Complementation method suitable for multi-type massive power user data

Abstract

The invention relates to a data completion method suitable for multiple types of massive power consumers. The method comprises the following steps: step 1, collecting power consumption data of multiple types of power consumers; 2, decomposing the power consumption data of different users into the sum of a trend term and a periodic term; step 3, cycle term verification is carried out; step 4, performing standard delay transformation on the power consumption data of the single user acquired in the step 1 and the time sequence vectors of the period term and the trend term decomposed in the step 2 to obtain a Hankel tensor; 5, establishing a time sequence smooth constraint and a periodic term constraint; step 6, establishing a comprehensive tensor completion objective function; and 7, solving the comprehensive tensor complementation objective function obtained in the step 6, and outputting complete data through an inverse MDT technology. The method can be suitable for one-dimensional and multi-dimensional data scenes, and the complementation effect is good.

Description

Technical Field

[0001] The present invention belongs to the technical field of power user data supplementation, and relates to a power user data supplementation method, in particular to a method for supplementing multi-type massive power user data. Background Art

[0002] Power data plays a vital role in modern society. It is not only the cornerstone of stable operation and efficient management of power grids, but also a key factor in promoting energy transformation and achieving green and low-carbon development. By collecting and analyzing power data in real time, we can accurately predict power demand, optimize resource allocation, reduce waste, and ensure safe and reliable power supply. At the same time, power data also provides a scientific basis for enterprises to save energy, reduce emissions, and control costs, helping them achieve sustainable development. In the construction of smart cities, power data is an important part of the smart energy system, promoting the formation of energy Internet and improving the efficiency and management level of urban energy utilization. Therefore, the accuracy, timeliness and security of power data are of immeasurable value to economic and social development.

[0003] At present, power data in the power system involves infrastructure data such as power grid lines, substations, switchgear, and the operating status and safety information of the power grid. These data are of great significance for the monitoring, dispatching and troubleshooting of the power grid. In the power market, they also include market transaction data, electricity price data, market subject information, etc. These data reflect the operating status and transaction situation of the power market, and are of great significance for the supervision and regulation of the power market.

[0004] However, due to the following reasons: aging and failure of data collection devices such as electric energy meters lead to data loss; communication is blocked during the transmission of electric power data or is attacked by malicious network or network failure occurs, resulting in data loss; the power time series data collection process is irreversible and has high real-time collection requirements. When the time series database writes data during high load periods, the database cannot handle all concurrent write request operations, resulting in data loss; when the system stores data, system hard disk failure, memory failure or other hardware failure may cause data writing failure, resulting in data loss. Data integrity cannot be guaranteed, resulting in poor performance of data-driven analysis, prediction, identification and other models.

[0005] In addition, the current data completion methods based on machine learning cannot effectively deal with the situation of missing high-dimensional data, and require a large amount of non-missing data pre-training. Data completion based on low-rank tensors and ordinary tensor decomposition models can effectively deal with high-dimensional data, but the repair effect is far lower than that based on machine learning when dealing with one-dimensional data repair problems.

[0006] Therefore, the present invention proposes a method for completing data of multiple types of massive power users, which can solve the above problems.

[0007] After searching, no public documents of the prior art identical or similar to the present invention were found. Summary of the invention

[0008] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a method for completing multi-type massive power user data, which can be applicable to one-dimensional and multi-dimensional data scenarios and has a good completion effect.

[0009] The present invention solves the practical problem by adopting the following technical solutions:

[0010] A method for completing data of multiple types of massive power users includes the following steps:

[0011] Step 1: Collect electricity consumption data of multiple types of electricity users, and extract electricity consumption data of any single user;

[0012] Step 2: Based on the single user power consumption data extracted in step 1, the power consumption data of different single users is decomposed into the sum of trend items and period items based on the time series power data decomposition method;

[0013] Step 3, using autocorrelation calculation and fast Fourier calculation to verify the periodic term data obtained in step 2;

