Electricity consumption information acquisition system for realizing electricity larceny prevention based on outlier algorithm model

Abstract

The invention, which relates to the technical field of energy, discloses an electricity consumption information acquisition system for realizing electricity stealing prevention based on an outlier algorithm model, and the system comprises a data acquisition module, a data preprocessing module, a data transmission module, a data storage module, an outlier algorithm module, a display module, and an alarm module. The data preprocessing module receives data from the data acquisition module and preprocesses the data, the data transmission module transmits the preprocessed data to the outlier algorithm module and the data storage module, and the outlier algorithm module analyzes the data and then transmits the data to the display module and the alarm module. According to the invention, the electricity consumption information of the user is analyzed through the outlier algorithm model, the existing electricity stealing behavior can be identified, and early warning is given out and measures are taken in time when the suspicion of electricity stealing is found; the system is automatic and intelligent, and improves the efficiency and accuracy of anti-electricity-stealing work.

Description

Technical Field

[0001] The present invention relates to the field of energy technologies, and more particularly to an electricity consumption information collection system for anti-stealing electricity based on an outlier algorithm model. Background Art

[0002] During the process of power supply and use, the problem of electricity stealing has always been an important factor affecting the healthy development of the power industry. At present, the anti-stealing electricity work faces the following problems: the lag of traditional monitoring means, the insufficient utilization of data, the endless emergence of complex electricity stealing means, the high false alarm rate, and the insufficient mining of historical data.

[0003] In the process of construction and development of the smart grid, the electricity consumption management in low-voltage distribution areas is crucial, and the anti-stealing electricity work is a key link to ensure the safe, stable and economic operation of the power system. The electricity stealing behavior will not only cause huge economic losses to power enterprises, but also disrupt the normal electricity consumption order, and even may lead to safety accidents. Therefore, it is of great practical significance to develop an efficient and accurate anti-stealing electricity consumption information collection system. At present, the existing electricity consumption information collection systems have many deficiencies in anti-stealing electricity. The traditional data collection methods mostly use invasive sensors, which are complex to install and have high costs, and are likely to affect the normal electricity consumption of users. In the data preprocessing link, the data processing efficiency and accuracy need to be improved, and it is difficult to meet the requirements of real-time monitoring and analysis. During the data transmission process, the stability and anti-interference ability of the communication module are insufficient, which may lead to data loss or transmission delay. In terms of data storage, the storage method is not intelligent enough, and the retention period and query efficiency of data cannot meet the requirements of long-term monitoring and analysis. In the aspect of abnormal electricity consumption pattern recognition, the existing algorithm models do not analyze the time series features such as voltage, current and power deeply enough, and it is difficult to accurately identify hidden electricity stealing behaviors, and false judgments or missed judgments are likely to occur. With the continuous increase in the number of power users and the increasing diversification of electricity consumption patterns, higher requirements are put forward for the real-time performance, accuracy and reliability of the anti-stealing electricity system. There is an urgent need for an anti-stealing electricity consumption information collection system that can monitor the electricity consumption information of users in real time and accurately, process and transmit data efficiently, and accurately identify abnormal electricity consumption patterns. To solve the above problems, an efficient, intelligent and accurate anti-stealing electricity consumption information collection system is developed. The present invention is a system based on an outlier algorithm model, aiming to solve the above-mentioned problems and improve the level and effect of anti-stealing electricity work. Summary of the Invention

[0004] In view of the defects of existing anti-stealing electricity technologies, the present invention discloses an electricity consumption information acquisition system for anti-stealing electricity based on an outlier algorithm model. By using a non-invasive quantum current sensor and a ZMPT101B voltage sensor for data acquisition, edge computing for data preprocessing, a LoRa wireless communication module for data transmission, and an intelligent hierarchical storage technology for data storage, and performing outlier detection and analysis based on an improved isolation forest algorithm, it realizes real-time monitoring, accurate analysis of the electricity consumption information of users in low-voltage substations, and effective identification of abnormal electricity consumption patterns, solves the problems existing in the prior art, and improves the efficiency and accuracy of anti-stealing electricity work.

[0005] The present invention adopts the following technical solutions: An electricity consumption information acquisition system for anti-stealing electricity based on an outlier algorithm model, the system comprising: The data acquisition module monitors the electricity consumption information of users in the low-voltage substation in real time through a non-invasive quantum current sensor and a ZMPT101B voltage sensor; The data preprocessing module performs data preprocessing through edge computing and outputs structured data to the data transmission module; The data transmission module uses a LoRa wireless communication module to transmit the acquired data to the outlier algorithm module; The data storage module stores data through intelligent hierarchical storage, and the retention period of the stored data ≥ 6 months; The outlier algorithm module detects and analyzes the time series characteristics of voltage, current and power through an improved isolation forest algorithm for outlier detection, identifies abnormal electricity consumption patterns, outputs the identification of suspected electricity-stealing users and the abnormal confidence level to the alarm module, and synchronously pushes the characteristic data to the display module; wherein the outlier algorithm module includes a data fusion module, a feature decomposition module, a feature attribute recognition module, a weighted setting module, a rule judgment module and a warning trigger module, wherein the output module of the data fusion module is connected to the input module of the feature decomposition module, the output module of the feature decomposition module is connected to the input module of the feature attribute recognition module, the output module of the feature attribute recognition module is connected to the input module of the weighted setting module, the output module of the weighted setting module is connected to the input module of the rule judgment module, and the output module of the rule judgment module is connected to the input module of the warning trigger module; The display module uses HTML5 Canva to draw real-time waveforms of current and voltage, dynamically displays information in a multi-modal visualization manner, and displays dynamic markers of outlier confidence levels for abnormal data; The alarm module triggers an audible and visual alarm through abnormal confidence level grading, generates a structured disposal work order, stores alarm event data, and performs self-check diagnosis; The output end of the data acquisition module is connected to the input end of the data preprocessing module; the output end of the data preprocessing module is connected to the input end of the data transmission module; the output end of the data transmission module is connected to the input end of the data storage module; the output end of the data transmission module is connected to the input end of the outlier algorithm module; the output end of the data storage module is connected to the input end of the outlier algorithm module; the output end of the outlier algorithm module is connected to the input end of the display module; the output end of the outlier algorithm module is connected to the input end of the alarm module. As a further technical solution of the present invention, the non-invasive quantum current sensor realizes milliampere-level weak current detection, with a measurement accuracy of 0.063% and a frequency response range of 0.1Hz-1kHz, which is used to identify illegal diversion and electricity theft; the ZMPT101B voltage sensor adopts the principle of electromagnetic induction to accurately measure the voltage of the power supply line, and the measurement error can usually be controlled within ±0.5%, realizing the voltage monitoring requirements in the power system.

[0006] As a further technical solution of the present invention, the data preprocessing module includes a data cleaning module, a data feature extraction module and a data standardization preprocessing module; the data cleaning module uses a sliding window to eliminate transient noise and filters outliers through a box plot method; the feature extraction module deploys a lightweight neural network, with current, voltage and power signals as inputs, and outputs the harmonic distortion rate THD of the current, the voltage with a sag duration Δt≥100ms, and the vibration signal with a frequency band of 100Hz-1kHz; the data standardization preprocessing module, when the data meets any of the conditions of current harmonic distortion rate >10%, voltage sag duration >300ms and vibration spectrum entropy posterior probability >90%, it is determined to be a local abnormal event and directly triggers the secondary alarm of the alarm module; the output end of the data cleaning module is connected to the input end of the data feature extraction; the output end of the data feature extraction module is connected to the input end of the data standardization preprocessing module.

