A power consumption information acquisition system and method for identifying faults of a power meter

Abstract

The application relates to the technical field of electric energy meter fault identification, and discloses a power consumption information acquisition system and method for identifying electric energy meter faults, a power consumption information acquisition system for identifying electric energy meter faults, which comprises a time-frequency fingerprint extraction module, a period characteristic construction module, a period characteristic analysis module, an instantaneous characteristic construction module and an electrical diagram construction module.The time-frequency fingerprint extraction module extracts a time-frequency fingerprint through three-layer discrete wavelet transformation; the period characteristic construction module constructs a period characteristic vector; the period characteristic analysis module obtains whether the electric energy meter has a fault through a first fault detection model; the instantaneous characteristic construction module constructs an instantaneous characteristic vector; the electrical diagram construction module constructs a sparse electrical diagram; and the electrical diagram analysis module obtains the fault type of the electric energy meter through a second fault detection model.The application screens out suspected faulty electric energy meters according to period characteristics, constructs a sparse electrical diagram according to the physical topological connection among the electric energy meters and the instantaneous characteristics, carries out correlation analysis on the sparse electrical diagram, identifies the fault type of the electric energy meter, and thus improves the overall detection efficiency and precision.

Description

Technical Field

[0001] This invention relates to the field of electricity meter fault identification technology, and more specifically, to an electricity consumption information collection system and method for identifying electricity meter faults. Background Technology

[0002] An electricity meter is an instrument used to measure and record electricity consumption. It is widely used in residential, commercial, and industrial electricity metering. In power systems, electricity meters are not only an important source of data for recording user electricity consumption behavior, but also key equipment in electricity billing and fault identification. Traditional electricity meter fault identification mainly relies on regular manual inspections and user-initiated repair reports. This method is not only inefficient, but also suffers from problems such as identification lag, incomplete coverage, and human error, making it difficult to meet the high requirements of real-time performance and accuracy in modern power systems.

[0003] With the development of big data and artificial intelligence technologies, the current method of collecting electricity consumption data from electricity meters to form time series data and using time series analysis models (such as LSTM and ARIMA) to model the operating status of electricity meters and detect anomalies has improved the automation of fault identification. However, this method generally ignores the possible electrical connections between electricity meters, resulting in unsatisfactory fault identification accuracy. For example, if multiple households' electricity meters are connected to the same line, when one household experiences an abnormal power consumption (such as a sudden power outage of high-power equipment or voltage fluctuations), it will affect the power parameter changes of other adjacent electricity meters. If only time series analysis of a single meter is used, it is easy to misjudge normal meters as abnormal. Summary of the Invention

[0004] This invention provides an electricity consumption information collection system and method for identifying electricity meter faults, thereby solving the technical problems mentioned in the background art.

[0005] This invention provides an electricity consumption information collection system for identifying electricity meter faults, comprising:

[0006] The time-frequency fingerprint extraction module is used to collect the active power of each energy meter to construct an active power sequence and extract the time-frequency fingerprint through a three-level discrete wavelet transform.

[0007] Time-frequency fingerprints consist of low-frequency energy values, mid-frequency energy values, and high-frequency energy values;

[0008] The periodic feature construction module is used to calculate the day-night power ratio of each energy meter and form a periodic feature vector with the time-frequency fingerprint;

[0009] The periodic feature analysis module is used to analyze the periodic feature vector of each energy meter through the first fault detection model. The output value indicates whether the energy meter has a fault, and the energy meters with faults are marked.

[0010] The instantaneous feature construction module is used to calculate the power factor, three-phase voltage imbalance and three-phase current mean of each energy meter at the current moment, and form an instantaneous feature vector with the current power consumption and active power.

[0011] The electrical graph construction module is used to construct a sparse electrical graph based on the instantaneous feature vector of each energy meter and the physical topology connection between the energy meters.

[0012] The electrical diagram analysis module is used to analyze sparse electrical diagrams using a second fault detection model. The output value indicates the fault type of the energy meter with fault markers.

