A remote monitoring device for intelligent electric energy meter

Abstract

The application provides a kind of intelligent electric energy meter remote monitoring device, belongs to intelligent electric energy meter technical field, the intelligent electric energy meter remote monitoring device proposed in the application includes electric meter end monitoring system and remote monitoring end system.Electric meter end uses electric energy acquisition unit, current sampling unit, voltage sampling unit and temperature sensing unit to collect data, carries out feature extraction through lightweight feature extraction module, realizes adaptive sampling based on abnormal characteristics.Remote monitoring end carries out deep analysis using hybrid neural network, including four levels of feature encoding, equation embedding, time series analysis and feature decoding, realizes the flow processing of multiple electric meter data through parallel processing module, and finally generates operation state evaluation result and early warning information.It solves the technical problems existing in the prior art that it is difficult to monitor multiple intelligent electric energy meter operation states in real time and efficiently and discover potential abnormalities.

Description

Technical Field

[0001] This invention belongs to the field of smart energy meter technology, and more specifically, relates to a remote monitoring device for smart energy meters. Background Technology

[0002] Smart meters are key devices in modern power systems for collecting and managing electricity consumption information. Traditional smart meter monitoring relies primarily on periodic data collection and offline analysis, employing a centralized data processing architecture. This involves periodically reading electricity consumption data, voltage and current data, and power factor data from the meters via acquisition terminals and transmitting them to the main station system for analysis. This monitoring method uses simple threshold judgment and statistical analysis methods to perform basic rationality checks and anomaly detection on the collected data, thereby monitoring the basic operating status of the smart meters.

[0003] However, with the continuous expansion of smart grids and the increasing complexity of electricity loads, traditional electricity meter monitoring technologies face numerous challenges. First, traditional periodic data collection methods use a fixed sampling frequency, failing to adaptively adjust sampling strategies based on meter operating conditions. This results in the inability to acquire high-frequency sampling data in a timely manner when anomalies occur, affecting the accuracy of anomaly analysis. Second, traditional centralized data processing methods require transmitting large amounts of raw data to the main station, putting pressure on communication bandwidth and incurring significant processing delays, making it difficult to meet real-time monitoring requirements. Third, traditional simple threshold judgment methods can only detect obvious anomalies, lacking effective means to identify complex anomalies such as gradual faults and multi-device coupled faults. Finally, traditional technologies lack the ability to comprehensively analyze the operating characteristics of a group of electricity meters, failing to uncover the correlations between multiple meters and hindering early warning of group anomalies.

[0004] Faced with the ever-increasing demand for electricity meter monitoring, traditional technologies struggle to address the challenge of efficient real-time monitoring of smart meters. How to achieve real-time monitoring of meter operating status, timely identification of abnormal features, multi-meter correlation analysis, and early warning has become a pressing technical challenge. Especially in scenarios involving large-scale smart meter deployments, a novel monitoring solution is needed that supports distributed feature extraction, adaptive sampling, and parallel processing to meet the requirements of real-time performance, accuracy, and scalability. In other words, existing technologies suffer from limitations in achieving efficient real-time monitoring of the operating status of multiple smart meters and detecting potential anomalies. Summary of the Invention

[0005] In view of this, the present invention provides a remote monitoring device for smart energy meters, which can solve the technical problem in the prior art that it is difficult to monitor the operating status of multiple smart energy meters in real time and efficiently and to detect potential anomalies.

[0006] This invention is implemented as follows: This invention provides a remote monitoring device for a smart electricity meter, comprising several meter-end monitoring systems and a remote monitoring terminal system. The meter-end monitoring system includes an 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, all installed inside the electricity meter. The remote monitoring terminal system includes a data receiving unit, a carrier resolution unit, a main control chip, a data storage unit, a display unit, and a deep analysis unit. The first feature extraction unit includes a lightweight feature extraction module, which comprises a feature extraction layer, a channel attention layer, a depthwise separable convolutional layer, an adaptive feature fusion layer, and a fully connected layer. The deep analysis unit includes a parallel processing module, which comprises a regular feature change equation, an abnormal feature propagation equation, a feature coupling equation, and a stability equation. The regular features consist of the voltage fundamental wave, the current fundamental wave, and the fundamental power factor. The abnormal features consist of harmonic distortion rate, transient disturbance, and temperature fluctuation. The feature deviation is the Euclidean distance between two adjacent sampled values ​​of the regular feature.

[0007] The power acquisition unit employs a precision sampling circuit design with a sampling frequency of 3200 Hz and a sampling accuracy of 0.2%. The current sampling unit utilizes a Hall sensor combined with a signal conditioning circuit, with the Hall sensor having a linear range of 0.1 amperes to 100 amperes. The voltage sampling unit employs a resistor voltage divider and isolation amplification method, with a voltage division ratio of 1000:1. The temperature sensing unit uses a digital temperature sensor with a measurement range of -40 degrees Celsius to 85 degrees Celsius and a measurement accuracy of ±0.5 degrees Celsius.

[0008] The first control chip is 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 using a power line carrier communication chip with a carrier frequency range of 3 kHz to 500 kHz; and the carrier parsing unit is implemented using a digital signal processor with a main frequency of 400 MHz.

[0009] The main control chip uses a quad-core processor with a clock speed of 1.5 GHz, the data storage unit uses a solid-state drive with a capacity of 256 gigabytes, the display unit uses a 7-inch LCD screen with a resolution of 1024 x 600 pixels, and the depth analysis unit uses a graphics processor with 384 stream processor cores.

[0010] The lightweight feature extraction module is used to perform the following steps: receiving the power pulse signal, voltage signal, current signal, and temperature signal output by the first control chip; performing a fast Fourier transform on the voltage signal and the current signal to extract the fundamental component and the first harmonic component, the second harmonic component, the third harmonic component, the fourth harmonic component, and the fifth harmonic component; using the improved lightweight neural network to extract features from the fundamental component and the harmonic components to obtain the regular features; calculating the feature deviation based on the regular features; generating the abnormal features when the feature deviation exceeds the preset threshold; and outputting the regular features and the abnormal features to the first control chip.

[0011] The inputs to the regularity feature change equation include historical regularity feature sequence, electricity meter running time, load change rate, ambient temperature change rate, and grid voltage fluctuation rate, and the output is the predicted value of the regularity feature at the next moment.

