Intelligent electric energy meter data acquisition and alarm device

Abstract

The invention provides an intelligent electric energy meter data acquisition and alarm device, which belongs to the technical field of intelligent electric energy meters, and comprises a plurality of electric energy meter end monitoring systems and a remote end monitoring system, the ammeter end monitoring system comprises an ammeter end electric energy acquisition unit, an ammeter end current sampling unit, an ammeter end voltage sampling unit, an ammeter end temperature sensing unit, an ammeter end humidity sensing unit, an ammeter end high-frequency current acquisition unit and an ammeter end control chip which are arranged in the electric energy meter; according to the device, Lora model parameters are optimized through an adaptive learning mechanism, end-far collaborative intelligent management and control are achieved, the accuracy and efficiency of safety management and control of the electric energy meters are effectively improved, and the technical problem that in the prior art, timely sensing of abnormal states of massive electric energy meters through a far end is insufficient is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent electricity meters, and more specifically, relates to an intelligent electricity meter data acquisition and alarm device. Background Art

[0002] As an important device for metering and settlement in the power system, the safety control of electricity meters is of great significance for ensuring the safe operation of the power grid and the order of electricity use. Traditional electricity meter management mainly relies on manual inspections and simple remote data acquisition to detect abnormal situations through regular meter reading and basic parameter monitoring. With the in-depth promotion of the construction of the intelligent power grid, the number of electricity meters has increased sharply, and the types of electricity loads have become increasingly complex. The traditional management method has been difficult to meet the safety control requirements of large-scale intelligent power grids.

[0003] The existing remote control systems for electricity meters mainly have the following problems: First, in the face of a large number of electricity meters, the traditional centralized management method is difficult to achieve real-time monitoring of each meter, resulting in abnormal situations that cannot be detected and processed in a timely manner; Second, the existing abnormal detection methods are mainly based on simple threshold judgments and cannot effectively identify complex abnormal electricity consumption behaviors, such as concealed illegal behaviors like electricity theft and unauthorized modification; Third, the operating status of electricity meters is affected by various factors, including power grid quality, environmental conditions, and load characteristics. The existing systems lack the ability to comprehensively analyze these factors and are difficult to accurately evaluate the health status of electricity meters. In addition, due to the wide distribution range and complex installation environment of electricity meters, the system needs to adapt to different communication conditions and operating environments, which poses a huge challenge to remote safety control.

[0004] Currently, the industry mainly tries to solve these problems in the following ways: One method is to increase the acquisition parameters and sampling frequencies, attempting to improve the accuracy of abnormal detection through more data. However, this method has led to a sharp increase in system burden and limited improvement in effectiveness; Another method is to deploy complex analysis models at the remote end. However, due to the lack of effective extraction of end-side features, the actual effects of the models are often not satisfactory. In practice, these methods have not fundamentally solved the problem of remote safety control of electricity meters. Especially in the face of large-scale intelligent power grids, the response speed, detection accuracy, and control efficiency of the system are difficult to meet the actual requirements. That is to say, there is a technical problem in the prior art of insufficient timely perception of the abnormal states of a large number of electricity meters through the remote end. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent electricity meter data acquisition and alarm device, which can solve the technical problem of insufficient timely perception of the abnormal states of a large number of electricity meters through the remote end in the prior art.

[0006] The present invention is implemented as follows: The present invention provides an intelligent electricity meter data acquisition and alarm device, which includes several meter-end monitoring systems and a remote-end monitoring system. The meter-end monitoring system includes a meter-end power 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 disposed inside the electricity 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 alarm unit. A data processing and reporting module is provided inside the meter-end control chip. The data processing and reporting module uses a lightweight neural network to extract features from the stable components and varying components of the voltage signal and current signal. The lightweight neural network performs feature extraction based on the Lora model parameters sent from the remote end. The remote-end 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 status evaluation layer, and is used to analyze the original data and feature data to generate power quality evaluation results, load feature evaluation results, and arc feature evaluation results.

[0007] Among them, the lightweight neural network includes a feature extraction layer, a channel attention layer, a depthwise separable convolution 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 convolution layer, and the output is the fused feature vector.

[0008] Among them, the Lora model decomposes the weight matrix of the lightweight neural network through 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 by training at the remote end, and the adaptation matrix is used to adjust the feature extraction process.

[0009] Among them, the stable components include the steady-state voltage value, the steady-state current value, and the 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 value of the phase difference between the steady-state voltage value and the steady-state current value.

[0010] Among them, the data processing and reporting module is also used to calculate historical load data, which is the data of the stable components in the previous 24 hours; calculate the load change rate, which is the ratio of the change amount between two adjacent sampling values of the stable components to the sampling time interval.

[0011] Among them, the meter - end control chip evaluates the operation state of the meter based on the meter state evaluation equation set, and the meter state evaluation equation set includes a load characteristic equation, a stability evaluation equation, a variability evaluation equation, and an abnormal state equation.

[0012] Among them, the load characteristic equation is used to evaluate the load operation state of the 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 state evaluation coefficient. The stability evaluation equation is used to evaluate the stability degree of power quality. The inputs include the voltage fluctuation amount, the current fluctuation amount, the temperature signal, and the preset threshold, and the output is the stability evaluation coefficient.

[0013] Among them, 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] Among them, when the remote - end model training module trains the Lora model, it uses the data parallel method, optimizes it using the synchronous stochastic gradient descent algorithm, the initial value of the learning rate is 0.001, adopts the cosine annealing scheduling strategy, the minimum learning rate is 0.00001, the batch size is 256, and the number of training rounds is 100 rounds.

[0015] Among them, the remote - end model training module calculates the autocorrelation coefficients at different time scales using the autocorrelation analysis method according to the time correlation of the historical abnormal record vector. The time scale ranges from 1 hour to 24 hours, the step size is 1 hour, and the time scale with the largest autocorrelation coefficient is selected as the characteristic time scale.

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

[0017] Further, the deep neural network includes a feature encoding layer, a power quality analysis layer, a load analysis layer, an arc analysis layer, and a status 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 status evaluation layer adopts an attention mechanism.

[0018] Optionally, the operation status evaluation report includes basic information, operation status, anomaly analysis, and warning prompt. The basic information includes the meter number, installation location, and user type. The operation status includes the values and levels of various evaluation indicators. The anomaly analysis includes the anomaly type, anomaly degree, and anomaly duration. The warning prompt includes the warning level, warning reason, and handling suggestion.

[0019] Further, the lightweight neural network is implemented using an improved MobileNet V2 structure. The standard convolution is decomposed into a depth convolution independent by channel and a pointwise convolution through depthwise separable convolution, and a residual connection structure is introduced to establish a shortcut path for identity mapping.

[0020] Optionally, a distributed training framework is used to train the Lora model. The training adopts a data parallelism method and is optimized using the synchronous stochastic gradient descent algorithm. The loss function adopts a weighted combination of mean squared error and cross entropy.

