Intermittent electricity stealing quantitative evaluation method based on double-model reconstruction error comparison

Abstract

The embodiment of the invention relates to an intermittent electricity larceny quantitative evaluation method based on dual-model reconstruction error comparison, and the method comprises the steps: obtaining the power utilization sequence data of multiple users, and constructing standard power utilization sequence data according to the power utilization sequence data of the multiple users; performing mask processing on the standard power utilization sequence data; training a normal user model and an electricity stealing user model by using the training set after mask processing; inputting the test set and the verification set after mask processing into the trained normal user model and the electricity stealing user model for reconstruction; according to the first normal reconstruction sequence and the first electricity larceny reconstruction sequence, determining an electricity larceny judgment threshold range based on the user tag of each user in the test set; and according to the second normal reconstruction sequence and the second electricity larceny reconstruction sequence, determining an electricity larceny suspicion index value of each user in the verification set, and based on the electricity larceny judgment threshold range and the electricity larceny suspicion index value, evaluating the electricity larceny suspicion level of each user to be tested in the verification set. According to the embodiment of the invention, misjudgment of electricity stealing behaviors can be effectively avoided.

Description

Technical Field

[0001] This disclosure relates to the field of electricity theft assessment technology, and in particular to an intermittent electricity theft quantitative assessment method based on dual-model reconstruction error comparison. Background Technology

[0002] Electricity, as an indispensable basic energy source in modern society, is directly related to national welfare, people's livelihood, and economic development through its stable supply and fair metering. Electricity theft, as an illegal means of obtaining electricity, not only causes huge economic losses to power companies but also poses a continuous threat to the safe and stable operation of the power grid. With the large-scale deployment of smart meters and electricity consumption information collection systems, power companies are now able to acquire massive amounts of user electricity consumption time-series data, laying the data foundation for data-driven electricity theft detection technology.

[0003] Current electricity theft detection technologies mainly fall into two categories: hardware detection and software analysis. Hardware detection primarily involves installing dedicated sensors or monitoring devices to directly identify anomalies in electrical parameters such as voltage and current in real time. This method is costly to deploy and has limited coverage. Software analysis, on the other hand, analyzes historical user electricity consumption data and applies algorithms such as machine learning and deep learning to automatically identify abnormal patterns. This approach has become the mainstream of current research.

[0004] However, when faced with the prominent challenge of intermittent electricity theft, existing technical approaches have the following shortcomings in detection efficiency: due to the non-continuous nature of intermittent electricity theft, existing technical approaches are insufficient in their ability to identify intermittent electricity theft, and are prone to false negatives; whether based on traditional feature engineering methods or some deep learning models, it is difficult to achieve high-precision separation between normal electricity consumption patterns and electricity theft behavior patterns when dealing with intermittent electricity theft, which affects the accuracy of electricity theft identification and leads to misjudgments.

[0005] Therefore, it is necessary to provide a new technical solution to improve one or more of the problems existing in the above solutions.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this disclosure is to provide an intermittent electricity theft quantitative assessment method based on dual-model reconstruction error comparison, which can accurately assess the electricity theft suspicion level of each user under test in the verification set and effectively avoid misjudgment of electricity theft behavior.

[0008] According to embodiments of this disclosure, a method for quantitative assessment of intermittent electricity theft based on dual-model reconstruction error comparison is provided, including: Obtain electricity consumption sequence data from multiple users, and construct standard electricity consumption sequence data based on the multi-user electricity consumption sequence data. The standard electricity consumption sequence data includes a training set, a test set, and a validation set. The standard electricity consumption sequence data is masked. Construct a normal user model and an electricity theft user model, and train the normal user model and the electricity theft user model respectively using the masked training set; The masked test set is input into the trained normal user model and the electricity theft user model respectively for reconstruction, to obtain the first normal reconstruction sequence and the first electricity theft reconstruction sequence. The first normal comprehensive fitting index and the first electricity theft comprehensive fitting index are determined based on the first normal reconstruction sequence and the first electricity theft reconstruction sequence, and the distribution of electricity theft suspicion index is determined based on the user tags of each user in the test set. The electricity theft judgment threshold range is determined based on the distribution of electricity theft suspicion index. The masked verification set is input into the trained normal user model and the electricity theft user model respectively for reconstruction, resulting in the second normal reconstruction sequence and the second electricity theft reconstruction sequence. The second normal comprehensive fitting index and the second electricity theft comprehensive fitting index are determined based on the second normal reconstruction sequence and the second electricity theft reconstruction sequence. The second normal comprehensive fitting index and the second electricity theft comprehensive fitting index are used to predict the electricity theft suspicion index value of each user in the verification set. The electricity theft suspicion level of each user in the verification set is evaluated based on the electricity theft judgment threshold range and the electricity theft suspicion index value.

[0009] In an exemplary embodiment of this disclosure, training the normal user model and the electricity theft user model using the masked training set includes: The masked training set is used as the original training sample mask sequence and input into the normal user model and the electricity theft user model respectively for reconstruction to obtain the third normal reconstruction sequence and the third electricity theft reconstruction sequence. The first mean square error loss corresponding to the normal user model is calculated based on the original training sample mask sequence and the third normal reconstruction sequence, and the second mean square error loss corresponding to the electricity theft user model is calculated based on the original training sample mask sequence and the third electricity theft reconstruction sequence. The normal user model and the electricity theft user model are trained by minimizing the first mean square error loss and the second mean square error loss, respectively.

[0010] In an exemplary embodiment of this disclosure, the step of inputting the masked test set into a trained normal user model and an electricity theft user model for reconstruction to obtain a first normal reconstruction sequence and a first electricity theft reconstruction sequence includes: The masked test set is used as the original test sample mask sequence and input into the normal user model and the electricity theft user model, respectively. The normal user model performs linear projection processing on the original test sample mask sequence to obtain the first normal reconstruction sequence; Furthermore, the electricity theft user model performs linear projection processing on the original test sample mask sequence to obtain the first electricity theft reconstruction sequence.

