A multi-dimensional data quality improvement method for non-intrusive load monitoring

Abstract

This invention relates to a multidimensional data quality improvement method for non-intrusive load monitoring, comprising: S1, selecting complete historical non-intrusive load monitoring data for temporal randomness, periodic correlation, and parametric correlation analysis; S2, constructing a suitable NILM tensor based on the analysis results and analyzing its low-rank property; S3, constructing an observation tensor for all non-intrusive load acquisition data; S4, performing regularized low-rank tensor completion on the observation tensor; and S5, calculating the relative recovery error evaluation index (RSE). This invention proposes a data recovery method based on low-rank tensor completion to address the problem of missing multiple measurement parameters in NILM. It not only compensates for the low accuracy of completion based solely on temporal correlation when continuous data is missing, but also makes the repaired multi-parameter measurement data more tightly consistent with electrical constraints, effectively improving the accuracy of non-intrusive load monitoring and identification, and has significant practical engineering implications.

Description

Technical Field

[0001] This invention belongs to the field of data quality improvement technology, specifically relating to a multi-dimensional data quality improvement method for non-intrusive load monitoring. Background Technology

[0002] Non-Intrusive Load Monitoring (NILM) is an important means of fine-grained sensing of the load side in new power systems. It is one of the technologies that enables highly flexible interaction between the user side and the power system. High-quality measurement data is the foundation of data-driven non-intrusive load monitoring technology.

[0003] However, non-intrusive load monitoring (NILM) data is susceptible to data loss during acquisition, transmission, and storage, impacting advanced smart grid applications such as demand response and user profiling. This patent addresses the data loss issue during NILM data acquisition and transmission by researching a method for repairing missing NILM data. This aims to improve NILM data quality, thereby enhancing the accuracy of non-intrusive load decomposition and contributing to the timely achievement of the "carbon peak and carbon neutrality" goals. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-dimensional data quality improvement method for non-intrusive load monitoring.

[0005] The technical problem solved by this invention is achieved through the following technical solution:

[0006] A method for improving the quality of multidimensional data in non-intrusive load monitoring, characterized by the following steps:

[0007] S1. Select historical complete non-intrusive load monitoring data for time series randomness, periodic correlation and parametric correlation analysis. The non-intrusive load monitoring data includes current (I), active power (P), power factor (λ) and voltage (U). Data loss may occur during the acquisition and transmission process.

[0008] S2. Based on the analysis results of step S1, construct a NILM tensor with a suitable structure. And analyze its low-rank property;

[0009] S3. Construct the observation tensor for all data acquired through non-intrusive load acquisition.

[0010] S4, observation tensor Perform regularized low-rank tensor completion;

[0011] S5. Calculate the relative recovery error evaluation index RSE to verify the data repair effect.

[0012] Moreover, step S1 specifically includes:

[0013] 1) Temporal randomness: Since the opening and closing of electrical equipment is random over a period of time, and the current on the power consumption side is related to the number of electrical equipment used and the start and stop time, the power factor is related to the electrical characteristics and working status of the electrical equipment. The entry and exit of equipment will also affect the power factor. The change trend of active power on the power consumption side follows the change of current. As the entry and exit (start and stop) of electrical equipment increases and decreases, since the opening and closing of electrical equipment is random due to user behavior, the sampling sequences of current, power factor and active power are random over a period of time. However, the voltage sampling sequence is almost horizontal because the voltage on the low-voltage power supply side in my country is stable at 220V.

[0014] 2) Periodic correlation analysis: Although electricity consumption data is random in time series, it shows good periodicity on a long time scale with a 24-hour period because user electricity consumption behavior is regular on a daily basis. Current, power factor and active power data show consistency with a 24-hour period, while voltage does not show periodicity but shows small fluctuations.

[0015] 3) Parameter correlation: In addition to the correlation in time sequence, the NILM measurement parameters (U, I, λ, P) have electrical correlation: Randomly extract 30 minutes of non-intrusive load monitoring data of current, active power and power factor, and align the time scale of each parameter data. When inductive electrical equipment starts, active power and current increase at the same time, and power factor decreases. Active power, current and power factor change at the same time.

