Non-intrusive electricity stealing detection method and system based on load identification

Abstract

The invention relates to the field of electricity larceny detection, in particular to a non-intrusive electricity larceny detection method and system based on load identification. According to the method, voltage and current signals of a total home-entry line of a user are collected, and a total load steady-state characteristic waveform is extracted to obtain total load characteristic data; detecting a load event to identify a switching moment and a characteristic variable quantity, and forming a load event sequence; matching an electrical appliance feature library to identify specific electrical appliance types and combinations, and determining actual electrical load composition; calculating a theoretical total electricity utilization power curve by combining typical power parameters of the electric appliance, and comparing the theoretical total electricity utilization power curve with an actually-measured curve to obtain a power difference rate; an adaptive detection threshold is dynamically set based on historical data, and when the difference rate continuously exceeds the threshold, electricity stealing is judged and an alarm is given. According to the invention, the problems of false alarm and missing alarm of electricity larceny detection caused by fixed threshold, one-sided feature extraction and insufficient load identification precision in the prior art are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of electricity theft detection, and in particular to a non-invasive electricity theft detection method and system based on load identification. Background Technology

[0002] With the rapid development of the power industry and the continuous growth of social electricity demand, electricity metering and anti-theft management have become key links in ensuring the safe and stable operation of the power grid and maintaining a fair order in the electricity market. Traditional methods of detecting electricity theft mostly rely on manual inspections or invasive monitoring equipment. The former is inefficient and costly, and it is difficult to achieve comprehensive coverage and real-time monitoring; the latter requires the installation of sensors at each electrical device, which not only increases hardware deployment and maintenance costs, but may also affect users' normal electricity use, and the load decomposition accuracy is limited in scenarios where multiple devices are operating simultaneously. Currently, non-intrusive load monitoring technology has gradually become a research hotspot in the field of electricity theft detection due to its advantage of only requiring the deployment of monitoring equipment on the user's main inlet line. However, existing technologies still have many shortcomings: On the one hand, most non-intrusive electricity theft detection methods rely on fixed thresholds to judge power differences, failing to fully consider individual differences in user electricity consumption behavior, parameter drift caused by appliance aging, and inherent fluctuations in electricity load at different times. This results in a lack of adaptability in threshold setting, easily leading to false alarms or missed alarms. On the other hand, existing technologies do not comprehensively extract the characteristics of the total load, often focusing only on steady-state power parameters and ignoring waveform details and the influence of harmonic components. This results in insufficient accuracy in identifying load switching events, which in turn affects the accuracy of identifying the actual electricity load composition, leading to a decrease in the reliability of the comparison results between theoretical and measured power. In addition, some methods have not designed effective combination identification mechanisms for scenarios where multiple appliances switch on and off simultaneously, making it difficult to accurately decompose complex load compositions, further exacerbating the error in judging power differences. Existing technology CN110412347B discloses a method for identifying electricity theft based on non-intrusive load monitoring. This method decomposes the total current and voltage to obtain the operating parameters of individual loads and compares them with meter data to determine electricity theft. However, this method does not involve setting a dynamic adaptive threshold and does not fully utilize waveform details and harmonic features to improve load identification accuracy. Its adaptability and accuracy in complex electricity usage scenarios need improvement. The core technical solution of this application aims to solve the technical problems of high false alarm rates and high false negative rates in existing non-intrusive electricity theft detection methods, such as lack of adaptability in threshold setting, incomplete load feature extraction, and insufficient accuracy in load composition identification. Summary of the Invention

[0003] This invention provides a non-intrusive electricity theft detection method and system based on load identification, aiming to solve the problems of false alarms and missed alarms in electricity theft detection caused by fixed thresholds, one-sided feature extraction, and insufficient load identification accuracy in the prior art.

[0004] To achieve the above objectives, the following technical solution is adopted.

[0005] A non-intrusive electricity theft detection method based on load identification includes the following steps: The system collects voltage and current signals from the user's main incoming line, extracts the steady-state characteristic waveform of the total load, and obtains the total load characteristic data. Load event detection is performed on the total load characteristic data to identify the load switching times and corresponding characteristic changes, resulting in a load event sequence. Based on the load event sequence, the specific types and combinations of appliances currently connected or disconnected are identified by matching appliance characteristic templates in a pre-established appliance characteristic library, thus obtaining the actual electricity load composition. According to the actual electricity load composition and combined with typical power parameters of each appliance type, a theoretical total power consumption curve is calculated. This theoretical total power consumption curve is then compared with the measured total power consumption curve directly calculated from the total load characteristic data to obtain the power difference rate. Based on historical statistical data of the power difference rate, an adaptive detection threshold is dynamically set. When the power difference rate continuously exceeds the adaptive detection threshold, electricity theft is detected, and an electricity theft alarm is generated.

[0006] Optionally, the step of collecting voltage and current signals from the user's main incoming line, extracting the steady-state characteristic waveform of the total load, and obtaining the total load characteristic data specifically includes: A synchronous sampling device is used to synchronously acquire voltage and current signals at a sampling rate higher than an integer multiple of the power frequency, so as to obtain the original voltage sampling data sequence and the original current sampling data sequence. The original voltage sampling data sequence and the original current sampling data sequence are subjected to power frequency extraction and harmonic separation using a digital filter bank to obtain the voltage fundamental component, the current fundamental component, the voltage harmonic component, and the current harmonic component. For each power frequency cycle, the RMS value of the fundamental current component, the phase difference between the fundamental voltage component and the fundamental current component, and the percentage of the amplitude of each current harmonic component relative to the amplitude of the fundamental current component are calculated to obtain a set of steady-state electrical parameters. The steady-state electrical parameters for multiple consecutive power frequency cycles are subjected to a moving average process to obtain a smoothed steady-state electrical parameter sequence. The derivative features of the fundamental current component waveform near the zero-crossing point and the distortion features of the fundamental voltage component waveform near the peak point are extracted and together constitute the waveform detail feature vector. The smoothed steady-state electrical parameter sequence is combined with the waveform detail feature vector to form the total load feature data.

[0007] Optionally, the step of performing load event detection on the total load characteristic data, identifying the load switching times and corresponding characteristic changes, and obtaining a load event sequence specifically includes: The rate of change of total active power calculated from the total load characteristic data is monitored in real time. When the rate of change of total active power exceeds a preset event trigger threshold, the moment is marked as a candidate event point. Centered on the candidate event point, a first preset time window is taken forward and a second preset time window is taken backward. The first statistical feature of the total load characteristic data within the first preset time window and the second statistical feature of the total load characteristic data within the second preset time window are calculated respectively. Calculate the difference between the first statistical feature and the second statistical feature. When the difference exceeds a preset feature change threshold, confirm the candidate event point as a valid load switching time. From the total load characteristic data of the first preset time window and the second preset time window corresponding to the confirmed valid load switching time, the change in the amplitude of the fundamental current wave, the change in the amplitude of each current harmonic, and the change in the phase difference are extracted to form the characteristic change of the load event. Record all valid load switching times and their corresponding characteristic changes in chronological order to form the load event sequence.

[0008] Optionally, based on the load event sequence, the step of identifying the specific types and combinations of appliances currently connected or disconnected by matching appliance feature templates in a pre-established appliance feature library to obtain the actual power load composition specifically includes: Extract the feature change of a load event from the load event sequence, calculate the similarity between the feature change and each electrical feature template in the electrical feature library, and obtain multiple electrical similarities. Select the top N appliance feature templates with the highest similarity from the multiple appliance similarities as candidate appliance templates, and record the appliance type and similarity score corresponding to each candidate appliance template; Determine whether the highest similarity score exceeds the preset single device matching threshold. If the highest similarity score exceeds the single device matching threshold, determine that this load event is a single appliance switching, and take the appliance type corresponding to the highest similarity score as the identification result. If the highest similarity score does not exceed the single device matching threshold, it is determined that the load event may be caused by the simultaneous switching of multiple appliances. The feature changes of the first N candidate appliance templates are linearly combined, and the minimum reconstruction error between the combined feature changes and the actual feature changes of the load event is calculated. When the minimum reconstruction error is lower than the preset combined device matching threshold, the load event is determined to be the combined switching of multiple identified appliances. Record all identified appliance types and their switching status in chronological order, and update the actual power load composition.