[0011] In an exemplary embodiment of this disclosure, determining the first normal comprehensive fitting index and the first electricity theft comprehensive fitting index based on the first normal reconstruction sequence and the first electricity theft reconstruction sequence includes: Calculate the first normal comprehensive fit index based on the first normal reconstruction sequence and the original test sample mask sequence; Furthermore, a first comprehensive fitting index for electricity theft is calculated based on the first electricity theft reconstruction sequence and the original test sample mask sequence.

[0012] In an exemplary embodiment of this disclosure, the step of determining the distribution of electricity theft suspicion indicators based on user tags of each user in the test set, and determining the electricity theft judgment threshold range based on the distribution of electricity theft suspicion indicators, includes: The electricity theft suspicion index value of each user in the test set is calculated using the first normal comprehensive fitting index and the first electricity theft comprehensive fitting index. The distribution of electricity theft suspicion indicators is determined based on the user tags of each user in the test set and the electricity theft suspicion index values ​​of each user; the tags include normal user tags and electricity theft user tags; The initial threshold is determined using formula (1): (1) in, Indicates the initial threshold. This represents the mean value of the electricity theft suspicion index for normal users in the test set under the threshold parameter θ. This represents the standard deviation of the electricity theft suspicion index values ​​for normal users in the test set under the threshold parameter θ. This represents the mean value of the electricity theft suspicion index for users in the test set, under the threshold parameter θ. J(θ) represents the standard deviation of the electricity theft suspicion index values ​​of electricity theft users in the test set under the threshold parameter θ; J(θ) represents the optimization objective function. Based on the initial threshold, the optimal threshold is adaptively determined using a gradient descent optimization algorithm; The range of the electricity theft judgment threshold is determined based on the optimal threshold and the dispersion of the distribution of the electricity theft suspicion index.

[0013] In an exemplary embodiment of this disclosure, an optimal threshold is adaptively determined using a gradient descent optimization algorithm based on the initial threshold, including: The gradient direction is calculated using the gradient descent optimization algorithm: (2) in, Indicates the gradient direction. This represents a very small disturbance. Indicates at the threshold The effect size calculated below, These represent the threshold values ​​respectively. The effect size was calculated below; The threshold is updated according to the gradient direction using formula (3): (3) in, This represents the threshold for iteration t+1. This represents the threshold after t iterations. Indicates the initial learning rate; Repeatedly calculate the gradient direction and update the threshold until the preset iteration condition or the maximum number of iterations is met, and obtain the optimal threshold.

[0014] In an exemplary embodiment of this disclosure, the step of inputting the masked verification set into the trained normal user model and the electricity theft user model for reconstruction to obtain a second normal reconstruction sequence and a second electricity theft reconstruction sequence includes: The masked verification set is used as the original verification sample mask sequence and input into the normal user model and the electricity theft user model. The normal user model performs linear projection processing on the original verification sample mask sequence to obtain the second normal reconstruction sequence; Furthermore, the electricity theft user model performs linear projection processing on the original verification sample mask sequence to obtain the second electricity theft reconstruction sequence.

[0015] In an exemplary embodiment of this disclosure, the step of determining a second normal comprehensive fitting index and a second electricity theft comprehensive fitting index based on the second normal reconstruction sequence and the second electricity theft reconstruction sequence, predicting the electricity theft suspicion index value of each test user in the verification set using the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index, and evaluating the electricity theft suspicion level of each test user in the verification set based on the electricity theft judgment threshold range and the electricity theft suspicion index value, includes: The second normal comprehensive fitting index is calculated based on the second normal reconstruction sequence and the original verification sample mask sequence, and the second electricity theft comprehensive fitting index is calculated based on the second electricity theft reconstruction sequence and the original verification sample mask sequence; The difference between the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index is processed to obtain the electricity theft suspicion index value of each user in the verification set. The electricity theft suspicion index value of each user in the verification set is compared with the electricity theft judgment threshold range to determine the electricity theft suspicion level of each user in the verification set.

[0016] In an exemplary embodiment of this disclosure, the masking process is a multi-block masking process.

[0017] In an exemplary embodiment of this disclosure, constructing standard electricity consumption sequence data based on the electricity consumption sequence data of the multiple users includes: The electricity consumption sequence data of the multi-users is preprocessed to obtain the standard electricity consumption sequence data.

[0018] The technical solution provided in this disclosure may include the following beneficial effects: In the embodiments of this disclosure, the normal user model focuses on learning the electricity consumption patterns of normal users during training, while the electricity theft user model focuses on learning typical electricity theft behavior characteristics during training. Using these two trained normal user models and electricity theft user models, the reconstructed sequence obtained from the masked test set is used to calculate the first normal comprehensive fitting index and the first electricity theft comprehensive fitting index. Combined with the user labels corresponding to each user in the test set, the electricity theft judgment threshold range is determined, which can ensure that the obtained electricity theft judgment threshold range has high reliability and accurately reflects the judgment criteria for intermittent electricity theft behavior. Furthermore, using these two trained normal user models and electricity theft user models, the masked validation set is reconstructed, and the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index are calculated based on the reconstruction sequence of these two models. By comprehensively evaluating the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index, the electricity theft suspicion index value of each test user in the validation set is generated. Using the quantified electricity theft judgment threshold range and electricity theft suspicion index value that can effectively identify intermittent electricity theft behavior, the electricity theft suspicion level of each test user in the validation set is accurately assessed, effectively avoiding misjudgment of electricity theft behavior.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0021] Figure 1 This diagram illustrates the steps of an intermittent quantitative assessment method for electricity theft based on dual-model reconstruction error comparison in an exemplary embodiment of this disclosure. Figure 2 This diagram illustrates the distribution of data tokens at different points in time in an exemplary embodiment of this disclosure. Detailed Implementation

[0022] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0023] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0024] This example implementation provides an intermittent quantitative assessment method for electricity theft based on dual-model reconstruction error comparison, referencing... Figure 1 As shown, the method may include steps S101 to S107.

[0025] Step S101: Obtain electricity consumption sequence data from multiple users, and construct standard electricity consumption sequence data based on the electricity consumption sequence data from multiple users. The standard electricity consumption sequence data includes a training set, a test set, and a validation set.

