Power data sparsity compression observation method for non-intrusive load monitoring

Abstract

This invention relates to a power data sparsity compression observation method for non-intrusive load monitoring, comprising the following steps: Step 1, acquiring power monitoring data and determining its electricity consumption behavior pattern; Step 2, initializing wavelet sparse basis for sample data based on simple power behavior patterns; Step 3, based on the determination results of the electricity consumption behavior pattern of the completed training samples and the results after wavelet basis initialization in Steps 1 and 2, performing improved K-SVD training and generating sparse basis; Step 4, training the sample dataset Y based on Step 3 to obtain a sparse dictionary D, and performing power monitoring data compression observation based on the improved sparse basis.

Description

Technical Field

[0001] This invention belongs to the field of data sparsity compression observation technology, and relates to a power data sparsity compression observation method, especially a power data sparsity compression observation method for non-intrusive load monitoring. Background Technology

[0002] As one of the largest carbon emitters, the power industry not only bears the heavy responsibility of ensuring national economic production but also needs to be a pioneer in promoting the low-carbon transformation of social industries. To this end, the State Grid Corporation of China has intensified its efforts in intelligent electricity management for the residential sector. Through advanced metering technologies such as non-intrusive load monitoring, it collects, aggregates, and analyzes residential electricity consumption information in real time. This helps relevant departments to arrange more rational power generation plans and formulate more reasonable marketing strategies. Simultaneously, it can guide electricity users' consumption behavior, reduce energy consumption, and promote the low-carbon transformation of the power grid.

[0003] Because non-invasive real-time monitoring technologies collect large amounts of data, related research uses compressed sensing technology to compress and observe the data at the sampling end, and then reconstructs and further analyzes it in data centers with relatively abundant computing power. However, compressed sensing technology requires data to be sparse in certain transform domains. Researchers have used Fourier transform to achieve effective sparsity for voltage and current data, but power consumption data does not possess sparsity properties under Fourier sparse bases. Other researchers have used wavelet sparse bases and K-SVD complete dictionaries to sparsify power consumption, but wavelet bases lack specificity and have poor sparsity effects. K-SVD improves sparsity to some extent by incorporating learning concepts, but its sparsity effect is often unsatisfactory when power consumption patterns are complex and the data volume is large, and it requires a significant amount of training time.

[0004] Electricity consumption data is an important object of non-intrusive load monitoring. However, current monitoring scenarios are diverse and electricity users have complex electricity consumption behavior patterns. Wavelet bases and K-SVD dictionaries are complete but cannot meet the sparsity requirements and real-time requirements of electricity consumption data. This seriously hinders the deployment of compressed sensing technology and also affects the management of residential electricity consumption and the promotion of low-carbon transformation. Therefore, solving the problem of effective sparsity of residential electricity consumption monitoring data is the key to promoting the low-carbon transformation of the power grid.

[0005] A search revealed no publicly available literature of the same or similar prior art as this invention. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose a power data sparsity compression observation method for non-intrusive load monitoring, which can ensure comprehensive real-time monitoring of residential electricity consumption in NILM scenarios.

[0007] The present invention solves its practical problem by adopting the following technical solution:

[0008] A power data sparsity compression observation method for non-intrusive load monitoring includes the following steps:

[0009] Step 1: Obtain power monitoring data and determine its power consumption behavior pattern;

[0010] Step 2: Initialize the wavelet sparse basis for the sample data based on the simple power behavior pattern;

[0011] Step 3: Based on the results of determining the electricity consumption behavior patterns of the completed training samples and the results after initializing the wavelet basis in Steps 1 and 2, perform improved K-SVD training and generate sparse basis.

[0012] Step 4: Based on the sample dataset Y trained in Step 3, a sparse dictionary D is obtained, and power monitoring data compression observation based on the improved sparse basis is performed.