[0021] Compared with the prior art, an intelligent electric energy meter data acquisition and alarm device provided by the present invention realizes efficient and safe control of a large number of electric energy meters through the organic combination of edge intelligence and remote management and control. At the edge side, the real-time extraction of the operation characteristics of the electric energy meter is realized through a lightweight neural network. The use of the improved MobileNet V2 structure significantly reduces the computational complexity while ensuring the accuracy of feature extraction. The introduction of the feature fusion layer and the channel attention mechanism enables the system to capture richer status information, providing a reliable data basis for the identification of abnormal behaviors.

[0022] The present invention establishes a complete status evaluation system, including evaluations in multiple dimensions such as load characteristics, stability, variability, and abnormal status. Through a hierarchical evaluation equation set, the system can comprehensively analyze the operation status of the electric energy meter and accurately identify various abnormal situations. The remote side uses a deep neural network for comprehensive analysis, combines historical data and time series characteristics, and realizes the accurate identification of abnormal electricity consumption behaviors. At the same time, the system has an adaptive learning ability and can continuously optimize the detection strategy according to historical abnormal records, improving the accuracy and efficiency of management and control.

[0023] Through the low-rank decomposition method of the Lora model, the present invention realizes the efficient update and distribution of model parameters, enabling the system to quickly adapt to the control requirements of different scenarios. The design of the multi-level alarm strategy ensures that abnormal situations can be processed in a timely manner, while the storage scheme based on the time series database provides data support for long-term analysis and optimization. The comprehensive application of these technologies enables the system to achieve efficient utilization of resources while ensuring the control effect. In summary, the present invention solves the technical problem of insufficient timely perception of abnormal states of a large number of electric energy meters through the remote end in the prior art. Brief Description of the Drawings

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

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

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

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

[0028] Figure 5 It is a graph showing the occurrence trend of various abnormal events during the 90-day test period in Example 2.

[0029] Figure 6 It is a graph showing the distribution of the load characteristic evaluation results in Example 2. Detailed Embodiments

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0031] As Figure 1 shown, it is a schematic diagram of the composition of an intelligent electric energy meter data acquisition and alarm device provided by the present invention. This device includes several meter-end monitoring systems and a remote-end monitoring system. The meter-end monitoring system includes a meter-end electric 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 arranged inside the electric energy meter;

[0032] 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 alarm unit;

[0033] The power acquisition unit at the meter end is electrically connected to the power meter measurement circuit, and is used to collect power pulse signals and transmit the power pulse signals to the control chip at the meter end;

[0034] The current sampling unit at the meter end is electrically connected to the current sampling circuit of the power meter, and is used to collect current signals, perform signal conditioning, and then transmit them to the control chip at the meter end;

[0035] The voltage sampling unit at the meter end is electrically connected to the voltage sampling circuit of the power meter, and is used to collect voltage signals, perform signal conditioning, and then transmit them to the control chip at the meter end;

[0036] The temperature sensing unit at the meter end is arranged inside the power 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 arranged inside the power 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 end is electrically connected to the current sampling circuit of the power meter, and is used to collect high-frequency current signals and transmit them to the control chip at the meter end;

[0039] A data processing and reporting module is arranged in the control chip at the meter end, and the data processing and reporting module 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. Decompose the voltage signal and the current signal, and extract the stable component and the variable component;

[0042] S13. Use the lightweight neural network to extract features from the stable component and the variable component, and the lightweight neural network extracts features through the Lora model parameters sent from the remote end;

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

[0044] S15. Calculate the load change rate, where the load change rate is the ratio of the change amount of two adjacent sampling values of the stable component to the sampling time interval;

[0045] S16. Evaluate the operating state of the meter based on the meter state evaluation equation set, and generate the comprehensive state evaluation result of the meter;

[0046] S17. Generate a data reporting strategy based on the comprehensive status evaluation result of the electricity meter, and select to report the original 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 the original data and feature data reported by multiple electricity meter monitoring systems;

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

[0050] S23. Analyze the original data and the feature data using the deep neural network to generate a power quality evaluation result, a load feature evaluation result, and an arc feature evaluation result;

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

[0052] S25. Generate an operation status evaluation report based on the power quality evaluation result, the load feature evaluation result, and the arc feature evaluation result;

[0053] S26. Determine whether to trigger an alarm signal according to the operation status evaluation report.

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

[0055] S31. Summarize the historical data of multiple electricity meter monitoring systems;

[0056] S32. Construct a Lora model training sample according to the historical data;

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

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

[0059] S35. Calculate the time scale parameter, which 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 parameter to the electricity meter monitoring system.

[0061] Among them, the lightweight neural network includes a feature extraction layer, a channel attention layer, a depthwise separable convolution 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 convolution layer, and the output is the fused feature vector;

[0062] The Lora model decomposes the weight matrix of the lightweight neural network through 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. The basis matrix is determined by training at the remote end, and the adaptation matrix is used to adjust the feature extraction process;

[0063] The stable components include the steady-state voltage value, the steady-state current value, and the 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 value of the phase difference between the steady-state voltage value and the steady-state current value;

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

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

[0066] The feature data includes the stable components, the variable components, and the load change rate;

[0067] The electricity meter state evaluation equations include a load characteristic equation, a stability evaluation equation, a variability evaluation equation, and an abnormal state equation;

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

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

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

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

[0072] The lightweight neural network is implemented using an improved MobileNet V2 structure. Specifically, it constructs an efficient and compact feature extraction module by deeply optimizing the inverted residual structure and depthwise separable convolution in the network architecture. This module first uses a 1×1 convolution for channel expansion operation to map low-dimensional features to a high-dimensional space to enhance the network's feature expression ability. Subsequently, it uses a 3×3 depthwise separable convolution for spatial feature extraction in the expanded high-dimensional feature space. The depthwise separable convolution significantly reduces the computational complexity by decomposing the standard convolution into two steps: depthwise convolution independent by channel and pointwise convolution. Finally, it uses a 1×1 convolution for channel compression to map the features back to the low-dimensional space. The whole process not only maintains good feature extraction ability but also greatly reduces the model parameter quantity and computational overhead. On this basis, the network also introduces a residual connection structure, establishing a shortcut path for identity mapping when the input and output channel numbers are the same and the stride is 1. This design not only helps the backpropagation of gradients to improve the model's training effect but also enhances the model's expression ability through feature reuse. In addition, the network adds a batch normalization layer and a ReLU6 activation function after each convolutional layer to enhance the model's non-linear expression ability and accelerate training convergence. The truncation design of ReLU6 can also provide better numerical stability for quantization deployment. To further improve the model performance, the improved version integrates a channel attention mechanism on the basis of the original MobileNet V2, enhancing effective features and suppressing redundant information by adaptively adjusting the importance of different channels. At the same time, a feature fusion path is established between different levels of the network, enriching the semantic expression of features and improving the detection ability of multi-scale targets. Finally, the network uses a global average pooling layer to replace the traditional fully connected layer, significantly reducing the model parameter quantity, and adds a dropout layer before the classification layer to effectively alleviate the overfitting phenomenon. The overall network not only maintains the lightweight feature but also achieves a good balance in accuracy and performance.

