CN113220671B - Power load missing data restoration method based on power utilization mode decomposition and reconstruction - Google Patents

Abstract

The invention discloses a power load missing data repair method based on power consumption mode decomposition and reconstruction, and relates to the field of power big data analysis and processing. The method first obtains the electricity load data of power users, and divides the data set into a complete load data set and a load data set to be repaired; Centrally extract the base vector dictionary matrix representing the user's electronic mode; then based on the base vector dictionary matrix, decompose and encode the load curve to be repaired, and determine that it is composed of the electronic mode; finally, based on the base vector dictionary matrix, according to the encoding of the load curve to be repaired. The vector reconstructs the load curve, and fills in and repairs the missing part of the power load data. The method of the present invention can be applied to the repair of missing load data in multiple days or in continuous periods.

Description

A method for repairing missing data of power load based on the decomposition and reconstruction of power consumption patterns

technical field

The invention relates to the field of power big data analysis and processing, in particular to a method for repairing missing data of power loads based on the decomposition and reconstruction of power consumption patterns.

Background technique

The widespread popularity of smart meters and the construction of electricity information collection systems provide a data basis for the research and analysis of user-side load big data. However, load data is incomplete due to issues such as meter failure or communication errors. Studying the repair method for missing load data can not only improve data quality, but also the premise of load data analysis, which is of great significance to smart grid and smart electricity consumption. Due to the randomness of users' electricity consumption and the start-stop characteristics of equipment, the data sequence of power load has the characteristics of rapid change and no fixed rules. At the same time, the missing load data can be divided into three missing types: isolated missing, continuous missing and total missing. Conventional interpolation algorithms are not suitable for repairing the continuous distribution of missing load data. Therefore, compared with geospatial data inpainting and image inpainting, load-missing data inpainting is more difficult.

User load data has two main characteristics: sparsity and diversity. Sparsity means that the daily load of a user can basically be linearly composed of several sub-patterns, for example, it can be decomposed into the electricity consumption curve of each device of the user; diversity means that a set of electronic patterns can be reconstructed into different patterns through different codes. Daily load curve. Based on the sparseness and diversity of power user loads, the daily load curve is decomposed into different load sub-patterns by sparse coding technology, and the different load curves are described as a linear combination of sub-patterns to realize load reconstruction, so as to reduce the load loss. data is repaired.

SUMMARY OF THE INVENTION

The technical problem to be solved and the technical task proposed by the present invention are to improve and improve the existing technical solutions, and to provide a power load missing data repair method based on the decomposition and reconstruction of the power consumption mode, so as to realize the continuous missing load data. Effective repair. Therefore, the present invention adopts the following technical solutions.

A method for repairing missing power load data based on the decomposition and reconstruction of power consumption patterns, which is characterized by comprising the steps of:

1) Obtain the electricity load data of power users from the electricity consumption information collection system, and divide the data set into a complete load data set and a to-be-repaired load data set according to whether the daily load data is collected completely;

2) Using the K singular value decomposition dictionary learning algorithm to extract the basis vector dictionary matrix representing the user's electronic mode from the complete load data set;

3) Decompose and encode the load curve to be repaired based on the base vector dictionary matrix, and determine that it is constituted by an electronic mode;

4) Based on the base vector dictionary matrix, the load curve is reconstructed according to the coding vector of the load curve to be repaired, and the missing part of the power load data is filled and repaired, that is, the power consumption data at the corresponding moment in the reconstructed load curve is regarded as the missing part of the load data. fix value.

This technical solution adopts the K-singular value decomposition dictionary learning algorithm, first obtains the electricity load data of power users, and divides the data set into a complete load data set and a to-be-repaired load data set according to whether the daily load data is collected completely. Based on the sparseness and diversity of power user loads, the K-singular value decomposition dictionary learning algorithm is used to extract the basis vector dictionary matrix representing the user's electronic mode from the complete load data set; then, based on the basis vector dictionary matrix, the load curve to be repaired is carried out. Decomposition and coding, determine that it is constituted by electronic mode; finally, based on the base vector dictionary matrix, the load curve is reconstructed according to the coding vector of the load curve to be repaired, and the missing part of the power load data is filled and repaired, that is, the corresponding moment in the reconstructed load curve will be reconstructed The electricity consumption data is used as the repair value for the missing partial load data. In this way, the continuous missing load data can be effectively repaired.