[0009] Optionally, the step of dynamically setting an adaptive detection threshold based on historical statistical data of the power difference rate specifically includes: In the initial stage, a conservative threshold based on the degree of power deviation of typical electricity theft behavior is used as the initial adaptive detection threshold. During system operation, the power difference rate of each detection cycle is continuously recorded, and the mean and standard deviation of the power difference rate within the sliding time window are calculated. The time series of the power difference rate is analyzed to identify the inherent fluctuation pattern of the power difference rate in different time periods and establish a time period benchmark model. The current power difference rate is compared with the historical mean power difference rate of the same time period, and its Z-score value is calculated. Combining the standard deviation of the power difference rate and the Z-score value, the adaptive detection threshold is dynamically adjusted, relaxing the adaptive detection threshold in periods with large inherent fluctuations and tightening the adaptive detection threshold in periods with small inherent fluctuations. When a continuous shift in the fluctuation pattern of the power difference rate is detected, a threshold learning mechanism is activated to adjust the parameters of the time period benchmark model to adapt to normal changes in user electricity consumption habits.

[0010] Optionally, after obtaining the actual power load composition, a load coordination mode analysis step can be further performed: From the actual electricity load composition, the time intervals and sequence of switching events for different electrical appliances within a preset time period are extracted to obtain an appliance usage sequence pattern. The appliance usage sequence pattern is then matched with a pre-stored database of typical normal electricity usage scenarios, and the matching confidence level between the current appliance usage sequence pattern and each typical normal electricity usage scenario is calculated. When the difference rate between the identified combined power of electrical appliances and the measured total power is high, and the matching confidence level between the current appliance usage sequence pattern and all typical normal electricity usage scenarios is lower than a preset abnormal pattern threshold, the determination of electricity theft is strengthened. The matching confidence level is used as auxiliary evidence and, together with the power difference rate, participates in the final decision on electricity theft.

[0011] Optional steps may also include environmental compensation procedures: Temperature and natural light intensity data of the user's environment are collected to obtain environmental parameters. Based on these environmental parameters, a pre-established environment-electricity consumption behavior correlation model is used to predict the theoretical expected power consumption of temperature-controlled appliances and lighting appliances in this environment, resulting in an environmental compensation power expectation value. This environmental compensation power expectation value is then incorporated into the calculation of the theoretical total power consumption curve to form a theoretical total power consumption curve compensated for environmental factors. The power difference rate is recalculated using the compensated theoretical total power consumption curve and the measured total power consumption curve to eliminate the impact of normal electricity consumption fluctuations caused by environmental changes on the electricity theft detection results.

[0012] Optionally, it also includes an online self-learning update step for the electrical appliance feature library: When the power difference rate remains below a preset learning trigger threshold and no electricity theft alarm is generated, the current period is determined to be a stable and normal power consumption period. During the stable and normal power consumption period, the characteristic changes of newly detected load events are temporarily stored as candidate new appliance features. Cluster analysis is performed on the candidate new appliance features to group load events with similar characteristics into the same category and calculate their cluster centers. When the number of samples in a certain cluster center exceeds a preset stability threshold, the cluster center feature is compared with all appliance feature templates in the existing appliance feature library. If the similarity is lower than the preset new device creation threshold, it is added as a new appliance feature template to the appliance feature library. For existing appliance feature templates, the feature changes of newly identified appliances during periods of normal power difference rate are used to update the appliance feature template with a weighted average, so that the appliance feature library can adapt to the actual situation of aging and new user appliances. A non-intrusive electricity theft detection system based on load identification includes: The system includes a data acquisition module, a feature extraction module, a load event detection module, a load identification module, a power analysis module, a threshold setting module, and an electricity theft detection module. The data acquisition module is configured to acquire the voltage and current signals of the user's main inlet line and send the voltage and current signals to the feature extraction module. The feature extraction module is configured to extract the steady-state characteristic waveform of the total load from the received voltage and current signals, obtain the total load characteristic data, and send the total load characteristic data to the load event detection module and the power analysis module. The load event detection module is configured to perform load event detection on the received total load feature data, identify the load switching time and the corresponding feature change, obtain the load event sequence, and send the load event sequence to the load identification module; The load identification module is configured to identify the specific types and combinations of electrical appliances currently in operation or switched off based on the received load event sequence by matching electrical feature templates in a pre-established electrical feature library, thereby obtaining the actual power load composition, and sending the actual power load composition to the power analysis module. The power analysis module is configured to calculate the theoretical total power consumption curve based on the received actual power load composition and the typical power parameters of each type of electrical appliance, and compare the theoretical total power consumption curve with the measured total power consumption curve directly calculated from the total load characteristic data to obtain the power difference rate, and send the power difference rate to the threshold setting module and the electricity theft determination module. The threshold setting module is configured to dynamically set an adaptive detection threshold based on the received historical statistical data of the power difference rate, and send the adaptive detection threshold to the electricity theft determination module. The electricity theft detection module is configured to compare the received power difference rate with the adaptive detection threshold. When the power difference rate continuously exceeds the adaptive detection threshold, it determines that there is electricity theft and generates an electricity theft alarm.

[0013] Optionally, the load event detection module includes: a power monitoring unit, an event marking unit, a feature calculation unit, and a sequence generation unit; The power monitoring unit is configured to monitor in real time the rate of change of total active power calculated from the total load characteristic data. When the rate of change of total active power exceeds a preset event trigger threshold, a trigger signal is sent to the event marking unit. The event marking unit is configured to mark the moment as a candidate event point when the trigger signal is received, and send the candidate event point information to the feature calculation unit; The feature calculation unit is configured to take the candidate event point as the center, take a first preset time window forward and a second preset time window backward, calculate the first statistical feature of the total load feature data in the first preset time window and the second statistical feature of the total load feature data in the second preset time window, respectively, and calculate the difference between the first statistical feature and the second statistical feature. When the difference exceeds a preset feature change threshold, the candidate event point is confirmed as a valid load switching time. From the total load feature data in the first preset time window and the second preset time window corresponding to the confirmed valid load switching time, the change in the fundamental current amplitude, the change in the amplitude of each current harmonic, and the change in the phase difference are extracted to form the feature change of the load event. The valid load switching time and its feature change are sent to the sequence generation unit. The sequence generation unit is configured to record all valid load switching times and their corresponding characteristic changes in chronological order to form the load event sequence.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This application collects voltage and current signals from the main incoming line, comprehensively extracts the steady-state characteristic waveforms of the total load to construct total load characteristic data, combines load event detection to identify load switching times and characteristic changes, accurately identifies the composition of electrical loads by matching the electrical appliance feature library, obtains the power difference rate by comparing theoretical and measured total power consumption curves, and dynamically sets adaptive detection thresholds based on historical statistical data. Finally, it realizes electricity theft alarm by continuously exceeding the threshold, effectively solving the problems of false alarms and missed alarms in electricity theft detection caused by fixed thresholds, one-sided feature extraction, and insufficient load identification accuracy in existing technologies, and significantly improving the accuracy and adaptability of electricity theft detection in complex electricity consumption scenarios. Further technical solutions include: synchronously acquiring signals at high sampling rates and separating fundamental and harmonic components; combining steady-state electrical parameter smoothing and waveform detail feature extraction to enrich the dimensions of total load characteristic data, providing more reliable basic data for subsequent load event detection and load identification, and improving the comprehensiveness and accuracy of feature extraction; another technical solution optimizes the identification process of load switching events by using power change rate-based event triggering, time window statistical feature comparison, and feature change extraction, improving the timeliness and effectiveness of event detection and providing accurate time nodes and feature basis for load composition identification; a similarity ranking and combination reconstruction error analysis mechanism designed for scenarios with multiple electrical appliances switching simultaneously effectively improves the identification accuracy of complex load compositions and avoids identification blind spots in single-device matching mode; and sliding window statistics, time period benchmark model construction, and Z-score are used to further enhance the identification accuracy of load switching events. The adaptive detection threshold intelligent optimization achieved through dynamic value adjustment allows the threshold to adapt to the different electricity consumption habits and time-period fluctuation characteristics of different users, further reducing the probability of false alarms and missed alarms. The analysis of appliance usage sequence patterns and matching with typical scenarios provide auxiliary evidence for electricity theft determination, enhancing the reliability of the determination results. The introduction of an environmental factor compensation mechanism eliminates the interference of normal electricity consumption fluctuations caused by environmental changes such as temperature and light on the judgment of power difference, improving the stability of detection results. The online self-learning update function of the appliance feature library enables the system to adapt to dynamic changes such as appliance aging and the addition of new equipment, extending the effective service life of the system and maintaining long-term detection accuracy. The corresponding system solution, through modular design, transforms the technical features of each method step into functional modules, ensuring the efficient implementation of the method and improving the scalability and maintainability of the system. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the control flow of an embodiment of a non-intrusive electricity theft detection method based on load identification according to the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0017] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0018] Example 1 like Figure 1 As shown, a non-intrusive electricity theft detection method based on load identification is presented in detail in this embodiment. By clarifying the core logic of each step, the hardware selection ideas and the algorithm application principles, the method is ensured to have good operability and wide applicability, and can be adapted to various common electricity use scenarios such as homes and small commercial places, so as to achieve accurate identification and timely alarm of electricity theft behavior.

