positive constant.

[0047] Specifically, the method combines two different dimensions of anomaly information to improve detection accuracy and robustness. The first is the reconstruction error of the deep autoencoder, which reflects the instantaneous deviation of the current power state in the feature space. The second is the negative log-likelihood calculated based on the normal state probability distribution learned from historical data, which reflects the probability of the current state occurring. Even if a state does not have a significant deviation in its own features, if its probability of occurrence is extremely low, it should be considered an anomaly.

[0048] Specifically, to form a single, clear basis for determining an anomaly, the reconstruction error and the negative log-likelihood probability are weighted and summed. For example, the weight coefficients w1 and w2 for the reconstruction error and negative log-likelihood probability can be set to 0.7 and 0.3, respectively. The comprehensive anomaly index S then equals 0.7 times the reconstruction error plus 0.3 times the negative log-likelihood probability. Assuming that at a certain moment, the normalized reconstruction error is 0.9, indicating a severe deviation, and the negative log-likelihood probability is 0.6, indicating a relatively rare state, the comprehensive anomaly index S is 0.7 times 0.9 plus 0.3 times 0.6, resulting in 0.63 plus 0.18, which equals 0.81. This comprehensive index is ultimately compared to a pre-set threshold, such as 0.75. If it exceeds the threshold, a system anomaly is determined. A timer is then started and the current time is recorded. If the comprehensive anomaly index remains above the threshold for a predetermined period of time, for example, five consecutive time points, a power usage event is officially confirmed. The start time of the event is the time when the threshold is first exceeded, and the duration of the event is the time from when the comprehensive abnormal indicator first exceeds the threshold to when it first falls below the threshold.

[0049] like Figure 3 The figure shows the process of event determination. The red part indicates that the comprehensive abnormality index exceeds the threshold (decision boundary).

[0050] The present invention also relates to an electric energy monitoring system for a remote electric energy meter, comprising a processor and a memory, wherein the memory stores a computer program. The processor can interact with the memory and call the computer program (for example, via a bus). The processor then executes the computer program. When the computer program is executed by the processor, the electric energy monitoring method for the remote electric energy meter of the above embodiment is implemented.

[0051] In the present invention, the aforementioned memory may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium.

[0052] This specification and the accompanying drawings are merely exemplary illustrations of the present invention as defined by the appended claims. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the scope of the present invention.

Claims

1. A method for monitoring electric energy of a remote electric energy meter, characterized in that: The following steps are involved: Obtain the original electrical parameter data of the remote electric energy meter, construct a time series factor graph model with voltage and current phasors as latent variables and the original electrical parameter data as observation nodes, and obtain the electric energy state sequence composed of the voltage and current phasors at each sampling point; Perform synchronous compression wavelet transform on the current phasor in the power state sequence to generate a time-frequency energy distribution diagram. The time-frequency energy distribution diagram is divided into fundamental band, low-order harmonic band, and high-order harmonic band. The energy centroid trajectory, energy kurtosis, and mutual correlation coefficient between sub-bands of each frequency band are calculated to form a monitoring feature vector sequence. The monitoring feature vector sequence is input into a deep autoencoder pre-trained on normal power consumption data to calculate the reconstruction error. The Lyapunov exponent of the monitoring feature vector sequence within a preset time window is calculated, including: based on a preset embedding dimension and time delay parameter, the data within the time window is reconstructed in phase space, and the maximum Lyapunov exponent is estimated based on the reconstructed phase space trajectory. The negative log-likelihood probability is obtained based on a pre-established Gaussian mixture model that describes the exponential distribution under normal operating conditions. The reconstruction error and the negative log-likelihood probability are weighted and fused to generate a comprehensive anomaly index. When the comprehensive anomaly index continuously exceeds the preset decision boundary for a predetermined period of time, it is determined to be an electricity consumption event, and its start time and duration are recorded. A deep autoencoder is a multi-layer neural network consisting of an encoder, a decoder, and a bottleneck layer connected between the two; the encoder is used to reduce the dimensionality of the monitoring feature vector sequence to generate a low-dimensional representation, and the decoder is used to reconstruct the monitoring feature vector sequence from the low-dimensional representation.

2. The method according to claim 1, characterized in that A time series factor graph model is constructed with voltage and current phasors as latent variables and the original electrical parameter data as observation nodes, including: setting the voltage and current phasors as latent variables containing amplitude and phase angle, and setting the three-phase voltage effective value, three-phase current effective value, three-phase active power and three-phase reactive power in the original electrical parameter data as observation nodes.

3. The method according to claim 2, characterized in that The method further includes reconstructing the electric energy state sequence composed of the voltage and current phasors at each sampling point by using a sum-product algorithm.

4. The method according to claim 1, wherein The time-frequency energy distribution diagram is divided into a fundamental band, a low-order harmonic band and a high-order harmonic band, including: dividing the area with a frequency of 45Hz to 55Hz in the time-frequency energy distribution diagram into a fundamental band; dividing the area with a frequency of 55Hz to 655Hz into a low-order harmonic band; and dividing the area with a frequency of 655Hz to 2500Hz into a high-order harmonic band.

5. The method according to claim 1, wherein The deep autoencoder is an autoencoder using a long short-term memory network.

6. The method according to claim 1, characterized in that Estimating the maximum Lyapunov exponent, including applying the Rosenstein algorithm to estimate the maximum Lyapunov exponent.

7. The method according to claim 1, characterized in that The comprehensive abnormality index is obtained by performing a weighted summation on the reconstruction error and the negative log-likelihood probability, and the weight coefficient is a preset positive constant.

8. An electric energy monitoring system for a remote electric energy meter, characterized in that: The invention comprises a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the electric energy monitoring method for a remote electric energy meter according to any one of claims 1 to 7.

Images