A smart energy meter data acquisition and alarm device

Abstract

This invention provides a smart energy meter data acquisition and alarm device, belonging to the field of smart energy meter technology. The invention includes several meter-end monitoring systems and a remote-end monitoring system. The meter-end monitoring system includes an energy acquisition unit, a current sampling unit, a voltage sampling unit, a temperature sensing unit, a humidity sensing unit, a high-frequency current acquisition unit, and a control chip installed inside the energy meter. The device optimizes LoRa model parameters through an adaptive learning mechanism, achieving intelligent control through remote collaboration, effectively improving the accuracy and efficiency of energy meter safety management, and solving the technical problem of insufficient timely detection of abnormal states of massive numbers of energy meters via remote means in existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of smart energy meter technology, and more specifically, relates to a smart energy meter data acquisition and alarm device. Background Technology

[0002] As crucial metering and billing devices in power systems, the safety management of electricity meters is of paramount importance for ensuring the safe operation of the power grid and the order of electricity consumption. Traditional electricity meter management mainly relies on manual inspections and simple remote data collection, identifying anomalies through periodic meter readings and basic parameter monitoring. With the deepening of smart grid construction, the number of electricity meters has increased dramatically, and the types of electricity loads have become increasingly complex. Traditional management methods are no longer sufficient to meet the safety management requirements of large-scale smart grids.

[0003] Existing remote monitoring and control systems for electricity meters suffer from the following main problems: First, when dealing with a massive number of electricity meters, traditional centralized management methods struggle to achieve real-time monitoring of each meter, leading to the inability to promptly detect and handle anomalies. Second, existing anomaly detection methods primarily rely on simple threshold judgments, failing to effectively identify complex abnormal electricity consumption behaviors, such as electricity theft and unauthorized modifications—activities that are highly concealed. Third, the operating status of electricity meters is affected by various factors, including grid quality, environmental conditions, and load characteristics; existing systems lack the ability to comprehensively analyze these factors, making it difficult to accurately assess the health status of electricity meters. Furthermore, due to the wide distribution and complex installation environments of electricity meters, the system needs to adapt to different communication conditions and operating environments, posing a significant challenge to remote security management and control.

[0004] Currently, the industry mainly attempts to address these issues through the following methods: One approach is to increase the acquisition parameters and sampling frequency, attempting to improve the accuracy of anomaly detection through more data. However, this method leads to a sharp increase in system load with limited improvement in effectiveness. Another approach is to deploy complex analysis models remotely, but due to the lack of effective extraction of end-side features, the actual performance of these models is often less than ideal. In practice, these methods have failed to fundamentally solve the problem of remote security management of electricity meters, especially when facing large-scale smart grids, where the system's response speed, detection accuracy, and management efficiency are insufficient to meet actual needs. In other words, existing technologies suffer from insufficient timely detection of abnormal states in a large number of electricity meters remotely. Summary of the Invention

[0005] In view of this, the present invention provides a smart energy meter data acquisition and alarm device, which can solve the technical problem in the prior art that the abnormal status of a large number of energy meters cannot be detected in a timely manner through remote means.

[0006] This invention is implemented as follows: This invention provides a smart energy meter data acquisition and alarm device. The smart energy meter data acquisition and alarm device includes several meter-end monitoring systems and a remote-end monitoring system. The meter-end monitoring system includes a meter-end energy acquisition unit, a meter-end current sampling unit, a meter-end voltage sampling unit, a meter-end temperature sensing unit, a meter-end humidity sensing unit, a meter-end high-frequency current acquisition unit, and a meter-end control chip, all installed inside the energy meter. The remote-end monitoring system includes a remote-end data receiving unit, a remote-end data storage unit, a remote-end analysis module, a remote-end model training module, and a remote-end... An alarm unit is included; the meter-side control chip is equipped with a data processing and reporting module, which uses a lightweight neural network to extract features from the stable and variable components of the voltage and current signals. The lightweight neural network extracts features using LoRa model parameters sent from a remote terminal. The remote terminal analysis module constructs a deep neural network structure, which includes a feature encoding layer, a power quality analysis layer, a load analysis layer, an arc analysis layer, and a state assessment layer. This structure is used to analyze the raw data and feature data to generate power quality assessment results, load characteristic assessment results, and arc characteristic assessment results.

[0007] The lightweight neural network includes a feature extraction layer, a channel attention layer, a depthwise separable convolutional layer, a feature fusion layer, and a fully connected layer. The input of the feature fusion layer is the multi-channel feature map output by the depthwise separable convolutional layer, and the output is the fused feature vector.

[0008] The LoRa model decomposes the weight matrix of the lightweight neural network using a low-rank decomposition method. The weight matrix is ​​decomposed into the product of a basis matrix and an adaptation matrix. The basis matrix is ​​determined during training at a remote location, and the adaptation matrix is ​​used to adjust the feature extraction process.

[0009] The stable components include a steady-state voltage value, a steady-state current value, and a steady-state power factor. The steady-state voltage value is the mean value of the voltage signal, the steady-state current value is the mean value of the current signal, and the steady-state power factor is the cosine of the phase difference between the steady-state voltage value and the steady-state current value.

[0010] The data processing and reporting module is also used to calculate historical load data, which is the data of the previous 24 hours of the stable component; and to calculate the load change rate, which is the ratio of the change in two adjacent sample values ​​of the stable component to the sampling time interval.

[0011] The meter control chip evaluates the meter's operating status based on a set of meter status evaluation equations, which includes load characteristic equations, stability evaluation equations, variability evaluation equations, and abnormal state equations.

[0012] The load characteristic equation is used to evaluate the load operation status of the electricity meter. The inputs include the steady-state voltage value, the steady-state current value, and the historical load data, and the output is the load status evaluation coefficient. The stability evaluation equation is used to evaluate the stability of power quality. The inputs include the voltage fluctuation, the current fluctuation, the temperature signal, and the preset threshold, and the output is the stability evaluation coefficient.

[0013] In the deep neural network of the remote end analysis module, the feature encoding layer adopts a fully connected neural network structure, the input dimension is the sum of the dimensions of the original data and the feature data, and the number of neurons in the hidden layer is 512, 256, and 128.

[0014] The remote model training module trains the LoRa model using a data-parallel approach, employing a synchronous stochastic gradient descent algorithm for optimization. The initial learning rate is 0.001, and a cosine annealing scheduling strategy is used. The minimum learning rate is 0.00001, the batch size is 256, and the number of training rounds is 100.

[0015] The remote model training module calculates the autocorrelation coefficients at different time scales based on the temporal correlation of historical anomaly record vectors using autocorrelation analysis. The time scales range from 1 hour to 24 hours, with a step size of 1 hour. The time scale with the largest autocorrelation coefficient is selected as the feature time scale.

[0016] Furthermore, the meter status assessment equation set includes load characteristic equation, stability assessment equation, variability assessment equation, and abnormal state equation. The load characteristic equation input includes the steady-state voltage value, the steady-state current value, and historical load data. The stability assessment equation input includes the voltage fluctuation, the current fluctuation, the temperature signal, and a preset threshold. The variability assessment equation input includes the steady-state power factor, the voltage fluctuation, the current fluctuation, and the load change rate. The abnormal state equation input includes the load status assessment coefficient, the stability assessment coefficient, the variability assessment coefficient, the historical abnormal record vector, and the time scale parameter.

[0017] Furthermore, the deep neural network includes a feature encoding layer, a power quality analysis layer, a load analysis layer, an arc analysis layer, and a state evaluation layer. The feature encoding layer adopts a fully connected neural network structure, the power quality analysis layer adopts a convolutional neural network structure, the load analysis layer adopts a long short-term memory network structure, the arc analysis layer adopts a residual neural network structure, and the state evaluation layer adopts an attention mechanism.

[0018] Optionally, the operation status assessment report includes basic information, operation status, anomaly analysis, and early warning prompts. The basic information includes the meter number, installation location, and user type. The operation status includes the values ​​and levels of various assessment indicators. The anomaly analysis includes the anomaly type, anomaly severity, and anomaly duration. The early warning prompts include the early warning level, early warning reason, and handling suggestions.

[0019] Furthermore, the lightweight neural network is implemented using an improved MobileNet V2 architecture. It decomposes standard convolution into channel-independent depthwise convolution and pointwise convolution through depthwise separable convolution, and introduces residual connection structures to establish a fast path for identity mapping.

[0020] Optionally, the LoRa model is trained using a distributed training framework. The training is performed in a data-parallel manner and optimized using a synchronous stochastic gradient descent algorithm. The loss function is a weighted combination of mean squared error and cross-entropy.

[0021] Compared with existing technologies, this invention provides a smart energy meter data acquisition and alarm device. This invention achieves efficient and secure management of a massive number of energy meters through the organic combination of edge-side intelligence and remote control. At the edge, a lightweight neural network is used to extract the energy meter's operating characteristics in real time. The improved MobileNet V2 architecture significantly reduces computational complexity while ensuring the accuracy of feature extraction. The introduction of a feature fusion layer and channel attention mechanism enables the system to capture richer state information, providing a reliable data foundation for identifying abnormal behavior.

