Intelligent electric energy meter remote monitoring device

Abstract

The invention provides a remote monitoring device for an intelligent electric energy meter, and belongs to the technical field of intelligent electric energy meters, and the remote monitoring device for the intelligent electric energy meter comprises an electric energy meter end monitoring system and a remote monitoring end system. The ammeter end adopts an electric energy acquisition unit, a current sampling unit, a voltage sampling unit and a temperature sensing unit to acquire data, performs feature extraction through a lightweight feature extraction module, and realizes adaptive sampling based on abnormal features. The remote monitoring end adopts a hybrid neural network to carry out deep analysis, the deep analysis comprises four levels of feature coding, equation embedding, time sequence analysis and feature decoding, streaming processing of multi-ammeter data is realized through a parallel processing module, and finally an operation state evaluation result and early warning information are generated. The technical problem that it is difficult to efficiently monitor the running states of multiple intelligent electric energy meters in real time and discover potential abnormalities in the prior art is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent electricity meters, and more particularly, relates to a remote monitoring device for intelligent electricity meters. Background Art

[0002] Intelligent electricity meters are key devices for realizing power consumption information collection and management in modern power systems. Traditional monitoring of intelligent electricity meters mainly relies on regular data collection and offline analysis methods, adopting a centralized data processing architecture. By means of the acquisition terminal, the power consumption data, voltage and current data, and power factor data of the electricity meter are read regularly and transmitted to the master station system for analysis. This monitoring method uses simple threshold judgment and statistical analysis methods to perform basic rationality checks and anomaly judgments on the collected data, so as to monitor the basic operating status of the electricity meter.

[0003] However, with the continuous expansion of the scale of the smart grid and the increasing complexity of the power consumption load, traditional electricity meter monitoring technologies face many challenges. First of all, the traditional regular collection method has a fixed sampling frequency and cannot adaptively adjust the sampling strategy according to the operating status of the electricity meter, resulting in the inability to obtain high-frequency sampling data in a timely manner when an anomaly occurs, which affects the accuracy of the analysis of abnormal operating conditions. Secondly, the traditional centralized data processing method needs to transmit a large amount of raw data to the master station, causing communication bandwidth pressure and large processing delays, making it difficult to meet the requirements of real-time monitoring. Thirdly, the traditional simple threshold judgment method can only detect obvious anomalies and lacks effective identification means for complex abnormal situations such as gradual faults and multi-device coupling faults. Finally, traditional technologies lack the overall analysis ability of the group operating characteristics of electricity meters, cannot excavate the correlation relationships between multiple electricity meters, and are difficult to achieve early warning of group anomalies.

[0004] Facing the growing demand for electricity meter monitoring, traditional technologies are difficult to solve the problem of efficient real-time monitoring of smart meters. How to achieve real-time monitoring of the operating status of electricity meters, timely identification of abnormal characteristics, multi-meter correlation analysis and early warning has become a technical problem to be solved urgently. Especially in the scenario of large-scale deployment of smart meters, a new monitoring solution that can support distributed feature extraction, adaptive sampling and parallel processing is needed to meet the requirements of real-time performance, accuracy and scalability. That is to say, there are technical problems in the prior art that it is difficult to perform real-time and efficient monitoring on the operating status of multiple intelligent electricity meters and discover potential anomalies. Summary of the Invention

[0005] In view of this, the present invention provides a remote monitoring device for intelligent electricity meters, which can solve the technical problem in the prior art that it is difficult to perform real-time and efficient monitoring on the operating status of multiple intelligent electricity meters and discover potential anomalies.

[0006] The present invention is implemented as follows: The present invention provides an intelligent electric energy meter remote monitoring device, which includes several electric meter end monitoring systems and a remote monitoring end system. The electric meter end monitoring systems include an electric energy acquisition unit, a current sampling unit, a voltage sampling unit, a temperature sensing unit, a first control chip, and a first feature extraction unit arranged inside the electric energy meter. The remote monitoring end system includes a data receiving unit, a carrier wave analysis unit, a main control chip, a data storage unit, a display unit, and a depth analysis unit. A lightweight feature extraction module is arranged inside the first feature extraction unit. The lightweight feature extraction module includes a feature extraction layer, a channel attention layer, a depthwise separable convolution layer, an adaptive feature fusion layer, and a fully connected layer. A parallel processing module is arranged inside the depth analysis unit. The parallel processing module includes a regular feature change equation, an abnormal feature propagation equation, a feature coupling equation, and a stability equation. The regular features are composed of a voltage fundamental wave, a current fundamental wave, and a fundamental power factor. The abnormal features are composed of a harmonic distortion rate, a transient disturbance, and a temperature fluctuation. The feature deviation is the Euclidean distance between two adjacent sampling values of the regular features.

[0007] Among them, the electric energy acquisition unit adopts a precision sampling circuit design with a sampling frequency of 3,200 Hz and a sampling accuracy reaching 0.2 level. The current sampling unit is realized by combining a Hall sensor with a signal conditioning circuit. The linear range of the Hall sensor is from 0.1 A to 100 A. The voltage sampling unit is realized by means of resistor voltage division and isolation amplification with a voltage division ratio of 1,000:1. The temperature sensing unit selects a digital temperature sensor with a measurement range from -40 °C to 85 °C and a measurement accuracy of ±0.5 °C.

[0008] Among them, the first control chip adopts a 32-bit microcontroller with a main frequency of 120 MHz. The first feature extraction unit is implemented based on a field programmable gate array with a clock frequency of 200 MHz. The data receiving unit is realized by using a power line carrier communication chip with a carrier frequency range from 3 kHz to 500 kHz. The carrier wave analysis unit is realized by using a digital signal processor with a main frequency of 400 MHz.

[0009] Among them, the main control chip adopts a quad-core processor with a main frequency of 1.5 GHz. The data storage unit adopts a solid state drive with a capacity of 256 GB. The display unit adopts a 7-inch liquid crystal display with a resolution of 1024×600 pixels. The depth analysis unit adopts a graphics processing unit with 384 stream processor cores.

[0010] Among them, the lightweight feature extraction module is used to perform the following steps: receive the electric energy pulse signal, the voltage signal, the current signal, and the temperature signal output by the first control chip, perform fast Fourier transform on the voltage signal and the current signal, extract the fundamental component, the first harmonic component, the second harmonic component, the third harmonic component, the fourth harmonic component, and the fifth harmonic component, use the improved lightweight neural network to extract features from the fundamental component and the harmonic components to obtain the regular features, calculate the feature deviation according to the regular features, generate the abnormal features when the feature deviation exceeds the preset threshold, and output the regular features and the abnormal features to the first control chip.

[0011] Among them, the input of the regular feature change equation includes the historical regular feature sequence, the operation duration of the electric energy meter, the load change rate, the environmental temperature change rate, and the power grid voltage volatility, and the output is the predicted value of the regular feature at the next moment.

[0012] Among them, the input of the abnormal feature propagation equation includes the abnormal feature vectors of adjacent electric meters, the line impedance parameters, the propagation delay coefficient, and the load similarity, and the output is the propagation influence range and the propagation influence intensity of the abnormal features.

[0013] Among them, the input of the feature coupling equation includes the regular feature vector, the abnormal feature vector, the coupling coefficient matrix, the feature correlation degree, the time scale parameter, and the perturbation intensity, and the output is the coupled comprehensive feature.

[0014] Among them, the input of the stability equation includes the singular value sequence of the feature matrix, the eigenvalue distribution, the matrix condition number, the time series stationarity index, and the state transition probability, and the output is the system stability evaluation index.

[0015] Among them, the adaptive feature fusion layer is used to dynamically adjust the weights of different feature channels and realize the adaptive fusion of multi-scale features. The input of the adaptive feature fusion layer is the multi-channel feature map output by the depthwise separable convolution layer, and the output is the fused feature vector.

[0016] Compared with the prior art, an intelligent electric energy meter remote monitoring device provided by the present invention proposes an intelligent electric energy meter remote monitoring scheme based on lightweight feature extraction and in-depth analysis. This scheme adopts a hierarchical architecture design, deploys a lightweight feature extraction module at the electric meter end, realizes the feature extraction of multi-source signals such as voltage, current, and temperature through an improved lightweight neural network, and adaptively adjusts the sampling frequency according to the abnormal features. At the monitoring end, a hybrid neural network structure is used for in-depth analysis, and parallel processing and correlation analysis of multi-electric meter features are realized through the multi-head self-attention mechanism, the equation embedding layer, and the time series analysis layer.