[0013] Furthermore, the length of the active power sequence is equal to the acquisition time period divided by the acquisition time interval, where the acquisition time period and the acquisition time interval are user-defined parameters; the active power sequence is subjected to wavelet transform through three-level discrete wavelet transform to obtain three-level detail coefficients, and the sum of the squares of the first-level detail coefficient, the second-level detail coefficient, and the third-level detail coefficient are calculated as the low-frequency energy value, the mid-frequency energy value, and the high-frequency energy value, respectively.

[0014] Furthermore, the day-night power ratio is equal to the ratio of the average power during the daytime period to the average power during the nighttime period, where the daytime period and the nighttime period are user-defined parameters.

[0015] Furthermore, the first fault detection model consists of a nonlinear transformation layer and a first classifier;

[0016] The non-linear transformation layer is used to transform the periodic feature vector into a first update vector, where the number of dimensions of the first update vector is a user-defined parameter;

[0017] The calculation formula for the nonlinear transformation layer is as follows:

[0018] Where 1≤i≤L, This represents the first update vector output by the nonlinear transformation layer, where L represents the maximum number of nonlinear transformation layers, and is a user-defined parameter. and These represent the first intermediate vectors output from the (i-1)th layer and the i-th layer, respectively. This represents the periodic feature vector of the input nonlinear transformation layer. and These represent the first weight parameter and the first bias parameter of the i-th layer of the nonlinear transformation layer, respectively. GELU represents the GELU activation function, and LaryerNorm represents the inter-layer normalization function.

[0019] The first classifier takes the first update vector as input and outputs a value indicating whether the electricity meter is faulty.

[0020] The first classifier is built based on a multilayer perceptron, and the corresponding activation function is the Sigmoid activation function.

[0021] Furthermore, during the training process of the first fault detection model, a reconstruction mapping layer is added after the nonlinear transformation layer;

[0022] The reconstruction mapping layer is used to transform the first update vector into a second update vector, wherein the second update vector has the same number of dimensions as the periodic feature vector;

[0023] The calculation formula for reconstructing the mapping layer is as follows:

[0024] Where 1≤j≤J, This represents the second update vector output by the reconstruction mapping layer, where J represents the maximum number of reconstruction mapping layers, J = L. and These represent the second intermediate vectors output from the (j-1)th and jth layers, respectively. This represents the first update vector of the input reconstruction mapping layer. and These represent the second weight parameter and the second bias parameter of the j-th layer of the reconstruction mapping layer, respectively. and σ and ⊙ represent the third weight parameter and the third bias parameter of the j-th layer of the reconstruction mapping layer, respectively; σ represents the Sigmoid activation function; ⊙ represents the Hadamard product.

[0025] The formula for calculating the loss function Loss of the first fault detection model is as follows:

[0026] Loss=λ rec ×Loss rec +λ clf ×Loss clf Loss rec Loss indicates reconstruction losses. clf Let λ represent the classification loss. rec λ represents the reconstruction loss coefficient. clf This indicates that the sum of the classification loss coefficient, reconstruction loss coefficient, and classification loss coefficient is 1.

[0027] Where 1 ≤ k ≤ K, and K represents the number of training samples. This represents the first update vector output by the non-linear transformation layer after inputting the sample data of the k-th training sample. This represents the second update vector output by the reconstructed mapping layer after inputting the sample data of the k-th training sample, where ||·||2 represents the L2 norm;

[0028] Where yk represents the sample label of the k-th training sample. This represents the output value of the first fault detection model after inputting the sample data of the kth training sample.

[0029] Furthermore, the power factor is equal to the phase difference between the line voltage and the line current;

[0030] Three-phase voltage imbalance U unb The calculation formula is as follows:

[0031] Among them U a U b and U c U represents the voltage of phase A, phase B, and phase C, respectively. avg This represents the average three-phase voltage. `max` represents the maximum value.