[0022] This invention establishes a complete state assessment system, encompassing assessments across multiple dimensions, including load characteristics, stability, variability, and abnormal states. Through a hierarchical set of assessment equations, the system can comprehensively analyze the operating status of electricity meters and accurately identify various abnormal situations. Remotely, a deep neural network is used for comprehensive analysis, combining historical data and time-series features to achieve precise identification of abnormal electricity consumption behavior. Simultaneously, the system possesses adaptive learning capabilities, continuously optimizing detection strategies based on historical anomaly records to improve the accuracy and efficiency of management and control.

[0023] By employing the low-rank decomposition method of the LoRa model, this invention achieves efficient updating and distribution of model parameters, enabling the system to quickly adapt to the management and control needs of different scenarios. The design of a multi-level alarm strategy ensures timely handling of abnormal situations, while the time-series database-based storage scheme provides data support for long-term analysis and optimization. The comprehensive application of these technologies allows the system to achieve efficient resource utilization while ensuring effective management and control. In summary, this invention solves the technical problem of insufficient timely detection of abnormal states of massive numbers of electricity meters via remote monitoring in existing technologies. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the device of the present invention.

[0025] Figure 2 This is a flowchart of the steps executed by the data processing and reporting module.

[0026] Figure 3 This is a flowchart of the steps performed by the remote analysis module.

[0027] Figure 4 A flowchart of the steps executed by the remote model training module.

[0028] Figure 5 This is a trend chart of various abnormal events during the 90-day trial period in Example 2.

[0029] Figure 6 This is a distribution diagram of the load characteristic assessment results in Example 2. Detailed Implementation

[0030] 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.

[0031] like Figure 1 The diagram shown is a schematic of the composition of a smart energy meter data acquisition and alarm device provided by the present invention. The device includes several meter-end monitoring systems and a remote monitoring system. The meter-end monitoring system includes an energy acquisition unit, a current sampling unit, a voltage sampling unit, a temperature sensing unit, a humidity sensing unit, a high-frequency current acquisition unit, and a control chip installed inside the energy meter.

[0032] The remote monitoring system includes a remote data receiving unit, a remote data storage unit, a remote analysis module, a remote model training module, and a remote alarm unit.

[0033] The power acquisition unit at the meter terminal is electrically connected to the metering circuit of the power meter, and is used to acquire power pulse signals and transmit the power pulse signals to the control chip at the meter terminal;

[0034] The current sampling unit at the meter terminal is electrically connected to the current sampling circuit of the energy meter, and is used to collect current signals and transmit them to the control chip at the meter terminal after signal conditioning.

[0035] The meter terminal voltage sampling unit is electrically connected to the electricity meter voltage sampling circuit, and is used to collect voltage signals and transmit them to the meter terminal control chip after signal conditioning.

[0036] The temperature sensing unit at the meter end is located inside the electricity meter and is used to collect temperature signals and transmit them to the control chip at the meter end.

[0037] The humidity sensing unit at the meter end is located inside the electricity meter and is used to collect humidity signals and transmit them to the control chip at the meter end.

[0038] The high-frequency current acquisition unit at the meter terminal is electrically connected to the current sampling circuit of the energy meter, and is used to acquire high-frequency current signals and transmit them to the control chip at the meter terminal.

[0039] The meter terminal control chip is equipped with a data processing and reporting module, which is used to perform the following steps:

[0040] S11. Receive the power pulse signal, the voltage signal, the current signal, the temperature signal, the humidity signal, and the high-frequency current signal;

[0041] S12. Perform signal decomposition on the voltage signal and the current signal to extract stable components and variable components;

[0042] S13. The lightweight neural network is used to extract features from the stable component and the variable component. The lightweight neural network extracts features using the LoRa model parameters sent from the remote terminal.

[0043] S14. Calculate historical load data, wherein the historical load data is the data of the previous 24 hours of the stable component;

[0044] S15. Calculate the load change rate, wherein the load change rate is the ratio of the change in two consecutive sampled values ​​of the stable component to the sampling time interval;

[0045] S16. Evaluate the meter's operating status based on the meter status evaluation equation set, and generate the meter's comprehensive status evaluation result.

[0046] S17. Generate a data reporting strategy based on the comprehensive status assessment results of the electricity meter, and select to report raw data or feature data according to the data reporting strategy.

[0047] The remote analysis module is used to perform the following steps:

[0048] S21. Receive raw data and feature data reported by multiple meter terminal monitoring systems;

[0049] S22. Construct a deep neural network structure, wherein the deep neural network includes a feature encoding layer, a power quality analysis layer, a load analysis layer, an arc analysis layer, and a state assessment layer;

[0050] S23. Analyze the original data and the feature data using the deep neural network to generate power quality assessment results, load characteristic assessment results, and arc characteristic assessment results;

[0051] S24. Store the comprehensive status assessment results of the electricity meter and generate a historical anomaly record vector;

[0052] S25. Generate an operation status assessment report based on the power quality assessment results, the load characteristic assessment results, and the arc characteristic assessment results;

[0053] S26. Determine whether to trigger an alarm signal based on the aforementioned operating status assessment report.

[0054] The remote model training module is used to perform the following steps:

[0055] S31. Summarize historical data from multiple meter monitoring systems;

[0056] S32. Construct LoRa model training samples based on the historical data;

[0057] S33. Train the LoRa model and optimize the LoRa model parameters;

[0058] S34. Calculate the preset threshold, which is determined through statistical analysis of the historical abnormal record vector;

[0059] S35. Calculate the time scale parameter, where the time scale parameter is the time correlation coefficient of the historical anomaly record vector;

[0060] S36. Send the Lora model parameters, the preset threshold, and the time scale parameters to the meter monitoring system.

[0061] The lightweight neural network includes a feature extraction layer, a channel attention layer, a depthwise separable convolutional layer, a feature fusion layer, and a fully connected layer. The input of the feature fusion layer is the multi-channel feature map output by the depthwise separable convolutional layer, and the output is the fused feature vector.

[0062] The LoRa model decomposes the weight matrix of the lightweight neural network using a low-rank decomposition method. The low-rank decomposition method includes: decomposing the weight matrix into the product of a basis matrix and an adaptation matrix, wherein the basis matrix is ​​determined during training at a remote location, and the adaptation matrix is ​​used to adjust the feature extraction process.

[0063] The stable components include steady-state voltage value, steady-state current value, and steady-state power factor. The steady-state voltage value is the mean value of the voltage signal, the steady-state current value is the mean value of the current signal, and the steady-state power factor is the cosine of the phase difference between the steady-state voltage value and the steady-state current value.

[0064] The variation components include voltage fluctuation and current fluctuation. The voltage fluctuation is the difference between the voltage signal and the steady-state voltage value, and the current fluctuation is the difference between the current signal and the steady-state current value.

[0065] The raw data includes the electrical energy pulse signal, the voltage signal, the current signal, the temperature signal, the humidity signal, and the high-frequency current signal;

[0066] The characteristic data includes the stable component, the variable component, and the load change rate;

[0067] The meter condition assessment equation set includes load characteristic equation, stability assessment equation, variability assessment equation, and abnormal state equation;

[0068] The load characteristic equation is used to evaluate the load operating status of the electricity meter. The inputs include the steady-state voltage value, the steady-state current value, and the historical load data. The output is the load status evaluation coefficient.

[0069] The stability evaluation equation is used to evaluate the stability of power quality. The inputs include the voltage fluctuation, the current fluctuation, the temperature signal, and the preset threshold. The output is the stability evaluation coefficient.

[0070] The variability assessment equation is used to evaluate the dynamic characteristics of the load. The inputs include the steady-state power factor, the voltage fluctuation, the current fluctuation, and the load change rate. The output is the variability assessment coefficient.

[0071] The abnormal state equation is used to comprehensively evaluate the operating status of the electricity meter. The inputs include the load status evaluation coefficient, the stability evaluation coefficient, the variability evaluation coefficient, the historical abnormal record vector, and the time scale parameter. The output is the comprehensive status evaluation result of the electricity meter.

[0072] The lightweight neural network is implemented using an improved MobileNet V2 architecture. Specifically, it constructs an efficient and compact feature extraction module by deeply optimizing the inverse residual structure and depthwise separable convolutions in the network architecture. This module first uses 1×1 convolutions to perform channel expansion operations to map low-dimensional features to a high-dimensional space to enhance the network's feature representation capabilities. Then, it uses 3×3 depthwise separable convolutions to extract spatial features in the expanded high-dimensional feature space. The depthwise separable convolutions significantly reduce computational complexity by decomposing the standard convolution into two steps: channel-independent depthwise convolutions and pointwise convolutions. Finally, it uses 1×1 convolutions to perform channel compression to map the features back to the low-dimensional space. The entire process not only maintains good feature extraction capabilities but also significantly reduces the number of model parameters and computational overhead. Building upon this foundation, the network introduces a residual connection structure. When the number of input and output channels is the same and the stride is 1, a shortcut path for identity mapping is established. This design not only facilitates gradient backpropagation, thus improving the model's training performance, but also enhances the model's expressive power through feature reuse. Furthermore, a batch normalization layer and a ReLU6 activation function are added after each convolutional layer to enhance the model's non-linear expressive power and accelerate training convergence. The ReLU6 truncation design also provides better numerical stability for quantization deployment. To further improve model performance, the improved version integrates a channel attention mechanism on top of the original MobileNet V2. This mechanism adaptively adjusts the importance of different channels to enhance effective features and suppress redundant information. Simultaneously, a feature fusion path is established between different network layers, enriching the semantic expression of features and improving the detection capability of multi-scale targets. Finally, the network uses a global average pooling layer instead of a traditional fully connected layer, significantly reducing the number of model parameters. A dropout layer is added before the classification layer to effectively alleviate overfitting. The overall network not only maintains its lightweight characteristics but also achieves a good balance between accuracy and performance.

