A method and system for identifying electricity theft behavior

Abstract

The present invention relates to the technical field of electricity theft behavior recognition, and discloses a method and system for electricity theft behavior recognition, including collecting electricity consumption behavior data and converting it into frequency domain components; performing frequency domain analysis on the frequency domain components according to the initial frequency band demarcation point to obtain initial frequency domain features, and adjusting the initial frequency band demarcation point according to the initial frequency domain features to obtain the frequency band demarcation point; performing frequency domain analysis on the frequency domain components according to the frequency band demarcation point to obtain frequency domain features, and integrating the frequency domain features into a frequency domain feature matrix; inputting the frequency domain feature matrix into a feature extraction model to obtain a convolutional feature vector, and performing data processing on the convolutional feature vector through an electricity theft behavior detection algorithm to obtain the electricity theft probability. The present invention improves the accuracy of frequency domain feature extraction by dynamically adjusting the frequency band demarcation point, extracts deep-level electricity consumption features through the feature extraction model, and calculates the electricity theft probability based on the eigenvalues after non-linear processing, improving the accuracy and efficiency of electricity theft behavior recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of electricity theft behavior recognition, and particularly to a method and system for recognizing electricity theft behavior. Background Art

[0002] With the wide application of power systems and the increasing growth of power demand, electricity theft behavior has become one of the important challenges faced by power companies. Electricity theft behavior not only directly affects the economic benefits of power companies, but also disrupts the normal power supply and demand balance and affects the operation stability of the power grid. To address this issue, power companies and related research institutions have proposed various electricity theft detection methods, including real-time monitoring methods based on smart meters, anomaly detection methods based on user electricity consumption data, and behavior recognition technologies based on machine learning. Among them, smart meters are widely used in the detection of electricity theft behavior because they can collect users' electricity consumption data in real time. Anomaly detection based on user electricity consumption data identifies abnormal electricity consumption patterns by analyzing historical data to discover potential electricity theft behavior. In recent years, with the development of artificial intelligence technology, deep learning and data mining methods have been introduced into the field of electricity theft detection. By constructing complex classification models to perform pattern recognition and classification on users' electricity consumption behaviors, the accuracy and response speed of electricity theft behavior detection have been improved.

[0003] However, there are still some deficiencies in the application of existing technologies. Rule-based anomaly detection methods rely on manually set thresholds and rules, which are difficult to adapt to complex user behaviors and changing electricity consumption patterns, often leading to misjudgments and missed detections. Machine learning-based electricity theft detection methods are often too simple in feature extraction and fail to fully utilize frequency domain information, making it difficult to effectively capture the periodic characteristics of users' electricity consumption behaviors. In the interactive processing of multi-dimensional features, existing methods lack fine non-linear interactive modeling, resulting in the neglect of the correlation between features, thus affecting the detection accuracy. Traditional electricity theft detection models usually have difficulty maintaining high computational performance when faced with a large amount of real-time data. Although deep learning models have certain advantages in feature recognition, there is still a trade-off between detection accuracy and computational complexity. How to construct an efficient electricity theft detection model that can effectively capture frequency domain features and combine deep learning and complex feature interactions remains a difficult point in the current technical field. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a method and system for recognizing electricity theft behavior, so as to solve the limitations of existing electricity theft behavior recognition methods in capturing the dynamic changes and periodic characteristics of electricity consumption behaviors, and achieve the effect of improving the efficiency and accuracy of electricity theft behavior recognition.

[0005] In a first aspect, the present invention provides a method for recognizing electricity theft behavior, the method comprising:

[0006] Collect the electricity consumption behavior data of the user to be identified, and convert the electricity consumption behavior data into frequency domain components through fast Fourier transform;

[0007] According to the preset initial frequency band demarcation point, perform frequency domain analysis on the frequency domain components to obtain initial frequency domain features, and adjust the initial frequency band demarcation point according to the initial frequency domain features to obtain the frequency band demarcation point;

[0008] According to the frequency band demarcation point, perform frequency domain analysis on the frequency domain components to obtain frequency domain features, and integrate the frequency domain features into a frequency domain feature matrix. The frequency domain features include high-frequency energy, low-frequency energy, and spectral entropy;

[0009] Input the frequency domain feature matrix into a preset feature extraction model to obtain a convolutional feature vector, and perform data processing on the convolutional feature vector through an electricity theft behavior detection algorithm to obtain the electricity theft probability of the user to be identified.

[0010] Further, the step of performing frequency domain analysis on the frequency domain components according to the preset initial frequency band demarcation point to obtain initial frequency domain features, and adjusting the initial frequency band demarcation point according to the initial frequency domain features to obtain the frequency band demarcation point includes:

[0011] Extract high-frequency energy and low-frequency energy from the frequency domain components according to the preset initial frequency band demarcation point, and calculate the energy ratio between the high-frequency energy and the low-frequency energy;

[0012] Calculate the spectral entropy according to the power spectral density of each frequency domain component;

[0013] Compare the energy ratio with the energy ratio threshold, and compare the spectral entropy with the spectral entropy threshold. According to the comparison results, adjust the initial frequency band demarcation point to obtain the frequency band demarcation point.