[0019] The system collects voltage and current signals from the main incoming power line to users, extracting the steady-state characteristic waveform of the total load to obtain total load characteristic data. In actual deployment, hardware devices with synchronous acquisition capabilities are selected as the core of data acquisition, paired with high-precision sensors that conform to the voltage and current range of the power consumption scenario to ensure the accuracy and stability of signal acquisition. The voltage sensor must be compatible with the range of residential AC voltage and be able to accurately sense subtle changes in line voltage; the current sensor adopts a non-contact installation design to avoid damage to the incoming power line, while meeting the current monitoring needs of different power appliances, ensuring that clear current signals can be captured under various power loads. The synchronous sampling device must ensure that the acquisition sequence of voltage and current signals is consistent, and the sampling rate is set to an integer multiple higher than the power frequency to fully preserve the waveform details and characteristic information of the signal and avoid feature loss due to insufficient sampling. During the sampling process, the sensor converts the line voltage and current signals into analog signals suitable for subsequent processing, and then the analog signals are converted into digital signals through an analog-to-digital converter. The digital signals are stored in designated storage units in chronological order to provide raw data support for subsequent feature extraction.

[0020] The acquired raw voltage and current digital signals are processed using a digital filter bank to extract power frequency components and separate harmonics, thus eliminating noise interference and separating the useful fundamental and harmonic components. The digital filter bank includes low-pass and band-pass filters. The low-pass filter extracts the fundamental components of voltage and current, filtering out high-frequency noise; the band-pass filter is designed specifically for each harmonic to ensure accurate separation of harmonic signals of different frequencies. Through filtering, the fundamental voltage component, fundamental current component, voltage harmonic component, and current harmonic component are obtained, laying the foundation for subsequent steady-state parameter calculations.

[0021] For each power frequency cycle, key electrical parameters characterizing the steady-state characteristics of the total load are calculated, including the effective value of the fundamental current component, the phase difference between the fundamental voltage component and the fundamental current component, and the proportion of the amplitude of each current harmonic component relative to the amplitude of the fundamental current component. This forms a set of parameters that reflects the steady-state characteristics of the load. Each power frequency cycle corresponds to a complete set of steady-state electrical parameters. By performing a moving average processing on the steady-state electrical parameters of multiple consecutive power frequency cycles, the influence of random noise on parameter stability is reduced, resulting in a smoothed steady-state electrical parameter sequence that better reflects the actual operating state of the load. During the moving average processing, for special cases at the sequence boundaries, a reasonable expansion method is used to supplement missing data, ensuring the continuity and smoothness of the entire parameter sequence.

[0022] In addition to steady-state electrical parameters, waveform detail features of the total load also need to be extracted to form the total load characteristic data. These waveform detail features include the derivative characteristics of the fundamental current component near zero-crossing points and the distortion characteristics of the fundamental voltage component near its peak value. Zero-crossing points are determined by detecting the sign change of the fundamental current component signal; each power frequency cycle contains two zero-crossing points. By calculating the rate-of-change correlation characteristics of the signal near the zero-crossing points, the variation pattern of the current waveform at these key locations is captured. Peak points are determined by comparing adjacent signal sample values; each power frequency cycle contains two peak points. By analyzing the offset of the peak points and the waveform slope changes, the distortion characteristics of the voltage waveform are extracted. Combining the smoothed steady-state electrical parameter sequence with the waveform detail feature vectors forms comprehensive and information-rich total load characteristic data, providing reliable data support for subsequent load event detection and load identification.

[0023] Load event detection is performed on total load characteristic data to identify load switching times and corresponding characteristic changes, forming a load event sequence. First, the change in total active power calculated based on the total load characteristic data is monitored in real time. When the rate of change of total active power exceeds a preset event trigger threshold, that moment is marked as a candidate event point. The event trigger threshold is determined based on historical data statistics of user electricity consumption scenarios to ensure effective differentiation between normal electricity fluctuations and actual load switching events.

[0024] Centered on the candidate event point, appropriate time windows of varying lengths are selected forward and backward. Statistical characteristics of the total load data within these two time windows are calculated, including parameters reflecting the overall data characteristics such as mean, variance, maximum, and minimum values. The difference in statistical characteristics between the two time windows is calculated to determine whether the candidate event point is a valid load switching time. If the difference exceeds a preset characteristic change threshold, the time is confirmed as a valid load switching time; if the difference does not reach the threshold, it is determined to be a normal power consumption fluctuation, and the candidate event point is discarded.

[0025] For confirmed valid load switching moments, key characteristic changes reflecting load variations are extracted from the total load characteristic data of the corresponding two time windows. These include changes in the fundamental current amplitude, the amplitude of each current harmonic, and the phase difference. These characteristic changes accurately characterize the signal changes before and after load switching. All valid load switching moments and their corresponding characteristic changes are recorded chronologically to form a load event sequence. Each entry in the sequence fully contains the time information and characteristic change information of a single load switching event, providing a clear analytical object for subsequent load identification.

[0026] Based on load event sequences, the system identifies the specific types and combinations of appliances currently connected or disconnected by matching appliance feature templates in a pre-established appliance feature library, thus obtaining the actual power load composition. The appliance feature library is the core foundation for load identification, pre-storing feature templates for various common appliances, covering air conditioners, refrigerators, washing machines, water heaters, induction cookers, lighting fixtures, and other appliances commonly used in daily life and small commercial scenarios. Each appliance's feature template includes typical characteristics of its feature changes during the switching process. These typical characteristics are obtained through multiple switching tests of the appliances under standard conditions and statistical analysis, ensuring accurate characterization of the switching characteristics of this type of appliance.