[0073] This improved network architecture retains the advantages of the original MobileNet V2 while further enhancing feature extraction capabilities through attention mechanisms and multi-scale feature fusion. This allows the model to better adapt to the visual task requirements of different scenarios. Its lightweight nature makes it particularly suitable for deployment on mobile devices and embedded platforms, while the scalability of the model structure provides flexible adjustment space for performance optimization in different application scenarios. In practical applications, the model size and performance can be balanced by adjusting the network's width factor, or the configuration of each module can be fine-tuned according to specific task requirements, thereby achieving optimal performance under resource constraints.

[0074] The specific implementation methods of the above steps are described in detail below. In the specific implementation method of the meter-end monitoring system, the meter-end energy acquisition unit adopts a high-precision pulse acquisition chip. This chip supports pulse signal acquisition from 1 Hz to 100 Hz, with a sampling accuracy of 0.1%. It uses opto-isolation technology to achieve signal acquisition, and has extremely strong anti-interference ability and stability.

[0075] The current sampling unit at the meter terminal uses a high-precision Hall current sensor with a range of 0 to 100 amperes, a sampling frequency of up to 10 kHz, and a sampling accuracy of 0.5%. It has a built-in signal conditioning circuit, including a low-pass filter and an operational amplifier, which can effectively filter out high-frequency interference and amplify the signal.

[0076] The meter terminal voltage sampling unit is implemented using a combination of resistor voltage divider and operational amplifier. The range is 0 to 380 volts, the sampling frequency can reach 10 kHz, the sampling accuracy is 0.5%, and it has built-in overvoltage protection circuit and signal conditioning circuit, which can effectively prevent voltage surges from damaging the sampling circuit.

[0077] The temperature sensing unit at the meter terminal uses a digital temperature sensor, with a measurement range of -40 degrees Celsius to 125 degrees Celsius, a measurement accuracy of 0.5 degrees Celsius, a sampling period of 1 second, and a digital filtering function that can effectively suppress temperature signal fluctuations.

[0078] The humidity sensing unit at the meter end uses a digital humidity sensor, with a measurement range of 0 to 100% relative humidity, a measurement accuracy of 3% relative humidity, a sampling period of 1 second, and a temperature compensation function to ensure measurement accuracy at different temperatures.

[0079] The high-frequency current acquisition unit at the meter terminal is implemented using a broadband current transformer with a bandwidth of 50 Hz to 100 kHz, a range of 0 to 100 amperes, a sampling frequency of up to 1 MHz, and a sampling accuracy of 1%. It is used to acquire the high-frequency components in the current signal.

[0080] The meter control chip is implemented using a 32-bit microcontroller with a main frequency of no less than 800 MHz. It has a built-in hardware multiplier and hardware divider, a hardware floating-point unit, and an on-chip storage capacity of no less than 1 megabyte, which is used to realize data acquisition, data processing, and data reporting functions.

[0081] The remote data receiving unit is implemented using a database cluster with a master-slave architecture. The master database is responsible for writing data, and the slave database is responsible for reading data. It can support 1 million data writes and 10 million data reads per second.

[0082] The remote data storage unit is implemented using a distributed file system with a storage capacity of up to 100 terabytes. It employs a multi-replica mechanism to ensure data reliability and supports compressed storage with a compression ratio of up to 1:10.

[0083] The remote analysis module is implemented using a GPU server cluster, with a single machine configuration of no less than 8 GPU cards and a video memory capacity of no less than 32 gigabytes per card, used to realize online inference of deep learning models.

[0084] The remote model training module is implemented using a GPU server cluster, with a single machine configuration of no less than 8 GPU cards and a video memory capacity of no less than 80 gigabytes per card, for the purpose of offline training of deep learning models.

[0085] The remote alarm unit is implemented using a message queue system, supporting multiple alarm methods, including SMS alarms, email alarms, and application push alarms, with an alarm delay of no more than 1 second.

[0086] The specific implementation of these components fully considers the characteristics and requirements of the electricity metering system, employing high-precision sensors and high-performance processors to ensure high system reliability and stability. Simultaneously, the use of a distributed architecture and cluster technology guarantees strong scalability and fault tolerance. In practical applications, the parameters of each component can be appropriately adjusted according to specific needs to meet the application requirements of different scenarios.

[0087] The specific steps performed by the data processing and reporting module are described below:

[0088] The specific implementation of step S11 is as follows: It acquires electrical pulse signals based on interrupt methods, acquires voltage and current signals using timer interrupts, acquires temperature and humidity signals using serial communication, and acquires high-frequency current signals using direct memory access. The sampling clock is provided by an external crystal oscillator with a frequency of 32.768 kHz. Sampling data is stored according to timestamps with a precision of 1 millisecond. Data storage uses a circular buffer with a size of 32 kilobytes. The main purpose of this step is to achieve efficient acquisition and temporary storage of various sensor data.

[0089] The specific implementation of step S12 involves using a Fast Fourier Transform (FFT) algorithm to perform frequency domain analysis on the voltage and current signals. The fundamental component is extracted as the stable component, and the harmonic components are extracted as the variable components. The FFT uses 1024 points, and a Hanning window function is used for preprocessing. Fast calculation is achieved through bit reversal and butterfly operations. The calculated result has a frequency resolution of 10 Hz, an amplitude accuracy of 0.1%, and a phase accuracy of 0.1 degrees. This step aims to decompose the complex voltage and current signals into stable and variable parts, facilitating subsequent analysis.

[0090] Step S13 is implemented based on an improved lightweight neural network structure. First, the input data is normalized using a min-max method. Then, preliminary feature extraction is performed through a feature extraction layer containing three convolutional layers with kernel sizes of 1x1, 3x3, and 1x1, and channel numbers of 32, 64, and 32 respectively. Each convolutional layer is followed by a batch normalization layer and an activation function layer using a modified linear unit (MLU). Next, a channel attention layer weights the channel dimensions of the feature map using global average pooling and two fully connected layers. Then, a depthwise separable convolutional layer enhances the features. Finally, a feature fusion layer fuses the multi-scale features using a weighted summation method, with weights optimized using backpropagation. The goal of this step is to extract key features from the input data, providing a foundation for subsequent analysis.

[0091] The specific implementation of step S14 involves establishing a sliding time window based on the stable component data. The time window length is 24 hours, and the sliding step size is 15 minutes. Statistical analysis is performed on the data within the window, calculating statistical characteristics such as maximum, minimum, average, and standard deviation. Simultaneously, the peak-to-valley difference and load factor of the load curve are calculated. The load factor is defined as the ratio of average load to maximum load, and the peak-to-valley difference is defined as the difference between maximum and minimum load. This step is used to obtain the statistical characteristics of historical loads, providing a basis for load analysis.

[0092] The specific implementation of step S15 involves performing differential calculations on the stable component, calculating the difference between two adjacent sampled values, and dividing it by the sampling time interval (1 minute) to obtain the rate of change. Simultaneously, the mean and standard deviation of the rate of change are calculated to establish a statistical model for the rate of change, used to determine whether load changes are abnormal. When the rate of change exceeds the mean plus or minus three times the standard deviation, it is determined to be an abnormal change. The purpose of this step is to monitor the dynamic characteristics of load changes.

[0093] The specific implementation of step S16 involves calculation based on the meter status assessment equation set. The equation set is established using a weighted coefficient method, with a weighted coefficient of 0.3 for the load characteristic equation, 0.3 for the stability assessment equation, 0.2 for the variability assessment equation, and 0.2 for the abnormal state equation. The calculation results of each equation are normalized and then weighted and summed to obtain the comprehensive assessment result. This step aims to comprehensively assess the meter's operating status.

[0094] The specific implementation of step S17 involves assessing the overall status of the electricity meter. If the assessment result is greater than 0.8, it is considered an abnormal state, and raw data is reported. If the assessment result is less than or equal to 0.8, it is considered a normal state, and characteristic data is reported. Data reporting is conducted wirelessly using a low-power wide-area network (LPWAN) protocol. The data packet size does not exceed 240 bytes, the transmission power does not exceed 20 milliwatts, and the communication distance can reach 10 kilometers. The purpose of this step is to achieve intelligent data reporting and reduce communication burden.