[0014] Further, the step of comparing the energy ratio with the energy ratio threshold, comparing the spectral entropy with the spectral entropy threshold, and adjusting the initial frequency band demarcation point according to the comparison results to obtain the frequency band demarcation point includes:

[0015] Judge whether the energy ratio is greater than the energy ratio threshold. If it is greater than the energy ratio threshold, then judge whether the spectral entropy is greater than the spectral entropy threshold. If it is greater than or equal to the spectral entropy threshold, then reduce the initial frequency band demarcation point. If it is less than the spectral entropy threshold, then keep the initial frequency band demarcation point unchanged;

[0016] If it is less than or equal to the energy ratio threshold, determine whether the spectral entropy is less than or equal to the spectral entropy threshold. If it is less than or equal to the spectral entropy threshold, increase the initial frequency band demarcation point. If it is greater than the spectral entropy threshold, keep the initial frequency band demarcation point unchanged.

[0017] Further, the high-frequency energy is calculated using the following formula:

[0018]

[0019] In the formula, represents the initial frequency band demarcation point, represents the maximum frequency, represents the y-th frequency component, represents the power spectral density of the y-th frequency component;

[0020] The low-frequency energy is calculated using the following formula:

[0021]

[0022] In the formula, represents the minimum frequency;

[0023] The spectral entropy is calculated using the following formula:

[0024]

[0025] In the formula, represents the spectral entropy of the power spectral density of the y-th frequency component, represents the normalized power spectral density of the y-th frequency component.

[0026] Further, the feature extraction model is constructed based on a deep convolutional neural network. The feature extraction model includes three sequentially connected convolutional layers, and residual connections are set between each convolutional layer.

[0027] Further, the steps of performing data processing on the convolutional feature vector through the electricity theft behavior detection algorithm to obtain the electricity theft probability of the user to be identified include:

[0028] Perform non-linear feature processing on each eigenvalue in the convolutional feature vector, and sum the processed eigenvalues to obtain a non-linear influence value;

[0029] Calculate the interaction value of each feature pair in the convolutional feature vector based on the processed eigenvalues, and sum the interaction values to obtain a non-linear interaction value;

[0030] Calculate the electricity theft probability of the user to be identified based on the non-linear influence value and the non-linear interaction value.

[0031] Further, the processed eigenvalue is represented by the following formula:

[0032]

[0033] where x i represents the i-th eigenvalue in the convolutional feature vector, and q represents the attenuation factor;

[0034] The non-linear influence value is calculated by the following formula:

[0035]

[0036] where n represents the total number of eigenvalues in the convolutional feature vector, and c represents the bias constant;

[0037] The interaction value is calculated by the following formula:

[0038]

[0039] where x j represents the j-th eigenvalue in the convolutional feature vector;

[0040] The non-linear interaction value is calculated by the following formula:

[0041]

[0042] where represents the adjustment parameter;

[0043] The electricity theft probability is calculated by the following formula:

[0044]

[0045] where represents the Sigmoid function.

[0046] Further, after the step of obtaining the electricity theft probability of the user to be identified, the method further includes:

[0047] Determining the electricity theft risk interval of the user to be identified according to the electricity theft probability and a preset electricity theft probability threshold;

[0048] Dynamically adjusting the electricity theft probability according to the real-time electricity consumption behavior data of the user to be identified, and giving an electricity theft early warning reminder according to the change trend of the corresponding electricity theft risk interval.

[0049] Further, after the step of giving the electricity theft early warning reminder, the method further includes:

[0050] Store and backup the electricity consumption behavior data and electricity theft analysis data, and generate an abnormal analysis report according to the electricity theft warning reminder and the electricity theft analysis data. The electricity theft analysis data includes the frequency domain characteristics and the electricity theft probability.

[0051] In a second aspect, the present invention provides an electricity theft behavior recognition system, which includes:

[0052] A data processing module, configured to collect the electricity consumption behavior data of a user to be recognized, and convert the electricity consumption behavior data into frequency domain components through fast Fourier transform;

[0053] A feature extraction module, configured to perform frequency domain analysis on the frequency domain components according to a preset initial frequency band demarcation point to obtain initial frequency domain characteristics, and adjust the initial frequency band demarcation point according to the initial frequency domain characteristics to obtain a frequency band demarcation point;

[0054] Perform frequency domain analysis on the frequency domain components according to the frequency band demarcation point to obtain frequency domain characteristics, and integrate the frequency domain characteristics into a frequency domain characteristic matrix. The frequency domain characteristics include high-frequency energy, low-frequency energy, and spectral entropy;

[0055] An electricity theft recognition module, configured to input the frequency domain characteristic matrix into a preset feature extraction model to obtain a convolutional feature vector, and perform data processing on the convolutional feature vector through an electricity theft behavior detection algorithm to obtain the electricity theft probability of the user to be recognized.