As a preferred technical means: in step 1), the power consumption load data of the power user is obtained from the power consumption information collection system, and the load data is divided into a complete load data set and a load data set to be repaired according to whether the daily load data is collected completely. , the complete daily load collection sample set X _N×M of a certain user can be expressed as:

为第j天的日负荷曲线，是一个N维特征向量；

为全部负荷曲线的第i个采集时刻的功率向量。对于待恢复的负荷曲线x＝[x₁,x₂,…,x_N]^T，

为空缺值，i∈Ω_nan＝{c₁,c₂,...,c_L}，c_l为第l个缺失点的序号，Ω_nan为采集缺失点的序号集合，L为负荷曲线采集缺失的数量。In the formula: N is the daily load collection points; M is the load collection days;

is the daily load curve of the jth day, which is an N-dimensional eigenvector;

is the power vector at the ith acquisition time of all load curves. For the load curve to be restored x=[x ₁ ,x ₂ ,...,x _N ] ^T ,

is the vacancy value, i∈Ω _nan ={c ₁ ,c ₂ ,...,c _L }, c _l is the serial number of the lth missing point, Ω _nan is the set of serial numbers of the missing points, and L is the load curve collection number of missing.

As a preferred technical means: in step 2), the K singular value decomposition dictionary learning algorithm is used to extract the basis vector dictionary matrix representing the user's electronic mode from the complete load data set, and the goal of dictionary learning is to learn a dictionary matrix B such that X _{N ×M} is approximately decomposed into:

X≈BZ

为单位化原子向量，同样为一个M维特征向量；Z＝[z₁,z₂,…,z_M]∈R^K×M为稀疏编码矩阵。在近似分解的同时要满足Z尽可能稀疏，则该近似分解问题的表达式为：In the formula: B∈R ^N×K is the dictionary matrix, K is the size of the dictionary, and each column of B

is a normalized atomic vector, which is also an M-dimensional feature vector; Z=[z ₁ , z ₂ ,...,z _M ]∈R ^K×M is a sparse coding matrix. In the approximate decomposition, Z should be as sparse as possible, the expression of the approximate decomposition problem is:

In the formula: ||·|| _F is the Frobenius norm, and its value is the square sum root of the _matrix elements, indicating the size of the reconstruction error _EB , the smaller the reconstruction error EB, the better the effect of dictionary learning; | |·|| ₀ is the 0 norm, and its value is the number of non-zero items in the matrix; T ₀ is the sparsity constraint threshold, which is used to constrain the number of non-zero items in the coding vector _zi . This formula can be matched by orthogonal matching The tracking algorithm solves.

As the preferred technical means: in step 3), on the basis of using K singular value decomposition algorithm for dictionary learning, the load curve to be repaired is decomposed and encoded based on the basis vector, and it is determined to be formed by electronic mode, and the load curve to be repaired is used to collect The successful load data part and the dictionary matrix of the corresponding time are encoded, and the encoding expression is:

x _/Ω = x-{x _i |i∈Ω _nan }

为B中第i维(行)特征向量；B_/Ω为完整的字典矩阵B去除采集缺失时刻对应特征行向量后的字典矩阵，

z^g为重构向量，为x_/Ω基于B_/Ω分解所得的稀疏编码向量，其值为基于采集成功的负荷数据确定的用电子模式构成，代表了待修复负荷曲线可能的用电模式。In the formula: x _/Ω is the load data collected successfully in the load curve x, and its length is NL;

is the i-th dimension (row) eigenvector in B; B _/Ω is the complete dictionary matrix B is the dictionary matrix after removing the corresponding characteristic row vector at the missing moment,

z ^g is the reconstruction vector, which is the sparse coding vector obtained by x _/Ω based on B _/Ω decomposition.

As a preferred technical means: in step 4), based on the base vector dictionary matrix, the load curve is reconstructed according to the coding vector of the load curve to be repaired, and its expression is:

x ^g = Bz ^g

where x ^g is the reconstruction load curve, which is reconstructed from the reconstruction vector z ^g and the complete dictionary matrix B. On this basis, fill in and repair the missing part of the power load data, that is, the power consumption data at the corresponding moment in the reconstructed load curve is used as the repair value of the missing part of the load data, and its expression is:

为重构负荷x^g中对应采集缺失时刻的负荷数据。in the formula

Collect the load data corresponding to the missing moment in the reconstructed load x ^g .