[0014] Step 4: Using the delay embedding transformation technology, the single user power consumption data collected in step 1 and the periodic item and trend item time series vector decomposed in step 2 are subjected to standard delay transformation to obtain a Hankel form tensor;

[0015] Step 5: Based on the trend item and period item data obtained in step 2, establish time series smoothing constraints and period item constraints;

[0016] Step 6: for the Hankel form tensor of the trend item data obtained in step 4, tensor parallel factor decomposition is used to obtain a tensor parallel factor decomposition model and a comprehensive tensor completion objective function is established in combination with the two constraints in step 5, which is suitable for different types of power users;

[0017] Step 7: Use alternating least squares (HALS) to solve the comprehensive tensor completion objective function obtained in step 6 and use the inverse MDT technology to output the complete data.

[0018] Moreover, the specific steps of step 2 include:

[0019] Decompose the user's electricity consumption data into the weighted sum of trend items and period items through moving average;

[0020] Perform additive decomposition on the user power consumption data extracted in step 1:

[0021]

[0022] Where: α = 2N + 1 is the smoothing order, x t The electricity consumption data of power users, s t is the trend term and e t is a periodic item.

[0023] Moreover, the specific method of step 3 is:

[0024] The periodic term data is verified by using autocorrelation calculation and fast Fourier calculation:

[0025]

[0026] Where: e t is the mean of the periodic terms after the series is decomposed, and T is the lag order.

[0027] Moreover, the specific method of step 4 is:

[0028] For single-user electricity consumption data: Assuming that one day is a basic cycle, select n days of data, the data length is L, and L×n vectors can be obtained, where the electricity consumption on the i-th (i=1,…,n) day can be expressed as x i ∈x L×1 ;

[0029] Then, the time series vector is subjected to standard delay transformation, i.e., Hankelization, by using the delay embedding transformation technique to obtain a Hankel form matrix;

[0030]

[0031] Among them, L is the length of the data vector, τ is the delay window size, and fold (τ,L-τ+1) :R τ(L-τ+1) →R τ×(L-τ+1) is a folding operator, δ∈{0,1} τ(L-τ+1)×L is a repeating matrix that satisfies:

[0032] vec(H τ (v))=δx (4)

[0033] By stacking the Hankel matrices established above in the third dimension, a Hankel tensor can be formed.

[0034] Moreover, the specific method of step 5 is:

[0035] By decomposing the trend item and period item data obtained in step 2, establish the time series smoothing constraint: And the periodic term constraint:

[0036] Moreover, the comprehensive tensor completion objective function and constraints of step 6 are:

[0037]

[0038]

[0039] Where: E to is all the values ​​in the period where the missing element is located, m is the number of delays, T is the period, and w is the weight coefficient (w t-mT +…+w t+mT =1), λ e is the regularization coefficient, represents the complete Hankel output tensor, Use 0 and 1 to represent missing and available elements, represents the Kronecker product, with missing terms in Hankel form represented by tensors To fill:

[0040]

[0041] in: ρ=[ρ (1) ,ρ (2) ,...,ρ (N) ] T is the smoothness parameter vector, L1∈R (In-1)×In , In∈R, is the number of elements in the factor vector. The first-order derivative operator is defined as:

[0042]

[0043] In the established objective function:

[0044] Represents observation items and reconstruct the tensor The root mean square error (RMSE) of the values ​​is used to estimate the difference between the two, where

[0045] Time series smoothness regularization term: After the first part of the decomposition of the power consumption data of power users participating in spot trading, the smooth trend part Combined with the special structure of the Hankel tensor, the time series smoothness constraint is added, that is, the values ​​of adjacent elements are close. The present invention adopts the quadratic variation (QV) constraint. The regularization term is a formal expression of the time series smoothness constraint, which is defined as:

[0046]

[0047] (3) It indicates that the data has a strong correlation with the cycle. However, due to the irregular noise in the electricity consumption data and the calculation of the autocorrelation coefficient, the data does not change strictly according to the periodic term. Therefore, a set of suitable parameters W is found according to the periodic value of each non-missing value of the data to fit the missing values ​​in the periodic term.