[0032] The formula for calculating the average three-phase current is as follows: Among them I a I b and I c These represent the currents of phase A, phase B, and phase C, respectively.

[0033] Furthermore, the sparse electrical graph consists of nodes and edges between nodes, where each node corresponds to an electricity meter. If there is a physical topological connection between electricity meters, an edge is constructed between the corresponding nodes. If the cosine similarity of the instantaneous feature vectors between electricity meters is greater than or equal to a preset similarity threshold, an edge is constructed between the corresponding nodes. The preset similarity threshold is a user-defined parameter.

[0034] Furthermore, the second fault detection model consists of a graph data analysis layer and a second classifier;

[0035] The graph data analysis layer is used to update the instantaneous feature vectors of all nodes in the sparse electrical graph and outputs a third update vector, where the number of dimensions of the third update vector is a user-defined parameter.

[0036] The second classifier takes the third update vector of the node corresponding to the energy meter with fault mark as input, and outputs the value representing the fault type of the energy meter with fault mark.

[0037] The second classifier is built based on a multilayer perceptron, and the corresponding activation function is the Softmax activation function.

[0038] Furthermore, the calculation formula for the graph data analysis layer includes:

[0039] Where 1 ≤ u ≤ U, and U represents the number of nodes in the sparse electrical graph. Let h represent the third update vector of the u-th node.u and h v Let N represent the instantaneous feature vectors of the u-th node and the v-th node, respectively. u Let W1 and W2 represent the first and second weight parameters, respectively, and b1 and b2 represent the first and second bias parameters, respectively.

[0040] W represents the correlation coefficient between the u-th node and the v-th node, W3 represents the third weight parameter, || represents the concatenation operation, T represents the transpose operation, and σ represents the Sigmoid activation function.

[0041] This invention provides a method for collecting electricity consumption information to identify electricity meter faults, comprising the following steps:

[0042] Step S201: Collect the active power of each electricity meter to construct an active power sequence, and extract the time-frequency fingerprint through three-level discrete wavelet transform;

[0043] Step S202: Calculate the day-night power ratio of each electricity meter and combine it with the time-frequency fingerprint to form a periodic feature vector;

[0044] Step S203: Analyze the periodic feature vector of each energy meter using the first fault detection model, and the output value indicates whether the energy meter has a fault, and mark the energy meters with faults.

[0045] Step S204: Calculate the power factor, three-phase voltage imbalance and three-phase current mean of each energy meter at the current moment, and form an instantaneous feature vector with the current power consumption and active power.

[0046] Step S205: Construct a sparse electrical graph based on the instantaneous feature vector of each electricity meter and the physical topology connection between the electricity meters;

[0047] Step S206: Analyze the sparse electrical diagram using the second fault detection model, and the output value represents the fault type of the energy meter with fault markers.

[0048] The beneficial effects of this invention are as follows: This invention first screens out suspected faulty energy meters based on periodic characteristics through a two-stage analysis, and then constructs a sparse electrical graph based on the physical topology connection and instantaneous characteristics between energy meters. Correlation analysis is then performed on the sparse electrical graph to identify the fault type of the energy meter. This avoids performing complex fault type analysis on all energy meters, reduces computational resource consumption, and also reduces the false alarm rate, thereby improving the overall detection efficiency and accuracy. Attached Figure Description

[0049] Figure 1This is a schematic diagram of an electricity consumption information collection system for identifying electricity meter faults according to the present invention;

[0050] Figure 2 This is a flowchart of a method for collecting electricity consumption information to identify electricity meter faults according to the present invention.

[0051] In the figure: Time-frequency fingerprint extraction module 101, periodic feature construction module 102, periodic feature analysis module 103, instantaneous feature construction module 104, electrical diagram construction module 105, electrical diagram analysis module 106. Detailed Implementation

[0052] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0053] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0054] like Figures 1-2 As shown, an electricity consumption information collection system for identifying electricity meter faults includes:

[0055] The time-frequency fingerprint extraction module 101 is used to collect the active power of each energy meter to construct an active power sequence and extract the time-frequency fingerprint through a three-level discrete wavelet transform.