[0017] The solution of the present invention effectively solves the problems existing in the traditional technology. First, by deploying a lightweight feature extraction module at the meter end, source compression and preprocessing of data are achieved, reducing the data transmission pressure. The improved lightweight neural network adopts depthwise separable convolution and channel attention mechanism, improving the efficiency and accuracy of feature extraction. Second, based on the adaptive sampling strategy for abnormal features, the system can automatically increase the sampling frequency when detecting abnormalities, ensuring the integrity of abnormal condition data. Third, a hybrid neural network structure is adopted for in-depth analysis, and the equation embedding layer is used to model the regular feature changes, abnormal feature propagation, feature coupling and system stability, realizing the effective identification of complex abnormal patterns. Finally, the parallel processing module supports the streaming processing of multi-meter data, and realizes the early warning of group abnormalities by constructing a state transition graph.

[0018] Through the above technical innovations, the present invention successfully solves the technical problem in the prior art that it is difficult to monitor the operation status of multiple intelligent electricity meters in real time and efficiently and discover potential abnormalities. This solution moves the feature extraction task forward to the meter end, adopts lightweight algorithms to reduce the computational overhead, and at the same time uses in-depth analysis methods at the monitoring end to mine multi-meter correlation features, ensuring both the real-time and accuracy of monitoring and realizing the scalability of the system. Especially, the introduction of physical model constraints in the equation embedding layer improves the interpretability and reliability of abnormal identification, providing strong support for the safe and stable operation of the smart grid. Description of the Drawings

[0019] Figure 1 It is a schematic structural diagram of the device of the present invention.

[0020] Figure 2 It is a flowchart of the steps executed by the lightweight feature extraction module.

[0021] Figure 3 It is a flowchart of the steps executed by the parallel processing module.

[0022] Figure 4 It is a 24-hour load characteristic curve graph of electricity meters in different regions in Embodiment 2.

[0023] Figure 5 It is a harmonic analysis result graph of voltage and current in Embodiment 2.

[0024] Figure 6 It is a stability evaluation result graph of the system during one-year operation in Embodiment 2. Detailed Embodiments

[0025] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention.

[0026] As Figure 1 shown in the figure, it is a schematic structural diagram of an intelligent electric energy meter remote monitoring device provided by the present invention. This device includes several meter-end monitoring systems and a remote monitoring-end system. The meter-end monitoring system includes an electric energy acquisition unit, a current sampling unit, a voltage sampling unit, a temperature sensing unit, a first control chip, and a first feature extraction unit arranged inside the electric energy meter;

[0027] The remote monitoring-end system includes a data receiving unit, a carrier wave analysis unit, a main control chip, a data storage unit, a display unit, and a deep analysis unit;

[0028] The electric energy acquisition unit is electrically connected to the electric energy meter measurement circuit, and is used for acquiring an electric energy pulse signal and transmitting the electric energy pulse signal to the first control chip;

[0029] The current sampling unit is electrically connected to the electric energy meter current sampling circuit, and is used for acquiring a current signal, performing signal conditioning, and then transmitting it to the first control chip;

[0030] The voltage sampling unit is electrically connected to the electric energy meter voltage sampling circuit, and is used for acquiring a voltage signal, performing signal conditioning, and then transmitting it to the first control chip;

[0031] The temperature sensing unit is arranged inside the electric energy meter, and is used for acquiring the internal temperature of the electric energy meter and transmitting it to the first control chip;

[0032] The first feature extraction unit is electrically connected to the first control chip, and is used for extracting features from the acquired signals and generating regular features and abnormal features;

[0033] The first control chip is used for adjusting the sampling frequency according to the abnormal features, and converting the regular features and the abnormal features into carrier signals for transmission through the power line;

[0034] The data receiving unit is used for receiving the carrier signals transmitted by multiple meter-end monitoring systems and transmitting them to the carrier wave analysis unit;

[0035] The carrier wave analysis unit is used for analyzing the carrier signals into digital signals and transmitting them to the main control chip;

[0036] The deep analysis unit is electrically connected to the main control chip, and is used for performing parallel stream processing on the regular features and the abnormal features of multiple electric meters.

[0037] Among them, a lightweight feature extraction module is arranged inside the first feature extraction unit. As Figure 2 shown in the figure, the lightweight feature extraction module is used to perform the following steps:

[0038] S11. Receive the power pulse signal, the voltage signal, the current signal, and the temperature signal output by the first control chip;

[0039] S12. Perform fast Fourier transform on the voltage signal and the current signal, and extract the fundamental component, the first harmonic component, the second harmonic component, the third harmonic component, the fourth harmonic component, and the fifth harmonic component;

[0040] S13. Use the improved lightweight neural network to extract features from the fundamental component, the first harmonic component, the second harmonic component, the third harmonic component, the fourth harmonic component, and the fifth harmonic component to obtain the regular features;

[0041] S14. Calculate the feature deviation according to the regular features, and generate the abnormal features when the feature deviation exceeds the preset threshold;

[0042] S15. Output the regular features and the abnormal features to the first control chip.

[0043] Among them, a parallel processing module is arranged in the in-depth analysis unit. As Figure 3 shown, the parallel processing module is used to execute the following steps:

[0044] S21. Receive the regular features and the abnormal features of multiple electric meters transmitted by the main control chip, and construct a feature tensor;

[0045] S22. Construct a hybrid neural network structure. The hybrid neural network includes a feature encoding layer, an equation embedding layer, a time series analysis layer, and a feature decoding layer; the feature encoding layer uses a multi-head self-attention mechanism to encode the input features and generate an attention-enhanced feature representation; the equation embedding layer contains 4 parallel equation processing units, and the equation processing unit includes a regular feature processing unit, an abnormal feature processing unit, a feature coupling processing unit, and a stability processing unit. The regular feature processing unit, the abnormal feature processing unit, the feature coupling processing unit, and the stability processing unit are respectively composed of a fully connected layer, a parameter-sharing recurrent neural network layer, and a dynamic weight layer; the time series analysis layer uses a bidirectional long short-term memory network to extract time series features and maintains long-term dependencies through residual connections; the feature decoding layer uses a deconvolution network to map the features back to the original space;

[0046] S23. Process the feature tensor using the equation embedding layer of the hybrid neural network. The parameters of the regular feature change equation are learned through the fully connected layer, and the input is the encoded regular feature sequence. The parameters of the abnormal feature propagation equation are dynamically updated through the recurrent neural network layer, and the input is the abnormal features of adjacent electricity meters. The coefficient matrix of the feature coupling equation is adaptively generated through the dynamic weight layer, and the input is the regular feature and the abnormal feature. The evaluation index of the stability equation is calculated through a multi-layer perceptron, and the input is the statistic of the feature matrix.

[0047] S24. Pass the output of the equation embedding layer through the time series analysis layer to extract the time series features and construct a state transition diagram.

[0048] S25. Use the feature decoding layer to convert the state transition diagram into the operation state evaluation results of each electricity meter.

[0049] S26. Generate a warning message according to the operation state evaluation results.

[0050] Among them, the parallel processing module includes:

[0051] The regular feature change equation is used to describe the evolution law of the regular feature over time. The input of the regular feature change equation includes the historical regular feature sequence, the operation duration of the electricity meter, the load change rate, the ambient temperature change rate, and the grid voltage fluctuation rate, and the output is the predicted value of the regular feature at the next moment.

[0052] The abnormal feature propagation equation is used to describe the propagation characteristics of the abnormal feature in the power grid. The input of the abnormal feature propagation equation includes the abnormal feature vector of adjacent electricity meters, the line impedance parameter, the propagation delay coefficient, and the load similarity, and the output is the propagation influence range and propagation influence intensity of the abnormal feature.