[0056] The present invention provides an electricity theft behavior recognition method and system. Through the dynamic adjustment mechanism of the frequency band demarcation point, the present invention can improve the accuracy of frequency domain feature extraction. Through a deep convolutional neural network, the present invention can capture deeper electricity consumption characteristics in the frequency domain features, and calculate the electricity theft probability based on the eigenvalues after nonlinear processing, which can improve the accuracy and efficiency of electricity theft behavior recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a schematic flowchart of the electricity theft behavior recognition method in an embodiment of the present invention;

[0058] Figure 2 is a schematic structural diagram of the electricity theft behavior recognition system in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0060] Please refer to Figure 1 , a method for identifying electricity theft behavior proposed in the first embodiment of the present invention, which includes steps S10 to S40:

[0061] Step S10, collect the electricity consumption behavior data of the user to be identified, and convert the electricity consumption behavior data into frequency domain components through fast Fourier transform;

[0062] Step S20, perform frequency domain analysis on the frequency domain components according to the preset initial frequency band demarcation point to obtain initial frequency domain features, and adjust the initial frequency band demarcation point according to the initial frequency domain features to obtain a frequency band demarcation point;

[0063] Step S30, perform frequency domain analysis on the frequency domain components according to the frequency band demarcation point to obtain frequency domain features, and integrate the frequency domain features into a frequency domain feature matrix, where the frequency domain features include high-frequency energy, low-frequency energy, and spectral entropy;

[0064] Step S40, input the frequency domain feature matrix into a preset feature extraction model to obtain a convolutional feature vector, and perform data processing on the convolutional feature vector through an electricity theft behavior detection algorithm to obtain the electricity theft probability of the user to be identified.

[0065] The present invention analyzes the electricity consumption behavior data of users to determine whether there is electricity theft behavior. Specifically, first, the electricity consumption behavior data of each user is collected. Data collection can be carried out by installing smart meters at the power entrances of each user and setting the collection frequency and collection period of the smart meters. The electricity consumption behavior data includes current, voltage, and power at different time points within the collection period. By installing smart meters at the power entrances of each user and setting reasonable collection frequencies and collection periods, the electricity consumption behavior data of users can be accurately collected, including the changes in current, voltage, and power at different time points. This refined data collection method can not only monitor the electricity consumption of users in real time but also provide high-precision basic data for subsequent anomaly detection and electricity theft behavior identification, significantly improving the system's analysis ability for complex electricity consumption patterns and the identification accuracy of abnormal behaviors.

[0066] Then, preprocess the collected electricity consumption behavior data. Here, the preprocessing includes using a statistical-based anomaly detection algorithm to identify and remove noise and incorrect electricity consumption behavior data. For the preprocessed electricity consumption behavior data, sort it according to the timestamp. Preferably, the Z-Score can be used to detect electricity consumption behavior data that significantly deviates from the normal range, so as to remove abnormal and greatly fluctuating noise data, use the linear interpolation method to complete the missing electricity consumption behavior data, and perform standardization processing on the completed electricity consumption behavior data. Of course, other data preprocessing methods can also be used to remove noise and complete missing values for the electricity consumption behavior data, and no excessive restrictions are imposed here.

[0067] In this embodiment, ensuring the integrity and consistency of the data through data preprocessing can improve the accuracy and reliability of subsequent analysis and model prediction. Sorting the data according to the timestamp further guarantees the continuity and effectiveness of time series analysis, which helps to capture the electricity consumption behavior patterns of users more accurately.

[0068] For the preprocessed electricity consumption behavior data, the present invention uses the frequency domain analysis method to extract the frequency domain features of the electricity consumption behavior data. Specifically, first, convert the electricity consumption behavior data into frequency domain components through the fast Fourier transform. The fast Fourier transform is a conventional signal processing tool used to obtain frequency information from the time domain. For the obtained frequency domain components, based on the preset initial frequency band demarcation points, extract the initial frequency domain features of the frequency domain components, and analyze the initial frequency domain features. According to the analysis results, dynamically adjust the initial frequency band demarcation points, and re-extract the frequency domain features according to the adjusted frequency band demarcation points, so as to improve the accuracy of feature extraction. The specific adjustment steps include:

[0069] According to the preset initial frequency band demarcation points, extract the high-frequency energy and low-frequency energy from the frequency domain components, and calculate the energy ratio between the high-frequency energy and the low-frequency energy;

[0070] Calculate the spectral entropy according to the power spectral density of each frequency domain component;

[0071] Compare the energy ratio with the energy ratio threshold, and compare the spectral entropy with the spectral entropy threshold. According to the comparison results, adjust the initial frequency band demarcation points to obtain the frequency band demarcation points.

[0072] In this embodiment, first, according to the obtained frequency domain components and the set initial frequency band demarcation points, calculate the high-frequency energy and low-frequency energy. The formulas are as follows:

[0073]

[0074]

[0075] Among them, Represents the initial frequency band demarcation point, which is used to divide high-frequency energy and low-frequency energy. The frequency band demarcation point is adaptively adjusted according to user behavior changes. Represents the maximum frequency of the current spectrum analysis. Represents the minimum frequency of the current spectrum analysis. Represents the y-th frequency component after fast Fourier transform. Represents the power spectral density of the y-th frequency component, that is, the power spectral density of the frequency component is represented by taking the modulus square of the frequency component. Represents the increment of each frequency band in the integral calculation.

[0076] According to the obtained high-frequency energy and low-frequency energy, calculate the energy ratio of high-frequency energy and low-frequency energy. The formula is:

[0077]

[0078] Then calculate the total energy of the entire frequency domain between the minimum frequency and the maximum frequency in the current spectrum analysis, that is, the sum of the power spectral densities of all frequency domain components. Normalize the power spectral density of each frequency domain component, and calculate the spectral entropy of each power spectral density. The formula is:

[0079]

[0080] In the formula, Represents the spectral entropy of the power spectral density of the y-th frequency component. Is the normalized power spectral density of the y-th frequency component.