[0027] The feature changes of each load event are extracted sequentially from the load event sequence, and their similarity is calculated with each appliance feature template in the appliance feature database. The similarity calculation employs an algorithm that effectively measures vector similarity; by comparing the degree of fit between the load event feature changes and the appliance feature templates, a similarity score is obtained for each template. Several appliance feature templates with the highest similarity scores are selected as candidate appliance templates, and the appliance type and similarity score corresponding to each candidate template are recorded.

[0028] The system determines whether the highest similarity score exceeds a preset single-device matching threshold. If it does, the load event is identified as a single appliance switching behavior, and the appliance type corresponding to the highest similarity score is taken as the identification result of this load event. If the highest similarity score does not exceed the single-device matching threshold, the load event may involve multiple appliances switching simultaneously. In this case, the feature changes of the first few candidate appliance templates are linearly combined, and the reconstruction error between the combined feature changes and the actual feature changes of the load event is calculated. When the reconstruction error is lower than the preset combined device matching threshold, the load event is determined to be the simultaneous switching of multiple appliances in that combination. If the reconstruction error is higher than the threshold, candidate templates are reselected for combination analysis until a matching appliance combination is found or it is marked as an unknown load switching.

[0029] All identified appliance types and their switching status are recorded chronologically, continuously updating the actual electrical load composition. This actual load composition clearly reflects the types and combinations of appliances operating at different times, providing accurate foundational information for subsequent calculations of the theoretical total power consumption curve. Events marked as unknown load switching are also recorded, providing a reference for subsequent updates and optimization of the appliance feature database.

[0030] Based on the actual electrical load composition and the typical power parameters of each appliance type, a theoretical total power consumption curve is calculated. This curve is then compared with the measured total power consumption curve directly calculated from the total load characteristic data to obtain the power difference rate. Typical power parameters for each appliance type are pre-stored in an appliance characteristic database, including key parameters such as the appliance's rated active power, reflecting the power consumption level of the appliance under normal operating conditions. The theoretical total power consumption curve is calculated at fixed time intervals. For each time point, all appliances in operation within the actual electrical load composition at that moment are statistically analyzed, and the typical power parameters of each appliance are summed to obtain the theoretical active power value for that time point. These values ​​are then arranged chronologically to form the theoretical total power consumption curve.

[0031] The measured total power consumption curve is directly calculated from the total load characteristic data. The calculation method is consistent with that for the total active power. The total load characteristic data corresponding to each time point is processed at the same time interval to obtain the measured active power value at that time point. These values ​​are then arranged in chronological order to form the measured total power consumption curve. The power difference rate is calculated based on a set statistical period. Within each statistical period, the difference in electricity consumption between the theoretical total power consumption curve and the measured total power consumption curve is calculated. The ratio of this electricity consumption difference to the theoretical total electricity consumption is then used as the power difference rate for that period. The power difference rate directly reflects the degree of deviation between theoretical and actual electricity consumption.

[0032] Based on historical statistical data of power difference rate, an adaptive detection threshold is dynamically set to adapt to the electricity consumption habits of different users and the fluctuation characteristics of electricity load at different times. In the initial stage of system operation, since sufficient historical data has not yet been accumulated, a conservative threshold based on the degree of power deviation of typical electricity theft behavior is used as the initial adaptive detection threshold to ensure that potential electricity theft is not missed in the initial stage. As the system continues to operate, power difference rate data for each statistical period is continuously accumulated to form a historical statistical data set.

[0033] Historical statistical data is analyzed to identify inherent fluctuation patterns in power difference rates across different time periods, establishing a time-period benchmark model. A day is divided into several typical time periods. For each time period, the historical mean and fluctuation range of the power difference rate are statistically analyzed and used as the benchmark parameter for that period. The current power difference rate is compared with the benchmark parameter for the same time period to calculate its deviation. Based on the fluctuation range of historical data, the adaptive detection threshold is dynamically adjusted. During periods with large inherent fluctuations in the power difference rate, the adaptive detection threshold is appropriately relaxed to avoid false alarms due to normal fluctuations; during periods with small inherent fluctuations, the adaptive detection threshold is tightened to improve the sensitivity of identifying electricity theft. When a persistent shift in the fluctuation pattern of the power difference rate is detected, a threshold learning mechanism is activated to update the parameters of the time-period benchmark model to adapt to normal changes in user electricity consumption habits, ensuring that the adaptive detection threshold remains at a reasonable level.

[0034] When the power difference rate consistently exceeds the adaptive detection threshold, electricity theft is detected and an alarm is generated. The criteria for determining whether the rate will consistently exceed the threshold are set based on the actual application scenario to ensure effective differentiation between short-term fluctuations and persistent electricity theft. Electricity theft alarms are divided into local and remote alarms. Local alarms are implemented through on-site audible and visual alarm devices to promptly alert users and on-site management personnel to abnormal power consumption. Remote alarms transmit alarm information to the power company's monitoring platform and the user's designated terminal device via a communication module. The alarm information includes key information such as alarm time, user identification, power difference rate, and load composition, facilitating timely monitoring and verification by power company staff, while also allowing users to understand their own power consumption status in real time.

[0035] After obtaining the actual electricity load composition, a load coordination mode analysis step is further performed. By analyzing the usage sequence patterns of electrical appliances, auxiliary evidence is provided for determining electricity theft, improving the reliability of the judgment results. From the actual electricity load composition, the time intervals and sequences of switching events of different electrical appliances within a set time period are extracted to form an appliance usage sequence pattern. This sequence pattern can reflect the user's electricity consumption habits and the coordinated operation rules of electrical appliances. Normal electricity consumption in different scenarios usually exhibits a relatively fixed sequence pattern.

[0036] The extracted appliance usage sequence patterns are matched against a pre-stored database of typical normal electricity usage scenarios. This database contains typical appliance usage sequence patterns for various common scenarios, such as weekday home scenarios, weekend home scenarios, and small shop operating scenarios. By calculating the matching confidence score between the current appliance usage sequence pattern and each typical normal electricity usage scenario, it is determined whether the current electricity usage pattern conforms to normal patterns. When the difference rate between the identified appliance combination power and the measured total power is high, and the matching confidence score between the current appliance usage sequence pattern and all typical normal electricity usage scenarios is lower than the preset abnormal pattern threshold, it indicates that the current electricity usage situation has both power deviation and does not conform to normal electricity usage patterns, further strengthening the judgment of electricity theft. At this time, the alarm level can be appropriately increased to remind relevant personnel to take priority action. The matching confidence score is used as auxiliary evidence and participates in the final electricity theft judgment decision together with the power difference rate. By comprehensively considering the two indicators, the probability of false judgment is effectively reduced and the accuracy of electricity theft identification is improved.

[0037] This method also includes an environmental factor compensation step to eliminate the impact of normal electricity consumption fluctuations caused by environmental changes on electricity theft detection results. Key environmental parameters of the user's environment are collected, primarily temperature and natural light intensity data. Data acquisition is achieved through adapted environmental sensors, with the sensor sampling interval set according to the rate of environmental change to ensure timely capture of changes in environmental parameters. Based on the collected environmental parameters, a pre-established environment-electricity consumption behavior correlation model is used to predict the expected power consumption of temperature-controlled appliances and lighting appliances under the given environmental conditions, yielding the expected value of environmental compensation power. The environment-electricity consumption behavior correlation model is established by analyzing the correspondence between a large number of environmental parameters and power consumption, accurately reflecting the impact of environmental changes on the power consumption of specific types of appliances.