[0073] While maintaining the advantages of the original MobileNet V2, this improved network structure further enhances the feature extraction ability through the attention mechanism and multi-scale feature fusion, enabling the model to better adapt to the visual task requirements in different scenarios. Its lightweight feature makes it particularly suitable for deployment on mobile devices and embedded platforms. At the same time, the scalability of the model structure also provides a flexible adjustment space for performance optimization in different application scenarios. In practical applications, the model size and performance can be balanced by adjusting the width factor of the network, or the configurations of each module can be finely tuned according to specific task requirements to obtain the optimal performance under resource constraints.

[0074] The specific implementation manners of the above steps will be described in detail below. In the specific implementation manner of the meter-end monitoring system, the meter-end power acquisition unit is implemented by using a high-precision pulse acquisition chip, which supports the acquisition of pulse signals from 1 Hz to 100 Hz, with a sampling accuracy of 0.1%. The signal acquisition is realized by using the optoelectronic isolation technology, which has extremely strong anti-interference ability and stability.

[0075] The meter-end current sampling unit is implemented by using 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-end voltage sampling unit is implemented by combining a resistor voltage divider and an operational amplifier, with a range of 0 to 380 volts, a sampling frequency of up to 10 kHz, and a sampling accuracy of 0.5%. It has a built-in overvoltage protection circuit and a signal conditioning circuit, which can effectively prevent the damage of the sampling circuit caused by voltage mutation.

[0077] The meter-end temperature sensing unit is implemented by using a digital temperature sensor, with a measurement range of -40 °C to 125 °C, a measurement accuracy of 0.5 °C, a sampling period of 1 second, and a digital filtering function, which can effectively suppress the fluctuation of the temperature signal.

[0078] The meter-end humidity sensing unit is implemented by using 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, which can ensure the measurement accuracy at different temperatures.

[0079] The meter-end high-frequency current acquisition unit is implemented by using a broadband current transformer, with a frequency band width 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%, which is used to acquire the high-frequency components in the current signal.

[0080] The meter end control chip is implemented using a 32-bit microcontroller with a main frequency of not less than 800 MHz, built-in hardware multiplier and hardware divider, and equipped with a hardware floating-point arithmetic unit. The on-chip storage capacity is not less than 1 megabyte, which is used to implement data acquisition, data processing, and data reporting functions.

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

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

[0083] The remote end analysis module is implemented using a GPU server cluster. The single-machine configuration is not less than 8 GPU cards, and the video memory capacity is not less than 32 gigabytes per card, which is used to implement online inference of deep learning models.

[0084] The remote end model training module is implemented using a GPU server cluster. The single-machine configuration is not less than 8 GPU cards, and the video memory capacity is not less than 80 gigabytes per card, which is used to implement offline training of deep learning models.

[0085] The remote end alarm unit is implemented using a message queue system, supporting multiple alarm methods, including SMS alarm, email alarm, and application push alarm, and the alarm delay does not exceed 1 second.

[0086] The specific implementation methods of these components fully consider the characteristics and requirements of the electric energy metering system, adopt high-precision sensors and high-performance processors to ensure the system has high reliability and stability. At the same time, the distributed architecture and cluster technology are adopted to ensure the system has strong scalability and fault tolerance. In actual applications, the parameters of each component can be appropriately adjusted according to specific needs to meet the application requirements in different scenarios.

[0087] The steps executed by the data processing and reporting module are specifically described as follows:

[0088] The specific implementation method of step S11 is to collect the electrical energy pulse signal based on the interrupt method, collect the voltage signal and current signal using the timer interrupt method, collect the temperature signal and humidity signal using the serial communication method, and collect the high-frequency current signal using the direct memory access method. The sampling clock is provided by an external crystal oscillator with a frequency of 32.768 kHz. The sampling data is stored according to the time stamp with a time stamp accuracy of 1 millisecond. The data storage adopts a circular buffer method with a buffer 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 is to perform frequency-domain analysis on the voltage signal and current signal using the fast Fourier transform algorithm, extract the fundamental component as the stable component, and extract the harmonic component as the variable component. The number of points for the fast Fourier transform is 1024, and the Hanning window function is used for preprocessing. Fast calculation is achieved through bit-reversal and butterfly operations. The frequency resolution of the calculation result is 10 Hz, the amplitude accuracy is 0.1%, and the phase accuracy is 0.1 degree. This step aims to decompose the complex voltage and current signals into two parts, stable and variable, for subsequent analysis.

[0090] The specific implementation of step S13 is based on an improved lightweight neural network structure. First, the input data is normalized using the maximum-minimum method. Then, preliminary feature extraction is performed through the feature extraction layer, which contains 3 convolutional layers with kernel sizes of 1×1, 3×3, and 1×1 respectively, and the number of channels is 32, 64, and 32 respectively. After each convolutional layer, a batch normalization layer and an activation function layer are connected. The activation function uses the rectified linear unit 6. Subsequently, the channel attention layer weights the channel dimension of the feature map, and the weights are calculated through global average pooling and a two-layer fully connected network. Then, feature enhancement is performed through the depthwise separable convolutional layer. Finally, the multi-scale features are fused through the feature fusion layer, and the fusion method is weighted summation, and the weights are optimized through the backpropagation algorithm. The goal of this step is to extract the key features of the input data and provide a basis for subsequent analysis.

[0091] The specific implementation of step S14 is to establish a sliding time window based on the stable component data. The length of the time window is 24 hours, and the sliding step is 15 minutes. Statistical analysis is performed on the data within the window, and statistical features such as the maximum value, minimum value, average value, and standard deviation are calculated. 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. This step is used to obtain the statistical features of historical loads and provide a basis for load analysis.

[0092] The specific implementation of step S15 is to perform difference calculation on the stable component, calculate the difference between two adjacent sampling values, and divide by the sampling time interval to obtain the change rate. The sampling time interval is 1 minute. At the same time, the mean and standard deviation of the change rate are calculated, and a statistical model of the change rate is established to determine whether the load change is abnormal. When the change rate exceeds the mean plus or minus 3 times the standard deviation, it is determined to be an abnormal change. The role of this step is to monitor the dynamic characteristics of load changes.

[0093] The specific implementation of step S16 is to calculate according to the equation set for evaluating the electricity meter status. The equation set is established using the weight coefficient method. The weight coefficient of the load characteristic equation is 0.3, the weight coefficient of the stability evaluation equation is 0.3, the weight coefficient of the variability evaluation equation is 0.2, and the weight coefficient of the abnormal status equation is 0.2. The calculation results of each equation are normalized and then weighted and summed to obtain the comprehensive evaluation result. This step aims to comprehensively evaluate the operation status of the electricity meter.

[0094] The specific implementation of step S17 is to, according to the comprehensive status evaluation result of the electricity meter, when the evaluation result is greater than 0.8, it is determined as an abnormal status and the original data is selected for reporting; when the evaluation result is less than or equal to 0.8, it is determined as a normal status and the characteristic data is selected for reporting. The data reporting uses a wireless communication method, the communication protocol uses 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. The purpose of this step is to achieve intelligent data reporting and reduce the communication burden.