[0012] The inputs to the abnormal feature propagation equation include the abnormal feature vectors of adjacent meters, line impedance parameters, propagation delay coefficients, and load similarity. The outputs are the propagation influence range and intensity of the abnormal features.

[0013] The inputs to the feature coupling equation include the regular feature vector, the abnormal feature vector, the coupling coefficient matrix, the feature correlation, the time scale parameter, and the perturbation intensity. The output is the coupled comprehensive feature.

[0014] The inputs to the stability equation include 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] The adaptive feature fusion layer is used to dynamically adjust the weights of different feature channels and achieve 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 convolutional layer, and the output is the fused feature vector.

[0016] Compared with existing technologies, this invention provides a remote monitoring device for smart energy meters, proposing a remote monitoring scheme for smart energy meters based on lightweight feature extraction and deep analysis. This scheme adopts a layered architecture design, deploying a lightweight feature extraction module at the meter end. An improved lightweight neural network is used to extract features from multi-source signals such as voltage, current, and temperature, and the sampling frequency is adaptively adjusted based on abnormal features. At the monitoring end, a hybrid neural network structure is used for deep analysis, achieving parallel processing and correlation analysis of multi-meter features through a multi-head self-attention mechanism, equation embedding layer, and time series analysis layer.

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

[0018] Through the aforementioned technological innovations, this invention successfully solves the technical problem in existing technologies of the difficulty in real-time and efficient monitoring of the operating status of multiple smart meters and the detection of potential anomalies. This solution moves the feature extraction task to the meter end, employs lightweight algorithms to reduce computational overhead, and uses deep analysis methods at the monitoring end to mine multi-meter correlation features, ensuring both real-time and accurate monitoring while achieving system scalability. In particular, the introduction of physical model constraints at the equation embedding layer improves the interpretability and reliability of anomaly identification, providing strong support for the safe and stable operation of the smart grid. Attached Figure Description

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

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

[0021] Figure 3 A flowchart illustrating the steps performed by the parallel processing module.

[0022] Figure 4 This is a 24-hour load characteristic curve of the electricity meters in different areas in Example 2.

[0023] Figure 5 The diagram shows the harmonic analysis results of voltage and current in Example 2.

[0024] Figure 6 This is a graph showing the stability evaluation results of the system during one year of operation in Example 2. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0026] like Figure 1 The diagram shown is a structural schematic of a remote monitoring device for a smart electricity meter provided by the present invention. The device includes several meter-end monitoring systems and a remote monitoring terminal system. The meter-end monitoring system includes an 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, all installed inside the electricity meter.

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

[0028] The power acquisition unit is electrically connected to the power meter metering circuit and is used to acquire power pulse signals and transmit the power pulse signals to the first control chip.

[0029] The current sampling unit is electrically connected to the current sampling circuit of the energy meter, and is used to collect current signals and transmit them to the first control chip after signal conditioning.

[0030] The voltage sampling unit is electrically connected to the voltage sampling circuit of the energy meter, and is used to collect voltage signals and transmit them to the first control chip after signal conditioning.

[0031] The temperature sensing unit is located inside the energy meter and is used to collect the internal temperature of the energy meter and transmit it to the first control chip.

[0032] The first feature extraction unit is electrically connected to the first control chip and is used to extract features from the acquired signal and generate regular and abnormal features.

[0033] The first control chip is used to adjust the sampling frequency according to the abnormal characteristics, and convert the regular characteristics and the abnormal characteristics into carrier signals for transmission through power lines;

[0034] The data receiving unit is used to receive carrier signals transmitted by multiple meter terminal monitoring systems and transmit them to the carrier parsing unit;

[0035] The carrier resolution unit is used to resolve the carrier signal into a digital signal and transmit it to the main control chip;

[0036] The deep analysis unit is electrically connected to the main control chip and is used to perform parallel streaming processing on the regular and abnormal features of multiple meters.

[0037] The first feature extraction unit includes a lightweight feature extraction module, such as... Figure 2 As shown, the lightweight feature extraction module is used to perform the following steps:

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

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

[0040] S13. Using the improved lightweight neural network, feature extraction is performed on the fundamental component and 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 based on the said regularity characteristics, and generate the abnormal feature when the feature deviation exceeds the preset threshold.

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

[0043] The depth analysis unit is equipped with a parallel processing module, such as... Figure 3 As shown, the parallel processing module is used to perform the following steps:

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

[0045] S22. Construct a hybrid neural network structure, wherein the hybrid neural network includes a feature encoding layer, an equation embedding layer, a temporal 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 attention-enhanced feature representations; the equation embedding layer contains four parallel equation processing units, each of which includes a regular feature processing unit, an anomaly feature processing unit, a feature coupling processing unit, and a stability processing unit, wherein the regular feature processing unit, the anomaly 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 temporal analysis layer uses a bidirectional long short-term memory network to extract temporal features and maintains long-term dependencies through residual connections; the feature decoding layer uses a deconvolutional network to map the features back to the original space;

[0046] S23. The feature tensor is processed 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 meters. The coefficient matrix of the feature coupling equation is adaptively generated through the dynamic weight layer, and the input is the regular features and the abnormal features. The evaluation index of the stability equation is calculated through a multilayer perceptron, and the input is the statistics 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 graph;

[0048] S25. The state transition diagram is converted into an operating status evaluation result for each meter using the feature decoding layer.

[0049] S26. Generate early warning information based on the operational status assessment results.

[0050] The parallel processing module includes:

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

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

[0053] The feature coupling equation between the regular features and the abnormal features has the following inputs: the regular feature vector, the abnormal feature vector, the coupling coefficient matrix, the feature correlation, the time scale parameter, and the perturbation intensity. 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. The output is the system stability evaluation index.

[0055] The regular characteristics consist of the fundamental voltage wave, the fundamental current wave, and the fundamental power factor; the abnormal characteristics consist of harmonic distortion rate, transient disturbance, and temperature fluctuation; the characteristic deviation is the Euclidean distance between two adjacent sampled values ​​of the regular characteristics.

[0056] The improved lightweight neural network includes a feature extraction layer, a channel attention layer, a depthwise separable convolutional 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 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 convolutional layer, and the output is the fused feature vector.

[0057] The specific implementation methods of the above steps are described in detail below.

[0058] In a specific implementation, the power acquisition unit of the power meter monitoring system adopts a precision 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 metering.