[0095] These steps fully utilize digital signal processing and deep learning technologies. Through multi-level data analysis and processing, intelligent collection and reporting of electricity metering data are achieved, which can effectively improve the system's operating efficiency and reliability.

[0096] The remote analysis module is used to perform the following steps:

[0097] The specific implementation of step S21 involves using a distributed message queue system to receive data. The message queue adopts a topic subscription model, establishing an independent data topic for each meter monitoring system. Topics are categorized according to geographical region and electricity consumption type. The maximum message latency is 100 milliseconds, the message reliability level is at least once, the message persistence strategy is asynchronous disk flushing, and the message storage time is 30 days. The system supports message replay and filtering, and message filtering can be performed based on timestamps and data types. This step ensures the reliability and real-time performance of data reception.

[0098] The specific implementation of step S22 involves constructing a multi-layer deep neural network. The feature encoding layer adopts a fully connected neural network structure, with the input dimension being the sum of the dimensions of the original data and the feature data. The number of neurons in the hidden layers is 512, 256, and 128, the activation function is a modified linear unit, and the dropout rate is 0.3. The power quality analysis layer adopts a convolutional neural network structure, containing 3 convolutional blocks. Each convolutional block contains 2 convolutional layers and 1 max pooling layer. The convolutional kernel size is 3x3, the stride is 1, and the padding method is the same. The pooling kernel size is 2x2, and the stride is 1. The length is 2; the load analysis layer adopts a long short-term memory network structure, containing two long short-term memory layers, each with a hidden state dimension of 128, the bias of the forget gate is initialized to 1, and the biases of the input and output gates are initialized to 0; the arc analysis layer adopts a residual neural network structure, containing four residual blocks, each containing two convolutional layers and one short-circuit connection, with a convolutional kernel size of 3x3 and 64 channels; the state evaluation layer adopts an attention mechanism, using a soft attention module to weight and fuse the output features of each layer, with the attention weights calculated through a parameterized similarity function. This step establishes a multi-task learning framework that can perform multiple analyses simultaneously.

[0099] The specific implementation of step S23 involves inputting the data into a deep neural network for forward propagation calculation. First, the input data undergoes dimensionality reduction and feature extraction through a feature encoding layer. Then, the extracted features are input into the power quality analysis layer, load analysis layer, and arc analysis layer for analysis. The power quality assessment results include indicators such as voltage deviation rate, current distortion rate, and power factor, with assessment thresholds of ±5%, 5%, and 0.9, respectively. The load characteristic assessment results include indicators such as load forecast value, load change trend, and load anomaly degree, with a prediction time window of 24 hours and a prediction step size of 15 minutes. The arc characteristic assessment results include indicators such as arc occurrence probability, arc duration, and arc energy, with judgment thresholds of 0.8, 100 milliseconds, and 10 joules, respectively. This step achieves a comprehensive analysis of the meter's operating status.

[0100] The specific implementation of step S24 involves using a time-series database to store the comprehensive status assessment results of the electricity meters. The database employs a time-sharded storage method, with a time slice size of one day and a data compression ratio of 1:10. Simultaneously, multi-level indexes are established to accelerate queries. These indexes include a time index, a meter number index, and an anomaly type index. Anomaly record vectors are constructed based on the assessment results. The vector dimension represents the number of anomaly types, and the vector elements represent the occurrence frequency of the corresponding anomaly type. Anomaly types include voltage anomalies, current anomalies, power factor anomalies, load anomalies, and arc anomalies. This step provides historical data support for anomaly analysis.

[0101] The specific implementation of step S25 involves generating an assessment report based on the various assessment results. The report includes four parts: basic information, operating status, anomaly analysis, and early warning prompts. Basic information includes the meter number, installation location, and user type. Operating status includes the values ​​and levels of each assessment indicator. Anomaly analysis includes the anomaly type, severity, and duration. Early warning prompts include the warning level, cause, and suggested handling. Warning levels are divided into general warnings, important warnings, and emergency warnings, with thresholds of 0.6, 0.8, and 0.9, respectively. This step transforms the analysis results into an understandable report format.

[0102] The specific implementation of step S26 involves determining whether an alarm needs to be triggered based on the operational status assessment report. A multi-level alarm strategy is adopted: general warnings are indicated by in-system notifications; important warnings are simultaneously sent via email; and emergency warnings are simultaneously sent via SMS and telephone. Alarm information includes alarm time, alarm reason, alarm level, and handling suggestions. The system supports alarm confirmation and alarm escalation functions. When the alarm duration exceeds a preset time and is not handled, the alarm level is automatically escalated. This step enables timely notification and handling of abnormal situations.

[0103] The implementation of these steps constructs a complete remote analysis system that uses deep learning technology to achieve intelligent analysis of electricity metering data. This system can promptly detect and handle anomalies, improving the system's intelligence level and management efficiency.

[0104] The remote model training module is used to perform the following steps:

[0105] The specific implementation of step S31 involves collecting historical data reported by the meter-end monitoring system through a distributed data acquisition system. The data sources include raw data and feature data, stored using a distributed file system. Data is fragmented according to time and geographical location, with each fragment not exceeding 1 gigabyte in size. A data quality control module preprocesses the data, including outlier detection, missing value handling, and noise filtering. Outlier detection uses the 3x standard deviation method, missing values ​​are filled using linear interpolation, and noise filtering uses median filtering with a filtering window size of 5. The purpose of this step is to establish a high-quality training dataset.

[0106] The specific implementation of step S32 involves constructing LoRa model training samples based on preprocessed historical data. These samples include input features and output labels. Input features include time-series data of voltage, current, temperature, humidity, and high-frequency current signals. Output labels include power quality assessment results, load characteristic assessment results, and arc characteristic assessment results. A sliding window method is used to construct the training samples, with a window length of 1 hour and a sliding step size of 15 minutes. The samples are normalized using a maximum-minimum normalization method, and data augmentation is performed simultaneously. Augmentation methods include time-series scaling, additive Gaussian noise, and random pruning, with an augmentation ratio of 3 times. This step provides sufficient training data for model training.

[0107] The specific implementation of step S33 involves training the LoRa model using a distributed training framework. Training is conducted in data parallelism, employing a training cluster of 8 GPU servers, each with 8 GPUs. Simultaneous stochastic gradient descent is used for optimization, with an initial learning rate of 0.001. A cosine annealing scheduling strategy is used, with a minimum learning rate of 0.00001, a batch size of 256, and 100 training epochs. An early stopping strategy is employed, stopping training when the validation set loss fails to decrease for 5 consecutive epochs. The loss function uses a weighted combination of mean squared error and cross-entropy with a 1:1 weight ratio. Gradient clipping is used to prevent gradient explosion, with a clipping threshold of 5. This step optimizes the model parameters.

[0108] The specific implementation of step S34 involves calculating a preset threshold based on historical anomaly record vectors. Using statistical analysis, the anomaly record vectors are first clustered using a density-based spatial clustering algorithm. The minimum number of samples for each cluster is 10, and the maximum distance parameter is 0.1, resulting in different types of anomaly clusters. Then, statistical characteristics, including mean, standard deviation, and quantiles, are calculated for each anomaly cluster. Based on these statistical characteristics, a preset threshold is determined: for voltage anomalies, the preset threshold is ±7% of the standard voltage; for current anomalies, the preset threshold is 150% of the rated current; and for power factor anomalies, the preset threshold is 0.85. This step provides a scientific standard for anomaly detection.

[0109] The specific implementation of step S35 involves analyzing the temporal correlation of historical anomaly record vectors. Autocorrelation analysis is used to calculate autocorrelation coefficients at different time scales, ranging from 1 hour to 24 hours, with a step size of 1 hour. The time scale with the highest autocorrelation coefficient is selected as the feature time scale. Simultaneously, cross-correlation coefficients are calculated to analyze the temporal correlation between different types of anomalies. A temporal correlation network is constructed, where network nodes represent anomaly types and edge weights represent cross-correlation coefficients. Time scale parameters are calculated based on the correlation network. This step reveals the temporal patterns of anomaly occurrence.

[0110] The specific implementation of step S36 involves using a remote parameter configuration system to distribute the trained model parameters to the electricity meter monitoring system. Parameter distribution is done in batches, with each batch distributing to 100 electricity meter monitoring systems. The distributed content includes LoRa model parameters, preset thresholds, and time scale parameters. An incremental update method is used, distributing only parameters that have changed. Parameter compression employs a sparse matrix compression method, achieving a compression ratio of 1:20. Transmission uses a secure channel and is encrypted using national cryptographic algorithms with a 256-bit key. Parameter consistency verification is also performed to ensure correct parameter distribution. This step achieves model deployment and updating.

[0111] These steps establish a complete model training and deployment process. Through distributed computing and optimization algorithms, efficient model training and updating are achieved. Simultaneously, scientific statistical analysis methods determine reasonable preset thresholds and time scale parameters, providing strong support for the intelligent operation of the system. In summary, this system, through deep learning technology and distributed computing, realizes intelligent monitoring and management of electricity metering equipment, possessing high practical and promotional value.