[0016] Moreover, step S2 specifically includes:

[0017] 1) As can be seen from the analysis in step S1, if the multidimensional NILM measurement data is organized into a third-order tensor structure, the temporal correlation of the data itself and the electrical correlation between data can be utilized simultaneously, which is beneficial to improving the data completion capability and accuracy. Therefore, this invention breaks through the traditional single-dimensional data repair method and constructs the high-dimensional NILM data into a third-order tensor structure based on the measurement parameter dimension, temporal dimension and periodic dimension of {U,I,λ,P}. Where T1 is the number of observation days, and T1≥14 is taken to utilize the correlation of electricity consumption behavior over the week; T2 is the number of sampling points within 24 hours, which depends on the sampling frequency and can be selected according to actual engineering needs; T3 is the number of measurement parameters.

[0018] 2) Analyze the third-order tensor Low-rank property:

[0019] ① For tensor expansion matrices Taking the number of days as the row vector, since the electricity consumption behavior of users shows similarity on a daily basis, through clustering transformation, the electricity consumption data for T1 days can be clustered into K categories, where K << T1; if the clustering result is weekdays and rest days, then K = 2; therefore, through clustering transformation, this matrix can be transformed into a low-rank matrix.

[0020] ② For the tensor unfolding matrix Taking the measurement parameters and the number of days as the column vectors, from the analysis in ① above, it can be seen that the data for T1 days can be transformed into K days through clustering transformation; since the measurement parameter T3 dimension is the product of U, I, and λ, the column vectors of P can be represented by the column vectors of U, I, and λ, so the matrix A can be completely represented. (2) The minimum number of column vectors is 3×K, that is, the unfolded matrix A (2) has a rank of 3×K;

[0021] ③ For the unfolded matrix Taking the four-dimensional electrical parameters of U, I, λ, and P as the row vectors, so the matrix A (3) can be completely represented by the row vectors of U, I, and λ, the row vector of P is a redundant vector, and the rank of the matrix A (3) is 3; therefore, the ranks of the unfolding matrices of each tensor are as follows: rank(A (1) ) = K, rank(A (2) ) = 3×K, rank(A (3) ) = 3;

[0022] Since there is a maximum value for the rank of any third-order tensor, which is taken as the minimum value of the product of any two dimensions of the tensor, so the maximum value of the tensor is:

[0023] rank max (A) = T3×T1

[0024] Then there is:

[0025] rank(A1), rank(A2), rank(A3) << rank max (A)

[0026] That is, the ranks of the unfolding matrices of the tensor are all much smaller than the maximum value of the rank of the tensor. Therefore, it can be seen that the NILM third-order tensor has low rank; and when there are missing data in the tensor , due to the destruction of the internal time-domain correlation and electrical correlation of the data, the rank of the tensor increases.

[0027] Moreover, the specific step S3 is: all the collected non-intrusive monitoring data are constructed into a third-order tensor Where T1 is the number of observation days, T2 is the number of sampling points within 24 hours, and T3 is the number of measurement parameters, T3 = 4; the set of locations of observable data is called Ω, and the set of locations of missing values ​​is called Ω. C .

[0028] Moreover, step S4 specifically includes:

[0029] 1) Analyze the measurement parameters {U,I,λ,P} of non-intrusive load monitoring. Since a household circuit is a one-port network, the total active power P of each household should conform to the one-port network power relationship with the terminal voltage and the effective values ​​of the total current U and I, as shown below:

[0030]

[0031] Where: T represents a time period, T>0;

[0032] P is the average power during the time period T, i.e., active power, and the unit is watts (W);

[0033] u is the instantaneous value of the terminal voltage, and the effective value U is 220V;

[0034] i represents the instantaneous value of the total current, in amperes (A).

[0035] λ is the power factor;

[0036] Since the multidimensional measurement parameters conform to the power relationship P=U×I×λ of the energy conversion process of the one-port network, by adding the power relationship of the one-port network as a regularization constraint, the spatial electrical correlation and temporal correlation of the solution are strengthened, the random and structured missing parts are repaired, and the results are more compactly in line with the electrical constraints.

[0037] 2) The regularized low-rank tensor completion model of this patent application is established as follows:

[0038]

[0039] in:

[0040] Ω C To observe the tensor The set of locations of missing values;

[0041] X ::m This is the m-th front slice of the completed tensor (m∈1,2,3,4);

[0042] The asterisk (*) represents the Hadamard product, defined as the element-wise multiplication of matrices of the same dimension, resulting in a matrix with the same dimension as the original matrix. Specifically, after missing elements are repaired, the product should satisfy the condition of the fourth frontal slice matrix (X) of the tensor. ::4) equals the first front slice (X) ::1 ) and the second front slice (X) ::2 ) and the third front slice (X) ::3 The Hadamard product is used to satisfy the power constraints of a one-port network.