[0059] The current sampling unit is implemented using a Hall sensor combined with a signal conditioning circuit. The linear range of the Hall sensor is 0.1 ampere to 100 ampere. The signal conditioning circuit includes a preamplifier, a bandpass filter, and an analog-to-digital converter. The preamplifier gain is adjustable, the cutoff frequency range of the bandpass filter is 45 Hz to 65 Hz, and the analog-to-digital converter has a sampling rate of 6400 Hz and a resolution of 16 bits.

[0060] The voltage sampling unit is implemented using resistor voltage division and isolation amplification, with a voltage division ratio of 1000 to 1. The isolation amplifier uses optocoupler, with a common-mode rejection ratio greater than 80 dB and an isolation voltage level of 4000 volts. The sampled signal is then subjected to 16-bit analog-to-digital conversion after anti-aliasing filtering.

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

[0062] The first control chip is a 32-bit microcontroller with a main frequency of 120 MHz. It has 2 megabytes of flash memory and 256 kilobytes of random access memory, and integrates a 12-channel 16-bit analog-to-digital converter, a hardware multiplier, and a floating-point arithmetic unit. It 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 MHz. 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 operation.

[0064] The data receiving unit is implemented using a power line carrier communication chip, which supports quadrature amplitude modulation and quadrature phase shift keying modulation. The carrier frequency range is 3 kHz to 500 kHz, 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 resolution unit is implemented using a digital signal processor with a main frequency of 400 MHz. It supports fast Fourier transform and digital filtering, and has an 8-channel direct memory access controller, enabling real-time data stream processing.

[0066] The main control chip uses a quad-core processor with a clock speed of 1.5 GHz, an 8-megabyte L3 cache, supports single instruction multiple data extended instruction set, and integrates a hardware encryption engine and secure boot mechanism.

[0067] The data storage unit uses a solid-state drive with a capacity of 256 gigabytes, employs triple-layered flash memory, supports power loss protection and data error correction, and achieves a write speed of 550 megabytes per second.

[0068] The display unit uses a 7-inch LCD screen with a resolution of 1024 x 600 pixels. It is a capacitive touch screen that supports multi-touch, adjustable brightness, and automatic dimming.

[0069] The deep analysis unit uses a graphics processor with 384 stream processor cores and a memory bandwidth of 128 gigabytes per second. It supports general-purpose computing units and can perform large-scale parallel computing.

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

[0071] The specific implementation of step S11 is as follows: When receiving power pulse signals, voltage signals, current signals, and temperature signals, a multi-buffer polling data reception mechanism is adopted. A 16-kilobyte circular buffer is allocated separately for each signal. Data is transmitted from the first control chip to the feature extraction unit through direct memory access. The data frame header contains 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, which is increased to 12800 Hz when a rapid change in the signal is detected. An idle frame interval is used between data frames, and the frame interval time is 125 microseconds.

[0072] The specific implementation of step S12 involves preprocessing the voltage and current signals before performing fast Fourier transforms. This preprocessing includes DC removal, trend removal, and windowing. The DC component is obtained by calculating and subtracting the signal mean. The trend term is eliminated by fitting with the least squares method. The Hanning window function is used to reduce spectral leakage, with a window length of 1024 points. To improve computational efficiency, a radix-4 fast Fourier transform algorithm is used, with in-situ calculations performed by a butterfly arithmetic unit. The calculated spectrum is normalized by amplitude. The frequency points of each harmonic are determined based on the fundamental frequency. The amplitude and phase information of the fundamental component and the first five harmonic components are extracted. The fundamental frequency is obtained through an 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] The specific implementation of step S13 involves using an improved lightweight neural network to extract features from the fundamental and harmonic components. The improved lightweight neural network employs a depthwise separable convolutional structure. First, a feature extraction layer extracts time-frequency features from the input signal. The convolutional kernel size is 3x3, the stride is 1, and the padding method is uniform, extracting 32 feature maps. Then, a channel attention layer is used, employing parallel processing of global average pooling and global max pooling to obtain channel descriptors. Channel weights are generated through two fully connected layers and a flexible maximum function, and the feature maps are then re-processed. After calibration, the features are then processed through depthwise separable convolutional layers. First, channel-wise depthwise convolution is performed, followed by pointwise convolution to reduce computational complexity. The convolutional features are then processed through an adaptive feature fusion layer, which employs a multi-scale feature pyramid structure. Spatial pyramid pooling is used to obtain contextual information at different scales, including 1x1, 2x2, and 4x4 pooling scales. Feature fusion is performed using a weighted summation method, with the weight coefficients adaptively learned through a soft attention mechanism. Finally, a fully connected layer maps the features to a regular feature space, yielding regular features such as voltage fundamental frequency, current fundamental frequency, and fundamental frequency power factor.

[0074] The specific implementation of step S14 is as follows: When calculating the feature deviation based on the regular characteristics, a sliding window method is used with a window length of 60 seconds and a sliding step size of 1 second. The Euclidean distance between the regular feature vectors of adjacent sampling points within the window is calculated to obtain the feature deviation sequence. 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 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 judged 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 for harmonic distortion rate is 5%, the warning threshold for transient disturbance is 20% of the rated value, and the warning threshold for temperature fluctuation is 5 degrees Celsius per minute.

[0075] The specific implementation of step S15 involves using a data compression and encrypted transmission mechanism when outputting regular and abnormal features to the first control chip. First, the feature data is compressed and encoded using a combination of run-length encoding and differential encoding. Regular features are represented by 16-bit fixed-point numbers, and abnormal features by 32-bit floating-point numbers. The compressed data is then encrypted using an Advanced Encryption Standard (AES) algorithm with a 256-bit key. The encrypted data is transmitted in blocks, each 4 kilobytes in size, containing the feature type, timestamp, compression flag, and checksum. This data is transmitted to the first control chip via a serial peripheral interface at a rate of 10 megabits per second. A stop-and-wait protocol ensures reliable data transmission. The receiving end verifies data integrity using cyclic redundancy check (CR). Data blocks that fail the check are retransmitted.