[0026] Step S102: Mask the standard power consumption sequence data.

[0027] Step S103: Construct a normal user model and an electricity theft user model, and train the normal user model and the electricity theft user model respectively using the masked training set.

[0028] Step S104: Input the masked test set into the trained normal user model and the electricity theft user model respectively for reconstruction to obtain the first normal reconstruction sequence and the first electricity theft reconstruction sequence.

[0029] Step S105: Determine the first normal comprehensive fitting index and the first electricity theft comprehensive fitting index based on the first normal reconstruction sequence and the first electricity theft reconstruction sequence, and determine the distribution of electricity theft suspicion index based on the user tags of each user in the test set, and determine the electricity theft judgment threshold range based on the distribution of electricity theft suspicion index.

[0030] Step S106: Input the masked verification set into the trained normal user model and the electricity theft user model respectively for reconstruction to obtain the second normal reconstruction sequence and the second electricity theft reconstruction sequence.

[0031] Step S107: Determine the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index based on the second normal reconstruction sequence and the second electricity theft reconstruction sequence. Use the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index to predict the electricity theft suspicion index value of each user to be tested in the verification set. Evaluate the electricity theft suspicion level of each user to be tested in the verification set based on the electricity theft judgment threshold range and the electricity theft suspicion index value.

[0032] In the embodiments of this disclosure, the normal user model focuses on learning the electricity consumption patterns of normal users during training, while the electricity theft user model focuses on learning typical electricity theft behavior characteristics during training. Using these two trained normal user models and electricity theft user models, the reconstructed sequence obtained from the masked test set is used to calculate the first normal comprehensive fitting index and the first electricity theft comprehensive fitting index. Combined with the user labels corresponding to each user in the test set, the electricity theft judgment threshold range is determined, which can ensure that the obtained electricity theft judgment threshold range has high reliability and accurately reflects the judgment criteria for intermittent electricity theft behavior. Furthermore, using these two trained normal user models and electricity theft user models, the masked validation set is reconstructed, and the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index are calculated based on the reconstruction sequence of these two models. By comprehensively evaluating the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index, the electricity theft suspicion index value of each test user in the validation set is generated. Using the quantified electricity theft judgment threshold range and electricity theft suspicion index value that can effectively identify intermittent electricity theft behavior, the electricity theft suspicion level of each test user in the validation set is accurately assessed, effectively avoiding misjudgment of electricity theft behavior.

[0033] The steps of the method described above in this example implementation will now be explained in more detail.

[0034] In one embodiment, in step S101, constructing standard electricity consumption sequence data based on the multi-user electricity consumption sequence data includes: The electricity consumption sequence data of the multi-users is preprocessed to obtain the standard electricity consumption sequence data.

[0035] It should be noted that the electricity consumption sequence data of multiple users can be obtained from the power company's marketing system and electricity consumption information collection system. The obtained multi-user electricity consumption sequence data includes electricity consumption sequence data of normal users and electricity consumption sequence data of users who steal electricity, with a sampling time interval of 15 minutes.

[0036] It should be noted that the standard electricity consumption sequence data is divided into a training set, a test set, and a validation set. The training set, test set, and validation set respectively contain electricity consumption sequence data from normal users and electricity consumption sequence data from users who steal electricity. Furthermore, the electricity consumption sequence data in the training set and test set are labeled with user tags. For example, the electricity consumption sequence data from normal users in the training set is labeled with a "normal user" tag, and the electricity consumption sequence data from users who steal electricity in the training set is labeled with a "electricity theft user" tag; similarly, the electricity consumption sequence data from normal users in the test set is labeled with a "normal user" tag, and the electricity consumption sequence data from users who steal electricity in the test set is labeled with a "electricity theft user" tag.

[0037] Furthermore, in step S101, the electricity consumption sequence data of the multiple users is preprocessed, including the following steps: The electricity consumption sequence data of multiple users is cleaned, and the cleaned multi-user electricity consumption sequence data is processed for missing values, outlier removal, and normalization to obtain standard electricity consumption sequence data.

[0038] Optionally, in step S101, electricity consumption sequences with consecutive missing times exceeding a preset threshold are removed, wherein the preset threshold is 7 days.

[0039] Optionally, in step S101, for electricity consumption sequence data missing individual data points, a mean-filling method based on local context is used to fill in the missing data points. For example, the average value is calculated using the measurements at the same time point for three days before and after the missing individual data point, and this average value is used to fill in the missing individual data points, so as to maintain the temporal continuity and waveform integrity of the electricity consumption sequence data to the greatest extent.

[0040] Optionally, in step S101, the three sigma criterion in statistics is used to identify outliers in the electricity consumption series data. The calculation method of the three sigma criterion is as follows: calculate the mean of the electricity consumption series. and standard deviation Based on the mean with standard deviation Set the preset range as [ -3 , +3 The method determines whether a data point in the electricity consumption sequence data falls outside a preset range: if a data point falls outside the preset range, it is considered an outlier and is removed. This method can effectively identify abnormal data caused by metering equipment failure or communication interference, effectively filter outliers, and improve the quality of electricity consumption sequence data.

[0041] Optionally, in step S101, the electricity consumption sequence data after missing value processing and outlier removal processing is normalized using the MAX-MIN scaling method, and the specific formula is as follows: (4) in, This represents the normalized electricity consumption sequence data for each user. This represents the i-th load data point in the user's electricity consumption sequence data. This represents the user's electricity consumption sequence data. It contains multiple load data points. This represents the minimum load data point in the user's electricity consumption sequence data x. This represents the largest load data point in the user's electricity consumption sequence data x.

[0042] It should be noted that the standard electricity consumption sequence data includes standard electricity consumption sequence data of C users, and the standard electricity consumption sequence data can be regarded as a multivariate time series; wherein, the standard electricity consumption sequence data corresponding to each user is a one-dimensional time series of length T.