[0043] 3) Solve the regularized low-rank tensor completion model, constructing an augmented Lagrangian function to transform the problem with equality constraints into an unconstrained optimization problem, as shown below:

[0044]

[0045] Where: M i (i = 1, 2, 3) is an auxiliary matrix, because X (i) (i = 1, 2, 3) The nuclear norms are interdependent and cannot be optimized individually, so they are introduced;

[0046] Y i (i = 1, 2, 3) is a Lagrange multiplier matrix;

[0047] μ is the penalty factor;

[0048] ν is a Lagrange multiplier;

[0049] 4) Solve iteratively using the alternating direction multiplier method (ADMM).

[0050] Furthermore, the formula for calculating the relative recovery error evaluation index RSE in step S5 is as follows:

[0051]

[0052] in: This is the original, complete tensor;

[0053] To complete the tensor;

[0054] ||·|| F This represents the F-norm.

[0055] The advantages and beneficial effects of this invention are as follows:

[0056] This invention addresses the issue of low load decomposition accuracy caused by missing data during the acquisition and transmission of non-intrusive load monitoring data. It proposes a data recovery method based on low-rank tensor completion to address the missing parameters of various NILM measurements. This method overcomes the limitation of using only one-dimensional time-series data for repair. By analyzing the temporal randomness, periodic autocorrelation, and cross-correlation between parameters of multiple parameter measurement data (voltage U, current I, power factor λ, and active power P) in electricity consumption data, a third-order NILM tensor with measurement parameter dimension, temporal dimension, and periodic dimension is constructed. The one-port network power relationship is used as a constraint condition for the solution space, simultaneously repairing the {U,I,λ,P} multi-parameter measurement data, and using the alternating direction multiplier method to accelerate the solution. It can not only make up for the problem of low accuracy when relying solely on time-series correlation to complete data when continuous data is missing, but also make the repair results of multi-parameter measurement data more closely conform to electrical constraints. This can effectively improve the accuracy of non-intrusive load monitoring and identification, and has good practical engineering significance. Attached Figure Description

[0057] Figure 1 This is a flowchart of the present invention;

[0058] Figure 2 This is a graph showing the NILM measurement data for seven consecutive days according to the present invention.

[0059] Figure 3 This is a graph showing the current, power factor, and active power data of the NILM of this invention over a certain period of time.

[0060] Figure 4 This is a flowchart of the iterative solution using the alternating direction multiplier method of the present invention;

[0061] Figure 5 This is a diagram showing the results of repairing continuously missing data in this invention and the comparison method;

[0062] Figure 6 This is a diagram illustrating the effect of repairing the continuity loss of current data in this invention.

[0063] Figure 7 This is a diagram showing the results of random data missing repair in this invention and the comparison method. Detailed Implementation

[0064] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0065] like Figure 1 As shown, a multi-dimensional data quality improvement method for non-intrusive load monitoring is innovative in that the method comprises the following steps:

[0066] S1. Select historical complete non-intrusive load monitoring data for time series randomness, periodic correlation and parametric correlation analysis. The non-intrusive load monitoring data includes current (I), active power (P), power factor (λ) and voltage (U). Data loss may occur during the acquisition and transmission process.

[0067] Step S1 specifically involves:

[0068] 1) Temporal randomness: Since the opening and closing of electrical equipment is random over a period of time, and the current on the power consumption side is related to the number of electrical equipment used and the start and stop time, the power factor is related to the electrical characteristics and working status of the electrical equipment. The entry and exit of equipment will also affect the power factor. The change trend of active power on the power consumption side follows the change of current. As the entry and exit (start and stop) of electrical equipment increases and decreases, since the opening and closing of electrical equipment is random due to user behavior, the sampling sequences of current, power factor and active power are random over a period of time. However, the voltage sampling sequence is almost horizontal because the voltage on the low-voltage power supply side in my country is stable at 220V.