[0053] The feature coupling equation of the regular feature and the abnormal feature. The input of the feature coupling equation includes the regular feature vector, the abnormal feature vector, the coupling coefficient matrix, the feature correlation degree, the time scale parameter, and the perturbation intensity, and the output is the coupled comprehensive feature.

[0054] The stability equation of the feature matrix is used to evaluate the overall stability of the operation state of the electricity meter group. The input of the stability equation includes the singular value sequence of the feature matrix, the eigenvalue distribution, the matrix condition number, the time series stationarity index, and the state transition probability, and the output is the system stability evaluation index.

[0055] The regular features are composed of the fundamental voltage, fundamental current, and fundamental power factor; the abnormal features are composed of the harmonic distortion rate, transient disturbance, and temperature fluctuation; the feature deviation is the Euclidean distance between two adjacent sampling values of the regular features.

[0056] The improved lightweight neural network includes a feature extraction layer, a channel attention layer, a depthwise separable convolution layer, an adaptive feature fusion layer, and a fully connected layer; the adaptive feature fusion layer is used to dynamically adjust the weights of different feature channels and achieve the adaptive fusion of multi-scale features. The input of the adaptive feature fusion layer is the multi-channel feature map output by the depthwise separable convolution layer, and the output is the fused feature vector.

[0057] The following describes in detail the specific implementation manners of the above steps.

[0058] In the specific implementation manner, the power acquisition unit of the watt-hour meter end monitoring system adopts a precise sampling circuit design, including a high-precision operational amplifier, an analog switch, and a sample-and-hold circuit. The sampling frequency is set to 3200 Hz, and the sampling accuracy reaches 0.2 level, which can accurately capture the rising edge and falling edge of the power pulse signal to ensure the accuracy of power measurement.

[0059] The current sampling unit is implemented by combining a Hall sensor with a signal conditioning circuit. The linear range of the Hall sensor is 0.1 A to 100 A. The signal conditioning circuit includes a preamplifier, a band-pass filter, and an analog-to-digital converter. The gain of the preamplifier is adjustable. The cut-off frequency range of the band-pass filter is 45 Hz to 65 Hz. The sampling rate of the analog-to-digital converter is 6400 Hz, and the resolution is 16 bits.

[0060] The voltage sampling unit is implemented by means of resistor voltage division and isolation amplification. The voltage division ratio is 1000:1. The isolation amplifier adopts an opto-coupling method, with a common-mode rejection ratio greater than 80 dB and an isolation voltage level reaching 4000 V. The sampled signal is subjected to anti-aliasing filtering and then 16-bit analog-to-digital conversion.

[0061] The temperature sensing unit selects a digital temperature sensor, with a measurement range of -40 °C to 85 °C, a measurement accuracy of ±0.5 °C, a sampling period of 1 s, and communicates with the first control chip through a serial interface.

[0062] The first control chip adopts a 32-bit microcontroller with a main frequency of 120 MHz, built-in 2 MB of flash memory and 256 KB of random access memory, integrated with a 12-channel 16-bit analog-to-digital converter, a hardware multiplier, and a floating-point arithmetic unit, and supports a real-time operating system.

[0063] The first feature extraction unit is implemented based on a field-programmable gate array, with a clock frequency of 200 megahertz. It includes 32 digital signal processing modules, 64 multiplier arrays, and 128 kilobytes of dual-port random access memory, supporting parallel data processing and pipelined operations.

[0064] The data receiving unit is implemented using a power line carrier communication chip, supporting quadrature amplitude modulation and quadrature phase shift keying modulation. The carrier frequency range is from 3 kilohertz to 500 kilohertz, and the communication rate can reach 250 kilobits per second. It has built-in forward error correction coding and cyclic redundancy check.

[0065] The carrier analysis unit is implemented using a digital signal processor, with a main frequency of 400 megahertz, supporting fast Fourier transform and digital filtering. It has an 8-channel direct memory access controller and has the ability to process real-time data streams.

[0066] The main control chip uses a quad-core processor, with a main frequency of 1.5 gigahertz, having an 8-megabyte level 3 cache, supporting the single instruction multiple data stream extended instruction set, and integrating a hardware encryption engine and a secure boot mechanism.

[0067] The data storage unit uses a solid-state drive with a capacity of 256 gigabytes. It uses triple-level cell flash memory, supports power-off protection and data error correction, and the write speed reaches 550 megabytes per second.

[0068] The display unit uses a 7-inch liquid crystal display with a resolution of 1024 by 600 pixels. It uses a capacitive touch screen, supports multi-touch, the display brightness is adjustable, and it has an automatic dimming function.

[0069] The deep analysis unit uses a graphics processing unit, with 384 stream processor cores, a video memory bandwidth of 128 gigabytes per second, supporting general computing units and capable of performing large-scale parallel computing.

[0070] Optionally, the lightweight neural network in the lightweight feature extraction module uses a gated recurrent unit structure. The number of neurons in the input layer is 32, the hidden layer uses a 3-layer structure, with 64 neurons in each layer, and the number of neurons in the output layer is 16. The activation function uses a rectified linear unit. The feature encoding layer uses an 8-head attention mechanism, with the dimension of each attention head being 64, the dimension of key-value pairs being 512, and the positional encoding using sine and cosine functions, and the output feature dimension being 512.

[0071] When receiving the electric energy pulse signal, voltage signal, current signal and temperature signal in the specific implementation of step S11, a multi-buffer polling data reception mechanism is adopted. A circular buffer with a size of 16 kilobytes is separately allocated for each signal. The data is transferred from the first control chip to the feature extraction unit through direct memory access. The data frame header includes the signal type identifier, timestamp and data length. After cyclic redundancy check, it is stored in the corresponding signal buffer. The sampled data is aligned and reordered according to the timestamp to ensure the synchronization of multiple signals. The sampling frequency can be adaptively adjusted according to the signal change rate. The basic sampling frequency is 6400 Hz, and it is increased to 12800 Hz when a rapid signal change is detected. An idle frame interval is adopted between data frames, and the frame interval time is 125 microseconds.

[0072] Before performing the fast Fourier transform on the voltage signal and current signal in the specific implementation of step S12, the signals are first preprocessed, including DC removal, detrending and windowing. The DC component is obtained by calculating the signal mean and subtracting it. The trend term is eliminated after being fitted by the least squares method. The Hanning window function is selected to reduce spectral leakage, and the window length is 1024 points. To improve the calculation efficiency, the radix-4 fast Fourier transform algorithm is adopted and calculated in-place through the butterfly operation unit. The calculated spectrum is processed by amplitude normalization. The frequency points of each harmonic are determined according to the fundamental frequency. The amplitude and phase information of the fundamental wave component and the first 5 harmonic components are extracted. The fundamental frequency is obtained through the interpolation-type frequency estimation algorithm, with a frequency resolution better than 0.1 Hz, an amplitude measurement accuracy better than 0.2%, and a phase measurement accuracy better than 0.2 degrees.

[0073] In the specific implementation of step S13, an improved lightweight neural network is used to extract features from the fundamental wave component and each harmonic component. The improved lightweight neural network adopts a depthwise separable convolution structure. First, the time-frequency domain features of the input signal are extracted through the feature extraction layer. The convolution kernel size is 3×3, the stride is 1, and the padding method is same padding. 32 feature maps are extracted. Then, through the channel attention layer, the global average pooling and global maximum pooling are processed in parallel to obtain the channel descriptor. The channel weights are generated through two fully connected layers and the softmax function to recalibrate the feature maps. Then, through the depthwise separable convolution layer, first the depthwise convolution of each channel is performed, and then the pointwise convolution is performed to reduce the calculation complexity. The convolved features pass through the adaptive feature fusion layer, which adopts a multi-scale feature pyramid structure. The context information of different scales is obtained through spatial pyramid pooling. The pooling scales include 1×1, 2×2, and 4×4. The feature fusion adopts the weighted sum method, and the weight coefficients are adaptively learned through the soft attention mechanism. Finally, through the fully connected layer, the features are mapped to the regular feature space to obtain regular features such as the voltage fundamental wave, current fundamental wave and fundamental wave power factor.