[0013] Furthermore, the specific steps of step 1 include:

[0014] (1) Collect and obtain the user power dataset Y for sparse basis learning training, where m is the length of each power sample sequence, and n is the number of samples;

[0015] (2) Set y1, y2, ... y n The power data used in the training are fed into the pattern detection algorithm proposed in this invention, which divides the power behavior pattern sample data into simple power behavior pattern sample data and complex power behavior pattern sample data.

[0016] (3) For the power samples y1, y2, ... y participating in the training n The pattern detection algorithm is executed sequentially.

[0017] (4) The pattern flag bit Bit_Y of the a-th sample can be obtained from the return value. a = 0 or 1, and the pattern flag Bit_Y = [Bit_Y1, Bit_Y2, ..., Bit_Y] of all samples can be obtained by looping through them. n If Bit_Y a If Bit_Y is 1, then the a-th power sample belongs to the complex power behavior pattern sample data. a If the value is 0, then the a-th power sample belongs to the simple power behavior pattern sample data.

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

[0019] (1) After obtaining the mode flag bit Bit_Y for each sample in step 1, filter out the samples whose corresponding flag bit is 0, i.e., the simple power behavior mode sample data, and denot them as Y_sample=[y1_sample, y2_sample, ..., y J _sample], where J is the number of sample data for the simple power behavior pattern;

[0020] (2) According to formula (1), the wavelet function of the Harr mother wave function Ψ(t) after scaling transformation and time shift is obtained, Ψ a,b (t):

[0021]

[0022] Where a is any non-zero real number, b is any real number, and according to the wavelet function Ψ a,b (t) can perform continuous wavelet transformation (CWT) on the continuous signal x(t), as shown in equation (2):

[0023]

[0024] (3) By discretizing the parameters of the scaling transform and displacement according to equation (2), the discrete form of the wavelet transform can be obtained, as shown in equation (3):

[0025]

[0026] (4) Arrange the discrete wavelet transform sequence according to certain rules to obtain the DWT sparse basis, denoted as D0; after obtaining the DWT sparse basis D0, use the formula Y_sample=D0α sample Sparsity is achieved for the sample data of each simple power behavior pattern, and α is calculated. sample The average sparsity, denoted as K. MEA And let K up =10*K MEA .

[0027] Furthermore, the specific steps of step 3 include:

[0028] (1) In the sparse coding stage of the K-SVD learning dictionary, before using the coding algorithm to sparsely encode each training sample, the Bit_Y calculated by the pattern detection algorithm in step 1 is used to determine the power behavior pattern of the training sample.

[0029] (2) If the sample data belongs to the complex power behavior pattern sample data, then the K obtained in step 2 will be... upThe maximum allowable sparsity of the encoding algorithm is calculated. If the sample data belongs to a simple power behavior pattern, the wavelet basis D0 obtained in step 2 is used to sparsely represent the sample, and the sparsity K of the sample under the wavelet basis is obtained. i DWT and K i DWT As the maximum allowed sparsity K down ;

[0030] (3) Through the above steps, a sparse dictionary D after learning can be obtained through training samples, thereby making the power observation data of power users sufficiently sparse.

[0031] Furthermore, the specific steps and methods of step 4 are as follows:

[0032] After training the sample dataset Y in step 3 to obtain the sparse dictionary D, let Ψ = D, then the actual collected user power monitoring data x can be analyzed using formula (4). real Compressed observations were performed to obtain the observed value y of the actual power monitoring data. real This significantly reduces the amount of data transmitted for NILM real-time monitoring.

[0033] y real =Φx real =ΦΨα (4)

[0034] In the formula, Φ∈R M×N For the observation matrix, the RIP condition must be satisfied, Ψ∈R N×N To obtain a sparse basis using the method proposed in this invention, α∈R N×1 x is the sparsity coefficient. real =Ψα∈R N×1 For real-time monitoring data of power consumption, y real ∈R M×1 These are compressed sensing observations of power monitoring data. Typically, M << N, therefore the amount of power data observed is much smaller than the actual amount of collected power consumption data.