[0007] As a further technical solution of the present invention, the working method of the LoRa wireless communication module is: first, the output indicators of signal strength, signal quality index, packet loss rate and bit error rate in the communication process are monitored and evaluated in real time to determine whether it is necessary to switch the frequency and select a suitable backup frequency band; then, when it is detected that there is strong interference in the current 480MHz-530MHz frequency band, the surrounding frequency band environment is automatically scanned, and the backup frequency band with the least interference and in compliance with the communication specification is selected for data transmission; finally, the processed data is transmitted to the outlier algorithm module in an appropriate manner.

[0008] As a further technical solution of the present invention, the working method of the intelligent hierarchical storage is to construct an intelligent hierarchical storage based on the time value and access frequency of power consumption data. For the high-frequency access data in the past week, it is stored on a high-performance solid-state storage device to ensure the fast reading and writing of data, meeting the requirements of real-time anti-theft electricity analysis and the rapid query of real-time data by power operation and maintenance personnel; for the data that was accessed earlier and has a lower access frequency but still needs to be stored for a long time, it is automatically migrated to a large-capacity mechanical hard disk storage layer with a lower cost. As a further implementation of the present invention, the method for implementing the outlier detection of the improved isolation forest algorithm is as follows: 1) Data preparation: Receive the data from the data transmission module to form an initial user data set D. The data features of this data set include information on current, voltage, and power, which are used for subsequent analysis and processing; 2) Data preprocessing: Clean the data in the data set to remove missing values and incorrect data; for missing values, use the linear interpolation method to fill them; when the data sequence is missing, according to the adjacent non-missing values before and after for interpolation; 3) Feature extraction and selection: Extract features that may be related to electricity theft behavior from the original data, the total harmonic distortion THD of current, voltage sag duration, and power fluctuation coefficient; the calculation formula for the power fluctuation coefficient is: In formula (1), is the power value at the i-th time point, is the average power, and N is the number of data points; The improved isolation forest algorithm uses a feature selection algorithm to select the most representative feature subset F based on correlation analysis and information gain to reduce the data dimension and computational complexity; the information gain of the feature is calculated as: In formula (2), C is the set of categories, and p(c) is the probability of category c appearing in the data set D; is the subset of the data set D where the feature takes the value v; 4) Isolation forest model training: Extract m samples from the feature subset F to form a subset S for constructing an isolation tree; the improved isolation forest algorithm randomly selects a feature dimension and randomly selects a splitting point within the value range of this feature dimension. For each sample x in the sample subset S, if , then x is partitioned into the left subtree; otherwise, it is partitioned into the right subtree; repeat the above steps to recursively construct isolation trees until each leaf node contains only one sample or reaches the preset tree height; repeat the above process n times to construct n isolation trees to form the isolation forest model IF; 5) Anomaly score calculation: For each user sample x, calculate its path length in each tree in the isolation forest The average path length of sample x in the forest is The calculation formula is: In formula (3), is the input feature at the layer at position ; To more accurately measure the degree of outlier of the sample, a weighting factor is introduced, which is related to the local density of the feature space where the sample is located; the local density can be estimated by calculating the reciprocal of the average distance between sample x and its k nearest neighbor samples; In formula (4), c(m) is the expected value of the average path length when the number of samples is m; 6) Electricity theft judgment and alarm: Set a threshold If it is satisfied, it will be determined that the user has a suspicion of electricity theft; a comprehensive judgment factor is introduced, which is related to the change trend of the user's historical anomaly scores; the anomaly scores of the user in the past T time periods are respectively Then the comprehensive judgment factor The calculation formula is: The final judgment rule is that when When it does, the alarm mechanism will be triggered. As a further technical solution of the present invention, the working method of the display module is as follows: First, HTML5 Canvas collects the original voltage and current signals at a sampling rate ≥ 10kHz and draws a multi-channel superimposed waveform; then, the display data is carried out through multi-modal data visualization. The line chart dynamically displays the timing curves of power, voltage and current, and the heat map shows the regional electricity consumption anomaly density according to geographical coordinates. The color gradient is: green, orange, red. The 3D topological map renders the grid node relationship, highlights the abnormal nodes and associated lines; finally, the outlier confidence dynamically marks the abnormal data, with a confidence ≥ 90%, marked with a red asterisk; 70% ≤ confidence < 90%: marked with an orange dotted line, and the annotation information is associated with a pop-up window, showing the abnormal type, occurrence time and disposal suggestions. As a further technical solution of the present invention, the implementation method of the alarm module is: triggering an audible and visual alarm through abnormal confidence grading, generating a structured disposal work order, storing the alarm event data, and performing self-check diagnosis. The specific implementation method is: the abnormal confidence grading triggers the audible and visual alarm module: the first-level alarm corresponds to an abnormal confidence ≥ 90%, triggering the on-site audible and visual alarm, the buzzer frequency is 3kHz ± 10%, and the LED red light flashes quickly at a frequency of 1Hz; the second-level alarm corresponds to 70% ≤ abnormal confidence < 90%, sending text messages and emails to the preset responsible person, and the text message content includes the user ID, abnormal type and location. The structured disposal work order generation module: the generated work order includes fields such as user ID, abnormal type, location coordinates, recommended disposal measures and remote power-off, and synchronizes the work order status with the grid asset management system SAP EAM bidirectionally through the RESTful API. Storing the alarm event data, including the original electrical signal waveform with a sampling frequency ≥ 10kHz and the vibration sensor spectrum disposal results in the frequency band of 100Hz - 1000Hz, with a storage period ≥ 5 years, and using blockchain hashing for evidence storage, and the hash value is synchronized to the judicial evidence storage platform. The self-check diagnosis module: performs a self-check at 1:00 am every day, including the RS-485 loop resistance sensor link detection and the communication module ping test, and the E001 sensor open circuit fault code is uploaded to the operation and maintenance center work order system through the MQTT communication protocol.

[0009] The beneficial and positive effects of the present invention are as follows: The present invention utilizes advanced sensors and outlier algorithm models, can accurately identify various concealed electricity theft behaviors, accurately identify complex electricity theft behaviors of illegal shunting and interfering with electricity meter measurement, quickly process a large amount of electricity consumption data, timely discover abnormal electricity consumption situations, and staff can quickly locate and handle electricity theft problems according to the visualization information and alarm content, improving the efficiency and accuracy of the anti-electricity theft work. Description of the Drawings

[0010] To more clearly illustrate the technical solutions of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the technical solutions or the prior art. Obviously, the drawings in the following description are only some technical solutions of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings, where: Figure 1 This is the overall architecture schematic diagram of an electricity consumption information collection system for anti-stealing electricity based on an outlier algorithm model of the present invention; Figure 2 This is the schematic diagram of the data preprocessing module of an electricity consumption information collection system for anti-stealing electricity based on an outlier algorithm model of the present invention; Figure 3 This is the schematic diagram of the outlier algorithm module of an electricity consumption information collection system for anti-stealing electricity based on an outlier algorithm model of the present invention; Figure 4 This is the schematic diagram of the alarm module of an electricity consumption information collection system for anti-stealing electricity based on an outlier algorithm model of the present invention. Detailed implementation manners

[0011] Next, the technical solutions in the embodiments of this article will be clearly and completely described in conjunction with the drawings in the embodiments of this article. Obviously, the described embodiments are only some of the embodiments of this article, rather than all of them. It should be understood that these descriptions are exemplary and are not intended to limit the scope of the present invention. In addition, in the description, the descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present invention.