[0048] R is the rank of the parallel factor decomposition of the trend term. The selection of R needs to fully balance the accuracy and computational complexity. Too low R value can reduce the computational complexity, but will produce a large truncation error and affect the model accuracy; too high R value will increase the computational complexity.

[0049] R is optimized by conducting a truncation error experiment. The specific steps are as follows:

[0050] (1) Upper limit of R max Related to the size of the tensor; For example, R max The value is shown in formula (9):

[0051] R max ≤min(I×J,I×K,J×K) (9)

[0052] Where I, J, and K represent the rows, columns, and number of tensors, respectively.

[0053] (2) Use equation (10) to determine the truncation error Q of the parallel factorization CP :

[0054]

[0055] Where: represents the non-missing index tensor reconstructed from r groups of CP decomposition components. Select the truncation error Q CP The rank size corresponding to the moment when it gradually becomes smaller and becomes relatively stable within a certain range is the rank of the tensor parallel factor decomposition in this paper.

[0056] Moreover, the specific method of step 7 is:

[0057] The alternating least squares method (HALS) is used to solve the optimization problem. According to the HALS algorithm, the following local objective function is considered to be minimized:

[0058]

[0059] in:

[0060] in:

[0061] Optimize the above formula in the following order, update the parameter matrix W, factor vector Adaptive factor g r , :

[0062]

[0063] in: is a subset of the unit vector;

[0064] (1) Update W: Optimize the weight coefficient W. For the purpose of simplifying the calculation and optimizing the parameters, the objective function F is p (W) is simplified to:

[0065]

[0066] Update the parameters, and the update rules are as follows:

[0067]

[0068] Where: η is the learning rate, is a function of W k gradient.

[0069] Find a set of suitable weight coefficients according to E t The expression calculates the missing value of the periodic term. The completed periodic sequence is transformed into a Hankel tensor by delayed embedding

[0070] (2) Update u: In order to prevent the occurrence of non-differentiable conditions during the optimization process, the present invention selects the quadratic variation constraint (QV) and adopts the gradient normalization update form. In order to simplify the symbols, u is used. k Indicates u r (n) The kth update, v k Represents the gradient of the objective function

[0071]

[0072] The above formula is considered to update the gradient on the hypersphere. In order to conveniently express the gradient of the objective function, can be simplified to:

[0073]

[0074] in:

[0075]

[0076] Then the objective function gradient is:

[0077]

[0078] (3) Update g r :The updating problem of the adaptive factor is an unconstrained quadratic optimization problem. The unique solution can be obtained analytically. The objective function can be simplified as follows:

[0079]

[0080] Then g r The update rule is:

[0081]

[0082] Use HALS to iterate the above three steps alternately until the error of the two updated calculations of each variable is less than the set threshold ε, which is generally set to 0.001. Finally, the completed Hankel tensor is obtained Output the complete data by delaying the inverse transform:

[0083]

[0084] in is the Moore–Penrose pseudo-inverse transform, unfold (τ,L-τ+1) fold (τ,L-τ+1) The reverse operation of .

[0085] Advantages and beneficial effects of the present invention:

[0086] 1. The present invention proposes a method for completing data of multiple types of massive power users. First, the data is decomposed into trend terms and periodic terms, and then transformed into Hankel tensors. The tensor parallel factor decomposition model is combined to perform single-dimensional or multi-dimensional data completion.