Beneficial effects:

The invention proposes a method for repairing missing data of power load based on the decomposition and reconstruction of the power consumption mode. First, the electricity load data of power users is obtained, and according to whether the daily load data is collected completely, the data set is divided into a complete load data set and a load data set to be repaired. Based on the sparseness and diversity of power user loads, the K-singular value decomposition dictionary learning algorithm is used to extract the basis vector dictionary matrix representing the user's electronic mode from the complete load data set; then, based on the basis vector dictionary matrix, the load curve to be repaired is carried out. Decomposition and coding, determine that it is constituted by electronic mode; finally, based on the base vector dictionary matrix, the load curve is reconstructed according to the coding vector of the load curve to be repaired, and the missing part of the power load data is filled and repaired, that is, the corresponding moment in the reconstructed load curve will be reconstructed The electricity consumption data is used as the repair value for the missing partial load data. In this way, the continuous missing load data can be effectively repaired. Considering the user's electricity consumption habits and the relatively fixed electricity consumption equipment, the invention divides the user load into several typical electricity consumption patterns based on historical load data, and repairs the missing load data based on the electricity consumption patterns. The electric power load data management personnel can apply the present invention to repairing the missing load data of multiple days or the missing load data of continuous periods according to their actual needs.

Description of drawings

Figure 1 is a flow chart of the present invention.

Figure 2 is the repair result of missing data of load curve 1;

Figure 3 is the repair result of missing data in load curve 2;

Figure 4 is the result of repairing missing data in load curve 3;

Fig. 5 is the actual coding and reconstruction coding result of load curve 1;

Fig. 6 is the actual coding and reconstruction coding result of load curve 2;

Fig. 7 is the actual coding and reconstruction coding result of load curve 3;

Fig. 8 is the corresponding basis vector of actual coding and reconstructed coding of load curve 1;

Fig. 9 is the corresponding basis vector of load curve 2 actual coding and reconstruction coding;

Fig. 10 shows the base vectors corresponding to actual coding and reconstructed coding in load curve 3.

Detailed ways

The technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings.

As shown in FIG. 1, FIG. 1 is the method flow of the present invention: firstly, the power consumption load data of the power user is obtained, and according to whether the daily load data is collected completely, the data set is divided into a complete load data set and a to-be-repaired load data set. Based on the sparseness and diversity of power user loads, the K-singular value decomposition dictionary learning algorithm is used to extract the basis vector dictionary matrix representing the user's electronic mode from the complete load data set; then, based on the basis vector dictionary matrix, the load curve to be repaired is carried out. Decomposition and coding, determine that it is constituted by electronic mode; finally, based on the base vector dictionary matrix, the load curve is reconstructed according to the coding vector of the load curve to be repaired, and the missing part of the power load data is filled and repaired, that is, the corresponding moment in the reconstructed load curve will be reconstructed The electricity consumption data is used as the repair value for the missing partial load data. In this way, the continuous missing load data can be effectively repaired. The specific steps are:

Step 1. Obtain the electricity load data of power users from the electricity consumption information collection system, and divide the load data into complete load data sets and load data sets to be repaired according to whether the daily load data is collected completely. The load collection sample set X _N×M can be expressed as:

为第j天的日负荷曲线，是一个N维特征向量；

为全部负荷曲线的第i个采集时刻的功率向量。对于待恢复的负荷曲线x＝[x₁,x₂,…,x_N]^T，

为空缺值，i∈Ω_nan＝{c₁,c₂,...,c_L}，c_l为第l个缺失点的序号，Ω_nan为采集缺失点的序号集合，L为负荷曲线采集缺失的数量。In the formula: N is the daily load collection points; M is the load collection days;

is the daily load curve of the jth day, which is an N-dimensional eigenvector;

is the power vector at the ith acquisition time of all load curves. For the load curve to be restored x=[x ₁ ,x ₂ ,...,x _N ] ^T ,

is the vacancy value, i∈Ω _nan ={c ₁ ,c ₂ ,...,c _L }, c _l is the serial number of the lth missing point, Ω _nan is the set of serial numbers of the missing points, and L is the load curve collection number of missing.

The K-singular value decomposition dictionary learning algorithm is used to extract the base vector dictionary matrix representing the user's electronic mode from the complete load data set.