[0076] The specific implementation of step S21 is as follows: When receiving the regular and abnormal features of multiple meters transmitted by the main control chip, a feature cache pool is first established, and an independent feature cache area is allocated to each meter. The cache area size is 32 megabytes, and a first-in-first-out (FIFO) strategy is adopted for management. The feature data is indexed according to the timestamp and meter identifier. When constructing the feature tensor, a three-dimensional tensor structure is adopted. The first dimension represents the meter number, the second dimension represents the time series, and the third dimension represents the feature dimension. The tensor size is the number of meters multiplied by the time step multiplied by the feature dimension. The time step is set to 144 points, corresponding to 24 hours of sampling data, and the sampling interval is 10 minutes. The feature dimension includes regular features and abnormal features, with a total dimension of 12. The tensor data type is a 32-bit floating-point number. Missing data is filled using a linear interpolation method, and outliers are corrected using a moving median method to ensure the integrity and continuity of the feature tensor.

[0077] The specific implementation of step S22 involves constructing a hybrid neural network structure. The feature encoding layer employs a multi-head self-attention mechanism with 8 attention heads, each with a dimension of 64. The input undergoes linear mapping to obtain a query matrix, key matrix, and value matrix, with a matrix dimension of 512. Attention weights are calculated using scaled dot product attention, and weight normalization and residual connections are employed. Position encoding uses sine and cosine functions, supporting a maximum sequence length of 1000. The regular feature processing unit in the equation embedding layer uses fully connected layers to extract feature representations. The hidden layer dimensions are 256, 128, and 64, and batch normalization is used. The system employs a unified and random deactivation regularization method. The anomaly feature processing unit uses a parameter-shared gated recurrent unit network with a hidden state dimension of 128. Gradient clipping is used to prevent gradient explosion. The dynamic weight layer of the feature coupling processing unit adaptively learns feature weights through an attention mechanism. The stability processing unit uses a multilayer perceptron to evaluate system stability. The temporal analysis layer uses a bidirectional long short-term memory network with a hidden layer dimension of 256. Residual connections are used to maintain long-term dependencies. The feature decoding layer uses a transposed convolutional network, containing 5 transposed convolutional layers with a kernel size of 3x3, a stride of 2, and the number of channels is halved layer by layer from 256.

[0078] The specific implementation of step S23 is as follows: When processing feature tensors using the equation embedding layer of a hybrid neural network, the regular feature change equation learns parameters through a fully connected layer. The input includes five variables: historical regular feature sequence, runtime, load change rate, ambient temperature change rate, and grid voltage fluctuation rate. Each variable goes through an independent feature extraction branch, and then feature fusion is performed to predict the regular feature value at the next moment. The prediction period is 10 minutes. The abnormal feature propagation equation dynamically updates parameters through a recurrent neural network layer. The input includes abnormal feature vectors 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 through a dynamic weight layer. The input includes regular feature vectors and abnormal feature vectors. Considering feature correlation, time scale parameters, and disturbance intensity, dynamic coupling of features is achieved. The evaluation index of the stability equation is calculated through a multilayer 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. The output is a stability evaluation score.

[0079] The specific implementation of step S24 involves using a bidirectional long short-term memory network to extract temporal features when the output of the equation embedding layer passes through the temporal analysis layer. The network contains hidden layers in both forward and backward directions, with a unidirectional hidden layer dimension of 256 and a time step of 144, corresponding to 24 hours of temporal data. Information flow is controlled through a gating mechanism, including an input gate, a forget gate, and an output gate. The gate parameters are learned through a backpropagation algorithm, and residual connections are used to prevent gradient vanishing. The extracted temporal features are used to construct a state transition graph, which adopts a directed weighted graph structure. Nodes represent the meter status, and edges represent state transition probabilities. The status includes four types: normal operation, mild anomaly, moderate anomaly, and severe anomaly. The state transition probabilities are obtained through a sliding time window with a window length of 24 hours and a sliding step of 1 hour.

[0080] The specific implementation of step S25 is as follows: When converting the state transition map into the operating status assessment result of each meter using the feature decoding layer, a 5-layer transposed convolutional network is used. Each layer contains transposed convolution, batch normalization, and activation function. The transposed convolution kernel size is 3x3, the stride is 2, and the number of channels is 256, 128, 64, 32, and 16 respectively. The same padding is used to maintain the feature map size. The activation function uses a modified linear unit. The last layer outputs the state probability distribution through a flexible maximum function. The state assessment result includes the operating status level, the abnormality score, and the state prediction confidence. The operating status level is divided into 4 levels, corresponding to normal, mild abnormality, moderate abnormality, and severe abnormality, respectively. The abnormality score ranges from 0 to 100. The state prediction confidence represents the reliability of the prediction result.

[0081] The specific implementation of step S26 is as follows: When generating early warning information based on the operational status assessment results, a multi-level early warning mechanism is adopted. The early warning level corresponds to the operational status level. Each early warning level is set with different triggering conditions and handling strategies. For minor anomalies, an early warning is triggered when the duration of the anomaly exceeds 1 hour or the anomaly severity score exceeds 60. The early warning information includes the anomaly type, occurrence time, and impact range. For moderate anomalies, an early warning is triggered when the duration of the anomaly exceeds 30 minutes or the anomaly severity score exceeds 75. The early warning information adds the anomaly development trend and handling suggestions. For severe anomalies, an early warning is triggered when the duration of the anomaly exceeds 10 minutes or the anomaly severity score exceeds 90. The early warning information adds an emergency handling plan and impact assessment. The early warning information is sent to the display unit through a message queue, using a real-time push mechanism, supporting SMS and email notifications. Early warning records are stored in the data storage unit, with a retention period of 180 days.

[0082] The derivation process of some of the equations is described below.

[0083] 1. Derivation of the Fast Fourier Transform Equation:

[0084] Starting with the basic form of the Discrete Fourier Transform:

[0085] The optimization steps include:

[0086] First, the N-point sequence is decomposed into two groups: odd and even.

[0087]

[0088] in, is the rotation factor.

[0089] By recursively decomposing, a radix-4 algorithm structure is finally formed, which can reduce the number of multiplication operations and improve computational efficiency.

[0090] 2. The process of constructing the equation for the change of regular characteristics: The regular characteristic matrix F is represented as:

[0091]

[0092] In the formula, F ij These represent the characteristic values ​​of the fundamental voltage, fundamental current, and power factor, respectively.

[0093] The regular eigenvector f is composed of column vectors of matrix F:

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

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

[0096] Parameter acquisition method: α, β, γ, δ, η are obtained by fitting historical data using the least squares method; T l Obtained through the cumulative runtime counter; L r Calculated based on load changes over 10 minutes; T r Calculated by sampling from a temperature sensor; V r Calculated by voltage sampling; the range of ∈ is ±0.02.