[0038] The expected environmental compensation power value is incorporated into the calculation of the theoretical total power consumption curve, resulting in a theoretical total power consumption curve compensated for environmental factors. The conversion process employs appropriate methods depending on the type of appliance to ensure that the compensated theoretical power curve more accurately reflects actual electricity consumption. The power difference rate is recalculated using the compensated theoretical total power consumption curve and the measured total power consumption curve. This compensation method effectively eliminates the impact of normal factors such as power fluctuations in temperature-controlled appliances due to temperature changes and power adjustments in lighting appliances due to changes in illumination on the power difference rate, making the electricity theft detection results more reliable and accurate.

[0039] To adapt to dynamic changes such as aging, replacement, and addition of user appliances, this method also includes an online self-learning update step for the appliance feature library. When the power difference rate remains below a preset learning trigger threshold and no electricity theft alarm is generated, the current period is determined to be a stable and normal power consumption period. At this time, the system is in a stable operating state, and the collected load event feature changes can reflect the appliance characteristics under normal power consumption conditions. During the stable and normal power consumption period, newly detected load event feature changes that are not recognized by the existing appliance feature library are temporarily stored as candidate new appliance features.

[0040] Cluster analysis is performed on candidate new appliance features to group load events with similar features into the same category, and the cluster center for each feature is calculated. This cluster analysis filters out potential new appliance types with stable features. When the number of samples in a certain cluster center exceeds a preset stability threshold, it indicates that the appliance corresponding to that feature frequently appears in the user's power consumption scenario and has stable operating characteristics. The cluster center feature is compared with all appliance feature templates in the existing appliance feature library. If the similarity is lower than the preset new device creation threshold, it is added as a new appliance feature template to the appliance feature library, enabling the identification of newly added appliances. For existing appliance feature templates, the template is updated using the newly identified feature changes during a stable and normal power consumption period. A weighted average is used to fuse the old and new features, ensuring that the appliance feature template accurately reflects the characteristic changes after appliance aging, and ensuring that the accuracy of load identification remains at a high level over the long term.

[0041] Example 2 A non-intrusive electricity theft detection system based on load identification is provided in this embodiment, corresponding to the aforementioned non-intrusive electricity theft detection method based on load identification. This system provides a clear structure and complete functions, and achieves efficient collaboration among various functional links through modular design, ensuring that the method can be stably implemented and operated, and meeting the electricity theft detection needs of various electricity consumption scenarios.

[0042] The system adopts a modular architecture, mainly including a data acquisition module, a feature extraction module, a load event detection module, a load identification module, a power analysis module, a threshold setting module, and an electricity theft detection module. Each module achieves data transmission and command interaction through standardized communication interfaces and a data bus, ensuring the system's high efficiency and reliability. The system's core control employs a dual-core architecture, one responsible for real-time data processing and the other for complex algorithm calculations, balancing the system's real-time response and computational capabilities to quickly complete large-scale data processing and analysis tasks.

[0043] The data acquisition module is primarily responsible for collecting voltage and current signals from the user's main incoming power line. After preliminary processing, the acquired signals are sent to the feature extraction module. This module consists of a voltage sensor, a current sensor, a signal conditioning circuit, a synchronous sampling circuit, and a data transmission interface. The voltage sensor is a high-precision sensor adapted to the civilian voltage range, capable of accurately sensing changes in line voltage and outputting a stable analog signal. The current sensor adopts a non-contact design, making installation convenient and not affecting the normal power supply of the line, and can accurately capture current signals under different power loads.

[0044] The signal conditioning circuit amplifies, filters, and isolates the analog signal output from the sensor, ensuring that the signal meets the input requirements of subsequent sampling circuits. The amplification circuit adjusts the amplification factor based on the amplitude of the sensor output signal to bring the signal to a suitable processing range. The filtering circuit employs appropriate filtering methods to remove high-frequency noise and interference signals from the signal, retaining useful voltage and current signals. The isolation circuit provides electrical isolation between the input and output signals, improving the system's anti-interference capability and operational safety, and preventing damage to the system due to line voltage fluctuations.

[0045] The synchronous sampling circuit is the core of the data acquisition module. It employs an acquisition chip with multi-channel synchronous sampling capabilities to ensure strict synchronization of the sampling timing of voltage and current signals. The sampling rate is set to an integer multiple higher than the power frequency to fully preserve the waveform characteristics of the signals. The synchronous sampling circuit, controlled by a core control chip, samples the conditioned voltage and current signals at set intervals, converting the analog signals into digital signals. It then performs preliminary formatting and verification of the digital signals to ensure data integrity. The data transmission interface uses a high-speed communication interface to quickly and stably send the processed digital signals to the feature extraction module, ensuring real-time data transmission and meeting the time requirements of subsequent feature extraction and analysis.

[0046] The feature extraction module is configured to extract the steady-state characteristic waveform of the total load from the received voltage and current signals, obtaining total load characteristic data, which is then sent to the load event detection module and the power analysis module, respectively. This module uses a high-performance microprocessor as its core processing unit, possessing powerful digital signal processing and data computation capabilities, enabling it to quickly complete complex feature extraction algorithm calculations.

[0047] The feature extraction module's software functions mainly include four parts: digital filter bank processing, steady-state parameter calculation, moving average processing, and waveform detail feature extraction. The digital filter bank, implemented through software programming, includes low-pass and band-pass filters, used to extract the fundamental component and separate harmonic components, respectively, filtering out noise interference in the signal. The steady-state parameter calculation module calculates key parameters such as the effective value of the fundamental current, phase difference, and the proportion of each harmonic for each power frequency cycle, forming a set of steady-state electrical parameters. The moving average processing module smooths the steady-state parameters for multiple consecutive power frequency cycles, reducing the impact of random noise. The waveform detail feature extraction module identifies zero-crossing points and peak points in the signal, extracts derivative and distortion features, and forms a waveform detail feature vector. The smoothed steady-state electrical parameter sequence is combined with the waveform detail feature vector to form complete total load characteristic data, which is sent to subsequent modules through an efficient data transmission interface, providing high-quality data support for load event detection and power analysis.

[0048] The load event detection module is configured to detect load events from the received total load characteristic data, identify the load switching times and corresponding characteristic changes, obtain a load event sequence, and send the load event sequence to the load identification module. This module adopts a multi-threaded processing architecture, including a power monitoring unit, an event marking unit, a feature calculation unit, and a sequence generation unit. These units work collaboratively to ensure the real-time performance and accuracy of load event detection.

[0049] The power monitoring unit receives total load characteristic data from the feature extraction module in real time, continuously calculates the total active power based on the total load characteristic data, and monitors the rate of change of total active power. When the rate of change of power exceeds a preset event trigger threshold, a trigger signal is immediately sent to the event marking unit to notify it to mark the candidate event point. The event trigger threshold is dynamically adjusted based on historical data accumulated during system operation to ensure effective differentiation between normal power consumption fluctuations and load switching events.

[0050] Upon receiving a trigger signal, the event marking unit immediately marks that moment as a candidate event point, recording the precise timestamp and corresponding total load characteristic data index of the candidate event point. The timestamp is obtained through the system's real-time clock module to ensure the accuracy of time recording; the data index is used to quickly locate the total load characteristic data corresponding to the candidate event point, facilitating subsequent feature calculations. The event marking unit quickly sends the candidate event point information to the feature calculation unit via an internal message queue to avoid data backlog.

[0051] After receiving candidate event point information, the feature calculation unit extracts total load characteristic data within two fixed-length time windows centered on the candidate event point. It calculates the statistical characteristics of the total load characteristic data within each time window, including parameters reflecting the overall data characteristics such as mean, variance, maximum, and minimum values. By calculating the difference in statistical characteristics between the two time windows, it determines whether the candidate event point is a valid load switching time. When the difference exceeds a preset characteristic change threshold, the time is confirmed as a valid load switching time, and characteristic changes such as the change in fundamental current amplitude, the change in amplitude of each current harmonic, and the change in phase difference are extracted from the corresponding time window data. If the difference does not reach the threshold, it is determined to be a normal power consumption fluctuation, and the candidate event point is discarded. The feature calculation unit sends the valid load switching time and its corresponding characteristic changes to the sequence generation unit.