[0112] The calculation of steady-state voltage and current values ​​is specifically expressed as follows:

[0113]

[0114] In the formula, V stable This is the steady-state voltage value; I stable This is the steady-state current value; V i Let I be the voltage value at the i-th sampling point; i is the current value at the i-th sampling point; N is the number of sampling points, ranging from 1000 to 10000.

[0115] The calculation of steady-state power factor is expressed as follows:

[0116] PF stable =cos(φ VI )=cos(∠V stable -∠I stable );

[0117] In the formula, PF stable φ is the steady-state power factor. VI The phase difference between voltage and current; ∠V stable The steady-state voltage phase angle; ∠I stable This represents the steady-state current phase angle.

[0118] The calculation of voltage fluctuations and current fluctuations is expressed as follows:

[0119] V fluctuation =V(t)-Vstable ;

[0120] I fluctuation =I(t)-I stable ;

[0121] In the formula, V fluctuation For voltage fluctuation; I fluctuation V(t) represents the current fluctuation; V(t) represents the voltage value at time t; and I(t) represents the current value at time t.

[0122] The load characteristic equation is expressed as follows:

[0123]

[0124] In the formula, F load V is the load condition assessment coefficient. n Rated voltage; I n Rated current; P load P represents the load power. n α1 represents the rated power; α2, α3 are weighting coefficients, determined by the least squares method; ε1 is the error term, ranging from 0 to 0.1.

[0125] The stability evaluation equation is expressed as follows:

[0126]

[0127] In the formula, F stability V is the stability evaluation coefficient; th Voltage fluctuation threshold; I th T is the current fluctuation threshold; T is the temperature value; T n Rated temperature; T th β1, β2, and β3 are the temperature deviation thresholds; β1, β2, and β3 are the weighting coefficients; ε2 is the error term, ranging from 0 to 0.1.

[0128] The variability assessment equation is expressed as follows:

[0129]

[0130] In the formula, F variation This is the coefficient for assessing variability; γ is the load change rate; k1 and k2 are the attenuation coefficients; γ1, γ2, γ3, and γ4 are the weighting coefficients; ε3 is the error term, ranging from 0 to 0.1.

[0131] The abnormal state equation is specifically represented as follows:

[0132]

[0133] In the formula, F abnormalThis is the result of the comprehensive condition assessment of the electricity meter; H abnormal t is the vector of historical anomaly records; t is the time variable; τ is the time scale parameter; δ1, δ2, δ3, and δ4 are weighting coefficients; ε4 is the error term, ranging from 0 to 0.1.

[0134] The computation of depthwise separable convolution is specifically represented as follows:

[0135]

[0136] In the formula, Y dwc Y is the output of a depthwise convolution; pwc For pointwise convolution output; K c For depthwise convolution kernels; X c is the input feature map; s is the stride; k is the kernel size; W c represents the pointwise convolution weights; M represents the number of input channels.

[0137] The specific representation of channel attention calculation is as follows:

[0138]

[0139] F excitation =σ(W2·ReLU(W1·F) squeeze ));

[0140] In the formula, F squeeze For global average pooling output; F excitatin σ is the channel weight; X is the input feature map; H is the feature map height; W is the feature map width; W1 and W2 are the weights of the fully connected layer; σ is the sigmoid function; ReLU is the corrected linear unit function.

[0141] The weight matrix decomposition of the Lora model is specifically represented as follows:

[0142] W = W0 + BA;

[0143] In the formula, W is the complete weight matrix; W0 is the basis matrix; B is the low-rank basis vector; and A is the fitness matrix.

[0144] The principles behind the construction of these equations mainly take into account the following aspects:

[0145] 1. The load characteristic equation adopts a normalized form, evaluates the load status by the ratio to the rated value, and uses an exponential function to reflect the nonlinear relationship;

[0146] 2. The stability assessment equation adopts an exponential decay form, reflecting the negative correlation between volatility and stability;

[0147] 3. The variability assessment equation comprehensively considers the power factor, fluctuation amount, and rate of change, and uses an exponential function to describe the impact of fluctuation amount;

[0148] 4. The abnormal state equation introduces a time decay term to reflect the characteristic that the influence of historical anomalies on the current state weakens over time.

[0149] 1. The derivation process of the equations for calculating steady-state voltage and current values ​​is described in detail below:

[0150] First, the original sampled data is segmented using the sliding window method, with a window length of 1000 sampling points and a sliding step size of 100 sampling points.

[0151] Then, outlier detection is performed on the data in each window, and outliers are removed using the 3σ criterion.

[0152] Finally, the arithmetic mean of the effective data is calculated as the steady-state value. This method can effectively suppress the influence of random fluctuations.

[0153] 2. The derivation of the steady-state power factor calculation equation is described in detail below:

[0154] First, perform a Fourier transform on the voltage and current signals to extract the fundamental component;

[0155] Then calculate the phase difference between the fundamental voltage and current;

[0156] Finally, the cosine of the phase difference is taken as the power factor, which avoids interference from higher harmonics.

[0157] 3. The derivation process of the load characteristic equation is described in detail below:

[0158] The load characteristic vector L is specifically represented as follows:

[0159]

[0160] The weight vector α is specifically represented as follows:

[0161]

[0162] The load characteristic equation can be expressed as:

[0163] F load =α T L+ε1;

[0164] The weight vector α is solved using the least squares method:

[0165] α=(L T L) -1 L T Y;

[0166] In the formula, Y is the load condition score marked by the expert.

[0167] 4. The derivation process of the stability evaluation equation is described in detail below:

[0168] The specific representation of the fluctuation eigenvector S is as follows:

[0169]

[0170] The weight vector β is specifically represented as follows:

[0171]

[0172] The stability evaluation equation can be expressed as:

[0173] F stability =β T S+ε2;

[0174] Using an exponential function allows the evaluation results to change smoothly between 0 and 1. 5. The derivation process of the variability evaluation equation is described in detail below:

[0175] The specific representation of the variable feature vector V is as follows:

[0176]

[0177] The weight vector γ is specifically represented as follows:

[0178]

[0179] The variability assessment equation can be expressed as:

[0180] F variation =γ T V+ε3;

[0181] The exponential decay term is used to describe the nonlinear effect of volatility on variability.

[0182] 6. Matrix representation of depthwise separable convolution:

[0183] Input feature map X c Specifically represented as an H×W dimensional matrix:

[0184]

[0185] Depth convolution kernel K c Specifically represented as a k×k dimensional matrix:

[0186]

[0187] Pointwise convolution weights W c Specifically represented as an M-dimensional vector:

[0188] W c =[w1 w2 … wM ].

[0189] 7. Specific representation of the weight matrix decomposition in the LoRa model:

[0190] The complete weight matrix W is an m×n dimensional matrix:

[0191]

[0192] The basis matrices W0 and W have the same dimension;

[0193] The low-rank basis vector B is an m×r dimensional matrix:

[0194]

[0195] The fitness matrix A is an r×n dimensional matrix:

[0196]

[0197] Where r is the rank, which is usually much smaller than m and n. This decomposition method can significantly reduce the number of parameters that need to be optimized.

[0198] Specifically, the principle of this invention is based on a hierarchical architecture and collaborative optimization design. On the edge, depthwise separable convolution decomposes the standard convolution operation into depthwise convolution and pointwise convolution, significantly reducing computational complexity. By introducing a channel attention mechanism, the system can adaptively adjust the importance of different feature channels, improving the effectiveness of feature extraction. The feature fusion layer enhances the system's ability to identify different types of anomalies through the combination of multi-scale features.

[0199] The design of the evaluation equations embodies the principle of multi-dimensional analysis. The load characteristic equation, through normalization and weight calculation, reflects the basic characteristics of load operation; the stability evaluation equation, using an exponential decay form, describes the impact of fluctuations on system stability; the variability evaluation equation comprehensively considers power factor, fluctuation, and rate of change, reflecting the dynamic characteristics of the system; and the abnormal state equation, by introducing a time decay term, reflects the impact of historical anomalies on the current state.

[0200] At the remote end, the deep neural network design employs a multi-layered analysis structure, including a feature encoding layer, a power quality analysis layer, a load analysis layer, and an arc analysis layer. This structure enables the analysis of the electricity meter's operating status from different perspectives, improving the accuracy of anomaly detection. The adaptive learning mechanism optimizes model parameters and evaluation thresholds by analyzing historical data, allowing the system to continuously improve its management and control effectiveness.

[0201] 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.

[0202] In this embodiment 1, the power acquisition unit at the meter end is implemented using a high-precision pulse acquisition chip ADE7953. This chip operates at a voltage of 3.3 volts and achieves signal isolation through an optocoupler 4N35. The sampling accuracy reaches 0.1%, and it supports pulse signal acquisition from 1 Hz to 100 Hz. A low-pass filter is used for signal preprocessing, with a cutoff frequency of 200 Hz. The circuit adopts a differential input method, with a common-mode rejection ratio greater than 80 dB, providing strong anti-interference capabilities. It is connected to the meter end control chip through a serial peripheral interface with a baud rate of 1 megabit per second.