[0056] Time-frequency fingerprints consist of low-frequency energy values, mid-frequency energy values, and high-frequency energy values;

[0057] The periodic feature construction module 102 is used to calculate the day-night power ratio of each electricity meter and form a periodic feature vector with the time-frequency fingerprint;

[0058] The periodic feature analysis module 103 is used to analyze the periodic feature vector of each energy meter through the first fault detection model, and the output value indicates whether the energy meter has a fault, and marks the energy meters with faults.

[0059] The instantaneous feature construction module 104 is used to calculate the power factor, three-phase voltage imbalance and three-phase current mean of each energy meter at the current moment, and form an instantaneous feature vector with the current power consumption and active power.

[0060] Electrical graph construction module 105 is used to construct a sparse electrical graph based on the instantaneous feature vector of each energy meter and the physical topological connection between the energy meters.

[0061] The electrical diagram analysis module 106 is used to analyze sparse electrical diagrams through a second fault detection model, and the output value indicates the fault type of the energy meter with fault markers.

[0062] It should be noted that the two-stage analysis of this invention first screens out the electricity meters suspected of being faulty, avoiding the need to perform complex fault type analysis on all electricity meters. This reduces computational resource consumption and also lowers the false alarm rate, thereby improving the overall detection efficiency and accuracy.

[0063] In one embodiment of the present invention, the length of the active power sequence is equal to the acquisition time period divided by the acquisition time interval, wherein the acquisition time period and the acquisition time interval are both custom parameters. Preferably, the acquisition time period is set to 7 days and the acquisition time interval is set to 15 minutes, then the length of the active power sequence is equal to 672. The active power sequence is subjected to wavelet transform through three-level discrete wavelet transform to obtain three-level detail coefficients, and the sum of the squares of the first-level detail coefficient, the second-level detail coefficient and the third-level detail coefficient are calculated as the low-frequency energy value, the mid-frequency energy value and the high-frequency energy value, respectively.

[0064] In one embodiment of the present invention, the day-night power ratio is equal to the ratio of the average power during the daytime period to the average power during the nighttime period, wherein the daytime period and the nighttime period are custom parameters. Preferably, the daytime period is set to 7:00 a.m. to 10:00 p.m. and the nighttime period is set to 10:00 p.m. to 7:00 a.m.

[0065] In one embodiment of the present invention, the first fault detection model consists of a nonlinear transformation layer and a first classifier;

[0066] The nonlinear transformation layer is used to convert the periodic feature vector into a first update vector, wherein the number of dimensions of the first update vector is a custom parameter, preferably, the number of dimensions of the first update vector is set to 16;

[0067] The calculation formula for the nonlinear transformation layer is as follows:

[0068] Where 1≤i≤L, This represents the first update vector output by the nonlinear transformation layer, and L represents the maximum number of nonlinear transformation layers, which is a user-defined parameter. Preferably, L is set to 3. and These represent the first intermediate vectors output from the (i-1)th layer and the i-th layer, respectively. This represents the periodic feature vector of the input nonlinear transformation layer. and These represent the first weight parameter and the first bias parameter of the i-th layer of the nonlinear transformation layer, respectively. GELU represents the GELU activation function, and LaryerNorm represents the inter-layer normalization function.

[0069] The first classifier takes the first update vector as input and outputs a value indicating whether the electricity meter is faulty.

[0070] The first classifier is built based on a multilayer perceptron, and the corresponding activation function is the Sigmoid activation function.

[0071] It should be noted that, based on the above, the first weight parameters of the three layers in the nonlinear transformation layer can be designed as matrices of size 4×8, 8×12, and 12×16, respectively. Then, the dimensions of the first intermediate vectors of the three layers are 8, 12, and 16, respectively. The first bias parameters of the three layers can be designed as vectors of size 1×8, 1×12, and 1×16, respectively. In addition, residual connections can be introduced to prevent gradient vanishing, thereby improving the stability of subsequent model training.