[0087] 2. The present invention proposes a method for completing data of multiple types of massive power users. Aiming at the missing problem of power consumption data of different power users, the method of time series power consumption data decomposition, delayed embedding transformation, and tensor parallel factor decomposition is used to achieve high-precision data completion of power consumption data of different power users. The present invention can be applied to the completion of different types of power users, not only for collecting data in the power system, but also for power consumption data of users in the power market. It has good expansion capability and low computational complexity. Specifically, the acquired data is first classified to determine the type of power user, and then the data is subjected to novel time series decomposition. Secondly, the decomposed one-dimensional sequence data is transformed into Hankel tensors respectively. For the trend items and period items after decomposition, they are also transformed into Hankel tensors accordingly, and corresponding constraints are established. A tensor parallel factor decomposition model is established for the trend items. Then, the trend items, period items constraints and the established tensor parallel factor decomposition model are combined to establish a comprehensive completion model suitable for different types of power users. Finally, the hierarchical alternating least squares solution is used and the complete data is output by reverse delay transformation. A set of evaluation indicators is used to compare the completion effect of the proposed method with that of the existing method. The results show that the present invention is applicable to one-dimensional and multi-dimensional data scenarios, and has a good completion effect, which proves the good engineering practical application significance of the present method. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 is a process flow chart of the present invention;

[0089] Figure 2 It is a data decomposition diagram of the present invention;

[0090] Figure 3 is a Hankel tensor form transformation diagram of the present invention;

[0091] Figure 4 This is a diagram showing the effect of completing the power consumption data of a single user according to the present invention.

[0092] Figure 5 This is a diagram showing the effect of the load aggregator (multi-user) electricity consumption complementation according to the present invention. DETAILED DESCRIPTION

[0093] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0094] A method for completing data of multiple types of massive power users is proposed. Through multiple steps such as data decomposition, periodic term constraint, smooth trend term constraint, and comprehensive objective function establishment, the missing values ​​of power consumption of different types of power users can be completed with high precision. The method includes the following steps:

[0095] Step 1: Collect electricity consumption data of multiple types of electricity users, and extract electricity consumption data of any single user;

[0096] The residential electricity consumption data, the electricity consumption data of coal plant users in a certain province in China, and the commercial electricity consumption data of a load participating in the Chinese electricity market obtained by the present invention respectively represent three different types of electricity users, namely ordinary residential users, industrial and commercial users, and comprehensive users including multiple entities.

[0097] Analysis of the characteristics of different types of users: Ordinary residential users have a small electricity consumption, with a certain periodicity, and the electricity consumption data has a small fluctuation; industrial and commercial users have a large electricity consumption, and due to the start-up and shutdown, input or withdrawal of large-power units, the electricity consumption data fluctuates greatly and has a strong periodicity; load aggregators can aggregate multiple small and medium-sized electricity users. An aggregator usually includes two or more electricity users, and the changes in electricity consumption data of each electricity user are similar to the characteristics of industrial and commercial users.

[0098] Step 2: Based on the single user power consumption data extracted in step 1, the power consumption data of different single users is decomposed into the sum of trend items and period items based on the time series power data decomposition method;

[0099] The specific method of step 2 is:

[0100] The electricity consumption data of individual users is decomposed into the sum of trend items and period items by using the moving average method;

[0101] In this embodiment, the specific steps of step 2 include:

[0102] A novel time series data decomposition method is proposed: the user power consumption data is decomposed into a weighted sum of trend term and period term through moving average. Figure 2 shown.

[0103] Perform additive decomposition on the user power consumption data extracted in step 1:

[0104]

[0105] Where: α = 2N + 1 is the smoothing order, x t The electricity consumption data of power users, s t is the trend term and e t is a periodic item.

[0106] Step 3, using autocorrelation calculation and fast Fourier calculation to verify the periodic term data obtained in step 2;

[0107] The step 3 can verify the periodic term in step 2 by autocorrelation calculation and fast Fourier transform. The specific method of step 3 is:

[0108] The periodic term data is verified by using autocorrelation calculation and fast Fourier calculation:

[0109]

[0110] Where: e t is the mean of the periodic terms after the series is decomposed, and T is the lag order.