[0035] Advantages and beneficial effects of the present invention:

[0036] 1. This invention addresses the need for compressed sensing technology in NILM scenarios to compress residential power monitoring data. It proposes an improved KSVD sparsity reduction method for residential power monitoring data. First, a pattern detection algorithm is proposed to identify and classify power data of different patterns into simple power behavior pattern data and complex power behavior pattern data. Then, wavelet transform is used to construct a DWT initial dictionary based on the good sparsity characteristics of simple power behavior pattern data, accelerating dictionary learning efficiency. This initial dictionary is then used to sparsify the simple power behavior pattern data, obtaining its average sparsity. Second, the DWT initial dictionary is used as the initial dictionary for the K-SVD algorithm. Based on the results of the pattern detection algorithm, different power behavior pattern data are differentiated during the K-SVD sparse coding stage, assigning them different sparsities to achieve efficient and accurate sparse coding. Finally, a complete dictionary is obtained after learning by the improved K-SVD algorithm, improving both dictionary learning efficiency and sparsity performance. Finally, the user power monitoring data is sparsed using the complete dictionary obtained from training, and compressed sensing technology is used to compress the observations. The observations are then sent to the data center to achieve accurate reconstruction of the residential power monitoring data and ensure comprehensive real-time monitoring of residential electricity consumption in the NILM scenario.

[0037] 2. This invention fully considers the differences in real-time power monitoring data, utilizes the good sparsity effect of wavelet transform on simple power behavior pattern data to construct a Haar wavelet sparse basis, and uses it as the initial dictionary of the K-SVD algorithm, which accelerates the learning and convergence speed of the dictionary and reduces the computational and time overhead of dictionary training.

[0038] 3. This invention identifies power monitoring data of different modes through the proposed mode detection algorithm and performs differentiated processing in the K-SVD sparse coding stage, assigning different sparsity to different training samples to achieve efficient and accurate sparse coding, improve the sparsity effect, further improve dictionary learning efficiency, and improve the reconstruction accuracy of compressed sensing reconstruction algorithm. Attached Figure Description

[0039] Figure 1 Here is a flowchart of the improved K-SVD complete dictionary generation algorithm;

[0040] Figure 2 This is a diagram illustrating the compressed observation process of residential power monitoring data;

[0041] Figure 3(a) is a graph of real-time monitoring data of residential electricity consumption;

[0042] Figure 3(b) shows the sparsity effect of the wavelet basis;

[0043] Figure 3(c) shows the sparsity effect of the K-SVD complete dictionary;

[0044] Figure 3(d) shows the sparse effect of the method of the present invention;

[0045] Figure 4(a) shows the compressed sensing reconstruction effect under wavelet sparse basis;

[0046] Figure 4(b) shows the compressed sensing reconstruction effect under the K-SVD complete dictionary;

[0047] Figure 4(c) is a diagram showing the compressed sensing reconstruction effect under the method of the present invention;

[0048] Figure 5 This is a graph showing the reconstruction error distribution results of different sparse algorithms;

[0049] Figure 6 This is a schematic diagram illustrating the adaptation of the method of the present invention to different observation matrices. Detailed Implementation

[0050] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings:

[0051] This invention addresses the challenges of effectively sparsifying residential power monitoring data using traditional methods and the difficulty in deploying compressed sensing technology. It proposes an improved sparsification and compressed observation method for KSVD power monitoring data that considers data characteristics. By combining the DWT sparse basis with the proposed pattern detection algorithm, effective sparsity of residential power monitoring data is achieved, ensuring efficient compression and accurate reconstruction of compressed sensing data.