[0081] The spectral entropy, as an index of signal complexity, can help determine whether the user's electricity consumption behavior is complex spike behavior or stable base load behavior. Then, based on the energy ratio of high-frequency energy and low-frequency energy and the spectral entropy of each power spectral density, by setting the energy ratio change threshold and spectral entropy threshold, adjust the frequency band demarcation point, and dynamically adjust the frequency band ranges of high frequency and low frequency. The specific adjustment steps include:

[0082] Judge whether the energy ratio is greater than the energy ratio threshold. If it is greater than the energy ratio threshold, then judge whether the spectral entropy is greater than the spectral entropy threshold. If it is greater than or equal to the spectral entropy threshold, then reduce the initial frequency band demarcation point. If it is less than the spectral entropy threshold, then keep the initial frequency band demarcation point unchanged;

[0083] If it is less than or equal to the energy ratio threshold, then judge whether the spectral entropy is less than or equal to the spectral entropy threshold. If it is less than or equal to the spectral entropy threshold, then increase the initial frequency band demarcation point. If it is greater than the spectral entropy threshold, then keep the initial frequency band demarcation point unchanged.

[0084] In this embodiment, first, the energy ratio is compared with the energy ratio threshold. When the energy ratio is greater than the energy ratio change threshold, it indicates that the energy proportion of the high-frequency part is too high, and there may be abnormal spike behavior. It is necessary to expand the analysis of the high frequency. Then, it is judged whether the spectral entropy of each power spectral density is greater than or equal to the spectral entropy threshold. If the spectral entropy is greater than or equal to the spectral entropy threshold, the value of the initial frequency band demarcation point is decreased; otherwise, the initial frequency band demarcation point remains unchanged. When the energy ratio is less than or equal to the energy ratio change threshold, it indicates that the low-frequency part is dominant, representing normal daily power consumption behavior. It is necessary to expand the analysis of the low frequency. Then, it is judged whether the spectral entropy of each power spectral density is less than or equal to the spectral entropy threshold. If the spectral entropy is less than or equal to the spectral entropy threshold, the value of the initial frequency band demarcation point is increased; otherwise, the initial frequency band demarcation point remains unchanged.

[0085] Based on the dynamically adjusted frequency band demarcation point, the frequency domain features of the frequency domain components are re-extracted, that is, the high-frequency energy, low-frequency energy, and spectral entropy, and the extracted frequency domain features are integrated into a frequency domain feature matrix F, and its expression is:

[0086]

[0087] where respectively represent the high-frequency energy, low-frequency energy, and spectral entropy of the user at time point t.

[0088] The high-frequency energy can reflect the user's spike power consumption behavior. For example, the startup of high-power equipment or power theft behavior is usually manifested as an increase in the energy of high-frequency signals;

[0089] The low-frequency energy generally corresponds to the user's daily basic load power consumption behavior. If the proportion of low-frequency energy relative to high-frequency energy decreases significantly, it may be a signal of abnormal behavior;

[0090] The spectral entropy is used to reflect the complexity of the signal. A high spectral entropy means that the signal distribution is relatively complex and may contain spike behavior and irregular loads, while a low spectral entropy indicates that the signal is relatively stable and usually represents the basic load.

[0091] In this embodiment, by dynamically adjusting the frequency band demarcation point, the accuracy and flexibility of electricity theft identification can be improved. Different from the fixed frequency band division scheme in the prior art, in this embodiment, by real-time monitoring the user's electricity consumption behavior and combining the change trends of high-frequency energy, low-frequency energy, high-low frequency energy ratio, and spectral entropy, the demarcation point between high frequency and low frequency is dynamically adjusted. When the high-frequency energy ratio exceeds the threshold, the frequency band demarcation point is automatically lowered to expand the analysis of the high-frequency part and capture peak loads and electricity theft behaviors. When the low-frequency energy dominates, the demarcation point is automatically raised to focus on daily basic load behaviors and reduce the excessive calculation of the high-frequency part, thus achieving the saving of computing resources. The introduction of spectral entropy realizes the judgment of the complexity and stability of the signal. High spectral entropy usually reflects complex peak behaviors, while low spectral entropy corresponds to stable basic loads. This adaptive adjustment mechanism ensures that the method provided by the present invention can flexibly adapt to the electricity consumption patterns of different users. Especially in the case of complex or rapidly changing electricity consumption behaviors, the dynamic adjustment mechanism can significantly improve the accuracy of feature extraction, effectively reduce misjudgments and missed judgments, and enhance the sensitivity and accuracy of anomaly detection.

[0092] Based on the above frequency domain feature matrix, the present invention performs convolutional feature extraction on the frequency domain feature matrix through a feature extraction model constructed by a deep convolutional neural network. Specifically, the feature extraction model consists of three sequentially connected convolutional layers. Preferably, each layer uses a convolutional kernel of 3×3, a stride of 1, and a padding method of "same", and the number of convolutional kernels in each layer is set to 32. The local features in the frequency domain feature matrix F are extracted through the first convolutional kernel, the output of the first convolution is input to the second convolutional kernel to extract deep features, and the output of the second convolution is input to the third convolutional layer. At the same time, residual connections are added between each layer of convolution. The input frequency domain feature matrix F is added to the output of the first convolution to obtain the first residual output. The first residual output is added to the output of the second convolution to obtain the second residual output, and the second residual output is added to the output of the third convolution to obtain the third residual output. After three convolutional operations and residual connections, the third residual output is subjected to global average pooling to convert the three-dimensional matrix into a one-dimensional feature vector, and the one-dimensional feature vector is used as the final output convolutional feature vector.