[0111] Step 4: Using the delay embedding transformation technology, the single user power consumption data collected in step 1 and the periodic item and trend item time series vector decomposed in step 2 are subjected to standard delay transformation, that is, Hankel transformation, to obtain a Hankel form tensor;

[0112] The specific method of step 4 is:

[0113] For the single user power consumption data collected in step 1 and the trend item and period item decomposed in step 2, the delayed embedding transformation technology (MDT) is used to transform them into Hankel tensors, such as Figure 3 As shown, the specific steps are as follows:

[0114] For single-user electricity consumption data: Assuming that one day is a basic cycle, select n days of data, the data length is L, and L×n vectors can be obtained, where the electricity consumption on the i-th (i=1,…,n) day can be expressed as x i ∈x L×1 ;

[0115] Then, the time series vector is subjected to a standard delay transformation, that is, Hankelization, using the delay embedding transformation technique to obtain a Hankel form matrix, which can make the data have good characteristics such as low rank or smoothness.

[0116]

[0117] Among them, L is the length of the data vector, τ is the delay window size, and fold (τ,L-τ+1) :R τ(L-τ+1) →R τ×(L-τ+1) is a folding operator, δ∈{0,1} τ(L-τ+1)×L is a repeating matrix that satisfies:

[0118] vec(H τ (v))=δx (4)

[0119] Whether it is residential electricity consumption data in low-voltage areas or electricity consumption data of industrial and commercial users on the medium and high voltage sides, the electricity consumption data of a single user has multiple periodicities and temporal and spatial correlations; the electricity consumption data of users contained in load aggregators in the power market has the characteristics of periodic changes, consistent collection frequency, and multiple subjects participating in power market transactions.

[0120] By stacking the Hankel matrices established above in the third dimension, a Hankel tensor can be formed.

[0121] Step 5: Based on the trend item and period item data obtained in step 2, establish time series smoothing constraints and period item constraints;

[0122] The specific method of step 5 is:

[0123] By decomposing the trend item and period item data obtained in step 2, establish the time series smoothing constraint: And the periodic term constraint:

[0124] Step 6: for the Hankel form tensor of the trend item data obtained in step 4, tensor parallel factor decomposition is used to obtain a tensor parallel factor decomposition model and a comprehensive tensor completion objective function is established in combination with the two constraints in step 5, which is suitable for different types of power users;

[0125] The comprehensive tensor completion objective function and constraint conditions of step 6 are:

[0126]

[0127] Where: E to is all the values ​​in the period where the missing element is located, m is the number of delays, T is the period, and w is the weight coefficient (w t-mT +…+w t+mT =1), λ e is the regularization coefficient, represents the complete Hankel output tensor, Use 0 and 1 to represent missing and available elements, represents the Kronecker product, with missing terms in Hankel form represented by tensors To fill:

[0128]

[0129] in: ρ=[ρ (1) ,ρ (2) ,...,ρ (N) ] T is the smoothness parameter vector, L1∈R (In-1)×In , In∈R, is the number of elements in the factor vector. The first-order derivative operator is defined as:

[0130]

[0131] In the established objective function:

[0132] Represents observation items and reconstruct the tensor The root mean square error (RMSE) of the values ​​is used to estimate the difference between the two, where

[0133] Time series smoothness regularization term: After the first part of the decomposition of the power consumption data of power users participating in spot trading, the smooth trend part Combined with the special structure of the Hankel tensor, the time series smoothness constraint is added, that is, the values ​​of adjacent elements are close. The patent of this invention adopts the quadratic variation (QV) constraint. The regularization term is a formal expression of the time series smoothness constraint, which is defined as:

[0134]

[0135] (3) It indicates that the data has a strong correlation with the cycle. However, due to the irregular noise in the electricity consumption data and the calculation of the autocorrelation coefficient, the data does not change strictly according to the periodic term. Therefore, a set of suitable parameters W is found according to the periodic value of each non-missing value of the data to fit the missing values ​​in the periodic term.

[0136] R is the rank of the parallel factor decomposition of the trend term. The selection of R needs to fully balance the accuracy and computational complexity. Too low R value can reduce the computational complexity, but will produce a large truncation error and affect the model accuracy; too high R value will increase the computational complexity.