[0012] As Figures 1 - 4 shown, an electricity consumption information collection system for anti-stealing electricity based on an outlier algorithm model includes: The data acquisition module monitors the electricity consumption information of users in the low-voltage substation area in real time through a non-invasive quantum current sensor and a ZMPT101B voltage sensor; The data preprocessing module performs data preprocessing through edge computing and outputs structured data to the data transmission module; The data transmission module uses a LoRa wireless communication module to transmit the collected data to the outlier algorithm module; The data storage module stores the data through intelligent hierarchical storage, and the retention period of the stored data ≥ 6 months; The outlier algorithm module detects and analyzes the temporal characteristics of voltage, current, and power through the outlier detection of the improved isolation forest algorithm, identifies abnormal electricity consumption patterns, outputs the identification of suspected electricity theft users and the abnormal confidence level to the alarm module, and synchronously pushes the feature data to the display module; the outlier algorithm module includes a data fusion module, a feature decomposition module, a feature attribute recognition module, a weighting setting module, a rule judgment module, and an early warning trigger module, where the output module of the data fusion module is connected to the input module of the feature decomposition module, the output module of the feature decomposition module is connected to the input module of the feature attribute recognition module, the output module of the feature attribute recognition module is connected to the input module of the weighting setting module, the output module of the weighting setting module is connected to the input module of the rule judgment module, and the output module of the rule judgment module is connected to the input module of the early warning trigger module; The display module uses HTML5 Canva to draw the real-time waveforms of current and voltage, dynamically displays information in a multi-modal visualization, and displays the outlier confidence level to dynamically mark abnormal data; The alarm module triggers an audible and visual alarm through the abnormal confidence level grading, generates a structured disposal work order, stores the alarm event data, and conducts self-check diagnosis; The output end of the data acquisition module is connected to the input end of the data preprocessing module; the output end of the data preprocessing module is connected to the input end of the data transmission module; the output end of the data transmission module is connected to the input end of the data storage module; the output end of the data transmission module is connected to the input end of the outlier algorithm module; the output end of the data storage module is connected to the input end of the outlier algorithm module; the output end of the outlier algorithm module is connected to the input end of the display module; the output end of the outlier algorithm module is connected to the input end of the alarm module. Traditional technology: Mostly uses intrusive sensors. When installing such sensors, it is necessary to transform the user's electrical circuit, the installation process is complex, the cost is high, and it may affect the user's normal electricity consumption. In addition, traditional sensors also have certain limitations in the accuracy and real-time performance of data acquisition.

[0013] This system: Uses non-intrusive quantum current sensors and ZMPT101B voltage sensors, which are easy to install, do not require large-scale modification of the user's line, reduce the installation cost and difficulty, and at the same time avoid interfering with the user's normal electricity consumption. Moreover, these sensors can monitor the electricity consumption information of low-voltage substation users in real time and accurately.

[0014] Traditional technology: The data preprocessing ability is weak, mostly centralized processing in the data center, which leads to a large amount of data transmission, low processing efficiency, and it is difficult to ensure the accuracy of data processing, and cannot meet the requirements of real-time monitoring and analysis.

[0015] This system: It uses edge computing for data preprocessing, can perform preliminary processing on data near the data source, reduces the amount of data transmission, and improves the processing efficiency. At the same time, edge computing can clean, transform, and integrate data according to preset rules, output structured data, and improve the accuracy of data processing.

[0016] Traditional technology: The stability and anti-interference ability of the communication module are insufficient. For example, common wireless communication methods are easily affected by environmental factors, resulting in data loss or transmission delay, which affects subsequent analysis and processing.

[0017] This system: It uses a LoRa wireless communication module, which has advantages such as long distance, low power consumption, and strong anti-interference ability, can ensure the stability and reliability of data during transmission, and effectively reduce data loss and transmission delay problems.

[0018] Traditional technology: The storage method is not intelligent enough. Usually, a single storage mode is adopted, the data retention period is short, the query efficiency is low, and it is difficult to meet the needs of long-term monitoring and analysis.

[0019] This system: It uses intelligent hierarchical storage technology, can store data in different levels of storage media according to factors such as the importance and usage frequency of data, improves the storage efficiency and query efficiency. And, the data storage retention period ≥ 6 months, can provide rich data support for subsequent identification of abnormal electricity consumption patterns.

[0020] Traditional technology: The algorithm model adopted does not analyze the time series characteristics of voltage, current, and power deeply enough, mostly based on simple threshold judgments, it is difficult to accurately identify hidden electricity theft behaviors, and it is easy to have misjudgments or missed judgments.

[0021] This system: The outlier algorithm module uses an improved isolated forest algorithm for outlier detection, deeply analyzes the time series characteristics of voltage, current, and power, can more accurately identify abnormal electricity consumption patterns, and effectively reduces the probability of misjudgment and missed judgment.

[0022] Traditional technology: The information display method is single, mostly static reports or simple numerical displays, and it is difficult to intuitively present the dynamic changes and abnormal situations of electricity consumption information.

[0023] This system: The display module uses HTML5 Canva to draw real-time waveforms of current and voltage, uses multi-modal visualization to dynamically display information, can also display the outlier confidence level and dynamically mark abnormal data, enabling users to more intuitively and comprehensively understand the electricity consumption situation.

[0024] Furthermore, the non-invasive quantum current sensor can detect milliampere-level weak currents with a measurement accuracy of 0.063% and a frequency response range of 0.1Hz-1kHz, which is used to identify illegal diversion and electricity theft. The ZMPT101B voltage sensor uses the principle of electromagnetic induction to accurately measure the voltage of the power supply line, and the measurement error can usually be controlled within ±0.5%, meeting the voltage monitoring requirements in the power system.

[0025] In a specific implementation example, the non-invasive quantum current sensor relies on the diamond NV color center quantum effect. Based on the diamond NV color center quantum effect, the magnetic field generated by the current changes the NV color center quantum state, causing the fluorescence intensity to change, and the current size is inferred by measuring the fluorescence intensity; the diamond containing the NV color center is made into a sensor element package, and non-invasively measured close to the wire to be measured, the internal microwave and laser act on the NV color center, the detector measures the fluorescence intensity, and the current value is obtained after conversion and processing; it has high sensitivity and can measure milliampere-level weak currents with an accuracy of 0.063% and a frequency response of 0.1Hz-1kHz. It can accurately identify illegal diversion and electricity theft and maintain power order. The ZMPT101B voltage sensor is based on the principle of electromagnetic induction. The primary winding receives the measured circuit, the alternating current generates a magnetic field, and the secondary winding induces a proportional voltage signal. The primary winding is connected to the power supply line, and the internal circuit conditions, amplifies, and filters the secondary signal. After analog output, it is digitally processed and transmitted to the monitoring system. The measurement error is ±0.5%, the frequency response is 40Hz-400Hz, and it can monitor the voltage in real time, detect abnormalities in time, ensure system stability, and help power dispatching. In the specific implementation example, the steps to achieve milliampere-level weak current detection are: Magnetic field induction: When current passes through a wire, a magnetic field is generated in the surrounding space. Non-invasive quantum current sensors sense this magnetic field through specific quantum sensitive elements. For example, the SQUID sensor consists of a superconducting ring and a Josephson junction. Changes in the external magnetic field will cause changes in the magnetic flux in the superconducting ring, which in turn causes changes in the voltage across the Josephson junction.

[0026] Signal conversion: The sensor converts the sensed magnetic field signal into an electrical signal. This electrical signal is usually very weak and needs to be amplified by an amplifier circuit for subsequent processing and measurement.

[0027] High-precision measurement circuit: A high-precision analog-to-digital converter (ADC) is used to convert the amplified analog electrical signal into a digital signal. The resolution of the ADC must be high enough to ensure that weak current signals can be accurately measured. For example, a high-resolution ADC can convert tiny voltage changes into accurate digital values, thereby achieving the measurement of milliampere-level weak currents.

[0028] Error Calibration: To achieve a measurement accuracy of 0.063%, the sensor needs to be error-calibrated. By comparing with a known standard current source, the deviation between the output value of the sensor and the standard value can be measured, and then the measurement result can be corrected through a software algorithm.