[0097] 3. The process of constructing the anomaly feature propagation equation:

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

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

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

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

[0102] The final equation is:

[0103] Parameter acquisition method: w ij Obtained through training with historical anomaly propagation data; τ ij Calculated based on the line distance and propagation speed; λ is determined experimentally, with a value ranging from 0.1 to 0.5; ξ ranges from ±0.05.

[0104] 4. The process of constructing the characteristic coupling equation:

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

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

[0107] The final equation is:

[0108] Parameter acquisition method: M ij The values ​​were obtained through neural network training; μ was determined through time series analysis, ranging from 0.01 to 0.1; ρ was determined through perturbation experiments, ranging from 0.1 to 0.3; and v ranged from ±0.03.

[0109] 5. The process of constructing the stability equation:

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

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

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

[0113]

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

[0115] The final equation is:

[0116] Parameter acquisition method: σ i The eigenvalues ​​were obtained through singular value decomposition of the feature matrix; κ was obtained through matrix operations; θ was optimized through experiments, with a value range of 0.5 to 2; and ω ranged from ±0.01.

[0117] Specifically, the principle of this invention is based on the concepts of hierarchical feature extraction and deep 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 convolutions, highlights important features using a channel attention mechanism, and achieves effective combination of multi-scale features through an adaptive feature fusion module. This design ensures the accuracy of feature extraction while meeting the limited computing resources required at the meter end. The adaptive sampling mechanism triggered by abnormal features adjusts the sampling frequency to achieve fine observation of abnormal operating conditions, providing a data foundation for accurate analysis of the causes of anomalies.

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

[0119] The technical solution of this invention is logically sound because it follows the design principles of "layered processing, proximity-based computation, and deep fusion." By employing adaptive algorithms at different levels, both processing efficiency and analytical depth are ensured. Simultaneously, the introduction of physical model constraints ensures the consistency of the analysis results with the actual operating patterns of the power grid.

[0120] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0121] In this embodiment 1, the power acquisition unit of the power meter terminal monitoring system adopts a precision 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 metering. The input impedance of the sampling circuit is greater than 10 megohms, 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 using a Hall sensor combined with a signal conditioning circuit. The linear range of the Hall sensor is 0.1 ampere to 100 ampere. The signal conditioning circuit includes a preamplifier, a bandpass filter, and an analog-to-digital converter. The preamplifier gain is adjustable, ranging from 1 to 1000 times. The cutoff frequency range of the bandpass filter is 45 Hz to 65 Hz, with an attenuation rate greater than 40 dB per decade. The analog-to-digital converter has a sampling rate of 6400 Hz, a resolution of 16 bits, and an integral nonlinearity error of less than ±2 least significant bits.

[0123] The voltage sampling unit is implemented using resistor voltage division and isolation amplification, with a voltage division ratio of 1000:1. The isolation amplifier uses optocoupler, with a common-mode rejection ratio greater than 80 dB and an isolation voltage level of 4000 volts. The sampled signal is subjected to 16-bit analog-to-digital conversion after anti-aliasing filtering. The cutoff frequency of the filter is 3200 Hz, the stopband rejection ratio is greater than 60 dB, and the passband ripple is less than 0.1 dB.

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

[0125] The first control chip is a 32-bit microcontroller with a main frequency of 120 MHz. It has 2 megabytes of flash memory and 256 kilobytes of random access memory. It integrates a 12-channel 16-bit analog-to-digital converter, a hardware multiplier, and a floating-point arithmetic unit. It supports a real-time operating system, has a watchdog timer and power-down protection function, and its operating temperature range is -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. It includes 32 digital signal processing modules, 64 multiplier arrays, and 128 kilobytes of dual-port random access memory. It supports parallel data processing and pipelined operation, with a processing latency of less than 100 microseconds and a power consumption of less than 2 watts.

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

[0128] The carrier resolution unit is implemented using a digital signal processor with a main frequency of 400 MHz. It supports fast Fourier transform and digital filtering, and has an 8-channel direct memory access controller with real-time data stream processing capability. The digital filter can reach the order of 128 and the filtering accuracy is better than 0.1 dB.

[0129] The main control chip uses a quad-core processor with a clock speed of 1.5 GHz, an 8-megabyte L3 cache, supports single instruction multiple data extended instruction set, integrates a hardware encryption engine and secure boot mechanism, consumes less than 10 watts of power, and supports dynamic frequency adjustment and sleep mode.

[0130] The data storage unit uses a solid-state drive with a capacity of 256 gigabytes. It employs triple-layered flash memory, supports power loss protection and data error correction, achieves a write speed of 550 megabytes per second, and has a random read / write performance of over 100,000 times per second. It also supports 256-bit encryption as an advanced encryption standard.

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

[0132] The deep analysis unit uses a graphics processor with 384 stream processor cores and a memory bandwidth of 128 gigabytes per second. It supports general-purpose computing units and can perform large-scale parallel computing, with a maximum computing power of 4 trillion floating-point operations per second. It supports single-precision and double-precision floating-point operations.

[0133] The lightweight neural network in the lightweight feature extraction module employs a gated recurrent unit (GRU) structure. The input layer has 32 neurons, the hidden layers have 3 layers (64 neurons per layer), and the output layer has 16 neurons. The activation function is a rectified linear unit (RCU). Training uses stochastic gradient descent with a learning rate of 0.001 and a batch size of 128. The feature encoding layer uses an 8-head attention mechanism, with each attention head having a dimension of 64 and key-value pairs having a dimension of 512. Position encoding uses sine and cosine functions, resulting in an output feature dimension of 512. Attention weights are calculated using a scaled dot product with a scaling factor equal to the reciprocal of the square root of 8.

[0134] The various units are interconnected via a high-speed serial bus with a clock frequency of 100 MHz. It uses differential signal transmission, has electromagnetic compatibility design, strong anti-interference ability, and a transmission distance of up to 10 meters.

[0135] The specific implementation of step S11 is as follows: When receiving power pulse signals, voltage signals, current signals, and temperature signals, the data is first transmitted from the first control chip to the input buffer of the feature extraction unit through direct memory access. A multi-buffer polling mechanism is adopted, and a 16-kilobyte circular buffer is allocated for each signal. The data frame header contains signal type identifier, timestamp, and data length information. Cyclic redundancy check is performed to ensure data integrity. The check code adopts a 32-bit cyclic redundancy check polynomial. When a data transmission error is detected, a retransmission mechanism is triggered, and the number of retransmissions does not exceed 3.