[0137] The present invention optimizes R by conducting a truncation error experiment, and the specific steps are as follows:

[0138] (1) Upper limit of R max Related to the size of the tensor; For example, R max The value is shown in formula (9):

[0139] R max ≤min(I×J,I×K,J×K) (9)

[0140] Where I, J, and K represent the rows, columns, and number of tensors, respectively.

[0141] (2) Use equation (10) to determine the truncation error Q of the parallel factorization CP :

[0142]

[0143] Where: represents the non-missing index tensor reconstructed from r groups of CP decomposition components. Select the truncation error Q CPThe rank size corresponding to the moment when it gradually becomes smaller and becomes relatively stable within a certain range is the rank of the tensor parallel factor decomposition in this paper.

[0144] Step 7, using alternating least squares (HALS) to solve the comprehensive tensor completion objective function obtained in step 6 and outputting the complete data using the inverse MDT technology;

[0145] The specific method of step 7 is:

[0146] The alternating least squares method (HALS) is used to solve the optimization problem. According to the HALS algorithm, the following local objective function is considered to be minimized:

[0147]

[0148] in:

[0149] in:

[0150] Optimize the above formula in the following order, update the parameter matrix W, factor vector Adaptive factor g r , :

[0151]

[0152] in: is a subset of the unit vector;

[0153] (1) Update W: Optimize the weight coefficient W. For the purpose of simplifying the calculation and optimizing the parameters, the objective function F is p (W) is simplified to:

[0154]

[0155] Update the parameters, and the update rules are as follows:

[0156]

[0157] Where: η is the learning rate, is a function of W k gradient.

[0158] Find a suitable set of weight coefficients according to E t The expression calculates the missing value of the periodic term. The completed periodic sequence is transformed into a Hankel tensor by delayed embedding

[0159] (2) Update u: In order to prevent the occurrence of non-differentiable conditions during the optimization process, the present invention selects the quadratic variation constraint (QV) and adopts the gradient normalization update form. In order to simplify the symbols, u is used. k Indicates u r (n) The kth update, v k Represents the gradient of the objective function

[0160]

[0161] The above formula is considered to update the gradient on the hypersphere. In order to conveniently express the gradient of the objective function, can be simplified to:

[0162]

[0163] in:

[0164]

[0165] Then the objective function gradient is:

[0166]

[0167] (3) Update g r :The updating problem of the adaptive factor is an unconstrained quadratic optimization problem. The unique solution can be obtained analytically. The objective function can be simplified as follows:

[0168]

[0169] Then g r The update rule is:

[0170]

[0171] Use HALS to iterate the above three steps alternately until the error of the two updated calculations of each variable is less than the set threshold ε, which is generally set to 0.001. Finally, the completed Hankel tensor is obtained Output the complete data by delaying the inverse transform:

[0172]

[0173] in is the Moore–Penrose pseudo-inverse transform, unfold (τ,L-τ+1) fold (τ,L-τ+1) The reverse operation of .

[0174] Step 8: Use relative recovery error (RRE) and mean absolute percentage error (MAPE) to verify the effectiveness of the data completion method proposed in the present invention.

[0175] The specific method of step 8 is:

[0176] Relative recovery error, mean absolute percentage error, symmetric mean absolute percentage error and root mean square error are used to evaluate the model recovery performance from different missing types and different missing ratios:

[0177]

[0178] Where: Complete the data of electricity consumption of missing electricity users; The actual data of power quality at the missing location. Figure 4 Give the effect diagram of completing the single user's electricity consumption data. Figure 5 The effect diagram when completing three electricity users (load aggregators) at the same time is given, which proves the effectiveness of the text method for different user types.

[0179] Table 1 shows the RRE results of missing value interpolation using the method of the present invention and tensor CP decomposition completion, low-rank matrix completion, and support vector machine regression.

[0180] Table 1 RRE comparison of different methods (data of a single industrial user)

[0181]

[0182]

[0183] Table 2 shows the MAPE results of missing value interpolation using the method of the present invention and tensor CP decomposition completion, low-rank matrix completion, and support vector machine regression.