[0074] The specific implementation of step S14 is that when calculating the feature deviation according to the regular features, the sliding window method is adopted. The window length is 60 seconds, and the sliding step is 1 second. Calculate the Euclidean distance between the regular feature vectors of adjacent sampling points within the window to obtain the feature deviation sequence. Conduct statistical analysis on the feature deviation sequence, calculate statistics such as mean, variance, skewness, and kurtosis, establish a probability distribution model of the feature deviation based on historical data, use the kernel density estimation method to obtain the probability density function of the deviation distribution, set the warning threshold to 3 times the standard deviation. When the feature deviation exceeds the warning threshold, it is determined as an abnormal state, and abnormal features are generated. The abnormal features include three indicators: harmonic distortion rate, transient disturbance, and temperature fluctuation. The warning threshold of the harmonic distortion rate is 5%, the warning threshold of the transient disturbance is 20% of the rated value, and the warning threshold of the temperature fluctuation is 5 degrees Celsius per minute.

[0075] The specific implementation of step S15 is that when outputting the regular features and abnormal features to the first control chip, a data compression and encrypted transmission mechanism is adopted. First, compress and encode the feature data. Adopt a combination of run-length encoding and differential encoding. Represent the regular features with 16-bit fixed-point numbers and the abnormal features with 32-bit floating-point numbers. The compressed data is encrypted through the Advanced Encryption Standard algorithm with a key length of 256 bits. The encrypted data is transmitted in blocks, and the size of each data block is 4 kilobytes. The data block contains the feature type, timestamp, compression flag, and checksum, and is transmitted to the first control chip through the Serial Peripheral Interface at a transmission rate of 10 megabits per second. Adopt the stop-and-wait protocol to ensure reliable data transmission. The receiving end verifies the data integrity through cyclic redundancy check, and requests retransmission for data blocks with failed checksums.

[0076] The specific implementation of step S21 is that when receiving the regular features and abnormal features of multiple electric meters transmitted by the main control chip, first establish a feature cache pool, allocate an independent feature buffer area for each electric meter, and the buffer area size is 32 megabytes. Manage it with the first-in-first-out strategy. The feature data is indexed according to the timestamp and electric meter identifier. When constructing the feature tensor, adopt a three-dimensional tensor structure. The first dimension represents the electric meter number, the second dimension represents the time series, and the third dimension represents the feature dimension. The tensor size is the number of electric meters multiplied by the time step multiplied by the feature dimension. The time step is set to 144 points, corresponding to the sampling data of 24 hours, and the sampling interval is 10 minutes. The feature dimension includes regular features and abnormal features, and the total dimension is 12. The tensor data type is 32-bit floating-point numbers. Complement the missing data using the linear interpolation method and correct the outliers using the moving median method to ensure the integrity and continuity of the feature tensor.

[0077] The specific implementation of step S22 is that when constructing a hybrid neural network structure, the feature encoding layer adopts a multi-head self-attention mechanism with 8 attention heads, and the dimension of each attention head is 64. The input is linearly mapped to obtain query, key, and value matrices with a dimension of 512. The attention weights are calculated through scaled dot-product attention, and weight normalization and residual connections are adopted. The position encoding uses sine and cosine functions and supports a maximum sequence length of 1000. The regular feature processing unit in the equation embedding layer uses a fully connected layer to extract feature representations, with hidden layer dimensions of 256, 128, and 64. Batch normalization and dropout regularization are used. The abnormal feature processing unit uses a gated recurrent unit network with parameter sharing, and the hidden state dimension is 128. Gradient clipping is used to prevent gradient explosion. The dynamic weight layer of the feature coupling processing unit adaptively learns feature weights through the attention mechanism. The stability processing unit uses a multi-layer perceptron to evaluate the system stability. The time series analysis layer uses a bidirectional long short-term memory network with a hidden layer dimension of 256, and residual connections are adopted to maintain long-term dependencies. The feature decoding layer uses a transposed convolutional network, which includes 5 transposed convolutional layers with a kernel size of 3×3, a stride of 2, and the number of channels is halved layer by layer from 256.

[0078] The specific implementation of step S23 is that when using the equation embedding layer of the hybrid neural network to process the feature tensor, the regular feature change equation learns parameters through a fully connected layer. The input includes the historical regular feature sequence, running duration, load change rate, ambient temperature change rate, and grid voltage volatility, a total of 5 input variables. Each variable passes through an independent feature extraction branch and then undergoes feature fusion to predict the regular feature value at the next moment, with a prediction period of 10 minutes. The abnormal feature propagation equation dynamically updates parameters through a recurrent neural network layer. The input includes the abnormal feature vector of adjacent meters, line impedance parameters, propagation delay coefficient, and load similarity. Considering the propagation characteristics of abnormal features in the power grid, the propagation influence range and propagation influence intensity are calculated. The coefficient matrix of the feature coupling equation is generated by the dynamic weight layer. The input is the regular feature vector and the abnormal feature vector. Considering the feature correlation, time scale parameter, and perturbation intensity, the dynamic coupling of features is realized. The evaluation index of the stability equation is calculated by a multi-layer perceptron. The input includes the singular value sequence of the feature matrix, eigenvalue distribution, matrix condition number, time series stationarity index, and state transition probability, and the stability evaluation score is output.

[0079] The specific implementation of step S24 is that when the output of the equation embedding layer passes through the time series analysis layer, a bidirectional long short-term memory network is used to extract time series features. The network includes hidden layers in both forward and backward directions. The dimension of the single hidden layer is 256, and the time step is 144, corresponding to 24-hour time series data. The information flow is controlled through a gating mechanism, including an input gate, a forget gate, and an output gate. The parameters of the gates are learned through the backpropagation algorithm. Residual connections are used to prevent gradient vanishing. The extracted time series features are used to construct a state transition graph. The state transition graph adopts a directed weighted graph structure. The nodes represent the electricity meter states, and the edges represent the state transition probabilities. The states include normal operation, mild anomaly, moderate anomaly, and severe anomaly. The state transition probabilities are obtained by statistical analysis using a sliding time window with a length of 24 hours and a sliding step of 1 hour.

[0080] The specific implementation of step S25 is that when using the feature decoding layer to convert the state transition graph into the operation state evaluation results of each electricity meter, a 5-layer transposed convolutional network is adopted. Each layer includes a transposed convolution, batch normalization, and an activation function. The size of the transposed convolution kernel is 3×3, the stride is 2, and the number of channels is 256, 128, 64, 32, and 16 in sequence. Same padding is used to maintain the size of the feature map. The activation function uses the rectified linear unit. The last layer outputs the state probability distribution through the softmax function. The state evaluation results include the operation state level, the anomaly degree score, and the state prediction confidence. The operation state level is divided into 4 levels, corresponding to normal, mild anomaly, moderate anomaly, and severe anomaly respectively. The range of the anomaly degree score is 0 to 100, and the state prediction confidence indicates the reliability of the prediction result.

[0081] The specific implementation of step S26 is that when generating warning information based on the operation state evaluation results, a multi-level warning mechanism is adopted. The warning level corresponds to the operation state level. Different triggering conditions and processing strategies are set for each warning level. For mild anomalies, a warning is triggered when the anomaly duration exceeds 1 hour or the anomaly degree score exceeds 60. The warning information includes the anomaly type, occurrence time, and affected range. For moderate anomalies, a warning is triggered when the anomaly duration exceeds 30 minutes or the anomaly degree score exceeds 75. The warning information adds the anomaly development trend and processing suggestions. For severe anomalies, a warning is triggered when the anomaly duration exceeds 10 minutes or the anomaly degree score exceeds 90. The warning information adds the emergency treatment plan and impact assessment. The warning information is sent to the display unit through a message queue, adopting a real-time push mechanism, supporting SMS and email notifications. The warning records are saved in the data storage unit with a storage period of 180 days.

[0082] The following provides a description of the derivation process of some equations.

[0083] 1. Derivation of the fast Fourier transform equation:

[0084] Starting from the basic form of the discrete Fourier transform:

[0085] The optimization steps include:

[0086] First, decompose the N-point sequence into odd and even groups:

[0087]

[0088] where is the rotation factor.

[0089] Through recursive decomposition, a radix-4 algorithm structure is finally formed, which can reduce the number of multiplication operations and improve the calculation efficiency.