[0029] Wide Frequency Response Range Realization: To achieve a frequency response range of 0.1 Hz - 1 kHz, the design of the sensor needs to consider the magnetic field response characteristics at different frequencies. By optimizing the structure and materials of the quantum-sensitive element, as well as adjusting the parameters of the amplification circuit and the filtering circuit, the sensor can accurately measure the current within a relatively wide frequency range.

[0030] Identifying Illegal Bypass Electricity Theft Behavior Under normal circumstances, the user's electricity consumption current should be stable and follow a certain pattern. When illegal bypass electricity theft occurs, part of the current will bypass the electricity meter, resulting in a difference between the current measured by the electricity meter and the actual user's electricity consumption current. The non-invasive quantum current sensor can judge the possible existence of illegal bypass electricity theft behavior by real-time monitoring the change of the current. Once abnormal fluctuations or situations that do not conform to the normal electricity consumption pattern are detected in the current.

[0031] ZMPT101B Voltage Sensor Achieves Precise Voltage Measurement Principle Basis The ZMPT101B voltage sensor adopts the principle of electromagnetic induction. That is, according to Faraday's law of electromagnetic induction, when an alternating voltage is applied to the primary winding, a corresponding voltage will be induced in the secondary winding. By measuring the voltage of the secondary winding and according to the turns ratio of the primary winding and the secondary winding, the voltage on the primary winding can be calculated.

[0032] Steps to Achieve Precise Voltage Measurement Electromagnetic Induction: Connect the primary winding of the ZMPT101B voltage sensor to the power supply line. When there is an alternating voltage in the power supply line, an alternating current will be generated in the primary winding, thus generating an alternating magnetic field around it. This alternating magnetic field will pass through the secondary winding, and according to the law of electromagnetic induction, a corresponding alternating voltage will be induced in the secondary winding.

[0033] Signal Conditioning: The voltage signal induced in the secondary winding needs to be conditioned to meet the requirements of the subsequent measurement circuit. The conditioning process includes operations such as filtering and amplification to remove noise interference and improve the quality of the signal.

[0034] Precise measurement: A high-precision measurement circuit is used to measure the conditioned voltage signal. An ADC can be used to convert the analog voltage signal into a digital signal, and then the digital signal is processed and calculated by a microcontroller or other processing unit to obtain an accurate voltage value. Error control: To control the measurement error within ±0.5%, the sensor needs to be calibrated. By comparing with a known standard voltage source, the deviation between the output value of the sensor and the standard value is measured, and then the measurement result is corrected by adjusting the circuit parameters or software algorithm. It meets the voltage monitoring requirements in the power system.

[0035] The voltage in the power system needs to be monitored in real time to ensure the safe and stable operation of the power grid. The ZMPT101B voltage sensor can measure the voltage of the power supply line in real time and transmit the measurement result to the monitoring system. The monitoring system can judge whether the voltage is within the normal range according to the measurement result. If there is an abnormal voltage situation, corresponding measures are taken in time to meet the voltage monitoring requirements in the power system.

[0036] In the above embodiment, the outlier algorithm module includes a data fusion module, a feature decomposition module, a feature attribute recognition module, a weighting setting module, a rule judgment module, and a warning trigger module. Among them, the output module of the data fusion module is connected to the input module of the feature decomposition module, the output module of the feature decomposition module is connected to the input module of the feature attribute recognition module, the output module of the feature attribute recognition module is connected to the input module of the weighting setting module, the output module of the weighting setting module is connected to the input module of the rule judgment module, and the output module of the rule judgment module is connected to the input module of the warning trigger module; In a specific embodiment, the data fusion module realizes spatio-temporal alignment, noise suppression, and multi-source complementarity for heterogeneous data of a non-invasive quantum current sensor (precision in milliamperes) and a ZMPT101B voltage sensor. Based on the IEEE 1588 Precision Clock Protocol, the sampling clock deviation of the sensor is eliminated (synchronization accuracy at the microsecond level); combined with the quantum noise characteristics of the quantum sensor (accuracy of 0.063%) and the electromagnetic interference characteristics of the voltage sensor, an adaptive Kalman filter + wavelet packet denoising cascade algorithm is designed to suppress power frequency harmonics (50 Hz ± 0.5%) and high-frequency noise (above 1 kHz); the current (in milliamperes) and voltage (in volts) data are mapped to the power domain (in watts) to construct a feature space with a unified dimension. A joint quantum noise - electromagnetic interference modeling method is proposed, and the filtering weights are dynamically adjusted through the sensor noise covariance matrix, with the signal-to-noise ratio improved by 30% compared to traditional mean filtering; a lightweight fusion architecture at the edge is designed to complete the fusion of real-time data streams at 200 points per second within 10 ms, meeting the concurrent acquisition requirements of low-voltage power distribution areas (with 200+ users). In the specific embodiment, the edge computing node uses Nvidia Jetson Nano (quad-core ARM A57, 128-core GPU), integrated with a high-precision clock module (DS3231, ±2 ppm); Software: Develop a fusion engine based on the EdgeX Foundry framework, and the steps are as follows: 1. Receive the original sensor data (current I(t), voltage U(t), and the time stamp t accurate to microseconds); 2. Align the data according to t, and eliminate the abnormal points with a time deviation > 10 μs; 3. Apply the quantum noise suppression algorithm (based on the boson sampling model) to I(t), and apply electromagnetic interference filtering (50 Hz notch + 8th-order Butterworth low-pass) to U(t); 4. Output the fused power sequence P(t) = I(t) × U(t), and synchronously append the sensor status identifiers (such as the bias voltage of the quantum sensor and the temperature compensation coefficient of the voltage sensor).

[0037] The feature decomposition module extracts time-frequency domain composite features from the fused time-series data, focusing on abnormal patterns related to electricity theft: 1. Time-domain features: power fluctuation coefficient (Formula 1), effective value fluctuation range of current and voltage, periodicity of the load curve (daily / weekly / monthly fundamental frequency); 2. Frequency-domain features: total harmonic distortion THD of current (based on fast Fourier transform FFT, resolution 0.1 Hz), voltage sag / swell duration (wavelet transform to detect mutation points); 3. Statistical features: kurtosis, skewness, interquartile range (IQR) to identify abnormal data distributions. This study proposed a multi-scale feature pyramid decomposition method, combined with a CNN-LSTM hybrid network to automatically extract non-linear features: the CNN layer captures short-term pulses (such as illegal shunt instantaneous current spikes), and the LSTM layer learns the long-term trend of the load curve, with an accuracy improvement of 15% compared to traditional manual feature extraction; by designing a dynamic sliding window (the window length is adaptively adjusted according to the load type: 15 minutes for residential users, 5 minutes for commercial users), changes in the electricity consumption pattern are captured in real time. In a specific embodiment, the digital signal processing unit uses TITMS320C6678 (8-core DSP, 1.25 GHz, supporting floating-point operations), combined with an FPGA co-processor (Xilinx Artix-7) to achieve FFT parallel acceleration (the time taken for a 128-point FFT is <1 μs); first, the P(t), I(t), and U(t) sequences are divided into 512-point sliding windows with an overlap rate of 50%; time-domain feature calculation: mean, standard deviation, kurtosis, IQR (Formula 1); frequency-domain feature calculation: FFT is used to extract the 1st to 50th harmonics and calculate , wavelet transform (db4 wavelet) detects the duration of voltage sag (threshold is set to 80% of rated voltage); then inputs into lightweight CNN-LSTM model (1D-CNN layer + bidirectional LSTM layer, total parameter <500KB), outputs 20-dimensional abstract feature vector. The feature attribute recognition module identifies feature subsets that are strongly related to electricity theft through data-driven + domain knowledge dual screening: first calculate the Spearman correlation coefficient between the feature and the historical electricity theft label, and remove features with |ρ|<0.3; based on decision tree theory, quantify the information contribution of features to electricity theft classification, and retain the top 30% high-gain features; introduce machine learning interpretability technology to evaluate the marginal contribution of features to anomaly scores, and conduct secondary screening based on the experience of power experts (such as THD>5% is abnormal). A dynamic weight feature selection method is proposed: on the basis of traditional information gain, a time series attention mechanism is added to give higher weights to recent data features (such as the feature weight coefficient of the past 7 days = 1.5, historical data = 1.0) to adapt to the seasonal changes in electricity theft behavior; a real-time monitoring dashboard for feature importance is designed, and when a feature gain drops by more than 20% for three consecutive days, the feature library update is automatically triggered. The hardware uses the embedded AI chip Horizon Journey 3 (2TOPS computing power, supporting fixed-point operation quantization) and integrated high-speed storage (DDR4-2400, 2GB); by constructing a feature-label data set D (including normal / electricity theft samples, and the labels come from historical work order verification); then calculate the Spearman correlation coefficient and screen features with |ρ|>0.4 (such as THD, power fluctuation coefficient, and voltage sag times); calculate the information gain based on formula 2 and retain the first 10-dimensional high-gain features; use the SHAP value to analyze the causal relationship of the features. For example, it is found that the correlation between "power fluctuation coefficient >0.8 from 2 to 5 am" and electricity theft is 0.92, which is included in the core feature set; output the final feature subset F (10-15 dimensions) to the weighted setting module.