[0203] The current sampling unit at the meter terminal uses a Hall current sensor ACS758, with a range of 0 to 100 amperes, a sampling frequency of 10 kHz, a sampling accuracy of 0.5%, a linearity better than 0.1%, a bandwidth of 120 kHz, and an output signal of 0 to 5 volts analog quantity. The built-in operational amplifier AD8554 forms a signal conditioning circuit with a gain of 2 times and a common-mode rejection ratio greater than 100 dB. It is connected to the meter terminal control chip through a 12-bit analog-to-digital converter AD7866, with a conversion time of less than 1 microsecond.

[0204] The meter terminal voltage sampling unit consists of a precision resistor voltage divider network and an AD8510 operational amplifier. The range is 0 to 380 volts, the voltage division ratio is 100:1, the voltage divider resistors are 0.1% precision metal film resistors with a temperature coefficient of less than 25 ppm per degree Celsius, the operational amplifier gain is 2 times, the bandwidth is 10 MHz, the offset voltage is less than 50 microvolts, and overvoltage protection is achieved through a transient suppression diode SMBJ380A with a protection voltage of 380 volts and a response time of less than 1 picosecond.

[0205] The temperature sensing unit at the meter end uses the DS18B20 digital temperature sensor, with a measurement range of -40 degrees Celsius to 125 degrees Celsius, a measurement accuracy of 0.5 degrees Celsius, a programmable resolution of 9 to 12 bits, corresponding to a temperature resolution of 0.5 to 0.0625 degrees Celsius, a sampling period of 1 second, and communication with the meter end control chip via a single-bus protocol, with a transmission rate of 15.4 kilobits per second and strong anti-interference capability.

[0206] The humidity sensing unit at the meter end uses the HI H8120 digital humidity sensor, which measures relative humidity from 0 to 100%, has a measurement accuracy of 3% relative humidity, a response time of less than 6 seconds, and an operating temperature range of -40 to 85 degrees Celsius. It supports automatic temperature compensation within the operating temperature range and communicates with the meter end control chip via an Inter-Integrated Circuit interface at a communication rate of 400 kHz.

[0207] The high-frequency current acquisition unit at the meter terminal is implemented using a broadband current transformer HCTS10 with a bandwidth of 50 Hz to 100 kHz, a range of 0 to 100 amperes, a sampling frequency of 1 MHz, a sampling accuracy of 1%, and an output signal of current type with a conversion ratio of 1000:1. The current signal is converted into a voltage signal through a transimpedance amplifier AD8429 with a gain of 1000 ohms and a bandwidth of 15 MHz.

[0208] The meter control chip is implemented using a 32-bit microcontroller STM32F407 with a main frequency of 168 MHz, 192 kilobytes of on-chip random access memory, 1 megabyte of on-chip flash memory, a hardware floating-point unit with a computing speed of up to 125 million floating-point operations per second, and an integrated DC-to-DC power management unit that supports multiple low-power modes.

[0209] The connections between these components are achieved using printed circuit boards with a 4-layer design. The signal layer has a wiring width of 8 mils, the power and ground layers use copper pour design, the via diameter is 0.3 mm, the minimum line spacing is 6 mils, the impedance matching error is less than 10%, all high-speed signals use differential routing, and the wiring length matching error is less than 0.5 mm.

[0210] The remote server is a rack-mount server equipped with dual Xeon 8352Y processors, each with 32 cores and 64 threads, a clock speed of 3.0 GHz, 256 gigabytes of memory, a solid-state drive array with a capacity of 20 terabytes, and a 10 Gigabit Ethernet network interface, supporting server cluster management and load balancing.

[0211] The installation location and connection method of all components must take electromagnetic compatibility requirements into account. Shielded cables should be used for analog signal transmission, and twisted-pair cables should be used for digital signal transmission. All metal housings must be reliably grounded with a grounding resistance of less than 4 ohms, and surge protectors and electromagnetic compatibility filters should be installed at the power input.

[0212] The meter terminal control chip is equipped with a data processing and reporting module, which is used to perform the following steps:

[0213] The specific implementation of step S11 involves acquiring various data through interrupts. The power pulse signal is acquired via an external interrupt, with the sampling clock provided by an external 32.768 kHz crystal oscillator. Temperature and humidity signals are acquired via a serial communication peripheral at a 1-second sampling period. Voltage and current signals are acquired via a timer interrupt at a 10 kHz sampling frequency. High-frequency current signals are acquired via direct memory access at a 1 MHz sampling frequency. The acquired data is stored in a circular buffer with timestamps. The buffer uses a double-buffered structure to avoid data read / write conflicts. The buffer size is 32 kilobytes, and the timestamp accuracy is 1 millisecond. The steady-state voltage value is calculated using the following formula: In the formula, N is the number of sampling points, ranging from 1000 to 10000, and V i Let be the voltage value at the i-th sampling point. The steady-state current value is calculated using the following formula: In the formula I i Let be the current value at the i-th sampling point.

[0214] The specific implementation of step S12 involves using a Fast Fourier Transform (FFT) algorithm to perform frequency domain analysis on the voltage and current signals. The number of transform points is 1024. A Hanning window function is used for preprocessing to reduce spectral leakage. Fast calculation is achieved through bit reversal and butterfly operations. The calculated result has a frequency resolution of 10 Hz, an amplitude accuracy of 0.1%, and a phase accuracy of 0.1 degrees. The fundamental component is extracted as the stable component, and the harmonic components are extracted as the variable components. The steady-state power factor is calculated using the following formula: PF stable =cos(φ VI )=cos(∠V stable -∠I stable ), where φ VI For the phase difference between voltage and current, ∠V stable For steady-state voltage phase angle, ∠I stable The steady-state current phase angle is given. Voltage fluctuations are calculated using the following formula: V fluctuation =V(t)-V stable In the formula, V(t) is the voltage value at time t. The current fluctuation is calculated using the following formula: I fluctuation =I(t)-I stable In the formula, I(t) is the current value at time t.

[0215] The specific implementation of step S13 is based on feature extraction using an improved lightweight neural network structure. First, the input data is normalized using a minimax method. Then, preliminary feature extraction is performed through a feature extraction layer containing three convolutional layers with kernel sizes of 1x1, 3x3, and 1x1, and channel numbers of 32, 64, and 32 respectively. Each convolutional layer is followed by a batch normalization layer and an activation function layer. The depthwise separable convolution is calculated using the following formula: In the formula Y dwc For depthwise convolution output, Y pwc For pointwise convolution output, K c X is a depthwise convolution kernel. c The input feature map is s, where stride is s, kernel size is k, and W is W. c Here, M represents the number of input channels, and M is the pointwise convolution weight. Channel attention is calculated using the following formula: In the formula F squeeze For global average pooling output, F excitation X is the channel weight, H is the input feature map, W is the feature map height, and W1 and W2 are the weights of the fully connected layer.

[0216] The specific implementation of step S14 is to establish a sliding time window based on the stable component data. The time window length is 24 hours and the sliding step size is 15 minutes. Statistical analysis is performed on the data within the window to calculate statistical characteristics such as the maximum value, minimum value, average value, and standard deviation. At the same time, the peak-to-valley difference and load factor of the load curve are calculated. The load factor is defined as the ratio of the average load to the maximum load, and the peak-to-valley difference is defined as the difference between the maximum load and the minimum load.

[0217] The specific implementation of step S15 is to calculate the load change rate, perform differential calculation on the stable component, calculate the difference between two adjacent sample values, and divide it by the sampling time interval to obtain the change rate. The sampling time interval is 1 minute. At the same time, calculate the mean and standard deviation of the change rate, establish a statistical model of the change rate, and determine it as an abnormal change when the change rate exceeds the mean plus or minus 3 times the standard deviation.

[0218] The specific implementation of step S16 involves calculation based on the meter condition assessment equation set, wherein the load characteristic equation is: The stability evaluation equation is: The variability assessment equation is: The abnormal state equation is:

[0219] The specific implementation of step S17 is as follows: based on the comprehensive status assessment result of the electricity meter, when the assessment result is greater than 0.8, it is determined to be an abnormal state and the original data is reported; when the assessment result is less than or equal to 0.8, it is determined to be a normal state and the characteristic data is reported. The data reporting adopts a low power wide area network protocol, the data packet size does not exceed 240 bytes, the transmission power does not exceed 20 milliwatts, and the communication distance can reach 10 kilometers.

[0220] The remote analysis module is used to perform the following steps:

[0221] The specific implementation of step S21 is to use a distributed message queue system to receive data, adopt a topic subscription mode, establish an independent data topic for each electricity meter monitoring system, classify topics according to geographical region and electricity consumption type, the message latency is less than 100 milliseconds, the message reliability level is at least once, the message persistence adopts an asynchronous disk flushing strategy, the message retention time is 30 days, and message filtering by timestamp and data type is supported. The system adopts a distributed architecture, supports horizontal scaling, and the throughput of a single node is not less than 100,000 messages per second.