[0052] An improved KSVD power monitoring data sparsification and compression observation method based on mode detection algorithm and DWT sparse basis initialization includes the following steps:

[0053] Step 1: Obtain power monitoring data and determine its power consumption behavior pattern;

[0054] The specific steps of step 1 include:

[0055] (1) Collect and obtain the user power dataset Y for sparse basis learning training, where m is the length of each power sample sequence, and n is the number of samples;

[0056] (2) Set y1, y2, ... y n The power data used in the training are fed into the pattern detection algorithm proposed in this invention, which divides the power behavior pattern sample data into simple power behavior pattern sample data and complex power behavior pattern sample data.

[0057] In this embodiment, the pattern detection algorithm is a behavior pattern identification algorithm proposed for residential power monitoring data. By detecting the rising and falling edges of abrupt changes and setting reasonable detection window width and error tolerance, it can distinguish between simple power behavior patterns and complex power behavior patterns, and can effectively eliminate the influence of abnormal power changes at individual moments.

[0058] (3) For the power samples y1, y2, ... y participating in the training n The pattern detection algorithm is executed sequentially.

[0059] In this embodiment, sample y a The algorithm execution flow is as follows:

[0060]

[0061]

[0062] (4) The pattern flag bit Bit_Y of the a-th sample can be obtained from the return value. a = 0 or 1, and the pattern flag Bit_Y = [Bit_Y1, Bit_Y2, ..., Bit_Y] of all samples can be obtained by looping through them. n If Bit_Y a If Bit_Y is 1, then the a-th power sample belongs to the complex power behavior pattern sample data. a If the value is 0, then the a-th power sample belongs to the simple power behavior pattern sample data.

[0063] Step 2: Initialize the wavelet sparse basis for the sample data based on the simple power behavior pattern;

[0064] The specific steps of step 2 include:

[0065] (1) After obtaining the mode flag bit Bit_Y for each sample in step 1, filter out the samples whose corresponding flag bit is 0, i.e., the simple power behavior mode sample data, and denot them as Y_sample=[y1_sample, y2_sample, ..., y J _sample], where J is the number of sample data for the simple power behavior pattern.

[0066] Since simple power behavior pattern sample data accounts for a large proportion of the total sample data, and the Haar wavelet sparse basis can achieve a good sparsity effect on such data, the Haar wavelet sparse basis is used as the initial dictionary of the K-SVD algorithm to accelerate the learning and convergence speed of the dictionary.

[0067] (2) According to formula (1), the wavelet function of the Harr mother wave function Ψ(t) after scaling transformation and time shift is obtained, Ψ a,b (t):

[0068]

[0069] Where a is any non-zero real number, b is any real number, and according to the wavelet function Ψ a,b (t) can perform continuous wavelet transformation (CWT) on the continuous signal x(t), as shown in equation (2):

[0070]

[0071] (3) By discretizing the parameters of the scaling transform and displacement according to equation (2), the discrete form of the wavelet transform can be obtained, as shown in equation (3):

[0072]

[0073] (4) Arrange the discrete wavelet transform sequence according to certain rules to obtain the DWT sparse basis, denoted as D0; after obtaining the DWT sparse basis D0, use the formula Y_sample=D0α sample Sparsity is achieved for the sample data of each simple power behavior pattern, and α is calculated. sample The average sparsity, denoted as K. MEA And let K up =10*K MEA .

[0074] Step 3: Based on the results of determining the electricity consumption behavior patterns of the completed training samples and the results after initializing the wavelet basis in Steps 1 and 2, perform improved K-SVD training and generate sparse basis.

[0075] After determining the power consumption behavior pattern of the training samples and initializing the wavelet basis in steps 1 and 2, the power sample data is trained using the improved K-SVD algorithm proposed in this invention to obtain the trained sparse basis D.

[0076] The specific steps of step 3 include:

[0077] (1) In the sparse coding stage of the K-SVD learning dictionary, before using the coding algorithm to sparsely encode each training sample, the Bit_Y calculated by the pattern detection algorithm in step 1 is used to determine the power behavior pattern of the training sample.