[0072] In one embodiment of the present invention, during the training process of the first fault detection model, a reconstruction mapping layer is added after the nonlinear transformation layer;

[0073] The reconstruction mapping layer is used to transform the first update vector into a second update vector, wherein the second update vector has the same number of dimensions as the periodic feature vector;

[0074] The calculation formula for reconstructing the mapping layer is as follows:

[0075] Where 1≤j≤J, This represents the second update vector output by the reconstruction mapping layer, where J represents the maximum number of reconstruction mapping layers, J = L. and These represent the second intermediate vectors output from the (j-1)th and jth layers, respectively. This represents the first update vector of the input reconstruction mapping layer. and These represent the second weight parameter and the second bias parameter of the j-th layer of the reconstruction mapping layer, respectively. and σ and ⊙ represent the third weight parameter and the third bias parameter of the j-th layer of the reconstruction mapping layer, respectively; σ represents the Sigmoid activation function; ⊙ represents the Hadamard product.

[0076] The formula for calculating the loss function Loss of the first fault detection model is as follows:

[0077] Loss=λ rec ×Loss rec +λ clf ×Loss clf Loss rec Loss indicates reconstruction losses. clf Let λ represent the classification loss. rec λ represents the reconstruction loss coefficient. clf The sum of the classification loss coefficient, reconstruction loss coefficient, and classification loss coefficient is 1. Preferably, the reconstruction loss coefficient is set to 0.3 and the classification loss coefficient is set to 0.7.

[0078] Where 1 ≤ k ≤ K, and K represents the number of training samples. This represents the first update vector output by the non-linear transformation layer after inputting the sample data of the k-th training sample. This represents the second update vector output by the reconstructed mapping layer after inputting the sample data of the k-th training sample, where ||·||2 represents the L2 norm;

[0079] Where y k This represents the sample label of the k-th training sample. This represents the output value of the first fault detection model after inputting the sample data of the kth training sample.

[0080] It should be noted that, based on the above, the second and third weight parameters of the three layers in the reconstruction mapping layer can be designed as matrices of size 16×12, 12×8, and 8×4, respectively. Then, the dimensions of the second intermediate vector of the three layers are 12, 8, and 4, respectively. The second and third bias parameters of the three layers can be designed as vectors of size 1×12, 1×8, and 1×4, respectively. In addition, the sigmoid activation function of the first classifier is used for binary classification, while the sigmoid activation function of the reconstruction mapping layer is used for information filtering, preventing noise propagation, enhancing the nonlinear expressive power of the model, and thus improving the robustness of the model.

[0081] It should be noted that a training sample consists of sample data and sample labels. The sample labels of the training samples used to train the first fault detection model are obtained through manual annotation. That is, the sample data is a periodic feature vector, and the sample label is whether the electricity meter is faulty. Furthermore, the periodic feature vector needs to be normalized before being input into the first fault detection model to eliminate the difference in units.

[0082] In one embodiment of the present invention, the power factor is equal to the phase difference between the line voltage and the line current;

[0083] Three-phase voltage imbalance U unb The calculation formula is as follows:

[0084] Among them U a U b and U c U represents the voltage of phase A, phase B, and phase C, respectively. avg This represents the average three-phase voltage. `max` represents the maximum value.

[0085] The formula for calculating the average three-phase current is as follows: Where I a I b and I c These represent the currents of phase A, phase B, and phase C, respectively.

[0086] It should be noted that line voltage refers to the voltage between two live wires, and line current refers to the current flowing through the load. The above calculation of power factor refers to the application in a symmetrical three-phase system. In an asymmetrical three-phase system, the power factor is equal to the ratio of the sum of the three-phase active power to the sum of the three-phase apparent power. The power factor can be directly obtained through a smart energy meter, that is, the phase difference is directly calculated by sampling the voltage and current waveforms at high frequency to output the power factor, which will not be elaborated here.