[0069] 2) Periodic correlation analysis: Although electricity consumption data is random in time series, it shows good periodicity on a long time scale with a 24-hour cycle because user electricity consumption behavior is regular on a daily basis. Figure 2 The graph shows the current, power factor, active power, and voltage data for a certain electricity user over seven consecutive days. The graph indicates that the current, power factor, and active power data curves exhibit consistency with a 24-hour cycle. Meanwhile, from... Figure 2 (d) It can be seen that the voltage does not have a periodicity, but exhibits small fluctuations; the current, power factor, and active power data show consistency with a 24-hour period, while the voltage does not have a periodicity, but exhibits small fluctuations.

[0070] 3) Parameter Correlation: In addition to temporal correlation, NILM measurement parameters (U, I, λ, P) exhibit electrical correlation: 30 minutes of non-intrusive load monitoring data for current, active power, and power factor are randomly selected, and the timescales of each parameter data are aligned. Figure 3 As can be seen within the dashed line, when an inductive electrical device starts up, the active power and current increase at the same time, while the power factor decreases. The active power, current, and power factor change simultaneously.

[0071] S2. Based on the analysis results of step S1, construct a NILM tensor with a suitable structure. And analyze its low-rank property;

[0072] Step S2 specifically involves:

[0073] 1) As analyzed in step S1, if the multi-dimensional NILM measurement data is organized into a third-order tensor structure, the temporal correlation of the data itself and the electrical correlation between the data can be utilized simultaneously, which is beneficial to improving the data completion ability and accuracy. Therefore, the present invention breaks through the traditional single-dimensional data repair method and constructs the NILM high-dimensional data into a third-order tensor structure based on the measurement parameter dimension, temporal dimension, and period dimension of {U, I, λ, P} where T1 is the number of observation days, and T1≥14 is taken to utilize the correlation of electricity consumption behavior between weeks; T2 is the number of sampling points within 24 hours, which depends on the sampling frequency and can be selected according to actual engineering requirements; T3 is the number of measurement parameters;

[0074] 2) Analyze the low-rank property of the third-order tensor :

[0075] ① For the tensor unfolding matrix Taking the number of days as the row vector, since the electricity consumption behavior of users shows similarity on a daily basis, after clustering transformation, the electricity consumption data of T1 days can be clustered into K categories, where K << T1; if the clustering result is weekdays and rest days, then K = 2; therefore, after clustering transformation, this matrix can be transformed into a low-rank matrix.

[0076] ② For the tensor unfolding matrix Taking the measurement parameters and the number of days as the column vectors, according to the analysis in ① above, the data of T1 days can be transformed into K days through clustering transformation; in the T3 dimension of the measurement parameters, since P is the product of U, I, and λ, the column vector of P can be represented by the column vectors of U, I, and λ, so the minimum number of column vectors that can completely represent the matrix A (2) is 3×K, that is, the rank of the unfolding matrix A (2) is 3×K;

[0077] ③ For the unfolding matrix Taking the four-dimensional electrical parameters of U, I, λ, and P as the row vectors, so the matrix A (3) can be completely represented by the row vectors of U, I, and λ, and the row vector of P is a redundant vector, and the rank of the matrix A (3) is 3; therefore, the ranks of the unfolding matrices of each tensor are as follows: rank(A (1) ) = K, rank(A (2) ) = 3×K, rank(A (3) ) = 3;

[0078] Since there is a maximum value for the rank of any third-order tensor, which is taken as the minimum value of the product of any two dimensions of the tensor, the maximum value of the tensor is:

[0079]

[0080] Then there is:

[0081] rank(A1),rank(A2),rank(A3)<<rank max (A)

[0082] tensor The rank average of the expanded matrix is ​​much smaller than the maximum rank of the tensor; therefore, it can be concluded that the NILM third-order tensor... It has low-rank property; while when tensor When data is missing, the rank of the tensor increases because the temporal and electrical correlations within the data are disrupted.

[0083] S3. Construct the observation tensor for all data acquired through non-intrusive load acquisition.

[0084] Step S3 specifically involves: constructing all the collected non-invasive monitoring data into a third-order tensor. Where T1 is the number of observation days, T2 is the number of sampling points within 24 hours, and T3 is the number of measurement parameters, T3 = 4; the set of locations of observable data is called Ω, and the set of locations of missing values ​​is called Ω. C .