[0090] 2. Construction process of the regular feature change equation: The regular feature matrix F is expressed as:

[0091]

[0092] In the formula, F ij respectively represent the eigenvalue of the fundamental voltage, fundamental current, and power factor.

[0093] The regular feature vector f is composed of the column vectors of the matrix F:

[0094] In the formula, f1, f2, and f3 respectively correspond to the features of voltage, current, and power factor.

[0095] The finally formed equation: F t = αF t-1 + βT l + γL r + δT r + ηV r + ∈.

[0096] Parameter acquisition method: α, β, γ, δ, and η are obtained by fitting historical data using the least squares method; T l is obtained through an accumulated running time counter; L r is calculated by the load change amount within 10 minutes; T r is calculated by sampling with a temperature sensor; V r is calculated by voltage sampling; the range of ∈ is ±0.02.

[0097] 3. Construction process of the abnormal feature propagation equation:

[0098] The abnormal feature matrix A is expressed as:

[0099] In the formula, A ij respectively represent the eigenvalue of the harmonic distortion rate, transient disturbance, and temperature fluctuation.

[0100] The line impedance matrix Z is expressed as:

[0101] In the formula, Z ij represents the line impedance value from node i to node j.

[0102] The finally formed equation:

[0103] Parameter acquisition method: w ij is obtained by training with historical anomaly propagation data; τ ij is calculated according to the line distance and propagation speed; λ is determined by experimental measurement, and the value range is 0.1 to 0.5; the range of ξ is ±0.05.

[0104] 4. Construction process of the feature coupling equation:

[0105] The coupling coefficient matrix M is expressed as:

[0106] In the formula, M ij represents the coupling coefficient between the i-th regular feature and the j-th anomaly feature.

[0107] The finally formed equation:

[0108] Parameter acquisition method: M ij is obtained by neural network training; μ is determined by time series analysis, and the range is 0.01 to 0.1; ρ is determined by perturbation experiment, and the range is 0.1 to 0.3; the range of v is ±0.03.

[0109] 5. Construction process of the stability equation:

[0110] Singular value decomposition of the feature matrix S: S = UΣV T ;

[0111] In the formula, Σ is a diagonal matrix of singular values.

[0112] The state transition probability matrix P is expressed as:

[0113]

[0114] In the formula, P ij represents the probability of transferring from state i to state j.

[0115] The finally formed equation:

[0116] Parameter acquisition method: σ i is obtained by singular value decomposition of the feature matrix; κ is obtained by matrix operation; θ is optimized by experiment, and the value range is 0.5 to 2; the range of ω is ±0.01.

[0117] Specifically, the principle of the present invention is as follows: The technical principle of the present invention is based on the idea of hierarchical feature extraction and in-depth analysis. At the meter end, an improved lightweight neural network is used for feature extraction. This network reduces the number of parameters through depthwise separable convolution, highlights important features using a channel attention mechanism, and effectively combines multi-scale features through an adaptive feature fusion module. This design not only ensures the accuracy of feature extraction but also meets the requirements of limited computing resources at the meter end. The adaptive sampling mechanism triggered by abnormal features realizes fine observation of abnormal working conditions by adjusting the sampling frequency, providing a data basis for accurately analyzing the reasons for abnormalities.

[0118] At the monitoring end, the design of the hybrid neural network combines the advantages of deep learning and physical models. The feature encoding layer uses a multi-head self-attention mechanism, which can capture the correlation between different features. The equation embedding layer introduces the physical laws of the energy meter operation into the model by constructing regular feature change equations, abnormal feature propagation equations, feature coupling equations, and stability equations, improving the interpretability of the analysis results. The time series analysis layer uses a bidirectional long short-term memory network and residual connections to effectively extract the long-term dependencies in the time series features. The feature decoding layer maps the features back to the original space through a deconvolution network for intuitive understanding of the results.

[0119] The technical solution of the present invention is logical because it follows the design principle of "hierarchical processing, local computing, and deep fusion". By adopting adaptive algorithms at different levels, it not only ensures the processing efficiency but also realizes in-depth analysis. At the same time, the introduction of the constraints of the physical model ensures the consistency between the analysis results and the actual operation rules of the power grid.

[0120] A specific Embodiment 1 of the present invention is provided below, and the specific implementation of each step in this Embodiment 1 is described in detail as follows.

[0121] In this Embodiment 1, the power acquisition unit of the energy meter end monitoring system adopts a precise sampling circuit design, which includes a high-precision operational amplifier, an analog switch, and a sample and hold circuit. The sampling frequency is set to 3200 Hz, and the sampling accuracy reaches 0.2 level. It can accurately capture the rising and falling edges of the power pulse signal to ensure the accuracy of power measurement. The input impedance of the sampling circuit is greater than 10 MΩ, the common mode rejection ratio is greater than 100 dB, and the signal-to-noise ratio is greater than 80 dB.

[0122] The current sampling unit is implemented by combining a Hall sensor with a signal conditioning circuit. The linear range of the Hall sensor is from 0.1 ampere to 100 amperes. The signal conditioning circuit includes a preamplifier, a band-pass filter, and an analog-to-digital converter. The gain of the preamplifier is adjustable, with a gain range of 1 to 1000 times. The cut-off frequency range of the band-pass filter is from 45 Hz to 65 Hz, and the attenuation rate is greater than 40 dB per decade. The sampling rate of the analog-to-digital converter is 6400 Hz, the resolution is 16 bits, and the integral non-linearity error is less than ±2 least significant bits.

[0123] The voltage sampling unit is implemented by means of resistor voltage division and isolation amplification. The voltage division ratio is 1000:1. The isolation amplifier uses an opto-coupling method, with a common-mode rejection ratio greater than 80 dB and an isolation voltage level reaching 4000 volts. The sampled signal is subjected to anti-aliasing filtering and then 16-bit analog-to-digital conversion. The cut-off frequency of the filter is 3200 Hz, the stop-band rejection is greater than 60 dB, and the pass-band ripple is less than 0.1 dB.

[0124] The temperature sensing unit selects a digital temperature sensor, with a measurement range from -40 degrees Celsius to 85 degrees Celsius, a measurement accuracy of ±0.5 degrees Celsius, a sampling period of 1 second, communicates with the first control chip through a serial interface, with a communication rate of 100 kHz, supports 16-bit temperature data output, has an over-temperature alarm function, and the alarm threshold can be set.

[0125] The first control chip uses a 32-bit microcontroller, with a main frequency of 120 MHz, built-in 2 megabytes of flash memory and 256 kilobytes of random access memory, integrated 12-channel 16-bit analog-to-digital converter, hardware multiplier, and floating-point arithmetic unit, supports a real-time operating system, has a watchdog timer and power-down protection function, and the operating temperature range is from -20 degrees Celsius to 70 degrees Celsius.

[0126] The first feature extraction unit is implemented based on a field-programmable gate array, with a clock frequency of 200 MHz, includes 32 digital signal processing modules, 64 multiplier arrays, and 128 kilobytes of dual-port random access memory, supports parallel data processing and pipelined operations, the processing delay is less than 100 microseconds, and the power consumption is less than 2 watts.

[0127] The data receiving unit is implemented using a power line carrier communication chip, supports quadrature amplitude modulation and quadrature phase shift keying modulation, the carrier frequency range is from 3 kHz to 500 kHz, the communication rate can reach 250 kbps, built-in forward error correction coding and cyclic redundancy check, the error correction ability is to correct 8-bit errors for every 1024-bit data, and has an adaptive channel equalization function.

[0128] The carrier analysis unit is implemented by a digital signal processor with a main frequency of 400 MHz, supports fast Fourier transform and digital filtering, has an 8-channel direct memory access controller built-in, has the ability to process real-time data streams, the order of the digital filter can reach 128 orders, and the filtering accuracy is better than 0.1 dB.

[0129] The main control chip uses a quad-core processor with a main frequency of 1.5 GHz, has an 8-MB level 3 cache built-in, supports the single instruction multiple data stream extended instruction set, integrates a hardware encryption engine and a secure boot mechanism, the power consumption of the processor is less than 10 W, and supports dynamic frequency adjustment and sleep mode.