[0087] In one embodiment of the present invention, the sparse electrical graph consists of nodes and edges between nodes, wherein each node corresponds to an electricity meter. If there is a physical topological connection between the electricity meters, an edge is constructed between the corresponding nodes. If the cosine similarity of the instantaneous feature vectors between the electricity meters is greater than or equal to a preset similarity threshold, an edge is constructed between the corresponding nodes. The preset similarity threshold is a custom parameter, preferably set to 0.6.

[0088] It should be noted that there may be multiple edges between nodes according to the above operation. For the sake of calculation, we directly perform the "pruning" operation, that is, only one edge is kept between nodes. Alternatively, the number of edges between nodes can also be used as a weight parameter in the calculation, which will not be elaborated here.

[0089] In one embodiment of the present invention, the second fault detection model consists of a graph data analysis layer and a second classifier;

[0090] The graph data analysis layer is used to update the instantaneous feature vectors of all nodes in the sparse electrical graph and output a third update vector. The number of dimensions of the third update vector is a custom parameter. Preferably, the number of dimensions of the third update vector is set to 32.

[0091] The second classifier takes the third update vector of the node corresponding to the energy meter with fault mark as input, and outputs the value representing the fault type of the energy meter with fault mark.

[0092] The second classifier is built based on a multilayer perceptron, and the corresponding activation function is the Softmax activation function.

[0093] In one embodiment of the present invention, the calculation formula for the graph data analysis layer includes:

[0094] Where 1 ≤ u ≤ U, and U represents the number of nodes in the sparse electrical graph. Let h represent the third update vector of the u-th node. u and h v Let N represent the instantaneous feature vectors of the u-th node and the v-th node, respectively. u Let W1 and W2 represent the first and second weight parameters, respectively, and b1 and b2 represent the first and second bias parameters, respectively.

[0095] W represents the correlation coefficient between the u-th node and the v-th node, W3 represents the third weight parameter, || represents the concatenation operation, T represents the transpose operation, and σ represents the Sigmoid activation function.

[0096] It should be noted that, according to the calculation formula of the graph data analysis layer, the first and second weight parameters should both be designed as 5×32 matrices, the first and second bias parameters should both be designed as 1×32 vectors, and the third weight parameter should be designed as a 1×10 vector.

[0097] It should be noted that the sample labels of the training samples used to train the second fault detection model can also be obtained through manual annotation. That is, the fault types of the electricity meter can include: abnormal power factor, three-phase voltage imbalance, uneven current, power surge, electricity theft, etc. Similarly, the instantaneous feature vectors of all nodes in the sparse electrical graph need to be normalized before being input into the second fault detection model to eliminate dimensional differences.

[0098] In one embodiment of the present invention, such as Figure 2 As shown, a method for collecting electricity consumption information to identify electricity meter faults includes the following steps:

[0099] Step S201: Collect the active power of each electricity meter to construct an active power sequence, and extract the time-frequency fingerprint through three-level discrete wavelet transform;

[0100] Step S202: Calculate the day-night power ratio of each electricity meter and combine it with the time-frequency fingerprint to form a periodic feature vector;

[0101] Step S203: Analyze the periodic feature vector of each energy meter using the first fault detection model, and the output value indicates whether the energy meter has a fault, and mark the energy meters with faults.

[0102] Step S204: Calculate the power factor, three-phase voltage imbalance and three-phase current mean of each energy meter at the current moment, and form an instantaneous feature vector with the current power consumption and active power.

[0103] Step S205: Construct a sparse electrical graph based on the instantaneous feature vector of each electricity meter and the physical topology connection between the electricity meters;

[0104] Step S206: Analyze the sparse electrical diagram using the second fault detection model, and the output value represents the fault type of the energy meter with fault markers.