[0052] The sequence generation unit records all valid load switching times and their corresponding characteristic changes in chronological order, forming a load event sequence. The sequence generation unit employs an efficient data storage structure to ensure rapid storage and retrieval of load event data. When the storage capacity reaches its limit, it automatically overwrites the oldest event data according to a first-in, first-out (FIFO) principle, ensuring the system can continuously record the latest load event information. The sequence generation unit sends the load event sequence to the load identification module through a standardized communication interface, ensuring the stability and integrity of data transmission.

[0053] The load identification module is configured to identify the specific types and combinations of appliances currently connected or disconnected based on the received load event sequence by matching appliance feature templates in a pre-established appliance feature library. This yields the actual power load composition, which is then sent to the power analysis module. This module, powered by a high-performance digital signal processor, possesses strong parallel computing capabilities, enabling it to quickly complete numerous similarity calculations and combination matching tasks.

[0054] The load identification module includes a storage unit, a similarity calculation unit, a candidate template selection unit, a single device matching unit, a combined device matching unit, and a load composition update unit. The storage unit uses high-capacity, high-reliability storage media to store information such as the electrical appliance feature library, historical load event data, and identification results. The electrical appliance feature library pre-stores feature templates for various common electrical appliances, covering commonly used appliances in multiple power consumption scenarios. Each appliance's feature template has undergone extensive experimental verification to ensure its accuracy and representativeness. Historical load event data is used for subsequent statistical analysis and updates to the electrical appliance feature library. Identification results are stored in real time for easy querying and traceability.

[0055] The similarity calculation unit sequentially extracts the feature changes of each load event from the load event sequence and calculates their similarity with each appliance feature template in the appliance feature database. The similarity calculation employs an efficient vector similarity algorithm, which can quickly measure the degree of fit between two feature vectors and obtain a similarity score for each appliance feature template. The candidate template selection unit selects several appliance feature templates with the highest scores from all similarity scores as candidate appliance templates, recording the appliance type and similarity score for each candidate template, providing alternatives for subsequent matching decisions.

[0056] The single-device matching unit compares the highest similarity score with a preset single-device matching threshold. If the highest similarity score exceeds the threshold, the load event is determined to be the switching behavior of a single appliance, and the corresponding appliance type is output as the identification result. If the highest similarity score does not exceed the single-device matching threshold, the load event may involve the simultaneous switching of multiple appliances, and the candidate appliance template information is sent to the combined-device matching unit. The combined-device matching unit linearly combines the feature changes of the candidate appliance templates and calculates the reconstruction error between the combined feature changes and the actual feature changes of the load event. When the reconstruction error is lower than the preset combined-device matching threshold, the load event is determined to be the simultaneous switching of multiple appliances in that combination. If the reconstruction error is higher than the threshold, candidate templates are reselected for combined analysis until a matching appliance combination is found or it is marked as an unknown load switching.

[0057] The load composition update unit records all identified appliance types and their switching status in chronological order, updating the actual electrical load composition in real time. The actual electrical load composition is stored in a structured data format, clearly reflecting the operating status of electrical equipment in different time periods. The load composition update unit sends the updated actual electrical load composition to the power analysis module via a high-speed data interface, ensuring that the power analysis module can perform calculations based on the latest load composition data.

[0058] The power analysis module is configured to calculate the theoretical total power consumption curve based on the received actual electrical load composition and typical power parameters of each appliance type. It then compares this theoretical curve with the measured total power consumption curve calculated directly from the total load characteristic data to obtain the power difference rate. Simultaneously, the power difference rate is sent to the threshold setting module and the electricity theft detection module. This module employs an embedded processor and real-time operating system architecture, possessing powerful data analysis and task scheduling capabilities, enabling it to efficiently complete power curve calculation and difference rate analysis tasks.

[0059] The core functions of the power analysis module are implemented by a theoretical power calculation unit, a measured power calculation unit, and a power difference rate calculation unit. The theoretical power calculation unit obtains actual electrical load composition data and typical power parameters for each type of appliance from the load identification module. It statistically analyzes the types and combinations of appliances operating at each time point at fixed time intervals, sums the typical power parameters of each appliance, and obtains the theoretical active power value for that time point. These values ​​are then arranged chronologically to form a theoretical total power consumption curve. The measured power calculation unit obtains total load characteristic data from the feature extraction module and calculates the measured total active power value for each time point at the same time interval as the theoretical power calculation, forming a measured total power consumption curve. This ensures that the time dimensions of the two curves are consistent, facilitating subsequent comparative analysis.

[0060] The power difference rate calculation unit calculates the electricity consumption of both the theoretical and measured total power consumption curves within each set statistical period. By calculating the ratio of the difference in electricity consumption to the theoretical consumption, the power difference rate for each statistical period is obtained. The power difference rate directly reflects the deviation between theoretical and actual electricity consumption and is a core indicator for determining whether electricity theft has occurred. The power analysis module transmits the calculated power difference rate in real time to the threshold setting module and the electricity theft detection module via a high-speed communication interface, ensuring that subsequent modules can obtain key data promptly and make accurate judgments.

[0061] The threshold setting module is configured to dynamically set an adaptive detection threshold based on historical statistical data of the received power difference rate, and then send the adaptive detection threshold to the electricity theft determination module. This module uses a low-power, high-performance microcontroller as its core processing unit, which can reduce overall system power consumption while ensuring calculation accuracy, making it suitable for long-term continuous operation scenarios.

[0062] The threshold setting module includes a historical data storage unit, a time-period benchmark model establishment unit, a deviation calculation unit, and a threshold adjustment unit. The historical data storage unit continuously receives power difference rate data from the power analysis module via a communication interface, using a cyclical storage method to retain historical data for the most recent period, ensuring sufficient sample support for threshold calculation. The time-period benchmark model establishment unit divides a day into several typical time periods, statistically analyzes the historical power difference rate data for each time period, calculates the mean and fluctuation range of the power difference rate for each time period, establishes a time-period benchmark model, and updates and stores the model parameters in real time, providing a basis for threshold adjustment.

[0063] The deviation calculation unit compares the current power difference rate with the baseline model parameters for the same time period, calculating the degree of deviation of the current power difference rate from the mean. This quantified deviation index reflects the degree of abnormality in current electricity consumption. The threshold adjustment unit dynamically adjusts the adaptive detection threshold based on the deviation index and the fluctuation range of the time-period baseline model. During periods with small deviations and large fluctuations, the adaptive detection threshold is appropriately relaxed; during periods with large deviations and small fluctuations, the adaptive detection threshold is tightened to ensure that the threshold adapts to the electricity consumption characteristics of different time periods. When a persistent shift in the fluctuation pattern of the power difference rate is detected, the threshold adjustment unit activates the threshold learning mechanism, recalculates the parameters of the time-period baseline model, and updates the adaptive detection threshold, ensuring that the threshold always remains at a reasonable level that matches users' electricity consumption habits. The threshold setting module periodically sends the adjusted adaptive detection threshold to the electricity theft detection module through a reliable data transmission interface, ensuring that the electricity theft detection module can use the latest threshold for judgment.

[0064] The electricity theft detection module is configured to compare the received power difference rate with an adaptive detection threshold. When the power difference rate continuously exceeds the adaptive detection threshold, electricity theft is detected and an alarm is generated. This module uses a highly reliable microcontroller as its core, possesses comprehensive alarm logic processing capabilities and multi-interface communication capabilities, and can respond and output alarm information in a timely manner.