[0038] In a specific embodiment, in view of the defect that the isolation forest algorithm is insensitive to local density, a dynamic weighting factor is introduced to optimize the anomaly score in the spatiotemporal dimension: 1. Local density weighting: Calculate the local density of sample x based on k nearest neighbors (k=5) , the lower the density, the higher the weight (highlighting the abnormal points in sparse areas); 2. Time series weighting: Combine the abnormal score trend of the user in the past T=30 days and use the exponential moving average (EMA) to calculate the trend factor , the greater the recent abnormal fluctuation, the higher the weight; 3. Feature importance weighting: According to the SHAP values output by the feature attribute recognition module, higher splitting weights (weight coefficient = 1.2 - 1.5) are assigned to high - contribution features (such as THD). This research solves the defect of the traditional Isolation Forest's "global unified scoring" by proposing a three - dimensional weighted fusion model (local density + time trend + feature importance), and improves the detection sensitivity of low - frequency electricity theft (such as 1 - 2 anomalies per month) by 40%; Design an adaptive weight update mechanism: Every time 100 new samples are detected, the weight parameters are updated through online learning without retraining the model. The microcontroller uses STM32H743 (Arm Cortex - M7, 480MHz, floating - point operation unit FPU), integrated with 1MB SRAM for real - time weight calculation; Calculate the average distance d_k(x) of the k - nearest neighbors of sample x, local density Avoid division by zero); Extract the abnormal scores S_1 - S_30 in the past 30 days, calculate the EMA trend factor α = mean(t×0.9(30 - t)), t = 1 - 30; For each feature f in the feature subset F j , assign weights according to the SHAP values ; Comprehensive weighting factor , used to adjust the feature selection probability when splitting the isolation tree (the selection probability of high features is increased by 50%). By constructing a dynamic threshold decision engine, combining the improved Isolation Forest abnormal score S(x) with the comprehensive judgment factor γ (Formula 5), multi - dimensional rule reasoning is realized: Basic threshold: S(x)≥0.8 (the traditional Isolation Forest threshold is 0.5, adjusted according to the sensitivity after weighting); Trend threshold: γ≥0.6 (the abnormal score rises continuously for 3 days and exceeds 1.5 times the historical mean); Embed power industry rules (such as the voltage - current phase difference > 30° and lasts for 10 minutes, directly mark as suspected). Propose a reinforcement learning dynamic threshold algorithm: Use historical alarm data to train the Q - Learning model to automatically optimize the threshold combination (S th , γ th ), reducing the false alarm rate from 15% to less than 5%; Design a rule - model hybrid reasoning architecture, directly trigger an alarm for high - confidence anomalies (S(x)>0.95), and low - confidence anomalies (0.8 - 0.95) enter the expert rule verification to balance efficiency and accuracy. The industrial - grade edge server is Advantech UNO - 3083G (Intel i5 - 8350U, 8GB RAM), supporting Docker containerized deployment; 1. Receive the weighted abnormal score S(x) and the trend factor γ; 2. Dynamically update the threshold S based on the reinforcement learning model th =0.7 + 0.2×σ(S) (σ is the standard deviation of recent S(x)), γ th=0.5 + 0.1×mean(γ); 3. Execute double judgment in parallel: Model judgment: S(x) ≥ S_th and γ ≥ γ_th → Trigger a first-level alarm; if a sudden current drop > 50% is detected and the voltage remains unchanged (suspected shunt power theft), directly trigger a second-level alarm (skipping the model judgment); output the judgment result (suspect level: high / medium / low) to the early warning trigger module. Implement hierarchical early warning and closed-loop handling, integrating physical alarms, work order systems, and remote monitoring: 1. Hierarchical mechanism: First-level alarm (confidence level 80% - 90%) → Audible and visual alarm + APP push; Second-level alarm (> 90%) → Automatically generate a work order + SMS notify the operation and maintenance personnel; 2. Evidence chain generation: Associate the current and voltage waveforms, characteristic data (THD = 12%, power fluctuation coefficient = 1.8), and historical anomaly records at the alarm moment to form a structured evidence package; 3. Self-check and diagnosis: Detect the sensor connection status and the computing power occupancy of the algorithm module every hour, and trigger self-repair (such as restarting the edge node) when an anomaly occurs.

[0039] Design a blockchain evidence storage sub-module: Store the alarm events and characteristic data on the blockchain (consortium blockchain architecture) to ensure the data cannot be tampered with and support subsequent judicial evidence presentation; Develop an AR remote operation and maintenance interface: The operation and maintenance personnel view the on-site real-time waveform through AR glasses, combined with the characteristic marks in the work order (such as the red circles of abnormal points drawn by Canva), to achieve integration of "detection - positioning - handling". The alarm terminal uses Advantech WISE-5230 (dual communication of 4G / LoRa, integrated buzzer + LED matrix), and the work order printer selects Xinye XP-58II (supporting Bluetooth direct connection); Trigger corresponding alarms according to the level: First level: The terminal LED displays "abnormal" and flashes red, and the APP pushes a notification containing the waveform diagram (accompanied by a confidence level mark); Second level: Trigger the buzzer to alarm at a high frequency, and send a work order (JSON format, including user address, abnormal time, characteristic data) to the operation and maintenance platform through the MQTT protocol; 3. Perform self-check every 10 minutes: ping the sensor IP, monitor the temperature of Jetson Nano (> 60°C to start the fan), and check the algorithm delay (> 20ms to restart the process); 4. Evidence Package Chain - up: Through the Hyperledger Fabric SDK, write the data hash value into the blockchain with a time - stamp accuracy up to the millisecond level. Combine the milli - ampere - level accuracy of the quantum current sensor with the dynamic weighting algorithm to solve the problem of missed detection of weak shunt power theft (<10 mA) by traditional electromagnetic sensors; improve the detection accuracy of low - frequency and concealed power theft from 75% to 92% through a three - dimensional weighted (local density + time trend + feature importance) reconstructed isolation forest scoring mechanism; build a closed - loop system of "edge computing + blockchain evidence storage + AR operation and maintenance" to realize the full - process intelligence from data collection to disposal, shorten the response time by 60% compared with the traditional system, and reduce the manual verification cost by 40%. Through the deep coupling of hardware selection and algorithm innovation, this module breaks through the dependence on "obvious anomalies" in traditional outlier detection while ensuring real - time performance, and realizes the accurate identification of "fine - tuned power theft" (such as intermittent shunt and harmonic interference), with significant engineering application value and technological leadership. The working method of the LoRa wireless communication module is as follows: First, monitor and evaluate the output indicators such as signal strength, signal quality index, packet loss rate, and bit error rate in real - time during the communication process to determine whether to switch frequencies and select a suitable backup frequency band; then, when strong interference is detected in the current 480 MHz - 530 MHz frequency band, automatically scan the surrounding frequency band environment and select the backup frequency band with the least interference and meeting the communication specifications for data transmission; finally, transmit the processed data to the outlier algorithm module in a suitable manner.