[0085] S4, observation tensor Perform regularized low-rank tensor completion;

[0086] Step S4 specifically involves:

[0087] 1) Analyze the measurement parameters {U,I,λ,P} of non-intrusive load monitoring. Since a household circuit is a one-port network, the total active power P of each household should conform to the one-port network power relationship with the terminal voltage and the effective values ​​of the total current U and I, as shown below:

[0088]

[0089] Where: T represents a time period, T>0;

[0090] P is the average power during the time period T, i.e., active power, and the unit is watts (W);

[0091] u is the instantaneous value of the terminal voltage, and the effective value U is 220V;

[0092] i represents the instantaneous value of the total current, in amperes (A).

[0093] λ is the power factor;

[0094] Since the multidimensional measurement parameters conform to the power relationship P=U×I×λ of the energy conversion process of the one-port network, by adding the power relationship of the one-port network as a regularization constraint, the spatial electrical correlation and temporal correlation of the solution are strengthened, the random and structured missing parts are repaired, and the results are more compactly in line with the electrical constraints.

[0095] 2) The regularized low-rank tensor completion model of this patent application is established as follows:

[0096]

[0097] in:

[0098] Ω C To observe the tensor The set of locations of missing values;

[0099] X ::m This is the m-th front slice of the completed tensor (m∈1,2,3,4);

[0100] The asterisk (*) represents the Hadamard product, defined as the element-wise multiplication of matrices of the same dimension, resulting in a matrix with the same dimension as the original matrix. Specifically, after missing elements are repaired, the product should satisfy the condition of the fourth frontal slice matrix (X) of the tensor. ::4 ) equals the first front slice (X) ::1 ) and the second front slice (X) ::2 ) and the third front slice (X) ::3 The Hadamard product is used to satisfy the power constraints of a one-port network.

[0101] 3) Solve the regularized low-rank tensor completion model, constructing an augmented Lagrangian function to transform the problem with equality constraints into an unconstrained optimization problem, as shown below:

[0102]

[0103] Where: M i (i = 1, 2, 3) is an auxiliary matrix, because X (i) (i = 1, 2, 3) The nuclear norms are interdependent and cannot be optimized individually, so they are introduced;

[0104] Y i (i = 1, 2, 3) is a Lagrange multiplier matrix;

[0105] μ is the penalty factor;

[0106] ν is a Lagrange multiplier;

[0107] 4) Solve iteratively using the alternating direction multiplier method (ADMM).

[0108] S5. Calculate the relative recovery error evaluation index RSE to verify the data repair effect. The formula for calculating the relative recovery error evaluation index RSE in step S5 is as follows:

[0109]

[0110] in: This is the original, complete tensor;

[0111] To complete the tensor;

[0112] ||·|| F This represents the F-norm.

[0113] This invention utilizes the publicly available non-intrusive load monitoring dataset iAWE to experimentally verify the data repair performance. The iAWE dataset collects parameters including voltage, current, active power, frequency, and power factor. The acquisition frequencies of the user's electricity meter (hereinafter referred to as the main meter) and the household appliance monitoring device are 1Hz and 6sec, respectively. This experiment selects 21 consecutive days of main meter voltage, current, power factor, and active power data from the iAWE dataset without missing values ​​as experimental sample data. Two modes are set: continuous missing values ​​and high-proportion random missing values, with different missing rates, to verify the effectiveness of the proposed data completion method in handling different data missing situations. Furthermore, KNN, low-rank matrix completion (LRMC), and classical low-rank tensor completion (LRTC) are selected as comparative methods to verify the effectiveness of the proposed method.

[0114] The specific experimental verification of data repair performance is as follows:

[0115] 1) In the continuous missing data recovery experiment, 5, 10, 15, 20, and 25 consecutive time periods with missing measurement parameter data were randomly selected within the total observation period. Each time period lasted 0.5 hours, and 50 sets of experiments were conducted each time. The average value of the missing data recovery error was calculated, and the RSE change results of the proposed method and the three comparative algorithms were obtained as follows: Figure 5 As shown.

[0116] Depend on Figure 5 Experimental results show that the proposed algorithm outperforms the three comparative algorithms in both low-proportion missing data scenarios and high-missing-rate scenarios. Specifically, with 5 consecutive missing data periods (2.5 hours in total), the proposed algorithm achieves a repair error rate of 8.92%. Even with a high missing rate of 12.5 hours (25 consecutive missing data periods), the proposed algorithm maintains an error rate of 27.19%, which is 11.38% lower than KNN, 15.63% lower than LRTC, and 33.81% lower than LRMC. This is further demonstrated by… Figure 6The figure shows the current data restoration effect of the proposed method when there are 10 consecutive missing data periods (a total of 5 hours). The dashed lines in the figure represent the restored data, which is equivalent to the original data. It can be seen that the restored data basically overlaps with the true original data values.