[0130] The data storage unit uses a solid-state drive with a capacity of 256 GB, uses triple-level stacked flash memory, supports power-off protection and data error correction, the write speed reaches 550 MB per second, the random read and write performance exceeds 100,000 times per second, and supports 256-bit Advanced Encryption Standard encryption.

[0131] The display unit uses a 7-inch liquid crystal display with a resolution of 1024 by 600 pixels, uses a capacitive touch screen, supports multi-touch, the display brightness is adjustable, the brightness range is 50 to 500 candela per square meter, the contrast ratio is 1000:1, the viewing angle is 170 degrees, and the response time is less than 25 milliseconds.

[0132] The depth analysis unit uses a graphics processor, has 384 stream processor cores, the video memory bandwidth is 128 GB per second, supports general computing units, can perform large-scale parallel computing, the maximum computing power reaches 4 trillion floating-point operations per second, and supports single-precision and double-precision floating-point operations.

[0133] The lightweight neural network in the lightweight feature extraction module uses a gated recurrent unit structure. The number of neurons in the input layer is 32, the hidden layer uses a 3-layer structure, each layer contains 64 neurons, the number of neurons in the output layer is 16, the activation function uses the rectified linear unit, the training uses the stochastic gradient descent method, the learning rate is 0.001, and the batch size is 128. The feature encoding layer uses an 8-head attention mechanism, the dimension of each attention head is 64, the dimension of the key-value pair is 512, the position encoding uses the sine-cosine function, the output feature dimension is 512, and the attention weight is calculated by scaled dot product, and the scaling factor is the reciprocal of the square root of 8.

[0134] Each unit is interconnected through a high-speed serial bus. The bus clock frequency is 100 MHz, uses differential signal transmission, has electromagnetic compatibility design, has strong anti-interference ability, and the transmission distance can reach 10 meters.

[0135] The specific implementation of step S11 is that when receiving the power pulse signal, voltage signal, current signal and temperature signal, first transfer the data from the first control chip to the input buffer of the feature extraction unit through the direct memory access method. Adopt the multi-buffer polling mechanism, allocate a circular buffer with a size of 16 kilobytes for each signal. The data frame header contains signal type identification, timestamp and data length information. Ensure data integrity through cyclic redundancy check. The check code uses a 32-bit cyclic redundancy check polynomial. When a data transmission error is detected, trigger the retransmission mechanism, and the maximum number of retransmissions does not exceed 3 times.

[0136] The specific implementation of step S12 is that before performing the fast Fourier transform on the voltage signal and current signal, first preprocess the signal, including removing the DC component, detrending and windowing. The DC component is obtained by calculating the signal mean and subtracting it. The trend term is eliminated after fitting by the least squares method. Select the Hanning window function to reduce spectral leakage. The window length is 1024 points. Then perform the fast Fourier transform:

[0137] In the formula, X(k) is the frequency-domain signal, x(n) is the time-domain signal, N is the number of sampling points, k is the frequency index, n is the time index, j is the imaginary unit. Perform in-place calculation through the radix-4 butterfly operation unit. The calculated spectrum is processed by amplitude normalization. Determine the frequency points of each harmonic according to the fundamental frequency, and extract the amplitude and phase information of the fundamental component and the first 5 harmonic components. The fundamental frequency is obtained by the interpolation-type frequency estimation algorithm, the frequency resolution is better than 0.1 Hz, the amplitude measurement accuracy is better than 0.2%, and the phase measurement accuracy is better than 0.2 degrees.

[0138] The specific implementation of step S13 is to use an improved lightweight neural network to extract features from the fundamental component and each harmonic component. First, perform time-frequency domain feature extraction on the input signal through the feature extraction layer. The convolution kernel size is 3×3, the stride is 1, and the padding method is same padding. Extract 32 feature maps. Then pass through the channel attention layer. Use global average pooling and global maximum pooling in parallel to obtain the channel descriptor. Generate the channel weight through two fully connected layers and the softmax function, and recalibrate the feature maps. Then pass through the depthwise separable convolution layer. First perform depthwise convolution for each channel, and then perform pointwise convolution to reduce the computational complexity. The convolved features pass through the adaptive feature fusion layer, which adopts a multi-scale feature pyramid structure. The calculation of adaptive feature fusion is as follows: W i +ζ;

[0139] In the formula, O is the fusion output, H i is the feature map of the i-th layer, W i is the corresponding weight, α iis the adaptive coefficient, L is the number of feature layers, ⊙ is the Hadamard product, ζ is the residual term. Different scales of context information are obtained through spatial pyramid pooling, and the pooling scales include 1×1, 2×2, and 4×4. Feature fusion adopts the weighted summation method, and the weight coefficients are adaptively learned through the soft attention mechanism.

[0140] The specific implementation of step S14 is that when calculating the feature deviation according to the regular features, the sliding window method is adopted. The window length is 60 seconds and the sliding step is 1 second. The calculation formula is as follows:

[0141] In the formula, d(t) is the feature deviation, f i (t) is the value of the i-th regular feature at time t, n is the dimension of the regular features. Statistical analysis is performed on the feature deviation sequence to calculate statistics such as mean, variance, skewness, and kurtosis. A probability distribution model of the feature deviation is established based on historical data, and the probability density function of the deviation distribution is obtained by using the kernel density estimation method. The warning threshold is set to 3 times the standard deviation. When the feature deviation exceeds the warning threshold, it is determined as an abnormal state, and abnormal features are generated. The abnormal features include three indicators: harmonic distortion rate, transient disturbance, and temperature fluctuation. The warning threshold of the harmonic distortion rate is 5%, the warning threshold of the transient disturbance is 20% of the rated value, and the warning threshold of the temperature fluctuation is 5 degrees Celsius per minute.

[0142] The specific implementation of step S15 is that when outputting the regular features and abnormal features to the first control chip, a data compression and encrypted transmission mechanism is adopted. First, the feature data is compressed and encoded. The combination of run-length encoding and differential encoding is used. The regular features are represented by 16-bit fixed-point numbers, and the abnormal features are represented by 32-bit floating-point numbers. The compressed data is encrypted through the Advanced Encryption Standard algorithm, and the key length is 256 bits. The encrypted data is transmitted in blocks, and the size of each data block is 4 kilobytes. The data block contains the feature type, timestamp, compression flag, and checksum, and is transmitted to the first control chip through the Serial Peripheral Interface at a transmission rate of 10 megabits per second. The stop-and-wait protocol is used to ensure reliable data transmission. The receiving end verifies the data integrity through cyclic redundancy check, and requests retransmission for the data blocks with checksum failure.

[0143] The specific implementation of step S21 is that when receiving the regular features and abnormal features of multiple electricity meters transmitted by the main control chip, first establish a feature cache pool, allocate an independent feature buffer area for each electricity meter, the buffer area size is 32 megabytes, managed by the first-in-first-out strategy, the feature data is indexed according to the timestamp and electricity meter identification, and a three-dimensional tensor structure is adopted when constructing the feature tensor. The first dimension represents the electricity meter number, the second dimension represents the time series, and the third dimension represents the feature dimension. The tensor size is the number of electricity meters multiplied by the time step multiplied by the feature dimension. The time step is set to 144 points, corresponding to the sampling data of 24 hours, and the sampling interval is 10 minutes. The feature dimension includes regular features and abnormal features. The regular feature matrix is expressed as:

[0144] In the formula, F ij respectively represent the eigenvalue of the fundamental voltage, fundamental current, and power factor. The abnormal feature matrix is expressed as:

[0145] In the formula, A ij respectively represent the eigenvalue of the harmonic distortion rate, transient disturbance, and temperature fluctuation. The total dimension is 12, and the tensor data type is 32-bit floating point. The missing data is filled by the linear interpolation method, and the outliers are corrected by the moving median method.