[0105] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0106] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A power consumption information collection system for identifying power meter faults, characterized in that, include: The time-frequency fingerprint extraction module is used to collect the active power of each energy meter to construct an active power sequence and extract the time-frequency fingerprint through a three-level discrete wavelet transform. Time-frequency fingerprints consist of low-frequency energy values, mid-frequency energy values, and high-frequency energy values; The periodic feature construction module is used to calculate the day-night power ratio of each energy meter and form a periodic feature vector with the time-frequency fingerprint; The periodic feature analysis module is used to analyze the periodic feature vector of each energy meter through the first fault detection model. The output value indicates whether the energy meter has a fault, and the energy meters with faults are marked. The instantaneous feature construction module is used to calculate the power factor, three-phase voltage imbalance and three-phase current mean of each energy meter at the current moment, and form an instantaneous feature vector with the current power consumption and active power. The electrical graph construction module is used to construct a sparse electrical graph based on the instantaneous feature vector of each energy meter and the physical topology connection between the energy meters. The electrical diagram analysis module is used to analyze sparse electrical diagrams using a second fault detection model, and the output value indicates the fault type of the energy meter with fault markers. The first fault detection model consists of a nonlinear transformation layer and a first classifier; The non-linear transformation layer is used to transform the periodic feature vector into a first update vector, where the number of dimensions of the first update vector is a user-defined parameter; The calculation formula for the nonlinear transformation layer is as follows: Where 1≤i≤L, This represents the first update vector output by the nonlinear transformation layer, where L represents the maximum number of nonlinear transformation layers, and is a user-defined parameter. and These represent the first intermediate vectors output from the (i-1)th layer and the i-th layer, respectively. This represents the periodic feature vector of the input nonlinear transformation layer. and These represent the first weight parameter and the first bias parameter of the i-th layer of the nonlinear transformation layer, respectively. GELU represents the GELU activation function, and LaryerNorm represents the inter-layer normalization function. The first classifier takes the first update vector as input and outputs a value indicating whether the electricity meter is faulty. The first classifier is built based on a multilayer perceptron, and the corresponding activation function is the Sigmoid activation function.

2. The electricity consumption information collection system for identifying electricity meter faults according to claim 1, characterized in that, The length of the active power sequence is equal to the acquisition time period divided by the acquisition time interval, where the acquisition time period and acquisition time interval are user-defined parameters. The active power sequence is subjected to wavelet transform through three-level discrete wavelet transform to obtain three-level detail coefficients, and the sum of the squares of the first-level detail coefficient, the second-level detail coefficient, and the third-level detail coefficient are calculated as the low-frequency energy value, the mid-frequency energy value, and the high-frequency energy value, respectively.

3. The electricity consumption information collection system for identifying electricity meter faults according to claim 1, characterized in that, The day-night power ratio is equal to the ratio of the average power during the daytime period to the average power during the nighttime period, where the daytime and nighttime periods are user-defined parameters.

4. The electricity consumption information collection system for identifying electricity meter faults according to claim 1, characterized in that, During the training of the first fault detection model, a reconstruction mapping layer is added after the nonlinear transformation layer; The reconstruction mapping layer is used to transform the first update vector into a second update vector, wherein the second update vector has the same number of dimensions as the periodic feature vector; The calculation formula for reconstructing the mapping layer is as follows: Where 1≤j≤J, This represents the second update vector output by the reconstruction mapping layer, where J represents the maximum number of reconstruction mapping layers, J=L. and These represent the second intermediate vectors output from the (j-1)th and jth layers, respectively. This represents the first update vector of the input reconstruction mapping layer. and These represent the second weight parameter and the second bias parameter of the j-th layer of the reconstruction mapping layer, respectively. and Let represent the third weight parameter and the third bias parameter of the j-th layer of the reconstruction mapping layer, respectively; let represent the Sigmoid activation function; and let ⊙ represent the Hadamard product. Loss function of the first fault detection model The calculation formula is as follows: ,in Indicates the losses incurred during reconstruction. Represents classification loss, Represents the reconstruction loss coefficient. This indicates that the sum of the classification loss coefficient, reconstruction loss coefficient, and classification loss coefficient is 1. Where 1≤k≤K, and K represents the number of training samples. This represents the first update vector output by the non-linear transformation layer after inputting the sample data of the k-th training sample. This represents the second update vector output by the reconstructed mapping layer after inputting the sample data of the k-th training sample. Represents the L2 norm; ,in This represents the sample label of the k-th training sample. This represents the output value of the first fault detection model after inputting the sample data of the kth training sample.