[0117] 2) In the random missing data recovery experiment, the random missing data rate varied from 5% to 60% in increments of 5%. For each missing data rate, 50 experiments were conducted, and the average recovery error was calculated. The RSE changes of the proposed method compared to LRTC, LRMC, KNN, and other similar algorithms are shown below. Figure 7 As shown.

[0118] Analysis of the experimental results shows that, compared with the three comparative algorithms, the proposed method still has the smallest repair error. Specifically, at a low missing rate of 5%, the RSE of the proposed algorithm and LRTC is 7.48%, while that of KNN is 7.65% and LRMC is 8.39%. At a high missing rate of 60%, the RSE of the proposed algorithm is only 25.33%, which is 15.63% lower than LRTC's 40.96%, 20.28% lower than KNN's 45.61%, and 32.59% lower than LRMC's 57.92%. This indicates that the proposed method is not only suitable for continuous missing data completion, but also has a lower completion error than the comparative methods when dealing with random missing data.

[0119] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A method for improving the quality of multidimensional data in non-intrusive load monitoring, characterized in that: The steps of the method are as follows: S1. Select complete historical non-intrusive load monitoring data for time-series randomness, periodic correlation, and parametric correlation analysis. The non-intrusive load monitoring data includes current. I Active power P Power factor λ and voltage U There is a probability of data loss during the collection and transmission process; S2. Based on the analysis results of step S1, construct a NILM tensor with a suitable structure and analyze its low-rank property. S3. Construct the observation tensor ℬ for all data acquired through non-intrusive load acquisition; S4. Perform regularization and low-rank tensor completion on the observation tensor ℬ; S5. Calculate the relative recovery error evaluation index (RSE) to verify the data repair effect; Step S1 specifically involves: 1) Temporal randomness: Since the opening and closing of electrical equipment is random over a period of time, and the current on the power consumption side is related to the number of electrical equipment used and the start and stop time, the power factor is related to the electrical characteristics and working status of the electrical equipment. The entry and exit of equipment will also affect the power factor. The change trend of active power on the power consumption side follows the change of current. As the entry and exit (start and stop) of electrical equipment increases and decreases, the current, power factor and active power sampling sequences are random over a period of time because the opening and closing of electrical equipment is random due to user behavior. However, the voltage sampling sequence shows a horizontal trend because the voltage on the low-voltage power supply side in my country is stable at 220V. 2) Periodic correlation analysis: Although electricity consumption data is random in time series, it shows good periodicity on a long time scale with a 24-hour period because user electricity consumption behavior is regular on a daily basis. Current, power factor and active power data show consistency with a 24-hour period, while voltage does not show periodicity but shows small fluctuations. 3) Parameter correlation: In addition to temporal correlation, NILM measurement parameters ( U , I , λ , P There is an electrical correlation between them: 30 minutes of non-intrusive load monitoring data of current, active power, and power factor are randomly selected and the time scales of each parameter data are aligned. When inductive electrical equipment starts, active power and current increase at the same time, and power factor decreases. Active power, current, and power factor change at the same time. Step S2 specifically involves: 1) As analyzed in step S1, if the multidimensional NILM measurement data is organized into a third-order tensor structure, both the temporal correlation of the data itself and the electrical correlation between data can be utilized simultaneously, which is beneficial to improving data completion capability and accuracy. Therefore, this invention breaks through the traditional single-dimensional data repair method and constructs the high-dimensional NILM data into a structure based on { U , I , λ , P The third-order tensor structure 𝒜∈} with measurement parameter dimension, temporal dimension, and periodic dimension. ,in T 1 represents the number of observation days. T 1≥14, to utilize the correlation of electricity consumption behavior across weeks; T 2 represents the number of sampling points within 24 hours, which depends on the sampling frequency and can be selected according to actual engineering needs; T 3 represents the measurement parameters; 2) Analyze the low-rank property of the third-order tensor 𝒜: ① For tensor expansion matrices A (1) ∈ Using days as row vectors, since