[0043] In one embodiment, in step S102, the standard electricity consumption sequence data is masked using a block masking method, as follows: The standard electricity consumption sequence data for each user is divided into N patches. Each patch has a fixed length of P; where, P is a hyperparameter, and the sequence fragment corresponding to user c for the i-th patch is defined as... The patch here covers consecutive time points, which helps the model learn local temporal dependencies.

[0044] Each patch for each user is mapped to a hidden dimension D using a linear layer formula, and an absolute position embedding is added to the patch to preserve the temporal order, thus constructing a set of electricity consumption sequence fragments for each user. The initial token embedding formula is defined below: (5) in, Indicates the initial token embedding. Represents the linear projection matrix. Indicates a patch of the original data. Indicates absolute position embedding, It is position embedding. It represents a set of vectors or points consisting of D real numbers.

[0045] This block-based approach reduces sequence length, improves computational efficiency, preserves temporal locality, and ensures the model is aware of the patch's position within the sequence. It's important to note that the absolute position embedding described above is a component used to provide temporal order information, in contrast to the linear projection representing the data content. Together, they generate the complete initial token embedding.

[0046] In addition, for each user's electricity consumption sequence fragment set, in order to adapt to the multi-granularity structure of the time series, function tokens are also used as special markers to aggregate information at different levels.

[0047] For example, a domain token can be placed at the beginning of the input sequence as a learnable embedding Z. [DOM] This domain token can be used by the model to capture global features of the entire time series samples, supporting domain classification tasks during pre-training. It should be noted that the input sequence refers to the entire user electricity consumption sequence. By placing the domain token at the beginning of the entire user electricity consumption sequence consisting of legitimate users and electricity thieves, the features of each dataset (legitimate and thieves) are learned.

[0048] For example, a separator token can be used as a variable separator, representing a learnable embedding Z. [VAR] For example, a [VAR] token is added to the end of each variable sequence to help the model distinguish between different variables and capture the relationships between them. It should be noted that this variable sequence refers to the electricity consumption sequence of each user.

[0049] For example, a shared learnable embedding Z can be used at the mask positions of the input sequence. [MASK] It is used for mask patch modeling during the training process.

[0050] The aforementioned function token design enables the model to simulate document structure, treating each variable as a sentence and multivariate time series as a document, thereby supporting multigranular representation learning.

[0051] It should be noted that the embedded tokens are flattened into a uniform sequence to handle arbitrary numbers of variables and time lengths. The input sequence constructed using the above functional token design is represented as follows: (6) in, The input sequence is a matrix containing (N+1)C+1 tokens, where N represents the number of patches for each user and C represents the number of users; [VAR] represents the separator token, [DOM] represents the domain token, and Z... [MASK]Indicates the mask location token. This represents the initial embedding vector for the Nth data patch of the Cth user.

[0052] Each line in the input sequence corresponds to a patch sequence for a variable, ending with a [VAR] token and beginning with a global [DOM] token. This design ensures that both the temporal patterns within variables and the correlations between variables are encoded.

[0053] It should also be noted that random masks disrupt the continuity of time series, making it difficult to capture anomalies in continuous time periods commonly seen in actual electricity theft. The block mask used in this application divides the sequence into semantically complete time blocks, which is more in line with the characteristic that electricity theft is concentrated in specific time periods in real-world scenarios.

[0054] Optionally, in step S102, N can be 24, using a fixed masking ratio of 25%. This means that 25% of the non-functional tokens in the input sequence are selected for masking; this ratio is chosen for intermittent electricity theft. Non-functional tokens are patch tokens, excluding functional tokens such as [DOM], [VAR], and [MASK].

[0055] In one embodiment, the masking process in step S102 is a multi-block masking process; wherein the specific steps of each block masking process refer to the block masking process described above.

[0056] Optionally, in step S102, the multiple block masking processes can be executed independently in four separate sessions, for example, in time slots 1, 2, 3, and 4. Time slot 1 covers time zones 1-7 (late night to early morning), time slot 2 covers time zones 7-13 (morning to noon), time slot 3 covers time zones 13-19 (afternoon to evening), and time slot 4 covers time zone 19-1 (night to midnight). This method can also be understood as multi-time-slot mask reconstruction, which ensures that it can cope with potential electricity theft during different time periods.

[0057] Example, reference Figure 2 The diagram shows the distribution of data tokens at different points in time. The horizontal axis represents the time series from time zone 1 to time zone 14, with each time zone corresponding to a data box. Data tokens for time period 1 are clearly marked with a dashed box, indicating that tokens for this time period have been masked; tokens for the remaining time periods are preserved.

[0058] It needs to be explained that the core characteristic of intermittent electricity theft is that illegal users steal electricity during specific time periods, resulting in electricity consumption data exhibiting an alternating pattern of "normal-abnormal-normal" over time. To effectively address this local anomaly, a masking strategy with multiple different locations is employed to simulate data loss scenarios.

[0059] In one embodiment, step S103, training the normal user model and the electricity theft user model using the masked training set, respectively, includes the following steps: Step S1031: The masked training set is used as the original training sample mask sequence and input into the normal user model and the electricity theft user model respectively for reconstruction to obtain the third normal reconstruction sequence and the third electricity theft reconstruction sequence.

[0060] Step S1032: Calculate the first mean square error loss corresponding to the normal user model based on the original training sample mask sequence and the third normal reconstruction sequence, and calculate the second mean square error loss corresponding to the electricity theft user model based on the original training sample mask sequence and the third electricity theft reconstruction sequence.

[0061] Step S1033: Train the normal user model and the electricity theft user model by minimizing the first mean square error loss and the second mean square error loss, respectively.

[0062] The above-mentioned masked training set is used as the original training sample mask sequence and input into the normal user model and the electricity theft user model respectively. The normal user model and the electricity theft user model are trained independently using the original training sample mask sequence, so that the normal user model focuses on learning the electricity consumption pattern of normal users during the training process, and the electricity theft user model focuses on learning the characteristics of typical electricity theft behavior during the training process.

[0063] For example, normal user models and electricity theft user models can be built based on the Transformer framework to achieve high-precision reconstruction of time series data by utilizing bidirectional contextual information. The Transformer framework is a deep learning model that uses a self-attention mechanism to replace recurrent neural networks and convolutional neural networks, enabling parallel computation for natural language processing tasks.