[0095] The implementation manners of these steps make full use of digital signal processing and deep learning technologies. Through multi-level data analysis and processing, the intelligent collection and reporting of power metering data are realized, which can effectively improve the operation efficiency and reliability of the system.

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

[0097] The specific implementation of step S21 is to receive data using a distributed message queue system. The message queue adopts the topic subscription mode, an independent data topic is established for each electricity meter terminal monitoring system, and the topics are classified according to geographical regions and electricity consumption types. The maximum delay time of the message is 100 milliseconds, the reliability level of the message is at least once, the persistence strategy of the message is asynchronous disk flushing, the storage time of the message is 30 days, and the system supports message replay and filtering, and message filtering can be performed according to the timestamp and data type. This step ensures the reliability and real-time nature of data reception.

[0098] 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 activation function uses the rectified linear unit, and the dropout rate is 0.3. The power quality analysis layer adopts a convolutional neural network structure, including 3 convolutional blocks. Each convolutional block contains 2 convolutional layers and 1 max-pooling layer. The convolutional kernel size is 3×3, the stride is 1, and the padding method is same padding. The pooling kernel size is 2×2, and the stride is 2. The load analysis layer adopts a long short-term memory network structure, including 2 long short-term memory layers. The hidden state dimension of each layer is 128. The bias of the forget gate is initialized to 1, and the biases of the input gate and the output gate are initialized to 0. The arc analysis layer adopts a residual neural network structure, including 4 residual blocks. Each residual block contains 2 convolutional layers and 1 short circuit connection. The convolutional kernel size is 3×3, and the number of channels is 64. The state evaluation layer adopts an attention mechanism, and the output features of each layer are weighted and fused through a soft attention module. The attention weights are calculated through a parameterized similarity function. This step establishes a multi-task learning framework and can perform analysis in multiple aspects simultaneously.

[0099] The specific implementation of step S23 is to input the data into the deep neural network for forward propagation calculation. First, the input data is dimensionally reduced and feature extracted by the feature encoding layer, and then the extracted features are respectively input into the power quality analysis layer, the load analysis layer, and the arc analysis layer for analysis. The power quality evaluation results include indicators such as voltage deviation rate, current distortion rate, and power factor, and the evaluation thresholds are ±5%, 5%, and 0.9 respectively. The load feature evaluation results include indicators such as load prediction value, load change trend, and load anomaly degree. The prediction time window is 24 hours, and the prediction step is 15 minutes. The arc feature evaluation results include indicators such as arc occurrence probability, arc duration, and arc energy, and the judgment thresholds are 0.8, 100 milliseconds, and 10 joules respectively. This step realizes a comprehensive analysis of the operation state of the electric meter.

[0100] The specific implementation of step S24 is to store the comprehensive state evaluation results of the electric meter in a time series database. The database adopts a time-sharded storage method, with the time slice size being 1 day and the data compression ratio being 1:10. At the same time, a multi-level index is established to accelerate querying. The indexes include time index, electric meter number index, and anomaly type index. Anomaly record vectors are constructed based on the evaluation results. The vector dimension is the number of anomaly types, and the vector elements are the occurrence times of the corresponding anomaly types. The anomaly types include voltage anomaly, current anomaly, power factor anomaly, load anomaly, arc anomaly, etc. This step provides historical data support for anomaly analysis.

[0101] The specific implementation of step S25 is to generate an evaluation report based on various evaluation results. The report includes four parts: basic information, operating status, anomaly analysis, and warning prompt. The basic information includes the electricity 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 warning prompt includes the warning level, warning reason, handling suggestions, etc. The warning levels are divided into general warning, important warning, and emergency warning, and the thresholds are 0.6, 0.8, and 0.9 respectively. This step converts the analysis results into an understandable report form.

[0102] The specific implementation of step S26 is to determine whether to trigger an alarm according to the operating status evaluation report, and adopt a multi-level alarm strategy. For general warnings, a system internal prompt is used. For important warnings, an email notification is sent simultaneously. For emergency warnings, a text message and a phone call are sent simultaneously. The alarm information includes the alarm time, alarm reason, alarm level, handling suggestions, etc. The system supports alarm confirmation and alarm escalation functions. When the alarm duration exceeds the preset time and is not processed, the alarm level is automatically increased. This step realizes the timely notification and handling of abnormal situations.

[0103] The implementation of these steps constructs a complete remote analysis system, which realizes the intelligent analysis of power metering data through deep learning technology, can timely detect and handle abnormal situations, and improves the intelligent level and management efficiency of the system.

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

[0105] The specific implementation of step S31 is to collect historical data reported by the electricity meter end monitoring system through a distributed data acquisition system. The data sources include raw data and feature data. The data is stored using a distributed file system, and the data is sharded and stored according to time and geographical location. The size of each shard does not exceed 1 gigabyte. The data is preprocessed through a data quality control module, including outlier detection, missing value processing, and noise filtering. The 3-sigma method is used for outlier detection, linear interpolation is used to supplement missing values, and median filtering is used for noise filtering. The size of the filtering window is 5. The purpose of this step is to establish a high-quality training dataset.

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

[0107] The specific implementation of step S33 is to train the Lora model using a distributed training framework. The training uses data parallelism, with 8 GPU servers forming a training cluster, and each server is configured with 8 GPUs. The synchronous stochastic gradient descent algorithm is used for optimization, with an initial learning rate of 0.001. The cosine annealing scheduling strategy is adopted, with a minimum learning rate of 0.00001, a batch size of 256, and 100 training epochs. The early stopping strategy is adopted, and training stops when the validation set loss does not decrease for 5 consecutive epochs. The loss function uses a weighted combination of mean squared error and cross entropy, with a weight ratio of 1:1. Gradient clipping is used to prevent gradient explosion, with a clipping threshold of 5. This step realizes the optimization of model parameters.

[0108] The specific implementation of step S34 is to calculate the preset threshold based on the historical anomaly record vector using statistical analysis methods. First, cluster analysis is performed on the anomaly record vector using the density-based spatial clustering algorithm. The minimum number of samples for clustering is 10, and the maximum distance parameter is 0.1 to obtain different types of anomaly clusters. Then, statistical features are calculated for each anomaly cluster, including mean, standard deviation, quantiles, etc. Based on these statistical features, the preset threshold is determined. For voltage anomalies, the preset threshold is plus or minus 7% of the standard voltage; for current anomalies, the preset threshold is 150% of the rated current; for power factor anomalies, the preset threshold is 0.85. This step provides a scientific judgment standard for anomaly detection.