[0222] The specific implementation of step S22 is to construct a multi-layer deep neural network. The feature encoding layer adopts a fully connected neural network structure, with the input dimension being the sum of the dimensions of the original data and the feature data. The number of neurons in the hidden layer is 512, 256, and 128. The output calculation formula of the feature encoding layer is: h i =ReLU(W i x i-1 +b i ), where h i For the output of the i-th layer, W i Let x be the weight matrix. i-1 b is the output of the previous layer. i The bias term is represented by ReLU, which is the modified linear unit activation function. The power quality analysis layer uses a convolutional neural network structure, containing three convolutional blocks. Each convolutional block contains two convolutional layers and one max-pooling layer. The convolutional kernel size is 3x3 with a stride of 1 and uniform padding. The pooling kernel size is 2x2 with a stride of 2. The load analysis layer uses a long short-term memory network structure, containing two long short-term memory layers. The hidden state dimension of each layer is 128. The forget gate calculation formula is: f t =σ(W f ·[h t-1 x t ]+b f The input gate calculation formula is: i t =σ(W i ·[h t-1 x t ]+b i The formula for calculating the output gate is: o t=W(W o ·[h t-1 x t ]+b o The cell state update formula is: c t =f t ·c t-1 +i t ·tanh(W c ·[h t-1 x t ]+b c The output state update formula is: h t =o t ·tanh(c t ), where σ is the sigmoid function and tanh is the hyperbolic tangent function.

[0223] The specific implementation of step S23 involves inputting the data into a deep neural network for forward propagation calculation. The input data is then subjected to dimensionality reduction and feature extraction through a feature encoding layer. The extracted features are then input into the power quality analysis layer, load analysis layer, and arc analysis layer for analysis. Power quality assessment uses the following equation: Q = w1·THD v +w2·THD i +w3·(1-PF)+w4·ΔV+ε q In the formula, THD v Total Harmonic Distortion (THD) of voltage i ρ is the total harmonic distortion of the current, PF is the power factor, ΔV is the voltage deviation, w1, w2, w3, w4 are weighting coefficients, and ε is the total harmonic distortion of the current. q This represents the error term. The load characteristic assessment uses the following equation: In the formula P avg For average load, P max The maximum load is LF, and the load factor is LF. α1, α2, α3, and α4 are the load change rate, ε is the weighting coefficient, and α is the load change rate. l This is the error term. The arc characteristic assessment uses the following equation: A = β1·E arc +β2·T arc +β3·f arc +ε a E in the formula arc For the energy of the electric arc, T arc f is the duration of the electric arc. arc β1, β2, and β3 are the arc generation frequency, β3 is the weighting coefficient, and ε is the arc generation frequency. a This is the error term.

[0224] The specific implementation of step S24 involves using a time-series database to store the comprehensive status assessment results of the electricity meters. The data is stored in time slices, with a time slice size of 1 day and a data compression ratio of 1:10. A multi-level index structure is established, including a time index, an electricity meter number index, and an anomaly type index. The formula for constructing the anomaly record vector is: H abnormal = [n1, n2, ..., n k ], where n i Let be the number of occurrences of the i-th type of exception, and k be the total number of exception types.

[0225] The specific implementation of step S25 is to generate an evaluation report based on the evaluation results. The evaluation report includes four parts: basic information, operating status, anomaly analysis, and early warning prompts. The basic information includes the meter number, installation location, user type, etc. The operating status includes the values ​​and levels of various evaluation indicators. The anomaly analysis includes the anomaly type, anomaly degree, anomaly duration, etc. The early warning prompts include the early warning level, early warning reason, and handling suggestions, etc. The threshold values ​​for judging the early warning level are 0.6, 0.8, and 0.9, respectively.

[0226] The specific implementation of step S26 is to determine whether an alarm is triggered based on the operation status assessment report, and adopt a multi-level alarm strategy. One optional approach is: for general warnings, use in-system notification; for important warnings, send email notifications simultaneously; and for emergency warnings, send SMS and telephone notifications simultaneously. The alarm information includes alarm time, alarm reason, alarm level, and handling suggestions, and supports alarm confirmation and alarm escalation functions. If the alarm duration exceeds the preset time and is not handled, the alarm level is automatically escalated. The multi-level alarm strategy can be set manually or based on experience using existing technologies.

[0227] The remote model training module is used to perform the following steps:

[0228] The specific implementation of step S31 involves collecting historical data reported by the meter-end monitoring system through a distributed data acquisition system. The data sources include raw data and feature data. A distributed file system is used to store the data, which is then fragmented according to time and geographical location. Each fragment is no larger than 1 gigabyte. Data quality control employs a 3x standard deviation method for outlier detection. The standard deviation calculation formula is: In the formula, σ is the standard deviation, N is the sample size, and x i Here, is the sample value, and μ is the sample mean. Missing values ​​are imputed using linear interpolation, with the following formula: In the formula, x is the point to be interpolated, x1 and x2 are the x-coordinates of adjacent known points, and y1 and y2 are the y-coordinates of adjacent known points. Median filtering is used for noise filtering, with a filter window size of 5.

[0229] The specific implementation of step S32 involves constructing LoRa model training samples based on preprocessed historical data. A sliding window method is used to construct the training samples, with a window length of 1 hour and a sliding step size of 15 minutes. Input features include time-series data of voltage, current, temperature, humidity, and high-frequency current signals. Output labels include power quality assessment results, load characteristic assessment results, and arc characteristic assessment results. The samples are then normalized to their maximum and minimum values ​​using the following formula: In the formula x norm The value is the normalized value, where x is the original value. min For the minimum value, x max This is the maximum value. Data augmentation methods include time series scaling, additive Gaussian noise, and random pruning, with an augmentation ratio of 3 times.

[0230] The specific implementation of step S33 involves training the LoRa model using a distributed training framework. Training is conducted in data parallelism, employing a training cluster of 8 GPU servers, each with 8 GPUs. The weight matrix decomposition of the LoRa model uses the following formula: W = W0 + BA, where W is the complete weight matrix, W0 is the basis matrix, B is the low-rank basis vector, and A is the adaptation matrix. Optimization is performed using the synchronous stochastic gradient descent algorithm, with a cosine annealing scheduling strategy for the learning rate. The learning rate calculation formula is: In the formula η t Let η be the current learning rate. min To minimize the learning rate, η max The maximum learning rate is given by t, the current iteration number, and T, the total number of iterations. The batch size is 256, the training epochs are 100, an early stopping strategy is used, and the loss function is a weighted combination of mean squared error and cross-entropy. The loss function formula is: L = λ₁L mse +λ2L ce In the formula L mse For mean square error loss, L ce λ1 and λ2 are the cross-entropy loss and the weighting coefficients.

[0231] The specific implementation of step S34 is to calculate a preset threshold based on the historical abnormal record vector, and then perform cluster analysis using a density-based spatial clustering algorithm. The minimum number of samples for each cluster is 10, the maximum distance parameter is 0.1, and the density calculation formula is: ρ i =∑ j χ(d ij -d c ), where ρ i Let d be the local density at point i. ij Let d be the distance between points i and j. cLet X(x) be the cutoff distance and X(x) be the indicator function. Statistical characteristics, including mean, standard deviation, and quantiles, are calculated for each outlier cluster, and a preset threshold is determined based on these statistical characteristics.

[0232] The specific implementation of step S35 involves analyzing the temporal correlation of historical anomaly record vectors using autocorrelation analysis. The formula for calculating the autocorrelation coefficient is as follows: In the formula r k Let x be the autocorrelation coefficient with a lag of k periods. t Here, is the time series value, and μ is the series mean. The cross-correlation coefficient is calculated at different time scales using the following formula: In the formula C xy (k) is the cross-covariance function, C xx (0), C yy (0) is the autocovariance function.

[0233] The specific implementation of step S36 involves using a remote parameter configuration system to distribute the trained model parameters to the electricity meter monitoring system. This is done in batches, with 100 electricity meter monitoring systems per batch. An incremental update method is used, and parameter compression employs a sparse matrix compression method, achieving a compression ratio of 1:20. Matrix compression utilizes singular value decomposition, with the decomposition formula: M=U∑V T In the formula, M is the original matrix, U and V are orthogonal matrices, and ∑ is a diagonal matrix. Transmission is encrypted using the national cryptographic algorithm with a 256-bit key length, and parameter consistency verification is performed simultaneously.

[0234] To better understand and implement this invention, the following is a specific application scenario example 2: A power company's R&D team is collecting data and triggering alarms on 5,000 smart meters in a region. This region includes 3,000 residential users, 1,500 commercial users, and 500 industrial users. First, the existing meters are upgraded by installing an improved data acquisition module inside each meter. The main parameters of the acquisition module are shown in Table 1.