users' electricity consumption behavior shows similarity on a daily basis, after clustering transformation, it can be... T Daily electricity consumption data aggregation K kind, K << T 1; If the clustering result is weekdays and rest days, then K =2; therefore, through clustering transformation, this matrix can be transformed into a low-rank matrix; ② For tensor expansion matrices A (2) ∈ Using the measured parameters and the number of days as column vectors, as analyzed in ① above, we know that... T One day's data can be transformed through clustering into K Heaven; Measurement parameters T 3 dimensions and because P for U , I , λ Multiply, therefore P The column vector can be used U , I , λ The column vector representation can therefore completely represent the matrix. A (2) The minimum number of column vectors is 3× K That is, the expanded matrix A (2) The rank is 3× K ; ③ For the expanded matrix A (3) ∈ ,by U , I , λ , P Four-dimensional electrical parameters are row vectors, therefore the matrix A (3) Available U , I , λ Complete representation of row vectors P The row vectors are redundant vectors, and the matrix A (3) The rank of the tensor is 3; therefore, the ranks of each expansion matrix of the tensor are as follows: rank ( A (1) )= K , rank ( A (2) ) = 3 × K , rank ( A (3) )=3; Since the rank of any third-order tensor has a maximum value, which is the minimum value of the product of any two dimensions of the tensor, the maximum value of tensor 𝒜 is: ； Then we have: ； That is, the rank of the tensor 𝒜 expansion matrix is ​​much smaller than the maximum rank of the tensor. Therefore, it can be seen that the NILM third-order tensor 𝒜 has low rank. When there is missing data in the tensor 𝒜, the rank of the tensor increases because the temporal and electrical correlations within the data are destroyed. Step S3 specifically involves: constructing a third-order tensor ℬ∈ from all the collected non-invasive monitoring data. ,in T 1 represents the number of observation days. T 2 represents the number of sampling points within 24 hours. T 3 represents the measurement parameters. T 3=4; and the set of locations of observable data is called Ω, and the set of locations of missing values ​​is called Ω. C ; Step S4 specifically involves: 1) Analysis of measurement parameters for non-intrusive load monitoring { U , I , λ, P Since a household circuit is a one-port network, the total active power of each household is... P With terminal voltage and total current RMS value U , I The power requirements for a single-port network should be as follows: ； in: T For a period of time, T >0; P for T The average power during this period, i.e., active power, is measured in watts (W). W ); u The instantaneous value and effective value of the terminal voltage. U It is 220V; i This is the instantaneous value of the total current, in amperes (A). A ); λ Power factor; Because the multidimensional measurement parameters conform to the power relationship of the one-port network energy conversion process. P = U × I × λ Therefore, by adding the power relationship of the one-port network as a regularization constraint, the electrical and temporal correlation of the solution space is strengthened, the random and structured missing parts are repaired, and the results are made more compact and in line with electrical constraints. 2) The regularized low-rank tensor completion model is established as follows: ； Where: ∑3 n =1 α n =1; Ω C For the set of locations of missing values ​​in the observation tensor ℬ; X ::m To complete the tensor 𝒳 of the first m A frontal slice ( m ∈1,2,3,4). " "This is the Hadamard product, defined as the element-wise multiplication of matrices of the same dimension, resulting in a matrix with the same dimension as the original matrix. In other words, after missing elements are repaired, it should satisfy the condition of the fourth frontal slice matrix of the tensor." X ::4 ) equals the first front slice ( X ::1 ) and the second front slice ( X ::2 ) and the third front slice ( X ::3 The Hadamard product is used to satisfy the power constraints of a one-port network. 3) Solve the regularized low-rank tensor completion model, constructing an augmented Lagrangian function to transform the problem with equality constraints into an unconstrained optimization problem, as shown below: ； in: M i ( i =1, 2, 3) is an auxiliary matrix, because X (i) ( i =1, 2, 3) The nuclear norms are interdependent and cannot be optimized individually, so they are introduced; Y i ( i =1, 2, 3) is a Lagrange multiplier matrix; μ As a penalty factor; 𝜈 For Lagrange multipliers; 4) Solve iteratively using the alternating direction multiplier method (ADMM).

2. The multidimensional data quality improvement method for non-intrusive load monitoring according to claim 1, characterized in that: The formula for calculating the relative recovery error evaluation index RSE in step S5 is as follows: ； Among them: 𝒜 origin This is the original, complete tensor; To complete the tensor; ||·|| F express F Norm.

Images