[0109] The specific implementation of step S35 is to analyze the time correlation of the historical anomaly record vector using the autocorrelation analysis method. Calculate the autocorrelation coefficients at different time scales, with the time scale ranging from 1 hour to 24 hours and a step of 1 hour. Select the time scale with the largest autocorrelation coefficient as the characteristic time scale. At the same time, calculate the cross-correlation coefficient to analyze the time correlation between different types of anomalies, and construct a time correlation network. The network nodes are anomaly types, and the edge weights are cross-correlation coefficients. Calculate the time scale parameter based on the correlation network. This step reveals the time law of anomaly occurrence.

[0110] The specific implementation of step S36 is to use a remote parameter configuration system to send the trained model parameters to the electricity meter monitoring system. The parameter sending is carried out in batches, with 100 electricity meter monitoring systems sent in each batch. The sent content includes Lora model parameters, preset thresholds, and time scale parameters. The incremental update method is adopted, and only the changed parameters are sent. The parameter compression uses the sparse matrix compression method, and the compression ratio can reach 1:20. The transmission uses a secure channel and is encrypted using the national cryptographic algorithm, with a key length of 256 bits. At the same time, parameter consistency verification is carried out to ensure the correct sending of parameters. This step realizes the deployment and update of the model.

[0111] The implementation of these steps establishes a complete model training and deployment process. Through distributed computing and optimization algorithms, the efficient training and update of the model are realized. At the same time, through scientific statistical analysis methods, reasonable preset thresholds and time scale parameters are determined, providing strong support for the intelligent operation of the system. Generally speaking, this system realizes the intelligent monitoring and management of electricity metering equipment through deep learning technology and distributed computing, and has high practical value and promotion value.

[0112] Among them, the calculation of the steady-state voltage value and current value is specifically expressed as follows:

[0113]

[0114] In the formula, V stable is the steady-state voltage value; I stable is the steady-state current value; V i is the voltage value at the i-th sampling point; I i is the current value at the i-th sampling point; N is the number of sampling points, and the value range is 1000 - 10000.

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

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

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

[0118] The calculation of the voltage fluctuation amount and current fluctuation amount is specifically expressed as follows:

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

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

[0121] Wherein, V fluctuation is the voltage fluctuation amount; I fluctuation is the current fluctuation amount; V(t) is the voltage value at time t; I(t) is the current value at time t.

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

[0123]

[0124] Wherein, F load is the load status evaluation coefficient; V n is the rated voltage; I n is the rated current; P load is the load power; P n is the rated power; α1, α2, α3 are weight coefficients determined by the least squares method; ε1 is the error term with a range of 0 to 0.1.

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

[0126]

[0127] Wherein, F stability is the stability evaluation coefficient; V th is the voltage fluctuation threshold; I th is the current fluctuation threshold; T is the temperature value; T n is the rated temperature; T th is the temperature deviation threshold; β1, β2, β3 are weight coefficients; ε2 is the error term with a range of 0 to 0.1.

[0128] The variability evaluation equation is specifically expressed as follows:

[0129]

[0130] Wherein, F variation is the variability evaluation coefficient; is the load change rate; k1, k2 are attenuation coefficients; γ1, γ2, γ3, γ4 are weight coefficients; ε3 is the error term with a range of 0 to 0.1.

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

[0132]

[0133] Wherein, F abnormalis the comprehensive status evaluation result of the electricity meter; H abnormal is the historical anomaly record vector; t is the time variable; τ is the time scale parameter; δ1, δ2, δ3, δ4 are the weight coefficients; ε4 is the error term, and its range is 0 to 0.1.

[0134] The calculation of depthwise separable convolution is specifically expressed as follows:

[0135]

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

[0137] The calculation of channel attention is specifically expressed as follows:

[0138]

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

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

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

[0142] W = W0 + BA;

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

[0144] The construction principles of these equations mainly consider the following aspects:

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

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

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

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

[0149] 1. The derivation process of the steady-state voltage value and current value calculation equations is described in detail as follows:

[0150] First, the sliding window method is used to segment the original sampling data. The window length is 1000 sampling points, and the sliding step is 100 sampling points;

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

[0152] Finally, the arithmetic mean of the valid 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 as follows:

[0154] First, Fourier transform is performed on the voltage and current signals to extract the fundamental components;

[0155] Then, the phase difference between the fundamental voltage and current is calculated;

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

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

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

[0159]

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

[0161]

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

[0163] F load =α T L+ε1;

[0164] Among them, the weight vector α is solved by the least squares method:

[0165] α=(L T L) -1 L T Y;

[0166] In the formula, Y is the load status score marked by experts.

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

[0168] The fluctuation eigenvector S is specifically represented 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 the exponential function form can make the evaluation result smoothly vary between 0 and 1. 5. The derivation process of the variability evaluation equation is described in detail as follows:

[0175] The variability eigenvector V is specifically represented as follows:

[0176]

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

[0178]

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

[0180] F variation = γ T V + ε3;

[0181] The exponential decay term is used to describe the non - linear influence of the fluctuation amount on the variability.

[0182] 6. The computational matrix representation of depth - wise separable convolution:

[0183] The input feature map X c is specifically represented as an H×W - dimensional matrix:

[0184]

[0185] The depth - wise convolution kernel K c is specifically represented as a k×k - dimensional matrix:

[0186]

[0187] The point - wise convolution weight W c is specifically represented as an M - dimensional vector:

[0188] W c = [w1 w2 … wM .

[0189] 7. Specific representation of the Lora model weight matrix factorization:

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

[0191]

[0192] The basis matrix W0 has the same dimension as W;

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

[0194]

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

[0196]

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

[0198] Specifically, the principle of the present invention is: The technical principle of the present invention is based on the design concept of hierarchical architecture and collaborative optimization. On the edge side, depthwise separable convolution is used to decompose the standard convolution operation into depth convolution and pointwise convolution, significantly reducing the computational complexity. By introducing the 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 equation system embodies the principle of multi-dimensional analysis. The load feature equation reflects the basic characteristics of load operation through normalization processing and weight calculation; the stability evaluation equation adopts an exponential decay form to describe the impact of the fluctuation amount on the system stability; the variability evaluation equation comprehensively considers the power factor, fluctuation amount, and change rate to reflect the dynamic characteristics of the system; the abnormal state equation reflects the impact of historical anomalies on the current state by introducing a time decay term.

[0200] On the remote side, the design of the deep neural network adopts a multi-level analysis structure, including a feature encoding layer, a power quality analysis layer, a load analysis layer, and an arc analysis layer. This structure can analyze the operating state of the electricity meter from different angles, improving the accuracy of anomaly detection. The adaptive learning mechanism optimizes the model parameters and evaluation thresholds by analyzing historical data, enabling the system to continuously improve the control effect.

[0201] The following provides a specific Embodiment 1 of the present invention. The specific implementation of each step in this Embodiment 1 is described in detail as follows.

[0202] In Embodiment 1, the electric energy acquisition unit at the meter end is implemented using the high-precision pulse acquisition chip ADE7953. The working voltage of this chip is 3.3 volts. Signal isolation is achieved through the optocoupler 4N35. The sampling accuracy reaches 0.1%, and it supports the acquisition of pulse signals from 1 Hz to 100 Hz. A low-pass filter is used for signal preprocessing, with a cut-off frequency of 200 Hz. The circuit adopts a differential input mode, with a common-mode rejection ratio greater than 80 dB, and has strong anti-interference ability. It is connected to the control chip at the meter end through a serial peripheral interface, with a baud rate of 1 megabit per second.