[0040] In a specific implementation example, the data cleaning module is used to remove noise and outliers because the power system is affected by external interference and the data contains transient noise and outliers; the data cleaning module combines the sliding window and box - plot method for processing; the sliding window moves a fixed window on the data sequence, calculates statistical features, and regards data points with large deviations as noise to be removed, which can be processed in real - time and the window size can be adjusted; the box - plot identifies outliers by calculating quartiles to further screen the data. The data feature extraction module uses a lightweight neural network to extract key features from the cleaned data; the network takes current, voltage, and power signals as inputs, learns the potential laws of the data, and outputs the current harmonic distortion rate THD, the voltage with a sag duration Δt≥100 ms, and the vibration signal in the range of 100 Hz - 1 kHz. The data standardization pre - processing module standardizes the feature data to unify the scale; judges anomalies based on the posterior probability of the current harmonic distortion rate, voltage sag duration, and vibration spectrum entropy; when the current harmonic distortion rate > 10%, the voltage sag duration > 300 ms, or the posterior probability of the vibration spectrum entropy > 90%, it is determined as a local anomaly and a secondary alarm is triggered.

[0041] Further, the working method of the intelligent hierarchical storage is to construct an intelligent hierarchical storage based on the time value and access frequency of power consumption data. For the high-frequency access data in the past week, it is stored on a high-performance solid-state storage device to ensure fast data reading and writing, meeting the requirements of real-time anti-stealing electricity analysis and the rapid query of real-time data by power operation and maintenance personnel; for the data that was accessed earlier and has a lower access frequency but still needs to be stored for a long time, it is automatically migrated to a large-capacity mechanical hard disk storage layer with lower cost.

[0042] In a specific implementation example, the basis for the layering of the intelligent hierarchical storage strategy is that there are significant differences in the time value and access frequency of power consumption data; recent data is crucial for real-time anti-stealing electricity analysis and operation and maintenance, and needs to be accessed quickly; although the timeliness of early data is low, it still has value in long-term analysis and needs to be stored for a long time; based on this, the system constructs an intelligent hierarchical storage mechanism. The high-performance solid-state storage layer of the intelligent hierarchical storage strategy is that the high-performance solid-state storage device SSD has fast reading and writing speeds and low latency, which is suitable for storing high-frequency data; the system will automatically store the power consumption data within the past week for real-time anti-stealing electricity analysis and rapid query by operation and maintenance personnel; these data can timely reflect power consumption behaviors and system states, helping to discover electricity-stealing behaviors and improve operation and maintenance efficiency. The large-capacity mechanical disk storage of the intelligent hierarchical storage strategy is that as time goes by, the access frequency of early data decreases, but it still has long-term value; the system will automatically migrate it to a mechanical hard disk HDD with low cost and large capacity; although the reading and writing speed of HDD is slow, due to the low access frequency of the data, the impact on performance is small; at the same time, the system establishes an indexing and management mechanism to facilitate the rapid positioning of low-frequency data. The intelligent hierarchical storage strategy has obvious advantages. High-frequency data is stored in SSD to meet real-time requirements and improve response and access performance; low-frequency data is stored in HDD to reduce costs and optimize resource allocation. The 6-month data retention period provides support for long-term analysis and decision-making, helping the scientific management and sustainable development of the power system. Further, the working method of the improved isolation forest algorithm is as follows: 1) Data preparation: Receive the data from the data transmission module to form an initial user data set D, and the data features of this data set include information on current, voltage, and power for subsequent analysis and processing; 2) Data preprocessing: Clean the data in the data set to remove missing values and incorrect data; for missing values, use the linear interpolation method to fill them; when the data sequence is missing, according to the adjacent non-missing values before and after for interpolation; 3) Feature extraction and selection: Extract features that may be related to electricity-stealing behaviors from the original data, the total harmonic distortion THD of current, the voltage sag duration, and the power fluctuation coefficient; the calculation formula of the power fluctuation coefficient is: In formula (1), is the power value at the i-th time point, is the average power, and N is the number of data points; The improved isolation forest algorithm uses a feature selection algorithm to select the most representative feature subset F based on correlation analysis and information gain to reduce the data dimension and computational complexity; the information gain The calculation formula is: In formula (2), C is the set of categories, and p(c) is the probability that category c appears in the data set D; is the subset of the data set D where the feature takes the value v; 4) Isolation forest model training: Extract m samples from the feature subset F to form a subset S for constructing an isolation tree; the improved isolation forest algorithm randomly selects a feature dimension and randomly selects a splitting point within the value range of this feature dimension For each sample x in the sample subset S, if , then x is divided into the left subtree; otherwise, it is divided into the right subtree; repeat the above steps to recursively construct the isolation tree until each leaf node contains only one sample or reaches the preset tree height; repeat the above process n times to construct n isolation trees to form the isolation forest model IF; 5) Anomaly score calculation: For each user sample x, calculate its path length in each tree in the isolation forest The average path length of sample x in the forest is The calculation formula is: In formula (3), is the input feature at the layer at position ; To more accurately measure the outlier degree of the sample, a weighted factor is introduced, which is related to the local density of the feature space where the sample is located; the local density can be estimated by calculating the reciprocal of the average distance between sample x and its k nearest neighbor samples; In formula (4), c(m) is the expected value of the average path length when the number of samples is m; 6) Electricity theft judgment and alarm: Set a threshold it will be determined that the user has a suspicion of electricity theft; a comprehensive judgment factor is introduced, which is related to the change trend of the user's historical anomaly scores; the anomaly scores of the user in the past T time periods are respectively then the comprehensive judgment factor The calculation formula is: The final judgment rule is that when When it does, the alarm mechanism will be triggered. In a specific embodiment, the improved isolation forest algorithm effectively integrates and analyzes various characteristics of user electricity consumption data through a unique calculation method, thereby accurately identifying electricity theft behavior, reducing the possibility of misjudgment, and thus improving the detection accuracy. Through a series of processes on the data characteristics of the initial user dataset D. During data preprocessing, by cleaning the data and linearly interpolating to fill in missing values, a high-quality data foundation is provided for subsequent analysis. In the feature extraction and selection stage, features such as the total harmonic distortion THD of the current, the voltage sag duration, and the power fluctuation coefficient, which may be related to electricity theft behavior, are refined from the original data. Using correlation analysis and information gain, the most representative feature subset F is selected, reducing the data dimension and computational complexity. During the training process of the isolation forest model, m samples are drawn from the feature subset F to construct an isolation tree. By randomly selecting the feature dimension f_k and the splitting point p_k, the isolation tree is recursively constructed, and this is repeated n times to form the isolation forest model IF. This method can quickly and effectively partition the user electricity consumption data and capture abnormal patterns in the data. When calculating the anomaly score, for each user sample x, first calculate its path length in each tree in the isolation forest and then obtain the average path length . To more accurately measure the degree of deviation of the sample from the group, a weighting factor related to the local density of the feature space where the sample is located is introduced The local density is estimated by calculating the reciprocal of the average distance between the sample x and its k nearest neighbor samples and then the final anomaly score is calculated through a series of formulas . In the electricity theft judgment and alarm link, set the threshold and the comprehensive judgment factor . When , it is determined that the user is suspected of electricity theft. At the same time, the comprehensive judgment factor is calculated by combining the anomaly scores of the user in the past T time periods. When , the alarm mechanism is triggered. Table 1 shows the parameter data of the improved isolation forest algorithm It can be seen from Table 1 that the setting of these parameters has an important impact on the performance of the improved isolation forest algorithm