[0093] For the extracted convolutional feature vector, in a preferred embodiment, the present invention provides an electricity theft behavior detection algorithm to process the convolutional feature vector, thereby obtaining the electricity theft probability of this user. The specific processing steps include:

[0094] Perform non-linear feature processing on each eigenvalue in the convolutional feature vector, and sum the processed eigenvalues to obtain a non-linear influence value;

[0095] Calculate the interaction value of each feature pair in the convolutional feature vector according to the processed eigenvalue, and sum the interaction values to obtain the non - linear interaction value;

[0096] Calculate the electricity theft probability of the user to be identified according to the non - linear influence value and the non - linear interaction value.

[0097] In this embodiment, non - linear feature processing is performed on each eigenvalue in the convolutional feature vector. The non - linear feature processing includes logarithmic smoothing processing, exponential decay non - linear processing, and polynomial combination. The eigenvalue after non - linear feature processing can be expressed as:

[0098]

[0099] where, θ(x i ) represents the i - th eigenvalue in the convolutional feature vector after passing through the non - linear feature function, x i represents the i - th eigenvalue in the convolutional feature vector, and q represents the decay factor.

[0100] In the non - linear feature processing of this embodiment, represents the natural logarithm transformation of the eigenvalue, which is used for smoothing processing to reduce the influence of extreme values on subsequent calculations. 0.1 is the smoothing coefficient to prevent the denominator from being zero and ensure calculation stability. represents the exponential decay term. By adjusting the decay factor, the small differences in features are enhanced to avoid the influence of too large or too small eigenvalue on the calculation. represents the square term of the eigenvalue, which is used as a non - linear enhancement term to improve the sensitivity of the feature to small fluctuations. Through the combination of the above three non - linear feature processing methods, the feature identification ability is effectively improved in terms of smoothing processing, sensitivity to small values, and non - linear enhancement, so that each feature is better represented at different scales, thus providing high - quality input for subsequent feature interaction.

[0101] Sum all the eigenvalues after non - linear feature processing and add the bias constant c to obtain the non - linear influence value A of all eigenvalues. The formula is:

[0102]

[0103] In the formula, n represents the total number of eigenvalues in the convolutional feature vector, and c represents the bias constant, which is used to adjust the output value range of the numerator part.

[0104] Based on the eigenvalues after non - linear feature processing, calculate the interaction value of each feature pair in the convolutional feature vector. The formula is:

[0105]

[0106] In the formula, represents the interaction value between the $i$-th eigenvalue and the $j$-th eigenvalue in the convolutional feature vector, represents the product of two non-linearly processed eigenvalues, which is used to introduce the linear combination of the two features, represents the absolute difference between two eigenvalues, which is used to measure the similarity between the two features, is used as a smoothing denominator to prevent the calculation from being unstable due to too small a difference value.

[0107] The above interaction formula enables the association of each pair of features to be captured more meticulously. This design allows the interaction value to not be limited to simple multiplication or addition, but rather to reflect the non-linear and differential relationships between feature pairs, making it particularly suitable for detecting multi-dimensional electricity theft behaviors.

[0108] By summing up the interaction values of all feature pairs, the non-linear interaction value $B$ between all feature pairs is calculated. Its formula is:

[0109]

[0110] where, represents a regulation parameter, which is used to control the weight of the interaction value as the denominator in the overall calculation during the subsequent electricity theft probability calculation. $n$ represents the total number of eigenvalues in the convolutional feature vector.

[0111] After obtaining the non-linear influence value and the non-linear interaction value through the above steps, the electricity theft probability $P$ of this user can be calculated based on these two values. Its formula is:

[0112]

[0113] In the formula, $P$ represents the electricity theft probability of the user, represents the Sigmoid function, which is used to convert the numerical value into the range of $[0,1]$.

[0114] A method for identifying electricity theft behavior proposed by the present invention, during the process of identifying electricity theft behavior, performs multi-layer convolution feature extraction on the frequency domain feature matrix through a deep convolutional neural network. The three convolutional layers are set with the same convolutional kernel size and stride, ensuring the coherence and consistency of feature extraction. The residual connections between each convolutional layer further enhance the learning ability of the model. By retaining the input feature information and gradually enhancing the depth of feature extraction, the loss of information during the convolution operation is minimized. This network structure design improves the feature retention and transmission ability of the model, enabling more detailed and deep-level electricity consumption features in the user's electricity consumption behavior to be captured during the detection of electricity theft behavior, especially being more sensitive to small abnormal changes in the data. After generating the convolutional feature vector, each eigenvalue is further optimized through three non-linear processing methods: logarithmic smoothing, exponential decay, and polynomial combination. This makes the features not only have the advantage of extreme value smoothing processing, reducing the impact of data extreme values on the results, but also being more sensitive to subtle differences through exponential decay, enabling small abnormal electricity consumption features to be effectively identified. The polynomial enhancement term further improves the ability to identify feature fluctuations, making the expression ability of each feature stronger at different scales, which is beneficial for identifying abnormal electricity consumption behavior in a complex electricity usage environment. The feature pair interaction value constructed based on the convolutional feature vector can carefully capture the non-linear and differential relationships between eigenvalues, avoiding the deficiencies of simple weighting or multiplication methods. Through the composite calculation in the interaction formula, the complex correlation relationships between feature pairs are more comprehensively expressed, enabling weak correlations between features to be identified and reflected. This multi-level interaction design of the present invention ensures the ability to handle multi-dimensional feature dependencies that may exist in electricity theft behavior, thereby further improving the accuracy of electricity theft identification. By calculating the electricity theft probability through a comprehensive electricity theft behavior detection algorithm and converting it to the range of [0, 1], the detection result can intuitively present the risk level of the user, facilitating subsequent real-time response and hierarchical processing, providing strong technical support for realizing efficient and accurate identification and processing of electricity theft behavior.