[0203] The current sampling unit at the meter end is implemented using the Hall current sensor ACS758, with a measurement 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. An internal operational amplifier AD8554 constitutes a signal conditioning circuit with a gain of 2 times. The common-mode rejection ratio of the operational amplifier is greater than 100 dB. It is connected to the control chip at the meter end through a 12-bit analog-to-digital converter AD7866, and the conversion time is less than 1 microsecond.

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

[0205] The temperature sensing unit at the meter end is implemented using the digital temperature sensor DS18B20, with a measurement range of -40 degrees Celsius to 125 degrees Celsius, a measurement accuracy of 0.5 degrees Celsius, a resolution programmable from 9 to 12 bits, corresponding to a temperature resolution of 0.5 to 0.0625 degrees Celsius, a sampling period of 1 second, and communicates with the control chip at the meter end using a single-wire protocol, with a transmission rate of 15.4 kbps and strong anti-interference ability.

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

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

[0208] The control chip at the meter end is implemented using the 32-bit microcontroller STM32F407, with a main frequency of 168 MHz, 192 KB of on-chip random access memory, 1 MB of on-chip flash memory, equipped with a hardware floating-point arithmetic unit, an operation speed of up to 125 million floating-point operations per second, integrated with a DC-to-DC power management unit, and supporting multiple low-power modes.

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

[0210] The remote server uses a rack-mounted server, configured with a dual Xeon 8352Y processor, 32 cores and 64 threads for each processor, a main frequency of 3.0 GHz, 256 GB of memory, a hard disk using a solid-state drive array with a capacity of 20 TB, and a network interface of 10 Gigabit Ethernet, supporting server cluster management and load balancing.

[0211] The installation positions and connection methods of all components need to consider the requirements of electromagnetic compatibility. Shielded cables are used for analog signals, twisted pairs are used for digital signals, all metal enclosures need to be reliably grounded, the grounding resistance is less than 4 ohms, and a surge protector and an electromagnetic compatibility filter are installed at the power input end.

[0212] A data processing and reporting module is set in the control chip at the meter end, and the data processing and reporting module is used to execute the following steps:

[0213] The specific implementation of step S11 realizes the acquisition of various types of data through interruption. Among them, the electrical energy pulse signal is collected through external interruption, the sampling clock is provided by an external 32.768 kHz crystal oscillator, the temperature signal and humidity signal are collected through the serial communication peripheral at a sampling period of 1 second, the voltage signal and current signal are collected through the timer interruption method at a sampling frequency of 10 kHz, and the high-frequency current signal is collected through the direct memory access method at a sampling frequency of 1 MHz. The collected data is stored in a circular buffer according to the time stamp. The buffer adopts a double-buffer structure to avoid data read-write conflicts. The buffer size is 32 KB, and the time stamp accuracy is 1 ms. The steady-state voltage value is calculated using the following formula: where N is the number of sampling points, and the value range is from 1000 to 10000, V i is the voltage value of the i-th sampling point. The steady-state current value is calculated using the following formula: where I i is the current value of the i-th sampling point.

[0214] The specific implementation of step S12 is to perform frequency-domain analysis on the voltage signal and current signal using the fast Fourier transform algorithm. The number of transformation points is 1024 points. The Hanning window function is used for preprocessing to reduce spectral leakage. Fast calculation is achieved through bit-reversal and butterfly operations. The frequency resolution of the calculation result is 10 Hz, the amplitude accuracy is 0.1%, and the phase accuracy is 0.1 degree. The fundamental wave component is extracted as the stable component, and the harmonic component is extracted as the variable component. The steady-state power factor is calculated using the following formula: PF stable = cos(φ VI ) = cos(∠V stable - ∠I stable ), where φ VI is the voltage-current phase difference, ∠V stable is the steady-state voltage phase angle, and ∠I stable is the steady-state current phase angle. The voltage fluctuation amount is calculated using the following formula: V fluctuation = V(t) - V stable , where V(t) is the voltage value at time t. The current fluctuation amount is calculated using the following formula: I fluctuation = I(t) - I stable , where I(t) is the current value at time t.

[0215] The specific implementation of step S13 is to extract features based on an improved lightweight neural network structure. First, the input data is normalized using the maximum-minimum method. Then, preliminary feature extraction is performed through a feature extraction layer, which contains 3 convolutional layers with kernel sizes of 1×1, 3×3, and 1×1 respectively, and the number of channels is 32, 64, and 32 respectively. After each convolutional layer, a batch normalization layer and an activation function layer are connected. The calculation of depthwise separable convolution uses the following formula: In the formula, Y dwc is the output of depth convolution, Y pwc is the output of pointwise convolution, K c is the depth convolution kernel, X c is the input feature map, s is the stride, k is the convolution kernel size, W c is the pointwise convolution weight, and M is the number of input channels. The calculation of channel attention uses the following formula: In the formula, F squeeze is the output of global average pooling, F excitation is the channel weight, X is the input feature map, H is the height of the feature map, W is the width of the feature map, and W1, W2 are the weights of the fully connected layers.

[0216] The specific implementation of step S14 is to establish a sliding time window based on the stable component data. The length of the time window is 24 hours, and the sliding step is 15 minutes. Statistical analysis is performed on the data within the window to calculate statistical features such as the maximum value, minimum value, average value, and standard deviation. At the same time, the peak-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-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. Differencing calculation is performed on the stable component, and the difference between two adjacent sampling values is calculated and divided by the sampling time interval to obtain the change rate. The sampling time interval is 1 minute. At the same time, the mean and standard deviation of the change rate are calculated, and a statistical model of the change rate is established. When the change rate exceeds the mean plus or minus 3 times the standard deviation, it is determined as an abnormal change.

[0218] The specific implementation of step S16 is to calculate according to the electricity meter status evaluation equations. Among them, the load characteristic equation is: The stability evaluation equation is: The variability evaluation equation is: The abnormal state equation is:

[0219] The specific implementation of step S17 is as follows: according to the comprehensive status evaluation result of the electricity meter, when the evaluation result is greater than 0.8, it is determined as an abnormal state and the original data is selected for reporting; when the evaluation result is less than or equal to 0.8, it is determined as a normal state and the feature data is selected for reporting. The data reporting uses 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 the topic subscription mode, establish an independent data topic for each electricity meter monitoring system, classify the topics according to geographical regions and electricity consumption types, the message delay time is less than 100 milliseconds, the message reliability level is at least once, the message persistence adopts the asynchronous disk flushing strategy, the message storage time is 30 days, support message filtering according to timestamps and data types, the system adopts a distributed architecture, supports horizontal expansion, 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, the input dimension is 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, and the output calculation formula of the feature encoding layer is: h i = ReLU(W i x i-1 + b i ), where h i is the output of the i-th layer, W i is the weight matrix, x i-1 is the output of the previous layer, b i is the bias term, and ReLU is the rectified linear unit activation function. The power quality analysis layer adopts a convolutional neural network structure, including 3 convolutional blocks, each convolutional block contains 2 convolutional layers and 1 max pooling layer, the convolutional kernel size is 3 by 3, the stride is 1, the padding method is same padding, the pooling kernel size is 2 by 2, and the stride is 2. The load analysis layer adopts a long short-term memory network structure, including 2 long short-term memory layers, the hidden state dimension of each layer is 128, and 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 output gate calculation formula is: o t= W(W o · [h t-1 , x t + b o ), the unit 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 is to input data into the deep neural network for forward propagation calculation, reduce the dimension and extract features of the input data through the feature encoding layer, and then input the extracted features into the power quality analysis layer, load analysis layer, and arc analysis layer for analysis. The power quality assessment uses the following equation: Q = w1·THD v + w2·THD i + w3·(1 - PF) + w4·ΔV + ε q , where THD v is the total harmonic distortion rate of voltage, THD i is the total harmonic distortion rate of current, PF is the power factor, ΔV is the voltage deviation, w1, w2, w3, w4 are weight coefficients, and ε q is the error term. The load characteristic assessment uses the following equation: where P avg is the average load, P max is the maximum load, LF is the load factor, is the load change rate, α1, α2, α3, α4 are weight coefficients, and ε l is the error term. The arc characteristic assessment uses the following equation: A = β1·E arc + β2·T arc + β3·f arc + ε a , where E arc is the arc energy, T arc is the arc duration, f arc is the arc occurrence frequency, β1, β2, β3 are weight coefficients, and ε a is the error term.

[0224] The specific implementation of step S24 is to store the comprehensive status evaluation results of the electricity meters in a time series database. The data is stored by time sharding, with the time slice size being 1 day and the data compression ratio being 1 to 10. A multi-level index structure including a time index, an electricity meter number index, and an anomaly type index is established. The construction formula for the anomaly record vector is: H abnormal =[n1, n2, …, n k , where n i is the occurrence times of the i-th type of anomaly, and k is the total number of anomaly types.

[0225] The specific implementation of step S25 is to generate an evaluation report based on each evaluation result. The evaluation report includes four parts: basic information, operating status, anomaly analysis, and early warning prompt. The basic information includes the electricity meter number, installation location, user type, etc. The operating status includes the values and levels of each evaluation index. The anomaly analysis includes the anomaly type, anomaly degree, anomaly duration, etc. The early warning prompt includes the early warning level, early warning reason, treatment suggestions, etc. The determination thresholds for the early warning levels are 0.6, 0.8, and 0.9 respectively.

[0226] The specific implementation of step S26 is to determine whether to trigger an alarm based on the operating status evaluation report, and a multi-level alarm strategy is adopted. An optional method is: for general early warnings, a system internal prompt is used; for important early warnings, an email notification is sent simultaneously; for urgent early warnings, a text message and a phone call notification are sent simultaneously. The alarm information includes the alarm time, alarm reason, alarm level, treatment suggestions, etc., and the alarm confirmation and alarm escalation functions are supported. When the alarm duration exceeds the preset time and is not processed, the alarm level is automatically increased; the multi-level alarm strategy can be set manually or by experience in the prior art.

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

[0228] The specific implementation of step S31 is to collect historical data reported by the electricity meter end monitoring system through a distributed data acquisition system. The data sources include raw data and feature data. The data is stored using a distributed file system, and the data is sharded according to time and geographical location, with each shard size not exceeding 1 gigabyte. The data quality control uses the 3-sigma method for outlier detection. The standard deviation calculation formula is: where σ is the standard deviation, N is the number of samples, x i is the sample value, and μ is the sample mean. Missing values are supplemented using the linear interpolation method. The interpolation formula is: where x is the point to be interpolated, x1 and x2 are the abscissas of adjacent known points, and y1 and y2 are the ordinates of adjacent known points. Noise filtering uses the median filtering method, and the filtering window size is 5.

[0229] The specific implementation of step S32 is to construct Lora model training samples based on the preprocessed historical data. The sliding window method is used to construct the training samples. The window length is 1 hour and the sliding step is 15 minutes. The input features include the time series data of voltage signals, current signals, temperature signals, humidity signals, and high-frequency current signals. The output labels include the power quality assessment results, load characteristic assessment results, and arc characteristic assessment results. The samples are normalized by the maximum and minimum values, and the normalization formula is: where x norm is the normalized value, x is the original value, x min is the minimum value, and x max is the maximum value. The data augmentation methods include the scale transformation of the time series, additive Gaussian noise, and random cropping, and the augmentation ratio is 3 times.

[0230] The specific implementation of step S33 is to train the Lora model using a distributed training framework. The training uses the data parallel method. An 8-GPU server training cluster is formed using 8 GPU servers, and each server is configured 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 base matrix, B is the low-rank base vector, and A is the adaptation matrix. The synchronous stochastic gradient descent algorithm is used for optimization, and the learning rate uses the cosine annealing scheduling strategy. The learning rate calculation formula is: where η t is the current learning rate, η min is the minimum learning rate, η max is the maximum learning rate, t is the current iteration number, and T is the total iteration number. The batch size is 256, and the number of training epochs is 100. The early stopping strategy is adopted, and the loss function uses a weighted combination of mean squared error and cross entropy. The loss function calculation formula is: L = λ1L mse + λ2L ce , where L mse is the mean squared error loss, L ce is the cross entropy loss, and λ1 and λ2 are the weight coefficients.

[0231] The specific implementation of step S34 is to calculate the preset threshold based on the historical anomaly record vector, and perform clustering analysis using the density-based spatial clustering algorithm. The minimum number of samples for clustering 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 is the local density of point i, d ij is the distance between point i and point j, and d cis the truncation distance, and X(x) is the indicator function. Calculate statistical features for each anomaly cluster, including mean, standard deviation, quantiles, etc., and determine the preset threshold based on these statistical features.

[0232] The specific implementation of step S35 is to analyze the temporal correlation of historical anomaly record vectors, using the autocorrelation analysis method. The formula for calculating the autocorrelation coefficient is: In the formula, r k is the autocorrelation coefficient at lag k, x t is the time series value, and μ is the sequence mean. Calculate the cross-correlation coefficients at different time scales. The formula for calculating the cross-correlation coefficient is: In the formula, C xy (k) is the cross-covariance function, C xx (0), C yy (0) are the autocovariance functions.

[0233] The specific implementation of step S36 is to use the remote parameter configuration system to send the trained model parameters to the electricity meter terminal monitoring system. Adopt a batch-by-batch method, with 100 electricity meter terminal monitoring systems sent in each batch. Adopt the incremental update method. Parameter compression uses the sparse matrix compression method, and the compression ratio can reach 1 to 20. Matrix compression uses the singular value decomposition method, and the decomposition formula is: M = U∑V T , where M is the original matrix, U and V are orthogonal matrices, and ∑ is a diagonal matrix. The transmission is encrypted using the national cryptographic algorithm, with a key length of 256 bits, and at the same time, parameter consistency verification is carried out.