[0064] It should be noted that the hidden dimension of both the normal user model and the electricity theft user model is D, and the number of layers is L. The Transformer framework serves as the backbone network of the model, handling the flattened token embedding. , The Transformer framework extends the one-dimensional format of word sequences to accommodate multivariate time series with arbitrary numbers of variables and time point lengths. Combined with a designed pre-training task, patch-level pattern features are aggregated on corresponding function tokens, ultimately forming a multi-granularity representation extraction capability.

[0065] It should be noted that step S1032 involves decoding the token output by the model into the original patch space in order to calculate the reconstruction error in step S1033. The first mean square error loss reflects the reconstruction error of the normal user model, and the second mean square error loss reflects the reconstruction error of the electricity theft user model.

[0066] It should be noted that the total number of mask patches for both the third normal reconstruction sequence and the third electricity theft reconstruction sequence is S=αNC, where N is the number of patches for each user's time series, C is the number of users, and α is the mask ratio.

[0067] For example, in step S1032, let This represents the token at the i-th position of the model output (corresponding to the mask position). The normal user model and the electricity theft user model each use their own linear layer to output the token. Projecting back into the patch space yields the reconstructed patch. : (7) in, Indicates a linear layer. ∈RD×P, where D represents the hidden dimension, P represents the patch length, and R represents the set of real numbers. The role of the linear layer is to map the high-dimensional representation back to the original time-space.

[0068] For example, the first mean square error loss and the second mean square error loss in steps S1033 and S1034 can both be expressed by the following formulas: (8) in, Represents the original training sample mask sequence; when When representing the first mean square error loss, This indicates a patch for rebuilding the normal user model; when When representing the second mean square error loss, This indicates a patch for rebuilding the electricity theft user model. The total number of mask patches is represented by P, and the patch length is represented by P. The loss is normalized to each time point to ensure scale consistency.

[0069] It should be noted that in step S103, mean squared error loss is used instead of cross-entropy because time series are continuous values, not discrete tokens, and mean squared error loss can better measure the reconstruction error of continuous data. Through mask reconstruction, the model learns to use bidirectional contextual information, i.e., past and future time points, to infer the missing parts, enhancing the model's deep understanding of time series and facilitating the capture of autocorrelation and periodic patterns.

[0070] In one embodiment, in step S104, the masked test set is input into the trained normal user model and the electricity theft user model for reconstruction, respectively, to obtain the first normal reconstruction sequence and the first electricity theft reconstruction sequence.

[0071] In one embodiment, step S104, which involves inputting the masked test set into the trained normal user model and the electricity theft user model for reconstruction to obtain a first normal reconstruction sequence and a first electricity theft reconstruction sequence, includes the following steps: Step S1041: The masked test set is used as the original test sample mask sequence and input into the normal user model and the electricity theft user model respectively.

[0072] Step S1042: The normal user model performs linear projection processing on the original test sample mask sequence to obtain the first normal reconstruction sequence; and the electricity theft user model performs linear projection processing on the original test sample mask sequence to obtain the first electricity theft reconstruction sequence.

[0073] Furthermore, in step S105, determining the first normal comprehensive fitting index and the first electricity theft comprehensive fitting index based on the first normal reconstruction sequence and the first electricity theft reconstruction sequence includes the following steps: Step S1051: Calculate the first normal comprehensive fitting index based on the first normal reconstruction sequence and the original test sample mask sequence; and calculate the first electricity theft comprehensive fitting index based on the first electricity theft reconstruction sequence and the original test sample mask sequence.

[0074] Furthermore, in step S105, determining the distribution of electricity theft suspicion indicators based on the user tags of each user in the test set, and determining the electricity theft judgment threshold range based on the distribution of electricity theft suspicion indicators, includes the following steps: Step S1052: Calculate the electricity theft suspicion index value for each user in the test set using the first normal comprehensive fitting index and the first electricity theft comprehensive fitting index.

[0075] Step S1053: Determine the distribution of electricity theft suspicion indicators based on the user tags of each user in the test set and the electricity theft suspicion indicator values ​​of each user; the user tags include normal user tags and electricity theft user tags.

[0076] Step S1054: Based on the principle of maximizing statistical effect size, determine the initial threshold using formula (1): (1) in, Indicates the initial threshold. This represents the mean value of the electricity theft suspicion index for normal users in the test set under the threshold parameter θ. This represents the standard deviation of the electricity theft suspicion index values ​​for normal users in the test set under the threshold parameter θ. This represents the mean value of the electricity theft suspicion index for users in the test set, under the threshold parameter θ. J(θ) represents the standard deviation of the electricity theft suspicion index values ​​of electricity theft users in the test set under the threshold parameter θ; J(θ) represents the optimization objective function, which is essentially to calculate the standardized distance between two distributions.

[0077] It should be noted that maximizing J(θ) ensures that the distribution of electricity theft suspicion indicators between normal users and electricity thieves has the greatest distinguishability, while suppressing intra-class variance and improving threshold robustness.

[0078] Step S1055: Based on the initial threshold, the optimal threshold is adaptively determined using the gradient descent optimization algorithm.

[0079] Step S1056: Determine the range of the electricity theft judgment threshold based on the optimal threshold and the dispersion of the electricity theft suspicion index distribution. , ].

[0080] Guided by the principle of maximizing statistical effect size, the above approach effectively ensures that the initial threshold has a sound theoretical basis. Furthermore, by adaptively determining the optimal threshold for the final decision and the range of thresholds for judging electricity theft, it ensures accurate assessment of the suspected electricity theft level for each user in the validation set, effectively avoiding misjudgments of electricity theft.

[0081] For example, in step S1055, a candidate value that maximizes J(θ) is selected as the initial threshold for gradient descent. This ensures that the gradient descent optimization algorithm begins its iterations from a region with a high probability of approaching the global optimum. Simultaneously, when adaptively determining the optimal threshold using the gradient descent optimization algorithm, the initial learning rate η = 0.01, the maximum number of iterations T = 1000, and the convergence tolerance δ = ... .