[0043] Domain feature extraction: New power domain features (total harmonic distortion THD, voltage sag duration, power fluctuation coefficient) are added, while traditional algorithms only use original electrical parameters (2) Intelligent feature selection: Dual-mechanism feature screening using correlation analysis (Pearson coefficient > 0.8) and information gain (threshold > 0.3) reduces the dimension by 40% compared to traditional random selection Current harmonic detection: The abnormal detection rate increases by 62% when THD > 8% (characteristics of electricity theft equipment); Voltage sag identification: The detection accuracy reaches 92% for a duration > 200 ms; Power fluctuation analysis: The sensitivity increases by 3 times when the standard deviation threshold σ of the fluctuation coefficient > 1.5. The performance comparison data is shown in Table 2. Combinations of different parameter values will cause changes in a series of processes such as feature extraction, model training, and anomaly judgment. By reasonably adjusting the magnitudes of these parameters, the improved isolation forest algorithm can achieve real-time and accurate detection of electricity theft behavior in user electricity consumption data. Changes in some parameters can also speed up the calculation speed of the algorithm, dynamically calculate the most reasonable electricity theft judgment scheme for the system, and improve the security and management efficiency of the power system. Further, the working method of the display module is as follows: First, HTML5 Canvas collects the original voltage and current signals at a sampling rate ≥ 10 kHz and draws a multi-channel superimposed waveform; then, it displays data through multi-modal data visualization. The line chart dynamically displays the time series curves of power, voltage, and current. The heat map shows the abnormal power consumption density of the area according to geographical coordinates, and the color gradient is: green, orange, red. The 3D topology map renders the power grid node relationship, highlights abnormal nodes and associated lines; finally, the outlier confidence dynamically marks abnormal data with a confidence level ≥ 90% marked with a red asterisk; 70% ≤ confidence level < 90%: marked with an orange dashed line, and the annotation information is associated with a pop-up window that displays the abnormal type, occurrence time, and disposal suggestions. In a specific implementation example, the display module collects the original voltage and current signals at a sampling rate ≥ 10 kHz and uses HTML5 Canvas to draw a multi-channel superimposed waveform, enabling operation and maintenance personnel to intuitively compare the changes between the two, promptly detect abnormal fluctuations in voltage and current, and judge potential system failures. Among them, the line chart dynamically displays the time series curves of power, voltage, and current, and operation and maintenance personnel can thereby understand the system operation status and change trends at different times; the heat map shows the abnormal power consumption density of the area with a color gradient of green, orange, and red according to geographical coordinates, and operation and maintenance personnel can quickly locate the areas with concentrated abnormalities and efficiently carry out inspection and troubleshooting; the 3D topology map renders the power grid node relationship, highlights abnormal nodes and associated lines, helping operation and maintenance personnel clarify the power grid structure and quickly formulate a fault handling plan; the module marks abnormal data based on the outlier confidence: when the confidence level ≥ 90%, it is marked with a red asterisk, and clicking on the pop-up window displays the abnormal type, time, and disposal suggestions; when 70% ≤ confidence level < 90%, it is marked with an orange dashed line, and relevant abnormal information can also be viewed by clicking. This marking method helps operation and maintenance personnel classify and process abnormal data, improving the processing efficiency and accuracy. Further, the implementation method of the alarm module is as follows: It triggers an audible and visual alarm through abnormal confidence level grading, generates a structured disposal work order, stores alarm event data, and conducts self-check diagnosis. The specific implementation method is: The abnormal confidence level grading triggers the audible and visual alarm module: The first-level alarm corresponds to an abnormal confidence level ≥ 90%, triggering the on-site audible and visual alarm, with the buzzer frequency at 3 kHz ± 10% and the LED red light flashing quickly at a frequency of 1 Hz; the second-level alarm corresponds to 70% ≤ abnormal confidence level < 90%, sending text messages and emails to the preset responsible person, and the text message content includes the user ID, abnormal type, and location. The module for generating a structured disposal work order: The generated work order includes fields such as ID, abnormal type, location coordinates, recommended disposal measures, and remote switch-off, and synchronizes the work order status bidirectionally with the power grid asset management system SAP EAM through the RESTful API. Storing alarm event data, including the original electrical signal waveform with a sampling frequency ≥ 10 kHz and the vibration sensor spectrum disposal results in the frequency band of 100 Hz - 1000 Hz, with a storage period ≥ 5 years, and using blockchain hashing for evidence storage, and the hash value is synchronized to the judicial evidence storage platform.Self-check diagnosis module: It performs self-check at 1:00 am every day, including the detection of the RS-485 loop resistance sensor link and the ping test of the communication module. The E001 sensor open circuit fault code is uploaded to the operation and maintenance center work order system through the MQTT communication protocol. Although the specific implementation manners of the present invention are described above, those skilled in the art should understand that these specific implementation manners are only examples. Without departing from the principle and essence of the present invention, those skilled in the art can make various omissions, substitutions and changes to the details of the above methods and systems. For example, combining the above method steps so as to perform substantially the same function in a substantially same method to achieve substantially the same result belongs to the scope of the present invention. Therefore, the scope of the present invention is only defined by the appended claims.

Claims

1. An electricity consumption information collection system for anti - electricity theft based on an outlier algorithm model, comprising a data collection module, a data pre - processing module, a data transmission module, a data storage module, an outlier algorithm module, a display module, and an alarm module; characterized in that: The data collection module monitors the electricity consumption information of users in the low - voltage area in real - time through a non - invasive quantum current sensor and a ZMPT101B voltage sensor; The data pre - processing module performs data pre - processing through edge computing and outputs structured data to the data transmission module; The data transmission module uses a LoRa wireless communication module to transmit the collected data to the outlier algorithm module; The data storage module stores data through intelligent hierarchical storage, and the retention period of the stored data is ≥6 months; The outlier algorithm module detects and analyzes the time - series characteristics of voltage, current, and power through an improved isolation forest algorithm for outlier detection, identifies abnormal electricity consumption patterns, outputs the identification of suspected electricity - theft users and the abnormal confidence level to the alarm module, and synchronously pushes the characteristic data to the display module; wherein the outlier algorithm module includes a data fusion module, a feature decomposition module, a feature attribute recognition module, a weighted setting module, a rule judgment module, and a warning trigger module. Among them, the output module of the data fusion module is connected to the input module of the feature decomposition module, the output module of the feature decomposition module is connected to the input module of the feature attribute recognition module, the output module of the feature attribute recognition module is connected to the input module of the weighted setting module, the output module of the weighted setting module is connected to the input module of the rule judgment module, and the output module of the rule judgment module is connected to the input module of the warning trigger module; The display module uses HTML5 Canva to draw real - time waveforms of current and voltage, dynamically displays information in a multi - modal visualization, and displays dynamic markers of outlier confidence levels for abnormal data; The alarm module triggers audible and visual alarms through abnormal confidence - level grading, generates a structured disposal work order, stores alarm event data, and conducts self - inspection and diagnosis; The output end of the data collection module is connected to the input end of the data pre - processing module; the output end of the data pre - processing module is connected to the input end of the data transmission module; the output end of the data transmission module is connected to the input end of the data storage module; the output end of the data transmission module is connected to the input end of the outlier algorithm module; the output end of the data storage module is connected to the input end of the outlier algorithm module; the output end of the outlier algorithm module is connected to the input end of the display module; the output end of the outlier algorithm module is connected to the input end of the alarm module.