[0065] The electricity theft detection module includes a threshold comparison unit, an alarm level determination unit, a local alarm unit, and a remote alarm unit. The threshold comparison unit receives the power difference rate from the power analysis module and the adaptive detection threshold from the threshold setting module in real time, comparing the two to determine if the power difference rate exceeds the threshold. When the power difference rate exceeds the threshold, a timer starts to record the duration. When the duration reaches a preset judgment standard, a preliminary judgment is made that electricity theft has occurred. If the power difference rate falls back below the threshold within the time limit, the preliminary judgment is canceled, and normal monitoring resumes.

[0066] The alarm level determination unit, combining the matching confidence results from load coordination mode analysis, classifies the severity of electricity theft into levels. A high-level alarm is triggered when the power difference rate significantly exceeds the threshold and the matching confidence is below the abnormal mode threshold; a medium-level alarm is triggered when the power difference rate slightly exceeds the threshold or the matching confidence is at a critical level; and a low-level alarm is triggered when only a short-term power difference rate exceedance occurs without other abnormal signs. Different alarm levels correspond to different processing priorities, facilitating appropriate response measures by relevant personnel based on the alarm level.

[0067] The local alarm unit connects to the audible and visual alarm device, outputting different alarm signals according to the alarm level. For high-level alarms, the device emits continuous audible and visual signals; for medium-level alarms, it emits intermittent audible and visual signals; and for low-level alarms, it only emits a light signal, ensuring that on-site personnel can quickly identify the alarm level. The remote alarm unit transmits alarm information to the power company's monitoring platform and user-designated terminal devices via various communication methods such as GPRS and WiFi. The alarm information includes detailed information such as alarm time, alarm level, user information, power difference rate data, load composition, and matching confidence level, facilitating power company staff to monitor suspected electricity theft in real time and arrange investigations. It also allows users to promptly understand their own abnormal electricity usage and cooperate with relevant inspections.

[0068] The power monitoring unit, event marking unit, feature calculation unit, and sequence generation unit in the load event detection module work closely together via an internal bus to form an efficient load event detection process. The power monitoring unit continuously monitors changes in total active power to ensure no potential load switching events are missed; the event marking unit quickly responds to trigger signals and accurately records candidate event point information; the feature calculation unit accurately filters out valid load switching moments and extracts characteristic changes through time window analysis and difference calculation; the sequence generation unit organizes valid event information in chronological order to form a clearly structured load event sequence, providing high-quality input data for subsequent load identification. The data flow between these units is optimized to minimize transmission delays and meet system real-time requirements.

[0069] The system employs a variety of standardized interface combinations for communication between its modules, selecting the appropriate communication method based on data transmission rate and distance requirements. Data transmission between short-distance modules utilizes high-speed parallel or serial interfaces to ensure high speed and reliability; data transmission between long-distance modules uses network or wireless communication interfaces to achieve remote data transmission. All communication interfaces are equipped with robust error detection and retransmission mechanisms to ensure no data loss or errors occur during transmission, guaranteeing smooth collaboration between system modules.

[0070] The system's power module features a wide voltage input design, adaptable to various power supply environments. It also incorporates multiple protection functions, including overvoltage protection, overcurrent protection, and short-circuit protection, ensuring system safety under voltage fluctuations or sudden faults. The power module outputs a stable DC voltage, providing reliable power support to all system modules and guaranteeing long-term stable operation. The system's enclosure is designed to be dustproof, waterproof, and resistant to electromagnetic interference, making it suitable for various indoor and outdoor installation environments. It also boasts excellent heat dissipation performance, ensuring that the system will not experience performance degradation or malfunction due to overheating during prolonged operation.

[0071] This system, through modular design and precise algorithms, can efficiently and accurately complete non-invasive electricity theft detection tasks, and is suitable for various electricity usage scenarios such as homes and small commercial establishments. The system not only effectively identifies common electricity theft behaviors, but also possesses good adaptability and scalability, continuously optimizing its performance as user electricity usage scenarios change and appliance types are updated. It provides reliable technical support for power companies' anti-electricity theft management, while also creating a safe and standardized electricity usage environment for users.

[0072] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A non-intrusive electricity theft detection method based on load identification, characterized in that, Includes the following steps: The system collects voltage and current signals from the user's main incoming line, extracts the steady-state characteristic waveform of the total load, and obtains the total load characteristic data. Load event detection is performed on the total load characteristic data to identify the load switching times and corresponding characteristic changes, resulting in a load event sequence. Based on the load event sequence, the specific types and combinations of appliances currently connected or disconnected are identified by matching appliance characteristic templates in a pre-established appliance characteristic library, thus obtaining the actual electricity load composition. According to the actual electricity load composition and combined with typical power parameters of each appliance type, a theoretical total power consumption curve is calculated. This theoretical total power consumption curve is then compared with the measured total power consumption curve directly calculated from the total load characteristic data to obtain the power difference rate. Based on historical statistical data of the power difference rate, an adaptive detection threshold is dynamically set. When the power difference rate continuously exceeds the adaptive detection threshold, electricity theft is detected, and an electricity theft alarm is generated.

2. The non-intrusive electricity theft detection method based on load identification according to claim 1, characterized in that, The steps of collecting voltage and current signals from the user's main incoming line, extracting the steady-state characteristic waveform of the total load, and obtaining the total load characteristic data specifically include: A synchronous sampling device is used to synchronously acquire voltage and current signals at a sampling rate higher than an integer multiple of the power frequency, so as to obtain the original voltage sampling data sequence and the original current sampling data sequence. The original voltage sampling data sequence and the original current sampling data sequence are subjected to power frequency extraction and harmonic separation using a digital filter bank to obtain the voltage fundamental component, the current fundamental component, the voltage harmonic component, and the current harmonic component. For each power frequency cycle, the RMS value of the fundamental current component, the phase difference between the fundamental voltage component and the fundamental current component, and the percentage of the amplitude of each current harmonic component relative to the amplitude of the fundamental current component are calculated to obtain a set of steady-state electrical parameters. The steady-state electrical parameters for multiple consecutive power frequency cycles are subjected to a moving average process to obtain a smoothed steady-state electrical parameter sequence. The derivative features of the fundamental current component waveform near the zero-crossing point and the distortion features of the fundamental voltage component waveform near the peak point are extracted and together constitute the waveform detail feature vector. The smoothed steady-state electrical parameter sequence is combined with the waveform detail feature vector to form the total load feature data.

3. The non-intrusive electricity theft detection method based on load identification according to claim 1, characterized in that, The steps of performing load event detection on the total load characteristic data, identifying the load switching times and corresponding characteristic changes, and obtaining the load event sequence specifically include: The rate of change of total active power calculated from the total load characteristic data is monitored in real time. When the rate of change of total active power exceeds a preset event trigger threshold, the moment is marked as a candidate event point. Centered on the candidate event point, a first preset time window is taken forward and a second preset time window is taken backward. The first statistical feature of the total load characteristic data within the first preset time window and the second statistical feature of the total load characteristic data within the second preset time window are calculated respectively. Calculate the difference between the first statistical feature and the second statistical feature. When the difference exceeds a preset feature change threshold, confirm the candidate event point as a valid load switching time. From the total load characteristic data of the first preset time window and the second preset time window corresponding to the confirmed valid load switching time, the change in the amplitude of the fundamental current wave, the change in the amplitude of each current harmonic, and the change in the phase difference are extracted to form the characteristic change of the load event. Record all valid load switching times and their corresponding characteristic changes in chronological order to form the load event sequence.