[0082] Specifically, step S1055 includes the following steps: The gradient direction is calculated using a gradient descent optimization algorithm. For example, the central difference method is used for numerical gradient approximation, and the threshold of the current iteration is considered. Its gradient direction The calculation formula is: (2) in, This represents a very small disturbance. Possible values 1%, Indicates at the threshold The effect size calculated below, These represent the threshold values ​​respectively. The effect size was calculated below.

[0083] Based on the calculated gradient direction The threshold is updated using formula (3): Update threshold: (3) in, This represents the threshold for iteration t+1. This represents the threshold after t iterations. This represents the initial learning rate.

[0084] Repeatedly calculate the gradient direction and update the threshold until the preset iteration condition or the maximum number of iterations t=T is met, and then set the threshold at convergence. Determined as the optimal threshold The optimal threshold is obtained. .

[0085] The preset iteration conditions are expressed as follows: | |<δ(9) Where δ represents the convergence tolerance.

[0086] It should be explained that, in order to address the problem of high false positive rates near the critical point with a single threshold, this application is based on an optimal threshold. Construct a dynamic dual-threshold decision interval, referring to step S1056.

[0087] For example, in step S1056, the threshold range for determining electricity theft is adaptively set based on the dispersion of the suspect index S distribution on the test set, and the calculation formula for the threshold range for determining electricity theft is as follows: (10) in, This represents the upper limit of the threshold range for determining electricity theft. This represents the lower limit of the threshold range for determining electricity theft. This represents the optimal threshold. Indicates the buffer coefficient. This represents the combined standard deviation.

[0088] For example, the pooled standard deviation is calculated using the following formula: (11) in, This represents the number of user samples marked as "normal". This represents the number of user samples marked as "electricity theft". This represents the standard deviation of the suspicious indicator values ​​for the normal user group. The standard deviation of the suspected index values ​​for electricity theft user groups.

[0089] It should be noted that the above-mentioned upper endpoint threshold is used. and lower endpoint threshold The system is designed with three levels of suspicion: a high-suspect electricity theft zone, an uncertain zone, and a normal zone. The high-suspect electricity theft zone is […]. ∞] indicates highly abnormal user behavior; the uncertain interval is [ , The value ] indicates that the user's behavior suggests uncertain suspicion of electricity theft, and manual intervention is recommended for verification; the normal range is [- This indicates that the user's behavior is in line with expectations.

[0090] In one embodiment, in step S106, the masked verification set is input into the trained normal user model and the electricity theft user model for reconstruction, respectively, to obtain the second normal reconstruction sequence and the second electricity theft reconstruction sequence, including the following steps: Step S1061: Use the masked verification set as the original verification sample mask sequence and input it into the normal user model and the electricity theft user model.

[0091] Step S1062: The normal user model performs linear projection processing on the original verification sample mask sequence to obtain the second normal reconstruction sequence; and the electricity theft user model performs linear projection processing on the original verification sample mask sequence to obtain the second electricity theft reconstruction sequence.

[0092] For example, in steps S104 and S106 above, the masked test set and validation set are input into the trained normal user model and electricity theft user model for reconstruction, and the corresponding reconstruction sequences are output respectively, such as the first normal reconstruction sequence and the first electricity theft reconstruction sequence, as well as the second normal reconstruction sequence and the second electricity theft reconstruction sequence. (Reference) Figure 2 As shown in the figure, 1' can be regarded as the first normal reconstruction sequence and the second normal reconstruction sequence output by the normal user model, and 1'' can be regarded as the first electricity theft reconstruction sequence and the second electricity theft reconstruction sequence output by the electricity theft user model.

[0093] Further, in step S107, the step of determining the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index based on the second normal reconstruction sequence and the second electricity theft reconstruction sequence, using the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index to predict the electricity theft suspicion index value of each user to be tested in the verification set, and evaluating the electricity theft suspicion level of each user to be tested in the verification set based on the electricity theft judgment threshold range and the electricity theft suspicion index value, includes the following steps: Step S1071: Calculate the second normal comprehensive fitting index based on the second normal reconstruction sequence and the original verification sample mask sequence, and calculate the second electricity theft comprehensive fitting index based on the second electricity theft reconstruction sequence and the original verification sample mask sequence.

[0094] Step S1072: Subtract the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index to obtain the electricity theft suspicion index value of each user to be tested in the verification set.

[0095] Step S1073: Compare the electricity theft suspicion index value of each user under test in the verification set with the electricity theft judgment threshold range to determine the electricity theft suspicion level of each user under test in the verification set.

[0096] It should be noted that the second normal comprehensive fitting index reflects the error between the original verification sample mask sequence and the second normal reconstruction sequence reconstructed from the normal user model, while the second electricity theft comprehensive fitting index reflects the error between the original verification sample mask sequence and the second electricity theft reconstruction sequence reconstructed from the electricity theft user model. By subtracting the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index, the obtained electricity theft suspicion index value reflects the comparison of the reconstruction errors of the two models.

[0097] For example, the calculation of the comprehensive fit index in steps S1051 and S1071 is done in the same way, as follows: First, calculate the coefficient of determination using the following formula: (12) in, The coefficient of determination is represented by the coefficient of determination. The closer to 1, the higher the reconstruction accuracy; n represents the number of mask points; Represents the true values, such as the original test sample mask sequence and the original verification sample mask sequence; The mean of the true values, such as the mean of the original test sample mask sequence and the mean of the original validation sample mask sequence; This represents the reconstruction sequence output by the model, such as the first normal reconstruction sequence, the first electricity theft reconstruction sequence, the second normal reconstruction sequence, and the second electricity theft reconstruction sequence.

[0098] Then, to improve the robustness of the evaluation, mean absolute percentage error (MASE) is introduced as an auxiliary verification indicator. MASE is used to measure the relative error level of the reconstructed values, and the calculation formula is as follows: (13) in, It represents the mean absolute percentage error.