[0136] The specific implementation of step S12 involves preprocessing the voltage and current signals before performing a Fast Fourier Transform (FFT). This preprocessing includes DC removal, trend removal, and windowing. The DC component is obtained by calculating and subtracting the signal mean. The trend term is eliminated by least-squares fitting. A Hanning window function is used to reduce spectral leakage, with a window length of 1024 points. Then, a Fast Fourier Transform is performed.

[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, and j is the imaginary unit. The calculation is performed locally through a radix-4 butterfly operation unit. The calculated spectrum is normalized by amplitude. The frequency points of each harmonic are determined according to the fundamental frequency. The amplitude and phase information of the fundamental component and the first 5 harmonic components are extracted. The fundamental frequency is obtained by an 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.

[0138] The specific implementation of step S13 involves using an improved lightweight neural network to extract features from the fundamental and harmonic components. First, a feature extraction layer extracts time-frequency features from the input signal. The convolution kernel size is 3x3, the stride is 1, and the padding method is the same, extracting 32 feature maps. Then, a channel attention layer is used, employing parallel processing of global average pooling and global max pooling to obtain channel descriptors. Channel weights are generated through two fully connected layers and a flexible maximum function to recalibrate the feature maps. Next, a depthwise separable convolutional layer is used, first performing channel-wise depthwise convolution, then pointwise convolution to reduce computational complexity. The convolved features are then processed by an adaptive feature fusion layer using a multi-scale feature pyramid structure. The calculation of adaptive feature fusion is as follows: W i +ζ;

[0139] In the formula, O represents the fused output, and H represents the fused output. i Let W be the feature map of the i-th layer. i For the corresponding weights, α iζ represents the adaptive coefficient, L represents the number of feature layers, ⊙ represents the Hadamard product, and ζ represents the residual term. Contextual information at different scales is obtained through spatial pyramid pooling, with pooling scales including 1x1, 2x2, and 4x4. Feature fusion adopts a weighted summation method, and the weight coefficients are adaptively learned through a soft attention mechanism.

[0140] The specific implementation of step S14 involves using a sliding window method when calculating the feature deviation based on the regular characteristics. The window length is 60 seconds, the sliding step size is 1 second, and the calculation formula is as follows:

[0141] In the formula, d(t) is the characteristic deviation, and f i (t) represents the value of the i-th regular feature at time t, and n represents the dimension of the regular feature. 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. 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 judged as an abnormal state and an abnormal feature is generated. The abnormal feature includes three indicators: harmonic distortion rate, transient disturbance, and temperature fluctuation. The warning threshold for harmonic distortion rate is 5%, the warning threshold for transient disturbance is 20% of the rated value, and the warning threshold for temperature fluctuation is 5 degrees Celsius per minute.

[0142] The specific implementation of step S15 involves using a data compression and encrypted transmission mechanism when outputting regular and abnormal features to the first control chip. First, the feature data is compressed and encoded using a combination of run-length encoding and differential encoding. Regular features are represented by 16-bit fixed-point numbers, and abnormal features by 32-bit floating-point numbers. The compressed data is then encrypted using an Advanced Encryption Standard (AES) algorithm with a 256-bit key. The encrypted data is transmitted in blocks, each 4 kilobytes in size, containing the feature type, timestamp, compression flag, and checksum. This data is transmitted to the first control chip via a serial peripheral interface at a rate of 10 megabits per second. A stop-and-wait protocol ensures reliable data transmission. The receiving end verifies data integrity using cyclic redundancy check (CR). Data blocks that fail the check are retransmitted.

[0143] The specific implementation of step S21 involves receiving the regular and abnormal features of multiple meters transmitted by the main control chip. First, a feature cache pool is established, allocating an independent feature cache area for each meter. The cache area size is 32 megabytes, managed using a first-in, first-out (FIFO) strategy. Feature data is indexed according to timestamps and meter identifiers. A three-dimensional tensor structure is used to construct the feature tensor: the first dimension represents the meter number, the second dimension represents the time series, and the third dimension represents the feature dimension. The tensor size is the number of meters multiplied by the time step multiplied by the feature dimension. The time step is set to 144 points, corresponding to 24 hours of sampling data, with a sampling interval of 10 minutes. The feature dimension includes regular and abnormal features. The regular feature matrix is ​​represented as follows:

[0144] In the formula, F ij The characteristic values ​​of the fundamental voltage, fundamental current, and power factor are represented respectively, and the abnormal characteristic matrix is ​​expressed as follows:

[0145] In the formula, A ij These represent the characteristic values ​​of harmonic distortion rate, transient disturbance, and temperature fluctuation, respectively. The total dimension is 12, and the tensor data type is a 32-bit floating-point number. Missing data is filled in using linear interpolation, and outliers are corrected using the moving median method.

[0146] The specific implementation of step S22 is as follows: When constructing the hybrid neural network structure, the feature encoding layer adopts a multi-head self-attention mechanism with 8 attention heads, each with a dimension of 64. The input is linearly mapped to obtain a query matrix, a key matrix, and a value matrix with a matrix dimension of 512. Attention weights are calculated by scaling dot product attention, and weight normalization and residual connections are used. Position encoding uses sine and cosine functions, supporting 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 random deactivation regularization are used, with a deactivation rate of 0.3. The abnormal feature processing unit adopts a parameter-shared gated recurrent unit network with a hidden state dimension of 128. Gradient clipping is used to prevent gradient explosion, with a clipping threshold of 5. The dynamic weight layer of the feature coupling processing unit adaptively learns feature weights through the attention mechanism. The feature coupling coefficient matrix is ​​represented as follows:

[0147] In the formula, M ij It 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 processing feature tensors using the equation embedding layer of a hybrid neural network, the regular feature change equation is expressed as: F t =αF t-1 +βT l+γL r +δT r +ηV r +∈;

[0149] In the formula, F t F represents the regular characteristics at the current moment. t-1 As a regular feature of the previous time step, T l L is the runtime. r T represents the load change rate. r V is the rate of temperature change. r Let be the voltage fluctuation rate, α, β, γ, δ, and η be the weighting coefficients, and ∈ be the random error term. The anomaly feature propagation equation is expressed as:

[0150] In the formula, A i Let (t) be the anomalous feature of node i at time t, and N(i) be the set of neighbors of node i. ij τ represents the influence weight between nodes. ij For propagation delay, Z ij Let λ be the line impedance, λ be the attenuation coefficient, and ξ be the noise term. The characteristic coupling equation is expressed as:

[0151]

[0152] In the formula, C(t) represents the coupling characteristic, 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] In the formula, S is the stability index, and σ i κ 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 involves using a bidirectional long short-term memory network to extract temporal features when the output of the equation embedding layer passes through the temporal analysis layer. The network contains hidden layers in both forward and backward directions, with a unidirectional hidden layer dimension of 256 and a time step size of 144, corresponding to 24 hours of time-series data. Information flow is controlled through a gating mechanism, including an input gate, a forget gate, and an output gate. The gate parameters are learned using a backpropagation algorithm, and residual connections are used to prevent gradient vanishing. The extracted temporal features are represented as follows:

[0155] In the formula, Let be the hidden state vector at time t, and n be the feature dimension. This vector is used to construct the state transition graph, which employs a directed weighted graph structure. Nodes represent the meter's state, and edges represent state transition probabilities. The state transition probability matrix is ​​expressed as:

[0156] In the formula, P ij This 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 through a sliding time window with a window length of 24 hours and a sliding step size of 1 hour.

[0157] The specific implementation of step S25 involves using a 5-layer transposed convolutional network when converting the state transition map into the operating status evaluation result for each meter using the feature decoding layer. Each layer contains transposed convolution, batch normalization, and an activation function. The transposed convolution kernel size is 3x3, the stride is 2, and the number of channels is 256, 128, 64, 32, and 16 respectively. The same padding is used to maintain the feature map size. The activation function uses a modified linear unit. The last layer outputs the state probability distribution through a flexible maximum function. The state probability vector is represented as follows:

[0158] In the formula, p i Let represent the probability value of the i-th state. The probability value is normalized by a flexible maximum function. The state assessment results include the operating state level, the degree of abnormality score, and the state prediction confidence. The operating state level is divided into 4 levels, corresponding to normal, mild abnormality, moderate abnormality, and severe abnormality, respectively. The degree of abnormality score ranges from 0 to 100, and the state prediction confidence ranges from 0 to 1.

[0159] The specific implementation of step S26 is to adopt a multi-level early warning mechanism when generating early warning information based on the operation status assessment results. The early warning level corresponds to the operation status level. Each early warning level is set with different triggering conditions and processing strategies. For minor anomalies, an early warning is triggered when the duration of the anomaly exceeds 1 hour or the anomaly severity score exceeds 60. The early warning information is represented as: W1={t1,α1,R1,L1}.

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

[0161] To better understand and implement this invention, the following is a specific application scenario example 2: In a smart grid research and development project of a power company, the research team implemented and tested a remote monitoring system for 10,000 smart meters within its jurisdiction for one year. The project selected smart meters distributed in urban commercial areas, residential areas, and industrial parks as monitoring targets, with 40% in commercial areas, 35% in residential areas, and 25% in industrial parks. The specific distribution of the meters is shown in Table 1:

[0162] Table 1. Statistical table of regional distribution of smart energy meters

[0163]

[0164] Figure 4 The 24-hour load characteristic curves of electricity meters in different regions are shown, with the horizontal axis representing time (hours) and the vertical axis representing the load factor (the ratio of actual power to rated power). During system implementation, the research team employed 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 for 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] This system uses Fast Fourier Transform (FFT) for signal processing. The basic transform formula is:

[0168]

[0169] The number of sampling points N was set to 1024, and the sampling frequency was 3200Hz. The system performed harmonic analysis on the voltage and current signals, extracting the fundamental frequency and the first five harmonic components. The harmonic measurement accuracy is shown in Table 3.

[0170] Table 3 Harmonic Measurement Accuracy Indicators

[0171] Harmonic number Amplitude accuracy Phase accuracy fundamental wave ±0.2％ ±0.2° Second harmonic ±0.3％ ±0.3° 3rd harmonic ±0.3％ ±0.3° 4th harmonic ±0.4％ ±0.4° 5th harmonic ±0.4％ ±0.4°

[0172] Figure 5 The harmonic analysis results for voltage and current are presented, including the percentage content of the fundamental frequency and the 2nd to 5th harmonics. During feature extraction, the system employs a regular feature variation 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:

[0175] Table 4. Parameter Table of Equation for Regular Characteristics and Changes

[0176] parameter Value illustrate α 0.85 Historical feature weighting coefficient β 0.03 Runtime weighting coefficient γ 0.05 Load change rate weighting coefficient δ 0.04 Temperature change rate weighting coefficient η 0.03 Voltage fluctuation weighting coefficient

[0177] During actual operation, the system modeled the propagation characteristics of abnormal features, using the following propagation equation:

[0178] During the one-year operation, the system detected various abnormal events, as shown in Table 5:

[0179] Table 5. Statistics of Abnormal Events

[0180] Exception types Number of occurrences Average duration Detection rate Voltage abnormality 856 15 minutes 99.2％ abnormal current 1243 22 minutes 98.7％ Harmonic exceedance 678 45 minutes 97.8％ Temperature anomaly 342 35 minutes 99.5％ Power factor abnormality 456 28 minutes 98.9％

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

[0182] Table 6. Warning Levels and Handling Strategies

[0183] Warning Level Triggering conditions Response time Processing strategy Level 1 Anomaly score >90 or lasting >10 minutes 5 minutes Emergency treatment Level 2 Anomaly score >75 or lasting >30 minutes 15 minutes Prioritize Level 3 Anomaly score >60 or lasting >60 minutes 30 minutes Standard processing Level 4 Anomaly score >45 or lasting >120 minutes 60 minutes Planned processing

[0184] During implementation, the research team focused on the stability assessment of the system, using 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. Parameters of the Stability Equation

[0187] parameter Range of values Optimization value k 5-10 8 κ 100-200 150 θ 0.5-2.0 1.5 ω ±0.01 0.008

[0188] Figure 6 The stability assessment results of the system during one year of operation are presented, including actual stability index curves and threshold lines. The system's performance indicators during actual operation are shown in Table 8.