4. The non-intrusive electricity theft detection method based on load identification according to claim 1, characterized in that, Based on the load event sequence, the steps of identifying the specific types and combinations of appliances currently connected or disconnected by matching appliance feature templates in a pre-established appliance feature library to obtain the actual electricity load composition specifically include: Extract the feature change of a load event from the load event sequence, calculate the similarity between the feature change and each electrical feature template in the electrical feature library, and obtain multiple electrical similarities. Select the top N appliance feature templates with the highest similarity from the multiple appliance similarities as candidate appliance templates, and record the appliance type and similarity score corresponding to each candidate appliance template; Determine whether the highest similarity score exceeds the preset single device matching threshold. If the highest similarity score exceeds the single device matching threshold, determine that this load event is a single appliance switching, and take the appliance type corresponding to the highest similarity score as the identification result. If the highest similarity score does not exceed the single device matching threshold, it is determined that the load event may be caused by the simultaneous switching of multiple appliances. The feature changes of the first N candidate appliance templates are linearly combined, and the minimum reconstruction error between the combined feature changes and the actual feature changes of the load event is calculated. When the minimum reconstruction error is lower than the preset combined device matching threshold, the load event is determined to be the combined switching of multiple identified appliances. Record all identified appliance types and their switching status in chronological order, and update the actual power load composition.

5. The non-intrusive electricity theft detection method based on load identification according to claim 1, characterized in that, The step of dynamically setting an adaptive detection threshold based on historical statistical data of the power difference rate specifically includes: In the initial stage, a conservative threshold based on the degree of power deviation of typical electricity theft behavior is used as the initial adaptive detection threshold. During system operation, the power difference rate of each detection cycle is continuously recorded, and the mean and standard deviation of the power difference rate within the sliding time window are calculated. The time series of the power difference rate is analyzed to identify the inherent fluctuation pattern of the power difference rate in different time periods and establish a time period benchmark model. The current power difference rate is compared with the historical mean power difference rate of the same time period, and its Z-score value is calculated. Combining the standard deviation of the power difference rate and the Z-score value, the adaptive detection threshold is dynamically adjusted, relaxing the adaptive detection threshold in periods with large inherent fluctuations and tightening the adaptive detection threshold in periods with small inherent fluctuations. When a continuous shift in the fluctuation pattern of the power difference rate is detected, a threshold learning mechanism is activated to adjust the parameters of the time period benchmark model to adapt to normal changes in user electricity consumption habits.

6. A non-intrusive electricity theft detection method based on load identification according to claim 4, characterized in that, After obtaining the actual power load composition, the load coordination mode analysis step is further performed: From the actual electricity load composition, the time intervals and sequence of switching events for different electrical appliances within a preset time period are extracted to obtain an appliance usage sequence pattern. The appliance usage sequence pattern is then matched with a pre-stored database of typical normal electricity usage scenarios, and the matching confidence level between the current appliance usage sequence pattern and each typical normal electricity usage scenario is calculated. When the difference rate between the identified combined power of electrical appliances and the measured total power is high, and the matching confidence level between the current appliance usage sequence pattern and all typical normal electricity usage scenarios is lower than a preset abnormal pattern threshold, the determination of electricity theft is strengthened. The matching confidence level is used as auxiliary evidence and, together with the power difference rate, participates in the final decision on electricity theft.

7. The non-intrusive electricity theft detection method based on load identification according to claim 1, characterized in that, It also includes steps for compensating for environmental factors: Temperature and natural light intensity data of the user's environment are collected to obtain environmental parameters. Based on these environmental parameters, a pre-established environment-electricity consumption behavior correlation model is used to predict the theoretical expected power consumption of temperature-controlled appliances and lighting appliances in this environment, resulting in an environmental compensation power expectation value. This environmental compensation power expectation value is then incorporated into the calculation of the theoretical total power consumption curve to form a theoretical total power consumption curve compensated for environmental factors. The power difference rate is recalculated using the compensated theoretical total power consumption curve and the measured total power consumption curve to eliminate the impact of normal electricity consumption fluctuations caused by environmental changes on the electricity theft detection results.

8. The non-intrusive electricity theft detection method based on load identification according to claim 1, characterized in that, It also includes the online self-learning update steps for the electrical appliance feature library: When the power difference rate remains below the preset learning trigger threshold and no electricity theft alarm is generated, the current period is determined to be a stable and normal power consumption period; during the stable and normal power consumption period, the characteristic changes of newly detected load events are temporarily stored as candidate new appliance characteristics. Cluster analysis is performed on the candidate new appliance features to group load events with similar features into the same category and calculate their cluster centers. When the number of samples in a certain cluster center exceeds a preset stability threshold, the features of that cluster center are compared with all appliance feature templates in the existing appliance feature library. If the similarity is lower than the preset new device creation threshold, it is added as a new appliance feature template to the appliance feature library. For existing appliance feature templates, the feature changes during periods of new identification and normal power difference rate are used to update the appliance feature template with a weighted average, so that the appliance feature library can adapt to the actual situation of aging and new user appliances.

9. A non-intrusive electricity theft detection system based on load identification, comprising a non-intrusive electricity theft detection method based on load identification according to any one of claims 1 to 8, characterized in that, include: The system includes a data acquisition module, a feature extraction module, a load event detection module, a load identification module, a power analysis module, a threshold setting module, and an electricity theft detection module. The data acquisition module is configured to acquire the voltage and current signals of the user's main inlet line and send the voltage and current signals to the feature extraction module. The feature extraction module is configured to extract the steady-state characteristic waveform of the total load from the received voltage and current signals, obtain the total load characteristic data, and send the total load characteristic data to the load event detection module and the power analysis module. The load event detection module is configured to perform load event detection on the received total load feature data, identify the load switching time and the corresponding feature change, obtain the load event sequence, and send the load event sequence to the load identification module; The load identification module is configured to identify the specific types and combinations of electrical appliances currently in operation or switched off based on the received load event sequence by matching electrical feature templates in a pre-established electrical feature library, thereby obtaining the actual power load composition, and sending the actual power load composition to the power analysis module. The power analysis module is configured to calculate the theoretical total power consumption curve based on the received actual power load composition and the typical power parameters of each type of electrical appliance, and compare the theoretical total power consumption curve with the measured total power consumption curve directly calculated from the total load characteristic data to obtain the power difference rate, and send the power difference rate to the threshold setting module and the electricity theft determination module. The threshold setting module is configured to dynamically set an adaptive detection threshold based on the received historical statistical data of the power difference rate, and send the adaptive detection threshold to the electricity theft determination module. The electricity theft detection module is configured to compare the received power difference rate with the adaptive detection threshold. When the power difference rate continuously exceeds the adaptive detection threshold, it determines that there is electricity theft and generates an electricity theft alarm.

10. A non-intrusive electricity theft detection system based on load identification according to claim 9, characterized in that, The load event detection module includes: a power monitoring unit, an event marking unit, a feature calculation unit, and a sequence generation unit; The power monitoring unit is configured to monitor in real time the rate of change of total active power calculated from the total load characteristic data. When the rate of change of total active power exceeds a preset event trigger threshold, a trigger signal is sent to the event marking unit. The event marking unit is configured to mark the moment as a candidate event point when the trigger signal is received, and send the candidate event point information to the feature calculation unit; The feature calculation unit is configured to take the candidate event point as the center, take a first preset time window forward and a second preset time window backward, calculate the first statistical feature of the total load feature data in the first preset time window and the second statistical feature of the total load feature data in the second preset time window, respectively, and calculate the difference between the first statistical feature and the second statistical feature. When the difference exceeds a preset feature change threshold, the candidate event point is confirmed as a valid load switching time. From the total load feature data in the first preset time window and the second preset time window corresponding to the confirmed valid load switching time, the change in the fundamental current amplitude, the change in the amplitude of each current harmonic, and the change in the phase difference are extracted to form the feature change of the load event. The valid load switching time and its feature change are sent to the sequence generation unit. The sequence generation unit is configured to record all valid load switching times and their corresponding characteristic changes in chronological order to form the load event sequence.

Images