Step 2. Use the K singular value decomposition dictionary learning algorithm to extract the base vector dictionary matrix representing the user's electronic mode from the complete load data set. The goal of dictionary learning is to learn a dictionary matrix B, so that X _N×M is approximately decomposed into:

X≈BZ

为单位化原子向量，同样为一个M维特征向量；Z＝[z₁,z₂,…,z_M]∈R^K×M为稀疏编码矩阵。在近似分解的同时要满足Z尽可能稀疏，则该近似分解问题的表达式为：In the formula: B∈R ^N×K is the dictionary matrix, K is the size of the dictionary, and each column of B

is a normalized atomic vector, which is also an M-dimensional feature vector; Z=[z ₁ , z ₂ ,...,z _M ]∈R ^K×M is a sparse coding matrix. In the approximate decomposition, Z should be as sparse as possible, the expression of the approximate decomposition problem is:

In the formula: ||·|| _F is the Frobenius norm, and its value is the square sum root of the _matrix elements, indicating the size of the reconstruction error _EB , the smaller the reconstruction error EB, the better the effect of dictionary learning; | |·|| ₀ is the 0 norm, and its value is the number of non-zero items in the matrix; T ₀ is the sparsity constraint threshold, which is used to constrain the number of non-zero items in the coding vector _zi . This formula can be matched by orthogonal matching The tracking algorithm solves.

Step 3. On the basis of using K singular value decomposition algorithm for dictionary learning, decompose and encode the load curve to be repaired based on the basis vector, determine that it is constituted by electronic mode, and use the load curve to be repaired to collect the successful load data part and the corresponding time The dictionary matrix of , encodes it, and the encoded expression is:

x _/Ω = x-{x _i |i∈Ω _nan }

为B中第i维(行)特征向量；B_/Ω为完整的字典矩阵B去除采集缺失时刻对应特征行向量后的字典矩阵，

z^g为重构向量，为x_/Ω基于B_/Ω分解所得的稀疏编码向量，其值为基于采集成功的负荷数据确定的用电子模式构成，代表了待修复负荷曲线可能的用电模式。In the formula: x _/Ω is the load data collected successfully in the load curve x, and its length is NL;

is the i-th dimension (row) eigenvector in B; B _/Ω is the complete dictionary matrix B is the dictionary matrix after removing the corresponding characteristic row vector at the missing moment,

z ^g is the reconstruction vector, which is the sparse coding vector obtained by x _/Ω based on B _/Ω decomposition.

Step 4. Based on the base vector dictionary matrix, the load curve is reconstructed according to the coding vector of the load curve to be repaired, and its expression is:

x ^g = Bz ^g

where x ^g is the reconstruction load curve, which is reconstructed from the reconstruction vector z ^g and the complete dictionary matrix B. On this basis, fill in and repair the missing part of the power load data, that is, the power consumption data at the corresponding moment in the reconstructed load curve is used as the repair value of the missing part of the load data, and its expression is:

为重构负荷x^g中对应采集缺失时刻的负荷数据。in the formula

Collect the load data corresponding to the missing moment in the reconstructed load x ^g .

The present invention is further described below with specific example:

1. Data sources

The example data is mainly derived from the daily load data of a resident user at 48:00 from May to October 2019. Three days of load curves are randomly selected to construct missing samples, and all of them are set as missing load data at 10 consecutive collection times.

Second, the load missing data repair results

The three load curves are repaired by the technical solution of the present invention, and 100 other complete daily load curves (ie M=100) collected by the user are selected as the training set for dictionary learning, and the dictionary size is set to 20, that is, K =20, set the sparsity constraint threshold T ₀ =5. The repair results of the three load curves are shown in Figure 2, Figure 3, and Figure 4, respectively. The sparse coding of the actual load curve and the reconstructed load curve are shown in Figure 5, Figure 6, and Figure 7, respectively. The corresponding basis vectors are encoded. As shown in Figure 8, Figure 9 and Figure 10, respectively.

It can be seen from Figure 2, Figure 3 and Figure 4 that the algorithm proposed in this chapter can better repair the missing load data regardless of whether the load is flat during the period of data missing or there is a large rise and fall. As can be seen from Figure 5, Figure 6 and Figure 7, for the load curve with missing data, after encoding based on the dictionary, the encoding result is similar to the actual complete load curve encoding result, and the part with larger encoding value is basically the same. , indicating that the missing load data curve can still obtain a code consistent with the actual complete load curve based on the dictionary matrix based on the part of the data collected successfully. Due to the consistency between the reconstructed coding and the original load decomposition coding, the reconstructed load curve based on the reconstructed coding and the complete dictionary is basically consistent with the actual complete load curve, so the missing load data can be repaired. It can be seen from Fig. 8, Fig. 9 and Fig. 10 that the basis vectors corresponding to the maximum coding (ie basis vector 13 in Fig. 8, basis vector 18 in Fig. 9 and basis vector 16 in Fig. 10) are relatively close to the actual complete load curve, while The basis vectors corresponding to the remaining codes are further patched and approximated, and finally the dictionary basis vectors can be used.