[0189] Table 8 System Performance Indicators

[0190] Performance indicators Design Goals Actual value reached Data collection accuracy ≥99.5％ 99.8％ Anomaly detection accuracy ≥95％ 97.2％ Early warning response time ≤5 minutes 3.5 minutes System stability ≥99％ 99.6％ Fault prediction accuracy ≥90％ 93.5％

[0191] Traditional remote monitoring of electricity meters suffers from the following technical problems: 1. It uses a fixed sampling frequency, which cannot be dynamically adjusted according to real-time status, resulting in large data volume and insufficient accuracy. 2. Feature extraction uses traditional statistical methods, which cannot effectively identify complex fault modes. 3. The anomaly propagation model is too simplistic and does not consider network topology and spatiotemporal correlation. 4. The early warning mechanism is singular and cannot achieve multi-level classification and accurate early warning. 5. The system stability assessment method is outdated and lacks theoretical support.

[0192] The solution in Example 2 has the following advantages over traditional technologies: 1. It adopts adaptive sampling technology, dynamically adjusting the sampling frequency based on anomaly characteristics, reducing data volume while ensuring data quality, and improving sampling efficiency by 45%. 2. It introduces an improved lightweight neural network for feature extraction, improving feature extraction accuracy by 15% and computational efficiency by 60% compared to traditional methods. 3. It establishes an anomaly propagation model considering network topology and spatiotemporal characteristics, increasing anomaly propagation prediction accuracy from 85% to 97%. 4. It implements a multi-level early warning mechanism based on multi-dimensional features, reducing false alarm rate by 75% and false negative rate by 85%. 5. It proposes a system stability assessment method based on singular value decomposition, improving assessment accuracy by 25%.

[0193] It should be noted that the variables involved in this invention are explained in detail in Table 9 below.

[0194] Table 9. Variable Explanation Table

[0195]

[0196] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A remote monitoring device for a smart energy meter, characterized in that, The system includes several meter-based monitoring systems and a remote monitoring system. The meter-based monitoring systems include an 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, all installed inside the energy meter. The remote monitoring system includes a data receiving unit, a carrier resolution unit, a main control chip, a data storage unit, a display unit, and a deep analysis unit. The first feature extraction unit contains a lightweight feature extraction module, which performs the following steps: S11. Receive the power pulse signal, voltage signal, current signal, and temperature signal output by the first control chip; S12. Perform a fast Fourier transform on the voltage and current signals to extract the fundamental component and the first 5 harmonic components. S13. Use an improved lightweight neural network to extract features from the fundamental component and the first 5 harmonic components to obtain regular features. The improved lightweight neural network includes a feature extraction layer, a channel attention layer, a depthwise separable convolutional layer, an adaptive feature fusion layer, and a fully connected layer. S14. Calculate the feature deviation based on the regularity characteristics. When the feature deviation exceeds the preset threshold, generate abnormal features. S15. Output regular and abnormal characteristics to the first control chip; The deep analysis unit is equipped with a parallel processing module, which performs the following steps: S21. Receive the regular and abnormal characteristics of multiple meters transmitted by the main control chip, and construct a feature tensor; S22. Construct a hybrid neural network structure, which includes a feature encoding layer, an equation embedding layer, a temporal 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 attention-enhanced feature representations. The equation embedding layer contains four parallel equation processing units, including a regular feature processing unit, an anomaly feature processing unit, a feature coupling processing unit, and a stability processing unit. Each equation processing unit consists of a fully connected layer, a parameter-sharing recurrent neural network layer, and a dynamic weight layer. The temporal analysis layer uses a bidirectional long short-term memory network to extract temporal features and maintains long-term dependencies through residual connections. The feature decoding layer uses a deconvolutional network to map the features back to the original space. S23. Process feature tensors using the equation embedding layer of a hybrid neural network; S24. Pass the output of the equation embedding layer through the time series analysis layer to extract time series features and construct a state transition graph; S25. Use the feature decoding layer to convert the state transition diagram into the operating status evaluation result of each meter; S26. Generate early warning information based on the operational status assessment results; The regularity features consist of the fundamental voltage wave, the fundamental current wave, and the fundamental power factor. The abnormal features consist of the harmonic distortion rate, transient disturbances, and temperature fluctuations. The feature deviation is the Euclidean distance between two adjacent sampled values ​​of the regularity features.

2. The remote monitoring device for smart energy meters according to claim 1, characterized in that, The power acquisition unit adopts a precision sampling circuit design with a sampling frequency of 3200 Hz and a sampling accuracy of 0.2%. The current sampling unit uses a Hall sensor combined with a signal conditioning circuit. The linear range of the Hall sensor is 0.1 amperes to 100 amperes. The voltage sampling unit is implemented by resistive voltage division and isolation amplification with a voltage division ratio of 1000:

1. The temperature sensing unit uses a digital temperature sensor with a measurement range of -40 degrees Celsius to 85 degrees Celsius and a measurement accuracy of ±0.5 degrees Celsius.

3. The remote monitoring device for smart energy meters according to claim 1, characterized in that, The first control chip is 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 using a power line carrier communication chip with a carrier frequency range of 3 kHz to 500 kHz. The carrier parsing unit is implemented using a digital signal processor with a main frequency of 400 MHz.

4. The remote monitoring device for smart energy meters according to claim 1, characterized in that, The main control chip uses 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 gigabytes. The display unit uses a 7-inch LCD screen with a resolution of 1024 x 600 pixels. The depth analysis unit uses a graphics processor with 384 stream processor cores.

5. The remote monitoring device for smart energy meters according to claim 1, characterized in that, The inputs to the regularity feature change equation of the regularity feature processing unit include historical regularity feature sequence, electricity meter running time, load change rate, ambient temperature change rate, and grid voltage fluctuation rate, and the output is the predicted value of the regularity feature at the next moment.

6. The remote monitoring device for smart energy meters according to claim 1, characterized in that, The inputs to the abnormal feature propagation equation of the abnormal feature processing unit include the abnormal features of adjacent meters, line impedance parameters, propagation delay coefficient, and load similarity. The outputs are the propagation influence range and propagation influence intensity of the abnormal features.

7. The remote monitoring device for smart energy meters according to claim 1, characterized in that, The inputs to the feature coupling equation of the feature coupling processing unit include the regular features, the abnormal features, the coupling coefficient matrix, the feature correlation, the time scale parameter, and the disturbance intensity. The output is the coupled comprehensive features.

8. The remote monitoring device for smart energy meters according to claim 1, characterized in that, The stability equation of the stability processing unit 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 as inputs, and outputs the system stability evaluation index.

9. The remote monitoring device for smart 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 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 convolutional layer, and the output is the fused feature vector.

Images