[0184] Table 2 Comparison of MAPE of different methods (residential user data set)

[0185]

[0186] Table 3 shows the SMAPE results of the missing value interpolation using the method of the present invention and tensor CP decomposition completion, low-rank matrix completion, and support vector machine regression.

[0187] Table 3 SMAPE comparison of different methods (residential user data set)

[0188]

[0189] Table 4 shows the RMSE results of missing value interpolation using the method of the present invention and tensor CP decomposition completion, low-rank matrix completion, and support vector machine regression.

[0190] Table 4 RMSE comparison of different methods (residential user data set)

[0191]

[0192] It can be seen from the experimental results in Table 1 to Table 4 that the method proposed in the present invention has a good completion effect.

[0193] It should be emphasized that the embodiments of the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation modes. Any other implementation modes derived by those skilled in the art based on the technical solutions of the present invention also fall within the scope of protection of the present invention.

Claims

1. A method for completing data of multiple types of massive power users, characterized by: The steps include: Step 1: Collect electricity consumption data of multiple types of electricity users, and extract electricity consumption data of any single user; Step 2: Based on the single user power consumption data extracted in step 1, the power consumption data of different single users is decomposed into the sum of trend items and period items based on the time series power data decomposition method; Step 3, using autocorrelation calculation and fast Fourier calculation to verify the periodic term data obtained in step 2; Step 4: Using the delay embedding transformation technology, the single user power consumption data collected in step 1 and the periodic item and trend item time series vector decomposed in step 2 are subjected to standard delay transformation to obtain a Hankel form tensor; Step 5: Based on the trend item and period item data obtained in step 2, establish time series smoothing constraints and period item constraints; Step 6: for the Hankel form tensor of the trend item data obtained in step 4, tensor parallel factor decomposition is used to obtain a tensor parallel factor decomposition model and a comprehensive tensor completion objective function is established in combination with the two constraints in step 5, which is suitable for different types of power users; Step 7: Use alternating least squares (HALS) to solve the comprehensive tensor completion objective function obtained in step 6 and use the inverse MDT technology to output the complete data.

2. A method for completing data of multiple types of massive power users according to claim 1, characterized in that: The specific steps of step 2 include: Decompose the user's electricity consumption data into the weighted sum of trend items and period items through moving average; Perform additive decomposition on the user power consumption data extracted in step 1: Where: α = 2N + 1 is the smoothing order, x t The electricity consumption data of power users, s t is the trend term and e t is a periodic item.

3. A method for completing data of multiple types of massive power users according to claim 1, characterized in that: The specific method of step 3 is: The periodic term data is verified by using autocorrelation calculation and fast Fourier calculation: Where: e t is the mean of the periodic terms after the series is decomposed, and T is the lag order.

4. A method for completing data of multiple types of massive power users according to claim 1, characterized in that: The specific method of step 4 is: For single-user electricity consumption data: Assuming that one day is a basic cycle, select n days of data, the data length is L, and L×n vectors can be obtained, where the electricity consumption on the i-th (i=1,…,n) day can be expressed as x i ∈x L×1 ; Then, the time series vector is subjected to standard delay transformation, i.e., Hankelization, by using the delay embedding transformation technique to obtain a Hankel form matrix; Among them, L is the length of the data vector, τ is the delay window size, and fold (τ,L-τ+1) :R τ(L-τ+1) →R τ×(L-τ+1) is a folding operator, δ∈{0,1} τ(L-τ+1)×L is a repeating matrix that satisfies: more(H τ (v))=δx (4) By stacking the Hankel matrices established above in the third dimension, a Hankel tensor can be formed.