[0115] In a preferred embodiment, after obtaining the electricity theft probability of the user to be identified, the present invention divides the risk interval according to the electricity theft probability, dynamically adjusts the electricity theft probability of the user, and issues a warning according to the risk interval. The specific steps include:

[0116] Determine the electricity theft risk interval of the user to be identified according to the electricity theft probability and the preset electricity theft probability threshold;

[0117] Dynamically adjust the electricity theft probability according to the real-time electricity consumption behavior data of the user to be identified, and issue an electricity theft warning reminder according to the change trend of the corresponding electricity theft risk interval.

[0118] In this embodiment, first, according to the calculated electricity theft probability P' of each user and the preset electricity theft probability thresholds R1 and R2, the electricity theft risk interval of the user is determined:

[0119] If , it is determined that the electricity theft probability of the user is in the low-risk interval;

[0120] If , it is determined that the electricity theft probability of the user is in the medium-risk interval;

[0121] If , it is determined that the electricity theft probability of the user is in the high-risk interval.

[0122] Then, according to the real-time electricity consumption behavior data, the electricity theft probability of the user is continuously adjusted. When the risk interval changes due to the change of the electricity theft probability, corresponding measures are taken. Preferably, if the electricity theft probability of the user remains in the low-risk interval, no warning measures are taken and normal monitoring is maintained. If the electricity theft probability of the user rises to the medium-risk interval, a first-level warning is triggered, the change of the user's electricity consumption behavior is automatically recorded, and a reminder notice is sent to the user.

[0123] If the electricity theft probability of the user remains in the medium-risk interval, the change of the user's electricity consumption behavior is automatically recorded, and a notice is sent to the user to remind that there is a certain risk in the user's behavior and to prevent it from escalating to a high risk. If the electricity theft probability of the user drops to the low-risk interval, the first-level warning is lifted and monitoring is continued. If the electricity theft probability of the user rises to the high-risk interval, a second-level warning is triggered, and on-site inspections are carried out by power management personnel, and automated restriction measures are taken, including restricting the user's electricity consumption power.

[0124] If the electricity theft probability of the user remains in the high-risk interval, the second-level warning state is maintained, and additional coercive measures are added, including power-off and reporting to a higher-level regulatory department. If the electricity theft probability of the user drops to the medium-risk interval, the warning is downgraded to a first-level warning, the change of the user's electricity consumption behavior is automatically recorded, and a reminder notice is sent to the user.

[0125] In this embodiment, by dynamically adjusting the electricity theft probability of the user and dividing the risk interval, precise monitoring and timely warning of electricity theft behavior are realized. According to the change of the user's electricity theft probability, hierarchical warnings and corresponding measures are taken, effectively reducing the risk of electricity theft behavior. Normal monitoring is maintained in the low-risk interval to avoid unnecessary interference. Reminders and records are made in the medium-risk interval to prevent risk escalation. Compulsory measures such as on-site inspections and power-off are taken in the high-risk interval to ensure power safety, and at the same time, the situation is reported to a higher-level regulatory department, improving the efficiency and safety of power management.

[0126] In another preferred embodiment, the present invention stores and backs up all data and generates an abnormal warning analysis report. The specific steps include:

[0127] Store and back up the electricity consumption behavior data and electricity theft analysis data, and generate an abnormal analysis report according to the electricity theft warning reminder and the electricity theft analysis data. The electricity theft analysis data includes the frequency domain characteristics and the electricity theft probability.

[0128] In this embodiment, a distributed database system is adopted, and a compression algorithm is used to compress and store all data. Here, the data includes the electricity consumption behavior data of users and the electricity theft analysis data generated in the electricity theft identification and analysis, such as data like frequency domain characteristics and electricity theft probability. At the same time, the database is backed up regularly, and an automatic feedback mechanism is established. After detecting the abnormal electricity theft behavior of users and triggering a warning, an abnormal analysis report is automatically generated. The report content includes the abnormal occurrence time, user ID, detected characteristic values, and the abnormal analysis report is sent to relevant management personnel.

[0129] In this embodiment, by adopting a distributed database system and a compression algorithm, all data is efficiently compressed and stored and backed up regularly to ensure the security and scalability of the data. The established automatic feedback mechanism can generate an abnormal analysis report containing detailed information in a timely manner after detecting the abnormal electricity theft behavior of users and triggering a warning, and automatically send it to relevant management personnel. This process not only improves the efficiency of data management, ensures the stability and reliability of the power system operation, but also realizes rapid response and precise analysis, enhances the monitoring and management capabilities of the power system, and helps to quickly take measures to prevent further losses.