5. The electricity consumption information collection system for identifying electricity meter faults according to claim 1, characterized in that, Power factor equals the phase difference between line voltage and line current; three-phase voltage imbalance. The calculation formula is as follows: ,in , and These represent the voltages of phase A, phase B, and phase C, respectively. This represents the average three-phase voltage. 'max' indicates taking the maximum value; the formula for calculating the average three-phase current is as follows: ,in , and These represent the currents of phase A, phase B, and phase C, respectively.

6. The electricity consumption information collection system for identifying electricity meter faults according to claim 1, characterized in that, A sparse electrical graph consists of nodes and edges between them. Each node corresponds to an electricity meter. If there is a physical topological connection between electricity meters, an edge is constructed between the corresponding nodes. If the cosine similarity of the instantaneous feature vectors between electricity meters is greater than or equal to a preset similarity threshold, an edge is constructed between the corresponding nodes. The preset similarity threshold is a user-defined parameter.

7. The electricity consumption information collection system for identifying electricity meter faults according to claim 6, characterized in that, The second fault detection model consists of a graph data analysis layer and a second classifier. The graph data analysis layer is used to update the instantaneous feature vectors of all nodes in the sparse electrical graph and outputs a third update vector, where the number of dimensions of the third update vector is a user-defined parameter. The second classifier takes the third update vector of the node corresponding to the energy meter with fault marking as input, and outputs the fault type of the energy meter with fault marking. The second classifier is built based on a multilayer perceptron and the corresponding activation function is the Softmax activation function.

8. The electricity consumption information collection system for identifying electricity meter faults according to claim 7, characterized in that, The calculation formulas for the graph data analysis layer include: Where 1 ≤ u ≤ U, and U represents the number of nodes in the sparse electrical graph. Let represent the third update vector of the u-th node. and Let represent the instantaneous feature vectors of the u-th node and the v-th node, respectively. Let represent the set of nodes that have an edge with the u-th node. and These represent the first weight parameter and the second weight parameter, respectively. and These represent the first bias parameter and the second bias parameter, respectively. , This represents the correlation coefficient between the u-th node and the v-th node. This represents the third weight parameter, || represents the concatenation operation, and T represents the transpose operation. This represents the Sigmoid activation function.

9. A method for collecting electricity consumption information to identify electricity meter faults, characterized in that, The electricity consumption information collection system for identifying electricity meter faults as described in any one of claims 1 to 8 includes the following steps: Step S201: Collect the active power of each electricity meter to construct an active power sequence, and extract the time-frequency fingerprint through three-level discrete wavelet transform; Step S202: Calculate the day-night power ratio of each electricity meter and combine it with the time-frequency fingerprint to form a periodic feature vector; Step S203: Analyze the periodic feature vector of each energy meter using the first fault detection model, and the output value indicates whether the energy meter has a fault, and mark the energy meters with faults. Step S204: Calculate the power factor, three-phase voltage imbalance and three-phase current mean of each energy meter at the current moment, and form an instantaneous feature vector with the current power consumption and active power. Step S205: Construct a sparse electrical graph based on the instantaneous feature vector of each electricity meter and the physical topology connection between the electricity meters; Step S206: Analyze the sparse electrical diagram using the second fault detection model, and the output value represents the fault type of the energy meter with fault markers.

Images