[0099] Finally, a comprehensive fit index is constructed based on the coefficient of determination and the mean absolute percentage error: (14) in, The Comprehensive Fit Index (CFI) is represented by the coefficient of determination. The algorithm is dominant and fine-tuned for low-confidence fits using the mean absolute percentage error (MAPE).

[0100] It should be noted that when MAPE is small, CFI ≈ When MAPE is large, the CFI value will be reduced appropriately, so as to more carefully evaluate the fitting quality, and to more robustly deal with intermittent electricity theft and avoid interference from the randomness of a single mask position.

[0101] It should also be noted that, for the normal user model, the set of CFI values ​​calculated across all masking periods is { }, all masked time periods are the aforementioned time periods 1, 2, 3, and 4. Select this set { The lower quartile Q1 of} is used as the preferred normal comprehensive fit index. Because Q1 is not sensitive to extremely low values, it can more stably reflect the minimum fit between user behavior and normal patterns.

[0102] It should also be noted that, for the electricity theft user model, the set of CFI values ​​calculated for all masked time periods { }, all masked time periods are the aforementioned time periods 1, 2, 3, and 4. Select this set { The upper quartile Q3 of} is used as the preferred comprehensive fitting index for electricity theft. Because Q3 is not sensitive to extremely high values, it can more stably reflect the maximum fit between user behavior and electricity theft patterns.

[0103] For example, the methods for calculating the suspected electricity theft index value in steps S1052 and S1072 are the same, as follows: Based on the optimized normal comprehensive fitting index And the optimal comprehensive fitting index for electricity theft Calculate the index value for suspected electricity theft: (15) in, This indicates the index value for suspected electricity theft.

[0104] It should be noted that this electricity theft suspicion index can effectively quantify the similarity between the electricity consumption pattern and the electricity theft pattern of the tested user. If S>0 and the value is large, it indicates that the electricity consumption pattern of the tested user is closer to the electricity theft pattern; if S<0 and the absolute value is large, it indicates that the electricity consumption pattern of the tested user is closer to the normal pattern.

[0105] The calculation method of the above comprehensive fitting index is based on the dynamic optimization strategy of statistical distribution, which effectively improves the signal-to-noise ratio and anti-overfitting ability of the suspected index.

[0106] To verify the effectiveness of the proposed method for quantitative assessment of intermittent electricity theft based on dual-model reconstruction error comparison, the following experimental data is provided. This experiment uses actual collected electricity consumption data from power users for systematic evaluation. The acquired electricity consumption sequence data includes daily load data from 500 power users, of whom 300 are verified normal users by the power company, and 200 are users confirmed to have engaged in intermittent electricity theft. Data was collected from each user every 15 minutes, recording 24 hours of electricity consumption behavior, resulting in 96 electricity consumption sequences at various time points. The standard electricity consumption sequence data was randomly divided into training, testing, and validation sets in an 8:1:1 ratio.

[0107] In this experiment, the specific parameters were set as follows: patch length was 8 (for 2 hours of data); hidden layer dimension was 128; mask ratio was 25%; and learning rate was 1× The training epochs were 20; the buffer coefficient α = 1.2. Model training was performed on an NVIDIA RTX4060 GPU using the AdamW optimizer.

[0108] Sure =0.35, =-0.28, the threshold range for judging electricity theft is [-0.28, 0.35].

[0109] The experimental results on the validation set are shown in Table 1: Table 1. Classification results of the validation set As shown in Table 1 above, the high-suspicion electricity theft category performed exceptionally well, with 18 electricity theft users correctly identified and only 2 legitimate users misidentified. In the legitimate user category, 26 legitimate users were accurately identified, and no electricity theft users were misidentified as legitimate. This demonstrates that the model achieves a low false positive rate while maintaining a high recall rate.

[0110] Furthermore, to illustrate the beneficial effects of the method proposed in this application, a comparison is made between the method proposed in this application and the traditional LSTM method.

[0111] Table 2 Classification results of the traditional LSTM method As can be seen from Table 2 above, the method proposed in this application is consistent with the traditional LSTM method in identifying the number of real electricity theft users. However, the number of normal users who are falsely identified in this application is significantly reduced. Furthermore, by classifying cases that are prone to false positives into uncertain categories, this application can effectively reduce the false alarm rate.

[0112] Regarding the system in the above embodiments, the specific manner in which each unit performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0113] It should be noted that although several units of the system for executing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units. Some or all of the units can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without any inventive effort.

[0114] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A method for quantitative assessment of intermittent electricity theft based on dual-model reconstruction error comparison, characterized in that, include: Obtain electricity consumption sequence data from multiple users, and construct standard electricity consumption sequence data based on the electricity consumption sequence data from multiple users. The standard electricity consumption sequence data includes a training set, a test set, and a validation set. The standard electricity consumption sequence data is masked. Construct a normal user model and an electricity theft user model, and train the normal user model and the electricity theft user model respectively using the masked training set; The masked test set is input into the trained normal user model and the electricity theft user model respectively for reconstruction, to obtain the first normal reconstruction sequence and the first electricity theft reconstruction sequence. The first normal comprehensive fitting index and the first electricity theft comprehensive fitting index are determined based on the first normal reconstruction sequence and the first electricity theft reconstruction sequence, and the distribution of electricity theft suspicion index is determined based on the user tags of each user in the test set. The electricity theft judgment threshold range is determined based on the distribution of electricity theft suspicion index. The masked verification set is input into the trained normal user model and the electricity theft user model respectively for reconstruction, to obtain the second normal reconstruction sequence and the second electricity theft reconstruction sequence. The second normal comprehensive fitting index and the second electricity theft comprehensive fitting index are determined based on the second normal reconstruction sequence and the second electricity theft reconstruction sequence. The second normal comprehensive fitting index and the second electricity theft comprehensive fitting index are used to predict the electricity theft suspicion index value of each user in the verification set. The electricity theft suspicion level of each user in the verification set is evaluated based on the electricity theft judgment threshold range and the electricity theft suspicion index value.