The method for repairing missing data of power load based on the decomposition and reconstruction of the power consumption mode shown in FIG. 1 above is a specific embodiment of the present invention, which has embodied the substantial features and progress of the present invention. Under the inspiration of the invention, equivalent modifications in terms of shape and structure are included in the protection scope of this scheme.

Claims (4)

1. A method for repairing missing data of power load based on the decomposition and reconstruction of power consumption mode, is characterized in that, comprises the steps: 1) Obtain the electricity load data of power users from the electricity consumption information collection system, and divide the data set into a complete load data set and a to-be-repaired load data set according to whether the daily load data is collected completely; 2) Using the K singular value decomposition dictionary learning algorithm to extract the basis vector dictionary matrix representing the user's electronic mode from the complete load data set; 3) Decompose and encode the load curve to be repaired based on the base vector dictionary matrix, and determine that it is constituted by an electronic mode; 4) Based on the base vector dictionary matrix, the load curve is reconstructed according to the coding vector of the load curve to be repaired, and the missing part of the power load data is filled and repaired, that is, the power consumption data at the corresponding moment in the reconstructed load curve is regarded as the missing part of the load data. repair value; Specifically, in step 2), the K singular value decomposition dictionary learning algorithm is used to extract the base vector dictionary matrix representing the user's electronic mode from the complete load data set. The goal of dictionary learning is to learn a dictionary matrix B such that X _N×M is approximately decomposed into: X≈BZ In the formula: B∈R ^N×K is the dictionary matrix, K is the size of the dictionary, and each column of B

is a unitized atomic vector, which is also an M-dimensional feature vector; Z=[z ₁ , z ₂ ,...,z _M ]∈R ^K×M is a sparse coding matrix; while approximate decomposition, it is necessary to satisfy Z as sparse as possible, Then the expression of the approximate decomposition problem is:

In the formula: ||·|| _F is the Frobenius norm, and its value is the square sum root of the _matrix elements, indicating the size of the reconstruction error _EB , the smaller the reconstruction error EB, the better the effect of dictionary learning; | |·|| ₀ is the 0 norm, and its value is the number of non-zero items in the matrix; T ₀ is the sparsity constraint threshold, which is used to constrain the number of non-zero items in the coding vector _zi . This formula can be matched by orthogonal matching The tracking algorithm solves. 2. A method for repairing missing data of power load based on power consumption pattern decomposition and reconstruction according to claim 1, characterized in that: in step 1), the power consumption load of the power user is obtained from the power consumption information collection system Data, according to whether the daily load data is collected completely, the load data is divided into a complete load data set and a load data set to be repaired, in which the user's complete daily load collection sample set X _N×M is:

In the formula: N is the daily load collection points; M is the load collection days;

is the daily load curve of the jth day, which is an N-dimensional eigenvector;

is the power vector at the ith acquisition time of all load curves; for the load curve to be restored x=[x ₁ , x ₂ ,...,x _N ] ^T ,

is the missing value, i∈Ω _nan ={c ₁ ,c ₂ ,…,c _L }, c _l is the serial number of the lth missing point, Ω _nan is the serial number set of the missing points collected, L is the missing point collected from the load curve quantity. 3. a kind of power load missing data repair method based on power consumption pattern decomposition and reconstruction according to claim 1, is characterized in that: in step 3), on the basis of adopting K singular value decomposition algorithm to carry out dictionary learning, based on The base vector decomposes and encodes the load curve to be repaired, determines that it is constituted by electronic mode, and uses the load data part of the load curve to be repaired successfully collected and the dictionary matrix at the corresponding time to encode it. The encoding expression is:

x _/Ω = x-{x _i |i∈Ω _nan }

In the formula: x _/Ω is the load data collected successfully in the load curve x, and its length is NL;

is the i-th dimension row eigenvector in B; B _/Ω is the complete dictionary matrix B is the dictionary matrix after removing the characteristic row vector corresponding to the missing moment,

z ^g is the reconstruction vector, which is the sparse coding vector obtained by x _/Ω based on B _/Ω decomposition. 4. A method for repairing missing data of power load based on power consumption pattern decomposition and reconstruction according to claim 1, characterized in that: in step 4), based on the base vector dictionary matrix, according to the coding vector of the load curve to be repaired The reconstructed load curve is expressed as: x ^g = Bz ^g In the formula, x ^g is the reconstructed load curve, which is reconstructed from the reconstructed vector z ^g and the complete dictionary matrix B; The electricity consumption data is used as the repair value of the missing partial load data, and its expression is:

in the formula

Collect the load data corresponding to the missing moment in the reconstructed load x ^g .

Images