[0078] (2) If the sample data belongs to the complex power behavior pattern sample data, then the K obtained in step 2 will be... upThe maximum allowable sparsity of the encoding algorithm is calculated. If the sample data belongs to a simple power behavior pattern, the wavelet basis D0 obtained in step 2 is used to sparsely represent the sample, and the sparsity of the sample under the wavelet basis is obtained. And As the maximum allowed sparsity K down .

[0079] In this embodiment, the improved K-SVD complete dictionary generation algorithm based on pattern detection algorithm and DWT sparse basis initialization is as follows: Figure 1 As shown:

[0080] The improved K-SVD complete dictionary generation algorithm flow is as follows:

[0081]

[0082] By following the above steps, a sparse dictionary D can be obtained from the training samples, thus making the power observation data of power users sufficiently sparse.

[0083] Step 4: Based on Step 3, train the sample dataset Y to obtain the sparse dictionary D, and perform power monitoring data compression observation based on the improved sparse basis.

[0084] The specific steps and methods for step 4 are as follows:

[0085] After training the sample dataset Y in step 3 to obtain the sparse dictionary D, let Ψ = D, then the actual collected user power monitoring data x can be analyzed using formula (4). real Compressed observations were performed to obtain the observed value y of the actual power monitoring data. real This significantly reduces the amount of data transmitted for NILM real-time monitoring, and the observation process is as follows: Figure 2 As shown:

[0086] y real =Φx real =ΦΨα (4)

[0087] In the formula, Φ∈R M×N For the observation matrix, the RIP condition must be satisfied, Ψ∈R N×N To obtain a sparse basis using the method proposed in this invention, α∈R N×1 x is the sparsity coefficient. real =Ψα∈R N×1 For real-time monitoring data of power consumption, y real ∈R M×1 These are compressed sensing observations of power monitoring data. Typically, M << N, therefore the amount of power data observed is much smaller than the actual amount of collected power consumption data.

[0088] In this embodiment, after acquiring the observed values ​​of real-time power monitoring data, the compressed power monitoring data y can be transmitted wirelessly. real Transmitted to the data center.

[0089] In the data center, y can be utilized according to formula (5). real The original sparse coefficients α are reconstructed, and then through x real =Ψα enables the reconstruction of residential power monitoring data for further analysis.

[0090] min||α||0s.t. y real =ΦΨα (5)

[0091] To verify the feasibility of the algorithm proposed in this invention in actual deployment, a test was conducted using actual household electricity consumption data as an example.

[0092] First, the acquired historical power monitoring data is used as training samples. Step 1 of the algorithm determines the power consumption behavior pattern for each sample, obtaining the pattern flag bit Bit_Y. Then, step 2 performs wavelet sparse basis initialization, obtaining the initialization dictionary D0. Combined with Bit_Y, simple power behavior pattern sample data Y_sample is selected. Finally, K is calculated using D0 and Y_sample. up .

[0093] After completing steps 1 and 2, use the obtained Bit_Y, D0, and K up Step 3 is performed, ultimately yielding the sparsed complete dictionary D. The sparsity effect of the proposed method is then tested using real-time power monitoring data, and compared with the sparsity effects of wavelet sparse bases and K-SVD complete dictionaries. The results are as follows: Figures 3(a)-3(d) The figure shown is a diagram illustrating the sparse effect of power monitoring data in this invention.

[0094] It can be seen that the method proposed in this invention can achieve a good sparsity effect on power monitoring data, and its sparsity performance is significantly superior to wavelet sparse basis and K-SVD complete dictionary.

[0095] Finally, based on the trained complete dictionary D, step 4 is performed: compressing the power monitoring data using compressed sensing technology on the basis of good data sparsity. Then, the data is reconstructed in the data center using the observation matrix Φ and the trained complete dictionary Ψ (i.e., D). The reconstruction results are compared with those of wavelet sparse basis and K-SVD complete dictionary. The results are as follows: Figures 4(a)-4(c) As shown:

[0096] Corresponding to the above reconstruction results, the reconstruction error distribution results of different sparse algorithms are as follows: Figure 5 As shown.