2. The electricity consumption information collection system for anti - electricity theft based on an outlier algorithm model according to claim 1, characterized in that: When the non-invasive quantum current sensor realizes milliampere-level weak current detection, the measurement accuracy is 0.063% and the frequency response range is 0.1Hz-1kHz, which is used to identify illegal diversion and electricity theft. The ZMPT101B voltage sensor adopts the principle of electromagnetic induction to accurately measure the voltage of the power supply line. The measurement error can usually be controlled within ±0.5%, meeting the voltage monitoring needs in the power system.

3. According to claim 1, a power consumption information collection system for anti-electricity theft based on an outlier algorithm model is characterized by: The data preprocessing module includes a data cleaning module, a data feature extraction module and a data standardization preprocessing module; the data cleaning module uses a sliding window to remove transient noise and filters outliers through a box plot method; the feature extraction module deploys a lightweight neural network, with current, voltage and power signals as inputs, and outputs the harmonic distortion rate THD of the current, the voltage with a sag duration Δt≥100ms, and a vibration signal with a frequency band of 100Hz-1kHz; the data standardization preprocessing module determines that the data is a local abnormal event when the data meets any of the conditions of current harmonic distortion rate >10%, voltage sag duration >300ms and vibration spectrum entropy posterior probability >90%, and directly triggers the secondary alarm of the alarm module; the output end of the data cleaning module is connected to the input end of the data feature extraction; the output end of the data feature extraction module is connected to the input end of the data standardization preprocessing module.

4. The power consumption information collection system for anti-electricity theft based on the outlier algorithm model according to claim 1 is characterized by: The working method of the LoRa wireless communication module is as follows: first, the output indicators of signal strength, signal quality index, packet loss rate and bit error rate in the communication process are monitored and evaluated in real time to determine whether it is necessary to switch the frequency and select a suitable backup frequency band; then, when it is detected that there is strong interference in the current 480MHz-530MHz frequency band, the surrounding frequency band environment is automatically scanned, and the backup frequency band with the least interference and in compliance with the communication specification is selected for data transmission; finally, the processed data is transmitted to the outlier algorithm module in an appropriate manner.

5. The power consumption information collection system for anti-electricity theft based on the outlier algorithm model according to claim 1 is characterized by: The working method of the intelligent tiered storage is to build an intelligent tiered storage based on the time value and access frequency of electricity consumption data. The high-frequency access data in the past week is stored on a high-performance solid-state storage device to ensure fast reading and writing of data, meet the real-time anti-electricity theft analysis and power operation and maintenance personnel's fast query needs for real-time data; for older and less frequently accessed data but still need to be stored for a long time, it is automatically migrated to a lower-cost large-capacity mechanical hard disk storage layer.

6. The electricity consumption information acquisition system for anti - electricity theft based on the outlier algorithm model according to claim 1, characterized in that: The working method of the improved isolation forest algorithm is: 1) Data preparation: receiving data from the data transmission module to form an initial user data set D, the data features of which include information on current, voltage and power for subsequent analysis and processing; 2) Data preprocessing: Clean the data in the dataset to remove missing values and incorrect data; for missing values, use linear interpolation to fill them; when the data sequence is missing, according to the non-missing values adjacent before and after to perform interpolation; 3) Feature extraction and selection: Extract features that may be related to electricity theft behavior from the original data, the total harmonic distortion THD of current, voltage sag duration, and power fluctuation coefficient; the calculation formula for the power fluctuation coefficient is: In formula (1), is the power value at the i-th time point, is the average power, and N is the number of data points; The improved isolation forest algorithm uses a feature selection algorithm to select the most representative feature subset F based on correlation analysis and information gain, so as to reduce the data dimension and computational complexity; the information gain calculation formula is: In formula (2), C is the set of categories, and p(c) is the probability that category c appears in the dataset D; is the subset of the dataset D where the feature takes the value v; 4) Isolation forest model training: Extract m samples from the feature subset F to form a subset S for constructing an isolation tree; the improved isolation forest algorithm randomly selects a feature dimension and randomly selects a splitting point within the value range of this feature dimension For each sample x in the sample subset S, if , then x is divided into the left subtree; otherwise, it is divided into the right subtree; repeat the above steps to recursively construct the isolation tree until each leaf node contains only one sample or reaches the preset tree height; repeat the above process n times to construct n isolation trees to form the isolation forest model IF; 5) Anomaly score calculation: For each user sample x, calculate its path length in each tree in the isolation forest The average path length of sample x in the forest is The calculation formula is: In formula (3), is the layer at position input feature; To more accurately measure the degree of outliers of the sample, a weighting factor is introduced which is related to the local density of the feature space where the sample is located; the local density can be estimated by calculating the reciprocal of the average distance between the sample x and its k nearest neighbor samples; In formula (4), c(m) is the expected value of the average path length when the number of samples is m; 6) Electricity theft judgment and alarm: Set a threshold it will be determined that the user has a suspicion of electricity theft; introduce a comprehensive judgment factor which is related to the change trend of the user's historical anomaly scores; the anomaly scores of the user in the past T time periods are respectively then the comprehensive judgment factor The calculation formula is: The final judgment rule is that when the alarm mechanism will be triggered.

7. The electricity consumption information acquisition system for anti - electricity theft based on the outlier algorithm model according to claim 1, characterized in that: The working method of the display module is: First, HTML5 Canvas collects the original voltage and current signals at a sampling rate of ≥10kHz and draws multi-channel superimposed waveforms; then, the data is displayed through multimodal data visualization. The line graph dynamically displays the time series curves of power, voltage and current, and the heat map shows the abnormal density of regional electricity consumption according to geographic coordinates. The color gradient is: green, orange, and red. The three-dimensional topology map renders the relationship between power grid nodes, highlights abnormal nodes and related lines; finally, the outlier confidence dynamically marks the abnormal data. Confidence ≥90% is marked with a red asterisk; 70%≤confidence<90%: marked with an orange dotted line, and the marked information is associated with a pop-up window to display the abnormal type, occurrence time and disposal suggestions.

8. The electricity consumption information collection system for anti - electricity theft based on the outlier algorithm model according to claim 1, characterized in that: The implementation method of the alarm module is: 1) Abnormal confidence level triggering sound and light alarm module: Level 1 alarm is abnormal confidence ≥ 90%, triggering on-site sound and light alarm, where the buzzer frequency is 3kHz±10%, and the LED red light flashes at a frequency of 1Hz; Level 2 alarm is 70%≤ abnormal confidence<90%, sending SMS and email to the preset responsible person, and the SMS content template is: user ID, abnormal type and location; 2) Generate a structured handling work order module, with fields including user ID, abnormality type, location coordinates, recommended handling measures and remote disconnection, and synchronize the work order status with the power grid asset management system SAP EAM in both directions through the RESTful API; store alarm event data, including the original electrical signal waveform with a sampling frequency ≥10kHz and the vibration sensor spectrum handling results with a frequency band of 100Hz-1000Hz, with a storage period of ≥5 years, and use blockchain hashing for evidence storage, and the hash value is synchronized to the judicial evidence storage platform; 3) Self-test diagnostic module: perform the following self-tests at 1:00 a.m. every day: RS-485 loop resistance sensor link detection and communication module ping test; E001 sensor open circuit fault code is uploaded to the operation and maintenance center work order system via the MQTT communication protocol.

Images