[0146] The specific implementation of step S22 is that when constructing the hybrid neural network structure, the feature encoding layer adopts the multi-head self-attention mechanism, the number of attention heads is 8, and the dimension of each attention head is 64. The input is linearly mapped to obtain the query matrix, key matrix, and value matrix, and the matrix dimension is 512. The attention weights are calculated by the scaled dot-product attention, and weight normalization and residual connection are adopted. The position encoding uses the sine and cosine functions, and the maximum sequence length supported is 1000. The regular feature processing unit in the equation embedding layer uses a fully connected layer to extract the feature representation, and the hidden layer dimensions are 256, 128, and 64. Batch normalization and dropout regularization are used, and the dropout rate is 0.3. The abnormal feature processing unit uses a gated recurrent unit network with parameter sharing, and the hidden state dimension is 128. Gradient clipping is used to prevent gradient explosion, and the clipping threshold is 5. The dynamic weight layer of the feature coupling processing unit adaptively learns the feature weights through the attention mechanism. The feature coupling coefficient matrix is expressed as:

[0147] In the formula, M ij represents the coupling coefficient between the i-th regular feature and the j-th abnormal feature.

[0148] The specific implementation of step S23 is that when using the equation embedding layer of the hybrid neural network to process the feature tensor, the regular feature change equation is expressed as: F t =αF t-1 +βT l+γL r +δT r +ηV r +∈;

[0149] Wherein, F t is the regular feature at the current moment, F t-1 is the regular feature at the previous moment, T l is the running duration, L r is the load change rate, T r is the temperature change rate, V r is the voltage volatility, α, β, γ, δ, η are weight coefficients, ∈ is the random error term, and the abnormal feature propagation equation is expressed as:

[0150] Wherein, A i (t) is the abnormal feature of node i at time t, N(i) is the neighbor set of node i, w ij is the influence weight between nodes, τ ij is the propagation delay, Z ij is the line impedance, λ is the attenuation coefficient, ξ is the noise term, and the feature coupling equation is expressed as:

[0151]

[0152] Wherein, C(t) is the coupling feature, μ is the time decay factor, D(t) is the perturbation intensity, i is the perturbation coefficient, v is the random fluctuation term, and the stability equation is expressed as:

[0153] Wherein, S is the stability index, σ i is the singular value, k is the number of singular values, κ is the condition number, P is the state transition probability, θ is the smoothing factor, and ω is the system noise.

[0154] The specific implementation of step S24 is that when the output of the equation embedding layer passes through the time series analysis layer, a bidirectional long short-term memory network is used to extract time series features. The network contains hidden layers in the forward and backward directions. The dimension of a single hidden layer is 256, and the time step is 144, corresponding to 24-hour time series data. The information flow is controlled through a gating mechanism, including an input gate, a forgetting gate, and an output gate. The parameters of the gates are learned through the backpropagation algorithm, and residual connections are used to prevent gradient disappearance. The extracted time series features are expressed as:

[0155] Wherein, is the hidden state vector at time t, n is the feature dimension, which is used to construct the state transition graph. The state transition graph adopts a directed weighted graph structure, where the nodes represent the meter states and the edges represent the state transition probabilities. The state transition probability matrix is expressed as:

[0156] In the formula, P ij represents the probability of transitioning from state i to state j. The states include four types: normal operation, mild anomaly, moderate anomaly, and severe anomaly. The state transition probability is obtained by statistical analysis of a sliding time window with a window length of 24 hours and a sliding step of 1 hour.

[0157] When the specific implementation of step S25 is to use the feature decoding layer to convert the state transition graph into the operation state evaluation results of each electric meter, a 5-layer transposed convolutional network is adopted. Each layer includes a transposed convolution, batch normalization, and an activation function. The size of the transposed convolution kernel is 3×3, the stride is 2, and the number of channels is 256, 128, 64, 32, and 16 in sequence. The same padding is used to maintain the size of the feature map. The activation function uses the rectified linear unit. The last layer outputs the state probability distribution through the softmax function. The state probability vector is expressed as:

[0158] In the formula, p i represents the probability value of the i-th state. The probability value is normalized through the softmax function. The operation state evaluation results include the operation state level, anomaly degree score, and state prediction confidence. The operation state level is divided into 4 levels, corresponding to normal, mild anomaly, moderate anomaly, and severe anomaly respectively. The anomaly degree score ranges from 0 to 100, and the state prediction confidence ranges from 0 to 1.

[0159] When the specific implementation of step S26 is to generate warning information based on the operation state evaluation results, a multi-level warning mechanism is adopted. The warning level corresponds to the operation state level. Different trigger conditions and processing strategies are set for each warning level. For mild anomalies, a warning is triggered when the anomaly duration exceeds 1 hour or the anomaly degree score exceeds 60. The warning information is expressed as: W1 = {t1, α1, R1, L1};

[0160] In the formula, t1 is the anomaly occurrence time, α1 is the anomaly type, R1 is the influence range, and L1 is the warning level. For moderate anomalies, a warning is triggered when the anomaly duration exceeds 30 minutes or the anomaly degree score exceeds 75. The warning information adds the anomaly development trend T2 and the processing suggestion S2. For severe anomalies, a warning is triggered when the anomaly duration exceeds 10 minutes or the anomaly degree score exceeds 90. The warning information adds the emergency treatment plan E3 and the impact assessment I3. The warning information is sent to the display unit through a message queue, and a real-time push mechanism is adopted, supporting SMS and email notifications. The warning records are saved in the data storage unit with a storage period of 180 days. Each warning record contains information such as the timestamp, device identifier, warning level, warning content, and processing status.

[0161] To better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: In the R & D project of the intelligent electricity meter remote monitoring system carried out by the intelligent grid research institute of a certain power company, the research team implemented and tested the remote monitoring system for 10,000 intelligent electricity meters in the jurisdiction for one year. The project selected the intelligent electricity meters distributed in urban commercial areas, residential areas, and industrial parks as the monitoring objects, among which the commercial area accounted for 40%, the residential area accounted for 35%, and the industrial park accounted for 25%. The specific distribution of the electricity meters is shown in Table 1:

[0162] Table 1 Statistical table of the regional distribution of intelligent electricity meters

[0163]

[0164] Figure 4 The 24-hour load characteristic curves of the electricity meters in different regions are shown. The horizontal axis is time (hours), and the vertical axis is the load rate (the ratio of the actual power to the rated power). During the system implementation process, the research team adopted a feature extraction method based on an improved lightweight neural network. The specific parameter configuration of this neural network is shown in Table 2:

[0165] Table 2 Parameter configuration table of the improved lightweight neural network

[0166] Network layer Parameter configuration Activation function Output dimension Input layer 32 neurons - 32 Hidden layer 1 64 neurons ReLU 64 Hidden layer 2 64 neurons ReLU 64 Hidden layer 3 64 neurons ReLU 64 Output layer 16 neurons Softmax 16

[0167] The system uses the fast Fourier transform for signal processing, and the basic transformation formula is:

[0168]

[0169] Among them, the number of sampling points N is set to 1024, and the sampling frequency is 3200 Hz. The system conducts harmonic analysis on voltage and current signals and extracts the fundamental wave and the first 5 harmonic components. The harmonic measurement accuracy is shown in Table 3:

[0170] Table 3 Harmonic measurement accuracy index table

[0171] Harmonic order Amplitude accuracy Phase accuracy Fundamental wave ±0.2％ ±0.2° Second harmonic ±0.3％ ±0.3° Third harmonic ±0.3％ ±0.3° Fourth harmonic ±0.4％ ±0.4° Fifth harmonic ±0.4％ ±0.4°

[0172] Figure 5 The harmonic analysis results of voltage and current are shown, including the percentage content of the fundamental wave and the 2nd - 5th harmonics. During the feature extraction process, the system adopts the regular feature change equation:

[0173] F t = αF t-1 + βF l + γL r + δT r + ηV r + ∈;

[0174] The optimized values of each parameter are shown in Table 4 as follows:

[0175] Table 4 Parameter Table of Regular Feature Variation Equation

[0176] Parameter Value Description α 0.85 Historical feature weight coefficient β 0.03 Running duration weight coefficient γ 0.05 Load change rate weight coefficient δ 0.04 Temperature change rate weight coefficient η 0.03 Voltage volatility weight coefficient

[0177] During the actual operation of the system, the propagation characteristics of abnormal features were modeled, and the following propagation equation was adopted:

[0178] During the one-year operation, the statistical data of various abnormal events detected by the system are shown in Table 5:

[0179] Table 5 Statistical Table of Abnormal Events

[0180] Abnormality type Occurrence times Average duration Detection rate Voltage abnormality 856 15 minutes 99.2％ Current abnormality 1243 22 minutes 98.7％ Harmonic exceeding standard 678 45 minutes 97.8％ Temperature abnormality 342 35 minutes 99.5％ Power factor abnormality 456 28 minutes 98.9％

[0181] The system adopts a multi-level early warning mechanism, and the triggering conditions and processing strategies for different early warning levels are shown in Table 6:

[0182] Table 6 Table of Early Warning Levels and Processing Strategies

[0183] Warning level Trigger condition Response time Processing strategy Level 1 Abnormality score > 90 or duration > 10 minutes 5 minutes Emergency processing Level 2 Abnormality score > 75 or duration > 30 minutes 15 minutes Priority processing Level 3 Abnormality score > 60 or duration > 60 minutes 30 minutes Routine processing Level 4 Abnormality score > 45 or duration > 120 minutes 60 minutes Scheduled processing

[0184] During the implementation process, the research team focused on the stability assessment of the system and adopted the following stability equation:

[0185] The optimized values of the parameters of this equation after actual operation are shown in Table 7:

[0186] Table 7 Parameter Table of Stability Equation

[0187] Parameter Value range Optimized value k 5-10 8 κ 100-200 150 θ 0.5-2.0 1.5 ω ±0.01 0.008

[0188] Figure 6 Shows the stability assessment results during the one-year operation of the system, including the actual stability index curve and the threshold line. The performance indicators of the system during actual operation are shown in Table 8:

[0189] Table 8 Table of System Performance Indicators

[0190] Performance indicator Design goal Actual achieved value Data acquisition accuracy ≥99.5％ 99.8％ Abnormality detection accuracy ≥95％ 97.2％ Warning response time ≤ 5 minutes 3.5 minutes System stability ≥99％ 99.6％ Fault prediction accuracy ≥90％ 93.5％

[0191] The traditional remote monitoring of electric energy meters mainly has the following technical problems: 1. Using a fixed sampling frequency, it cannot be dynamically adjusted according to the real-time state, resulting in a large amount of data and insufficient accuracy. 2. Traditional statistical methods are used for feature extraction, and complex fault patterns cannot be effectively identified. 3. The abnormal propagation model is too simple and does not consider the network topology and spatio-temporal correlation. 4. The early warning mechanism is single and cannot achieve multi-level classification and accurate early warning. 5. The system stability evaluation method is backward and lacks theoretical support.

[0192] The solution in this Embodiment 2 has the following improvements compared with the traditional technology: 1. Adaptive sampling technology is adopted to dynamically adjust the sampling frequency according to abnormal features, reducing the amount of data while ensuring data quality, and the sampling efficiency is increased by 45%. 2. An improved lightweight neural network is introduced for feature extraction. Compared with the traditional method, the feature extraction accuracy is increased by 15%, and the calculation efficiency is increased by 60%. 3. An abnormal propagation model considering network topology and spatio-temporal characteristics is established, and the abnormal propagation prediction accuracy is increased from the original 85% to 97%. 4. A multi-level early warning mechanism based on multi-dimensional features is realized, the false alarm rate is reduced by 75%, and the missed alarm rate is reduced by 85%. 5. A system stability evaluation method based on singular value decomposition is proposed, and the evaluation accuracy is increased by 25%.

[0193] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Table 9 below.

[0194] Table 9 Variable Explanation Table

[0195]

[0196] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention.

Claims

1. A remote monitoring device for intelligent electric energy meters, characterized in that, It includes several electricity meter end monitoring systems and a remote monitoring end system. The electricity meter end monitoring system includes an electric energy acquisition unit, a current sampling unit, a voltage sampling unit, a temperature sensing unit, a first control chip, and a first feature extraction unit disposed inside the electricity meter. The remote monitoring end system includes a data receiving unit, a carrier wave analysis unit, a main control chip, a data storage unit, a display unit, and a depth analysis unit. A lightweight feature extraction module is provided inside the first feature extraction unit. The lightweight feature extraction module includes a feature extraction layer, a channel attention layer, a depthwise separable convolution layer, an adaptive feature fusion layer, and a fully connected layer. A parallel processing module is provided inside the depth analysis unit. The parallel processing module includes a regular feature change equation, an abnormal feature propagation equation, a feature coupling equation, and a stability equation. The regular feature consists of a voltage fundamental wave, a current fundamental wave, and a fundamental power factor. The abnormal feature consists of a harmonic distortion rate, a transient disturbance, and a temperature fluctuation. The feature deviation is the Euclidean distance between two adjacent sampling values of the regular feature.

2. The remote monitoring device for intelligent electric energy meters according to claim 1, characterized in that The electric energy acquisition unit adopts a precision sampling circuit design with a sampling frequency of 3200 Hz and a sampling accuracy reaching 0.2 level. The current sampling unit is implemented by combining a Hall sensor with a signal conditioning circuit. The linear range of the Hall sensor is from 0.1 A to 100 A. The voltage sampling unit is implemented by means of resistor voltage division and isolation amplification with a voltage division ratio of 1000:

1. The temperature sensing unit selects a digital temperature sensor with a measurement range from -40 °C to 85 °C and a measurement accuracy of ±0.5 °C.

3. The remote monitoring device for intelligent electricity meters according to claim 1, characterized in that The first control chip adopts a 32-bit microcontroller with a main frequency of 120 MHz. The first feature extraction unit is implemented based on a field programmable gate array with a clock frequency of 200 MHz. The data receiving unit is implemented by using a power line carrier communication chip with a carrier frequency range from 3 kHz to 500 kHz. The carrier wave analysis unit is implemented by using a digital signal processor with a main frequency of 400 MHz.

4. The remote monitoring device for intelligent electric energy meters according to claim 1, wherein The main control chip adopts a quad-core processor with a main frequency of 1.5 GHz. The data storage unit uses a solid state drive with a capacity of 256 GB. The display unit uses a 7-inch liquid crystal display with a resolution of 1024×600 pixels. The depth analysis unit uses a graphics processing unit with 384 stream processor cores.

5. The remote monitoring device for intelligent electric energy meters according to claim 1, characterized in that, The lightweight feature extraction module is used to perform the following steps: receive the power pulse signal, the voltage signal, the current signal, and the temperature signal output by the first control chip, perform fast Fourier transform on the voltage signal and the current signal, extract the fundamental component, the first harmonic component, the second harmonic component, the third harmonic component, the fourth harmonic component, and the fifth harmonic component, use the improved lightweight neural network to extract features from the fundamental component and the harmonic components to obtain the regular features, calculate the feature deviation according to the regular features, generate the abnormal features when the feature deviation exceeds the preset threshold, and output the regular features and the abnormal features to the first control chip.

6. The remote monitoring device for intelligent electric energy meters according to claim 1, characterized in that, The input of the regular feature change equation includes the historical regular feature sequence, the operation duration of the electricity meter, the load change rate, the ambient temperature change rate, and the grid voltage volatility, and the output is the predicted value of the regular features at the next moment.

7. The remote monitoring device for intelligent electric energy meters according to claim 1, characterized in that, The input of the abnormal feature propagation equation includes the abnormal feature vectors of adjacent electric meters, the line impedance parameters, the propagation delay coefficient, and the load similarity, and the output is the propagation influence range and the propagation influence intensity of the abnormal features.

8. The remote monitoring device for intelligent electric energy meters according to claim 1, characterized in that, The input of the feature coupling equation includes the regular feature vector, the abnormal feature vector, the coupling coefficient matrix, the feature correlation degree, the time scale parameter, and the perturbation intensity, and the output is the coupled comprehensive feature.

9. The remote monitoring device for intelligent electricity meters according to claim 1, wherein The input of the stability equation includes the singular value sequence of the feature matrix, the eigenvalue distribution, the matrix condition number, the time series stationarity index, and the state transition probability, and the output is the system stability evaluation index.

10. The remote monitoring device for intelligent electric energy meters according to claim 1, characterized in that, The adaptive feature fusion layer is used to dynamically adjust the weights of different feature channels and achieve the adaptive fusion of multi-scale features. The input of the adaptive feature fusion layer is the multi-channel feature map output by the depthwise separable convolution layer, and the output is the fused feature vector.

Images