[0235] Table 1 Data Acquisition Module Parameter Configuration Table

[0236] Parameter type Parameter value Accuracy requirements Voltage sampling frequency 10kHz 0.5％ Current sampling frequency 10kHz 0.5％ High-frequency current sampling frequency 1MHz 1.0％ Temperature sampling period 1s 0.5℃ Humidity sampling period 1s 3％ Data cache size 32KB - processor clock speed 168MHz - Communication rate 1Mbps -

[0237] A lightweight neural network is deployed on the edge side, and the network structure parameters are shown in Table 2:

[0238] Table 2 Lightweight Neural Network Configuration Parameters

[0239]

[0240]

[0241] The parameter configurations for the state evaluation equations are shown in Table 3:

[0242] Table 3 Parameter settings for the state assessment equation system

[0243] Equation type Weighting coefficient Threshold parameter Error range Load characteristic equation 0.3,0.3,0.4 0.8 0.1 Stability evaluation equation 0.4,0.3,0.3 0.7 0.1 Variability assessment equation 0.3,0.2,0.3,0.2 0.75 0.1 Abnormal state equations 0.3,0.3,0.2,0.2 0.8 0.1

[0244] Eight GPU servers were deployed remotely, each configured with eight GPUs and 32GB of video memory per card. A deep neural network was built for data analysis, and the network training parameters are shown in Table 4.

[0245] Table 4 Training parameters for deep neural networks

[0246] Parameter type Parameter value illustrate Batch size 256 Number of training samples per batch Learning rate 0.001 Initial learning rate Number of training rounds 100 Total number of training rounds Early stopping threshold 5 Number of consecutive rounds without improvement Loss weights 0.5,0.5 MSE and CE weights

[0247] The system ran for 90 days, collecting a total of 108MB of data. The statistics of various abnormal events identified are shown in Table 5.

[0248] Table 5. Statistics of Abnormal Events

[0249] Exception types Number of occurrences Accurately identify the number of times Recognition rate Voltage abnormality 856 825 96.4％ abnormal current 923 894 96.9％ Power factor abnormality 467 442 94.6％ abnormal load 1258 1196 95.1％ Arc fault 89 86 96.6％ Temperature anomaly 245 236 96.3％ Communication error 156 148 94.9％

[0250] Figure 5 The data shows the occurrence trends of various abnormal events during the 90-day trial period. It can be seen that the overall trend of abnormal events is decreasing, indicating that the system's control effect is gradually becoming apparent. The system's key performance indicators are as follows: edge-side feature extraction time less than 100 milliseconds, data compression ratio of 1:10, communication latency less than 200 milliseconds, remote analysis response time less than 1 second, and average anomaly identification accuracy of 95.8%. Figure 6 The distribution of load characteristic assessment results during actual operation is shown, exhibiting an approximately normal distribution with values ​​concentrated around 0.75, indicating good stability of the system's assessment results. Compared to traditional remote monitoring methods for electricity meters, this invention has significant advantages in the following aspects: First, traditional methods use fixed threshold judgments, achieving an accuracy rate of only about 85%, while this invention, through lightweight neural networks and multi-dimensional evaluation equations, improves the accuracy rate to over 95%. Second, traditional methods require the transmission of all raw data, resulting in a heavy communication burden, while this invention reduces data transmission volume by 90% through end-side feature extraction and adaptive reporting strategies. Third, traditional methods generally have a response time of over 3 seconds, while this invention, through distributed architecture and optimization algorithms, controls the response time to within 1 second. Fourth, traditional methods struggle to identify complex anomaly patterns, while this invention, through deep learning and multi-dimensional analysis, can accurately identify various anomalies. Experimental results show that this invention effectively solves the technical problems of remote security management of electricity meters, significantly improving the system's intelligence level and management efficiency.

[0251] It should be noted that the variables involved in this invention are explained in detail in Table 6 below.

[0252] Table 6. Variable Explanation Table

[0253]

[0254] 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 smart energy meter data acquisition and alarm device, characterized in that, The system includes several meter-level monitoring systems and a remote monitoring system. The meter-level monitoring systems include a meter-level energy acquisition unit, a meter-level current sampling unit, a meter-level voltage sampling unit, a meter-level temperature sensing unit, a meter-level humidity sensing unit, a meter-level high-frequency current acquisition unit, and a meter-level control chip, all installed inside the meter. The remote monitoring system includes a remote-level data receiving unit, a remote-level data storage unit, a remote-level analysis module, a remote-level model training module, and a remote-level alarm unit. The meter-level control chip contains a data processing and reporting module, which performs the following steps: S11, receives power pulse signals, voltage signals, current signals, temperature signals, humidity signals, and high-frequency current signals; S12. Perform signal decomposition on voltage and current signals to extract stable and variable components; S13. Use a lightweight neural network to extract features from stable and variable components. The lightweight neural network extracts features using LoRa model parameters sent by the remote model training module. S14. Calculate historical load data, which is the data of the previous 24 hours of the stable component. S15. Calculate the load change rate, which is the ratio of the change in two consecutive sampled values ​​of the stable component to the sampling time interval. S16. Evaluate the operating status of the electricity meter based on the set of electricity meter status evaluation equations, and generate a comprehensive electricity meter status evaluation result. S17. Generate a data reporting strategy based on the comprehensive status assessment results of the electricity meter, and select to report raw data or feature data according to the data reporting strategy. The remote analysis module is used to perform the following steps: S21. Receive raw data or feature data reported by multiple meter terminal monitoring systems; S22. Construct a deep neural network structure, which includes a feature encoding layer, a power quality analysis layer, a load analysis layer, an arc analysis layer, and a state assessment layer. S23. Utilize deep neural networks to analyze the raw data and feature data, and generate power quality assessment results, load characteristic assessment results, and arc characteristic assessment results; S24. Store the comprehensive status assessment results of the electricity meter and generate a vector of historical anomaly records; S25. Generate an operational status assessment report based on the power quality assessment results, load characteristic assessment results, and arc characteristic assessment results; S26. Determine whether to trigger an alarm signal based on the operational status assessment report; The meter-side energy acquisition unit is electrically connected to the metering circuit and is used to acquire energy pulse signals and transmit them to the meter-side control chip. The meter-side current sampling unit is electrically connected to the current sampling circuit and is used to acquire current signals, perform signal conditioning, and transmit them to the meter-side control chip. The meter-side voltage sampling unit is electrically connected to the voltage sampling circuit and is used to acquire voltage signals, perform signal conditioning, and transmit them to the meter-side control chip. The meter-side high-frequency current acquisition unit is electrically connected to the current sampling circuit and is used to acquire high-frequency current signals and transmit them to the meter-side control chip. The meter-side temperature sensing unit and humidity sensing unit are located inside the meter and are used to acquire temperature and humidity signals and transmit them to the meter-side control chip. The remote data receiving unit is implemented using a database cluster, the remote data storage unit is implemented using a distributed file system, the remote analysis module is implemented using a GPU server cluster, the remote model training module is implemented using a GPU server cluster, and the remote alarm unit is implemented using a message queue system.

2. The smart energy meter data acquisition and alarm device according to claim 1, characterized in that, The lightweight neural network includes a feature extraction layer, a channel attention layer, a depthwise separable convolutional layer, a feature fusion layer, and a fully connected layer. The input of the feature fusion layer is the multi-channel feature map output by the depthwise separable convolutional layer, and the output is the fused feature vector.

3. The smart energy meter data acquisition and alarm device according to claim 1, characterized in that, The LoRa model decomposes the weight matrix of the lightweight neural network using a low-rank decomposition method. The weight matrix is ​​decomposed into the product of a basis matrix and an adaptation matrix. The basis matrix is ​​determined during training at a remote location, and the adaptation matrix is ​​used to adjust the feature extraction process.

4. The smart energy meter data acquisition and alarm device according to claim 1, characterized in that, The stable components include steady-state voltage value, steady-state current value, and steady-state power factor. The steady-state voltage value is the mean value of the voltage signal, the steady-state current value is the mean value of the current signal, and the steady-state power factor is the cosine of the phase difference between the steady-state voltage value and the steady-state current value.

5. The smart energy meter data acquisition and alarm device according to claim 4, characterized in that, The meter condition assessment equation set includes load characteristic equation, stability assessment equation, variability assessment equation, and abnormal state equation.

6. The smart energy meter data acquisition and alarm device according to claim 5, characterized in that, The load characteristic equation is used to evaluate the load operation status of the electricity meter. The inputs include the steady-state voltage value, the steady-state current value, and the historical load data. The output is the load status evaluation coefficient. The stability evaluation equation is used to evaluate the stability of power quality. The inputs include voltage fluctuation, current fluctuation, temperature signal, and preset threshold. The output is the stability evaluation coefficient.

7. The smart energy meter data acquisition and alarm device according to claim 6, characterized in that, In the deep neural network of the remote analysis module, the feature encoding layer adopts a fully connected neural network structure, the input dimension is the sum of the dimensions of the original data and the feature data, and the number of neurons in the hidden layer is 512, 256, and 128.

8. The smart energy meter data acquisition and alarm device according to claim 7, characterized in that, The remote model training module trains the LoRa model using a data-parallel approach, optimized with a synchronous stochastic gradient descent algorithm, an initial learning rate of 0.001, a cosine annealing scheduling strategy, a minimum learning rate of 0.00001, a batch size of 256, and 100 training epochs.

9. The smart energy meter data acquisition and alarm device according to claim 8, characterized in that, The remote model training module calculates the autocorrelation coefficient at different time scales based on the time correlation of historical anomaly record vectors using autocorrelation analysis. The time scale ranges from 1 hour to 24 hours, with a step size of 1 hour. The time scale with the largest autocorrelation coefficient is selected as the feature time scale.

Images