[0130] Please refer to Figure 2 , based on the same inventive concept, a electricity theft behavior identification system proposed in the second embodiment of the present invention includes:

[0131] A data processing module 10, configured to collect the electricity consumption behavior data of the user to be identified, and convert the electricity consumption behavior data into frequency domain components through fast Fourier transform;

[0132] A feature extraction module 20, configured to perform frequency domain analysis on the frequency domain components according to a preset initial frequency band demarcation point to obtain initial frequency domain characteristics, and adjust the initial frequency band demarcation point according to the initial frequency domain characteristics to obtain a frequency band demarcation point;

[0133] Perform frequency domain analysis on the frequency domain components according to the frequency band demarcation point to obtain frequency domain characteristics, and integrate the frequency domain characteristics into a frequency domain characteristic matrix. The frequency domain characteristics include high-frequency energy, low-frequency energy, and spectral entropy;

[0134] The electricity theft identification module 30 is configured to input the frequency-domain feature matrix into a preset feature extraction model to obtain a convolutional feature vector, and perform data processing on the convolutional feature vector through an electricity theft behavior detection algorithm to obtain the electricity theft probability of the user to be identified.

[0135] The technical features and technical effects of the electricity theft behavior identification system proposed in the embodiments of the present invention are the same as those of the method proposed in the embodiments of the present invention, and will not be described in detail here. Each module in the above-mentioned electricity theft behavior identification system can be implemented in whole or in part by software, hardware, and their combinations. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0136] In summary, an electricity theft behavior identification method and system proposed in the embodiments of the present invention, the method collects the electricity consumption behavior data of the user to be identified, and converts the electricity consumption behavior data into frequency-domain components through fast Fourier transform; according to a preset initial frequency band boundary point, perform frequency-domain analysis on the frequency-domain components to obtain initial frequency-domain features, and adjust the initial frequency band boundary point according to the initial frequency-domain features to obtain a frequency band boundary point; according to the frequency band boundary point, perform frequency-domain analysis on the frequency-domain components to obtain frequency-domain features, and integrate the frequency-domain features into a frequency-domain feature matrix, the frequency-domain features include high-frequency energy, low-frequency energy, and spectral entropy; input the frequency-domain feature matrix into a preset feature extraction model to obtain a convolutional feature vector, and perform data processing on the convolutional feature vector through an electricity theft behavior detection algorithm to obtain the electricity theft probability of the user to be identified. The present invention improves the accuracy of frequency-domain feature extraction through a dynamic adjustment mechanism of the frequency band boundary point, extracts deeper electricity consumption features in the frequency-domain features through a deep convolutional neural network, and calculates the electricity theft probability based on the non-linearly processed feature values, improving the accuracy and efficiency of electricity theft behavior identification.

[0137] Each embodiment in this specification is described in a progressive manner. For the parts that are the same or similar in each embodiment, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0138] The above-described embodiments merely represent several preferred embodiments of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and substitutions can be made, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the protection scope of the claims.

Claims

1. A method for identifying electricity theft behavior, characterized in that, Including: Collecting the electricity consumption behavior data of the user to be identified, and converting the electricity consumption behavior data into frequency domain components through fast Fourier transform, where the electricity consumption behavior data includes current, voltage, and power at different time points within the collection period; Performing frequency domain analysis on the frequency domain components according to the preset initial frequency band demarcation point to obtain initial frequency domain features, and adjusting the initial frequency band demarcation point according to the initial frequency domain features to obtain the frequency band demarcation point; Performing frequency domain analysis on the frequency domain components according to the frequency band demarcation point to obtain frequency domain features, and integrating the frequency domain features into a frequency domain feature matrix, where the frequency domain features include high-frequency energy, low-frequency energy, and spectral entropy; Inputting the frequency domain feature matrix into a preset feature extraction model to obtain a convolutional feature vector, and performing data processing on the convolutional feature vector through an electricity theft behavior detection algorithm to obtain the electricity theft probability of the user to be identified; Among them, the step of performing frequency domain analysis on the frequency domain components according to the preset initial frequency band demarcation point to obtain initial frequency domain features, and adjusting the initial frequency band demarcation point according to the initial frequency domain features to obtain the frequency band demarcation point includes: Extracting high-frequency energy and low-frequency energy from the frequency domain components according to the preset initial frequency band demarcation point, and calculating the energy ratio between the high-frequency energy and the low-frequency energy; Calculating the spectral entropy according to the power spectral density of each frequency domain component; Comparing the energy ratio with the energy ratio threshold, and comparing the spectral entropy with the spectral entropy threshold, and adjusting the initial frequency band demarcation point according to the comparison results to obtain the frequency band demarcation point; The step of performing data processing on the convolutional feature vector through an electricity theft behavior detection algorithm to obtain the electricity theft probability of the user to be identified includes: Performing non-linear feature processing on each eigenvalue in the convolutional feature vector, and summing the processed eigenvalues to obtain a non-linear influence value; Calculating the interaction value of each feature pair in the convolutional feature vector according to the processed eigenvalues, and summing the interaction values to obtain a non-linear interaction value; Calculating the electricity theft probability of the user to be identified according to the non-linear influence value and the non-linear interaction value; The processed eigenvalue is represented by the following formula: where x i represents the i-th eigenvalue in the convolutional feature vector, and q represents the decay factor; The non-linear influence value is calculated by the following formula: In the formula, n represents the total number of eigenvalues in the convolutional feature vector, and c represents the bias constant; The interaction value is calculated by the following formula: where x j represents the j-th eigenvalue in the convolutional feature vector; The non-linear interaction value is calculated by the following formula: In the formula, represents the adjustment parameter; The electricity theft probability is calculated by the following formula: In the formula, represents the Sigmoid function.