5. A method for completing data of multiple types of massive power users according to claim 1, characterized in that: The specific method of step 5 is: By decomposing the trend item and period item data obtained in step 2, establish the time series smoothing constraint: And the periodic term constraint:

6. A method for completing data of multiple types of massive power users according to claim 1, characterized in that: The comprehensive tensor completion objective function and constraints of step 6 are: Where: E to is all the values ​​in the period where the missing element is located, m is the number of delays, T is the period, and w is the weight coefficient (w t-mT +…+w t+mT =1), λ e is the regularization coefficient, represents the complete Hankel output tensor, Use 0 and 1 to represent missing and available elements, represents the Kronecker product, with missing terms in Hankel form represented by tensors To fill: in: ρ=[ρ (1) ,ρ (2) ,...,ρ (N) ] T is the smoothness parameter vector, L1∈R (In -1)×In , In∈R, is the number of elements in the factor vector. The first-order derivative operator is defined as: In the established objective function: Represents observation item and reconstruct the tensor The root mean square error (RMSE) of the values ​​is used to estimate the difference between the two, where Time series smoothness regularization term: After the first part of the decomposition of the power consumption data of power users participating in spot trading, the smooth trend part Combined with the special structure of the Hankel tensor, the time series smoothness constraint is added, that is, the values ​​of adjacent elements are close. The present invention adopts the quadratic variation (QV) constraint. The regularization term is a formal expression of the time series smoothness constraint, which is defined as: (3) Indicates the strong correlation of the data cycle, but because there is irregular noise in the electricity consumption data, and after the autocorrelation coefficient calculation, the data does not strictly change according to the periodic term, so according to the periodic value of each non-vacant value of the data, a set of suitable parameters W is found to fit the missing values ​​in the periodic term; R is the rank of the parallel factor decomposition of the trend term. The selection of R needs to fully balance the accuracy and computational complexity. Too low R value can reduce the computational complexity, but will produce a large truncation error and affect the model accuracy; too high R value will increase the computational complexity. R is optimized by conducting a truncation error experiment. The specific steps are as follows: (1) Upper limit of R max Related to the size of the tensor; For example, R max The value is shown in formula (9): R max ≤min(I×J,I×K,J×K) (9) Where I, J, and K represent the rows, columns, and number of tensors, respectively; (2) Use equation (10) to determine the truncation error Q of the parallel factorization CP : Where: represents the non-missing index tensor reconstructed from r groups of CP decomposition components; the truncation error Q is selected CP The rank size corresponding to the moment when it gradually becomes smaller and becomes relatively stable within a certain range is the rank of the tensor parallel factor decomposition in this paper.

7. A method for completing data of multiple types of massive power users according to claim 1, characterized in that: The specific method of step 7 is: The alternating least squares method (HALS) is used to solve the optimization problem. According to the HALS algorithm, the following local objective function is considered to be minimized: in: in: Optimize the above formula in the following order, update the parameter matrix W, factor vector Adaptive factor g r , : in: is a subset of the unit vector; (1) Update W: Optimize the weight coefficient W. For the purpose of simplifying the calculation and optimizing the parameters, the objective function F is p (W) is simplified to: Update the parameters, and the update rules are as follows: Where: η is the learning rate, is a function of W k The gradient of Find a suitable set of weight coefficients according to E t The expression calculates the missing value of the periodic term; the completed periodic sequence is transformed into a Hankel tensor by delayed embedding (2) Update u: In order to prevent the occurrence of non-differentiable conditions during the optimization process, the present invention selects the quadratic variation constraint (QV) and adopts the gradient normalization update form. In order to simplify the symbols, u is used. k express The kth update, v k Represents the gradient of the objective function The above formula is considered to update the gradient on the hypersphere. In order to conveniently express the gradient of the objective function, can be simplified to: in: Then the objective function gradient is: (3) Update g r :The updating problem of the adaptive factor is an unconstrained quadratic optimization problem. The unique solution can be obtained analytically. The objective function can be simplified as follows: Then g r The update rule is: Use HALS to iterate the above three steps alternately until the error of the two updated calculations of each variable is less than the set threshold ε, which is generally taken as 0.001; finally, the completed Hankel tensor is obtained Output the complete data by delaying the inverse transform: in is the Moore–Penrose pseudo-inverse transform, unfold (τ,L-τ+1) fold (τ,L-τ+1) The reverse operation of .

Images