2. The intermittent electricity theft quantitative assessment method based on dual-model reconstruction error comparison according to claim 1, characterized in that, The process of training the normal user model and the electricity theft user model using the masked training set includes: The masked training set is used as the original training sample mask sequence and input into the normal user model and the electricity theft user model respectively for reconstruction to obtain the third normal reconstruction sequence and the third electricity theft reconstruction sequence. The first mean square error loss corresponding to the normal user model is calculated based on the original training sample mask sequence and the third normal reconstruction sequence, and the second mean square error loss corresponding to the electricity theft user model is calculated based on the original training sample mask sequence and the third electricity theft reconstruction sequence. The normal user model and the electricity theft user model are trained by minimizing the first mean square error loss and the second mean square error loss, respectively.

3. The intermittent electricity theft quantitative assessment method based on dual-model reconstruction error comparison according to claim 1, characterized in that, The process of inputting the masked test set into the trained normal user model and the electricity theft user model for reconstruction, respectively, to obtain the first normal reconstruction sequence and the first electricity theft reconstruction sequence, includes: The masked test set is used as the original test sample mask sequence and input into the normal user model and the electricity theft user model, respectively. The normal user model performs linear projection processing on the original test sample mask sequence to obtain the first normal reconstruction sequence; Furthermore, the electricity theft user model performs linear projection processing on the original test sample mask sequence to obtain the first electricity theft reconstruction sequence.

4. The intermittent electricity theft quantitative assessment method based on dual-model reconstruction error comparison according to claim 3, characterized in that, The step of determining the first normal comprehensive fitting index and the first electricity theft comprehensive fitting index based on the first normal reconstruction sequence and the first electricity theft reconstruction sequence includes: Calculate the first normal comprehensive fit index based on the first normal reconstruction sequence and the original test sample mask sequence; Furthermore, a first comprehensive fitting index for electricity theft is calculated based on the first electricity theft reconstruction sequence and the original test sample mask sequence.

5. The intermittent quantitative assessment method for electricity theft based on dual-model reconstruction error comparison according to any one of claims 1-4, characterized in that, The step of determining the distribution of electricity theft suspicion indicators based on user tags of each user in the test set, and determining the electricity theft judgment threshold range based on the distribution of electricity theft suspicion indicators, includes: The electricity theft suspicion index value of each user in the test set is calculated using the first normal comprehensive fitting index and the first electricity theft comprehensive fitting index. The distribution of electricity theft suspicion indicators is determined based on the user tags of each user in the test set and the electricity theft suspicion index values ​​of each user; the tags include normal user tags and electricity theft user tags; The initial threshold is determined using formula (1): （1） in, Indicates the initial threshold. This represents the mean value of the electricity theft suspicion index for normal users in the test set under the threshold parameter θ. This represents the standard deviation of the electricity theft suspicion index values ​​for normal users in the test set under the threshold parameter θ. This represents the mean value of the electricity theft suspicion index for users in the test set, under the threshold parameter θ. J(θ) represents the standard deviation of the electricity theft suspicion index values ​​of electricity theft users in the test set under the threshold parameter θ; J(θ) represents the optimization objective function. Based on the initial threshold, the optimal threshold is adaptively determined using a gradient descent optimization algorithm; The range of the electricity theft judgment threshold is determined based on the optimal threshold and the dispersion of the distribution of the electricity theft suspicion index.

6. The intermittent electricity theft quantitative assessment method based on dual-model reconstruction error comparison according to claim 5, characterized in that, Based on the initial threshold, the optimal threshold is adaptively determined using a gradient descent optimization algorithm, including: The gradient direction is calculated using the gradient descent optimization algorithm: （2） in, Indicates the gradient direction. This represents a very small disturbance. Indicates at the threshold The effect size calculated below, These represent the threshold values ​​respectively. The effect size was calculated below; The threshold is updated according to the gradient direction using formula (3): （3） in, This represents the threshold for iteration t+1. This represents the threshold after t iterations. Indicates the initial learning rate; Repeatedly calculate the gradient direction and update the threshold until the preset iteration condition or the maximum number of iterations is met, and obtain the optimal threshold.

7. The intermittent electricity theft quantitative assessment method based on dual-model reconstruction error comparison according to claim 1, characterized in that, The process of inputting the masked verification set into the trained normal user model and the electricity theft user model for reconstruction, respectively, yields a second normal reconstruction sequence and a second electricity theft reconstruction sequence, including: The masked verification set is used as the original verification sample mask sequence and input into the normal user model and the electricity theft user model. The normal user model performs linear projection processing on the original verification sample mask sequence to obtain the second normal reconstruction sequence; Furthermore, the electricity theft user model performs linear projection processing on the original verification sample mask sequence to obtain the second electricity theft reconstruction sequence.

8. The intermittent electricity theft quantitative assessment method based on dual-model reconstruction error comparison according to claim 7, characterized in that, The step of determining a second normal comprehensive fitting index and a second electricity theft comprehensive fitting index based on the second normal reconstruction sequence and the second electricity theft reconstruction sequence, using the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index to predict the electricity theft suspicion index value of each test user in the verification set, and evaluating the electricity theft suspicion level of each test user in the verification set based on the electricity theft judgment threshold range and the electricity theft suspicion index value, includes: The second normal comprehensive fitting index is calculated based on the second normal reconstruction sequence and the original verification sample mask sequence, and the second electricity theft comprehensive fitting index is calculated based on the second electricity theft reconstruction sequence and the original verification sample mask sequence; The difference between the second normal comprehensive fitting index and the second electricity theft comprehensive fitting index is processed to obtain the electricity theft suspicion index value of each user in the verification set. The electricity theft suspicion index value of each user in the verification set is compared with the electricity theft judgment threshold range to determine the electricity theft suspicion level of each user in the verification set.

9. The intermittent electricity theft quantitative assessment method based on dual-model reconstruction error comparison according to claim 1, characterized in that, The masking process is a multi-block masking process.

10. The intermittent quantitative assessment method for electricity theft based on dual-model reconstruction error comparison according to claim 1, characterized in that, The step of constructing standard electricity consumption sequence data based on the electricity consumption sequence data of the multiple users includes: The electricity consumption sequence data of the multiple users is preprocessed to obtain the standard electricity consumption sequence data.

Images