[0234] To better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: The R & D team of a certain power enterprise collects data and alarms for 5000 smart electricity meters in the region. This region includes 3000 residential users, 1500 commercial users, and 500 industrial users. First, upgrade the hardware of the existing electricity meters, and install an improved data collection module inside each electricity meter. The main parameters of the collection module are shown in Table 1:

[0235] Table 1 Parameter Configuration Table of Data Collection Module

[0236] Parameter type Parameter value Accuracy requirement 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 main frequency 168MHz - Communication rate 1Mbps -

[0237] Deploy a lightweight neural network on the edge side. The network structure parameters are shown in Table 2:

[0238] Table 2 Configuration Parameters of Lightweight Neural Network

[0239]

[0240]

[0241] The parameter configuration of the state evaluation equation set is shown in Table 3 as follows:

[0242] Table 3 Parameter Settings of the State Evaluation Equation Set

[0243] Equation type Weight 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 evaluation equation 0.3,0.2,0.3,0.2 0.75 0.1 Abnormal state equation 0.3,0.3,0.2,0.2 0.8 0.1

[0244] Deploy 8 GPU servers at the remote end, with each configured with 8 GPUs and a video memory of 32 GB per card, to build a deep neural network for data analysis. The network training parameters are shown in Table 4 as follows:

[0245] Table 4 Deep Neural Network Training Parameters

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

[0247] The system runs for 90 days, and a total of 108 MB of data is collected. The statistics of various types of abnormal events identified are shown in Table 5 as follows:

[0248] Table 5 Abnormal Event Statistics Table

[0249] Abnormal type Occurrence times Number of accurate identifications Recognition rate Voltage anomaly 856 825 96.4％ Current anomaly 923 894 96.9％ Power factor anomaly 467 442 94.6％ Load anomaly 1258 1196 95.1％ Arc fault 89 86 96.6％ Temperature anomaly 245 236 96.3％ Communication anomaly 156 148 94.9％

[0250] Figure 5 It shows the occurrence trends of various types of abnormal events during the 90-day test period. It can be seen that the overall abnormal events show a downward trend, indicating that the control effect of the system is gradually emerging. The key performance indicators of the system reach: the end-side feature extraction time is less than 100 milliseconds, the data compression ratio reaches 1:10, the communication delay is less than 200 milliseconds, the remote analysis response time is less than 1 second, and the average accuracy rate of abnormal identification reaches 95.8%. Figure 6 It shows the distribution of the load feature evaluation results during the actual operation process, presenting the characteristics of an approximate normal distribution, concentrated around 0.75, indicating that the evaluation results of the system have good stability. Compared with the traditional remote monitoring method of electric energy meters, the present invention has the following significant advantages: First, the traditional method uses a fixed threshold for judgment, and the recognition accuracy rate is only about 85%, while the present invention uses a lightweight neural network and a multi-dimensional evaluation equation set to increase the accuracy rate to more than 95%. Second, the traditional method needs to transmit all the original data, resulting in a large communication burden, while the present invention reduces the data transmission volume by 90% through end-side feature extraction and an adaptive reporting strategy. Third, the response time of the traditional method is generally more than 3 seconds, while the present invention controls the response time within 1 second through a distributed architecture and an optimized algorithm. Fourth, the traditional method is difficult to identify complex abnormal patterns, while the present invention can accurately identify various abnormal situations through deep learning and multi-dimensional analysis. The test results show that the present invention effectively solves the technical problems of remote security control of electric energy meters, and significantly improves the intelligent level and control efficiency of the system.

[0251] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Table 6 as follows.

[0252] Table 6 Variable Explanation Table

[0253]

[0254] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should be covered within the protection scope of the present invention.

Claims

1. An intelligent electric energy meter data acquisition and alarm device, characterized in that, It includes several meter-end monitoring systems and a remote-end monitoring system. The meter-end monitoring system includes a meter-end power 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 disposed inside the electric 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 alarm unit. A data processing and reporting module is provided inside the meter-end control chip. The data processing and reporting module uses a lightweight neural network to extract features from the stable components and variable components of the voltage signal and current signal. The lightweight neural network performs feature extraction using the Lora model parameters sent from the remote end. The remote-end analysis module constructs a deep neural network structure. 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, and is used to analyze the original data and feature data to generate power quality evaluation results, load feature evaluation results, and arc feature evaluation results.

2. The intelligent electric 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 convolution 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 convolution layer, and the output is the fused feature vector.

3. The intelligent electric energy meter data acquisition and alarm device according to claim 1, wherein The Lora model decomposes the weight matrix of the lightweight neural network by 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 by training at the remote end, and the adaptation matrix is used to adjust the feature extraction process.

4. The intelligent electric energy meter data acquisition and alarm device according to claim 1, characterized in that The stable components include the steady-state voltage value, the steady-state current value, and the 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 value of the phase difference between the steady-state voltage value and the steady-state current value.

5. The intelligent electric energy meter data acquisition and alarm device according to claim 1, characterized in that, The data processing and reporting module is also used to calculate historical load data, which is the data of the stable components in the previous 24 hours, and calculate the load change rate, which is the ratio of the change amount between two adjacent sampling values of the stable components to the sampling time interval.

6. The intelligent electric energy meter data acquisition and alarm device according to claim 1, characterized in that The meter-end control chip evaluates the operating state of the electric meter based on an electric meter state evaluation equation set, which includes a load feature equation, a stability evaluation equation, a variability evaluation equation, and an abnormal state equation.

7. The intelligent electric energy meter data acquisition and alarm device according to claim 6, wherein, The load feature equation is used to evaluate the operating state of the electric meter load. The inputs include the steady-state voltage value, the steady-state current value, and the historical load data, and the output is a load state evaluation coefficient. The stability evaluation equation is used to evaluate the stability degree of the power quality. The inputs include the voltage fluctuation amount, the current fluctuation amount, the temperature signal, and the preset threshold, and the output is a stability evaluation coefficient.

8. The intelligent electric energy meter data acquisition and alarm device according to claim 1, wherein, 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 layers is 512, 256, and 128.

9. The intelligent electric energy meter data acquisition and alarm device according to claim 1, characterized in that When the remote - end model training module trains the Lora model, it uses the data - parallel method, optimizes it using the synchronous stochastic gradient descent algorithm. The initial value of the learning rate is 0.001, adopts the cosine annealing scheduling strategy, the minimum learning rate is 0.00001, the batch size is 256, and the number of training epochs is 100.

10. The intelligent electric energy meter data acquisition and alarm device according to claim 1, characterized in that The remote - end model training module calculates the autocorrelation coefficients at different time scales from 1 hour to 24 hours with a step size of 1 hour using the autocorrelation analysis method according to the time correlation of the historical anomaly record vectors, and selects the time scale with the largest autocorrelation coefficient as the feature time scale.

Images