2. The electricity theft behavior recognition method according to claim 1, characterized in that The step of comparing the energy ratio with the energy ratio threshold, and comparing the spectral entropy with the spectral entropy threshold, and adjusting the initial frequency band demarcation point according to the comparison results to obtain the frequency band demarcation point includes: Judging whether the energy ratio is greater than the energy ratio threshold. If it is greater than the energy ratio threshold, then judging whether the spectral entropy is greater than the spectral entropy threshold. If it is greater than or equal to the spectral entropy threshold, then reducing the initial frequency band demarcation point. If it is less than the spectral entropy threshold, then keeping the initial frequency band demarcation point unchanged; If it is less than or equal to the energy ratio threshold, then determine whether the spectral entropy is less than or equal to the spectral entropy threshold. If it is less than or equal to the spectral entropy threshold, increase the initial frequency band demarcation point. If it is greater than the spectral entropy threshold, keep the initial frequency band demarcation point unchanged.

3. The electricity theft behavior recognition method according to claim 1, characterized in that Calculate the high-frequency energy using the following formula: In the formula, represents the initial frequency band demarcation point, represents the maximum frequency, represents the y-th frequency component, represents the power spectral density of the y-th frequency component; Calculate the low-frequency energy using the following formula: In the formula, represents the minimum frequency; Calculate the spectral entropy using the following formula: In the formula, represents the spectral entropy of the power spectral density of the y-th frequency component, represents the power spectral density of the y-th frequency component after normalization.

4. The electricity theft behavior recognition method according to claim 1, characterized in that The feature extraction model is constructed based on a deep convolutional neural network. The feature extraction model includes three consecutively connected convolutional layers, and residual connections are set between each convolutional layer.

5. The electricity theft behavior recognition method according to claim 1, wherein After the step of obtaining the electricity theft probability of the user to be identified, it further includes: Determine the electricity theft risk interval of the user to be identified according to the electricity theft probability and a preset electricity theft probability threshold; Dynamically adjust the electricity theft probability according to the real-time electricity consumption behavior data of the user to be identified, and give an electricity theft warning reminder according to the change trend of the corresponding electricity theft risk interval.

6. The electricity theft behavior identification method according to claim 5, wherein After the step of giving the electricity theft warning reminder, it further includes: Store and back up the electricity consumption behavior data and the electricity theft analysis data, and generate an abnormal analysis report according to the electricity theft warning reminder and the electricity theft analysis data. The electricity theft analysis data includes the frequency domain features and the electricity theft probability.

7. A power theft behavior recognition system, characterized in that, It includes: A data processing module, configured to collect the electricity consumption behavior data of the user to be identified, and convert the electricity consumption behavior data into frequency domain components through fast Fourier transform. The electricity consumption behavior data includes current, voltage, and power at different time points within the collection period; A feature extraction module, configured to perform frequency domain analysis on the frequency domain components according to a preset initial frequency band demarcation point to obtain initial frequency domain features, and adjust the initial frequency band demarcation point according to the initial frequency domain features to obtain a frequency band demarcation point, including: Extract high-frequency energy and low-frequency energy from the frequency domain components according to a preset initial frequency band demarcation point, and calculate the energy ratio between the high-frequency energy and the low-frequency energy; Calculate the spectral entropy according to the power spectral density of each frequency domain component; Compare the energy ratio with the energy ratio threshold, and compare the spectral entropy with the spectral entropy threshold. According to the comparison results, adjust the initial frequency band demarcation point to obtain a frequency band demarcation point; Perform frequency domain analysis on the frequency domain components according to the frequency band demarcation point to obtain frequency domain features, and integrate the frequency domain features into a frequency domain feature matrix. The frequency domain features include high-frequency energy, low-frequency energy, and spectral entropy; An electricity theft identification module, configured to input the frequency domain feature matrix into a preset feature extraction model to obtain a convolutional feature vector, and perform data processing on the convolutional feature vector through an electricity theft behavior detection algorithm to obtain the electricity theft probability of the user to be identified, including: Perform non-linear feature processing on each eigenvalue in the convolutional feature vector, and sum the processed eigenvalues to obtain a non-linear influence value; Calculate the interaction value of each feature pair in the convolutional feature vector according to the processed eigenvalues, and sum the interaction values to obtain a non-linear interaction value; Calculate the electricity theft probability of the user to be identified according to the non-linear influence value and the non-linear interaction value; The processed eigenvalue is represented by the following formula: where x i represents the i-th eigenvalue in the convolutional feature vector, and q represents the decay factor; Calculate the non-linear influence value by the following formula: In the formula, n represents the total number of eigenvalues in the convolutional feature vector, and c represents the bias constant; Calculate the interaction value by the following formula: where x j represents the j-th eigenvalue in the convolutional feature vector; Calculate the non-linear interaction value by the following formula: In the formula, represents the adjustment parameter; Calculate the electricity theft probability by the following formula: In the formula, represents the Sigmoid function.

Images