[0097] Calculations show that the reconstruction error MSE of the proposed algorithm after data normalization is 0.0028, while the reconstruction errors MSE of the K-SVD complete dictionary and wavelet sparse basis are 0.0091 and 0.0119, respectively. It can be seen that compared to the K-SVD complete dictionary and wavelet sparse basis, the proposed algorithm achieves better sparsity for real-time power monitoring data, while also improving the reconstruction accuracy of compressed sensing algorithms.

[0098] After testing the sparsity and reconstruction effects of the power monitoring data, this invention tested the performance of the proposed algorithm from two aspects: efficiency and generality. First, training was performed using 1000, 3000, and 5000 measured power sample data respectively. The time required for the proposed algorithm and the K-SVD algorithm to complete one round of dictionary training was tested, and the results are shown in Table 3-1:

[0099] Table 3-1 Comparison of training time between K-SVD algorithm and the algorithm of this invention

[0100]

[0101] It can be seen that, compared with the K-SVD algorithm, the dictionary learning efficiency of the algorithm proposed in this invention has been significantly improved, and the improvement in dictionary training efficiency is more obvious with the number of training samples.

[0102] Finally, to verify the versatility of the proposed algorithm in practical deployment, different observation matrices were paired with the sparse basis trained by the proposed method to perform compressed observation and reconstruction of power monitoring data. The reconstruction results are as follows: Figure 6 As shown.

[0103] The results above show that the present invention has good adaptability to various observation matrices, especially deterministic observation matrices, with a reconstruction error of only 4.27×10-3.

[0104] In summary, the algorithm proposed in this invention improves the execution efficiency of the algorithm to the greatest extent by differentiating the power data of different behavior patterns and fully considering the characteristics of the data. At the same time, it improves the sparsity effect and reconstruction accuracy of real-time power monitoring data, ensures real-time monitoring of user power data in NILM scenarios, and helps the power industry to achieve low-carbon transformation.

[0105] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A power data sparsity compression observation method for non-intrusive load monitoring, characterized in that: Includes the following steps: Step 1: Obtain power monitoring data and determine its power consumption behavior pattern; Step 2: Initialize the wavelet sparse basis for the sample data based on the simple power behavior pattern; Step 3: Based on the results of determining the electricity consumption behavior patterns of the completed training samples and the results after initializing the wavelet basis in Steps 1 and 2, perform improved K-SVD training and generate sparse basis. Step 4: Based on the sample dataset Y trained in Step 3, a sparse dictionary D is obtained, and power monitoring data compression observation based on the improved sparse basis is performed. The specific steps of step 1 include: (1) Collect and obtain user power datasets for sparse basis learning training. Y ,in , m The length of each power sample sequence, n The number of samples; (2) The power data used in the training are fed into the pattern detection algorithm proposed in this invention, which divides the power data into simple power behavior pattern sample data and complex power behavior pattern sample data. (3) Power samples used in training The pattern detection algorithm is executed sequentially. (4) Based on the return value, the first... a The pattern flag bit_Y of each sample a =0 or 1, obtain the pattern flag bit Bit_Y of all samples by looping through them. Bit_Y = [Bit_Y1, Bit_Y2, ..., Bit_Y] n If Bit_Y a For the 1st case a Each power sample belongs to a complex power behavior pattern sample data, if Bit_Y a If it is 0, then the first a Each power sample belongs to a simple power behavior pattern sample data; The specific steps of step 2 include: (1) After obtaining the mode flag bit Bit_Y for each sample through step 1, filter out the samples whose corresponding flag bit is 0, that is, the simple power behavior mode sample data, and record it as Y _sample=[y1_sample, y2_sample, ..., yJ_sample], where J is the number of sample data for the simple power behavior pattern; (2) The Haar mother wave function is obtained according to formula (1). Ψ ( t Wavelet functions after scaling and time shifting, Ψ a,b ( t ): （1） in, a For any nonzero real number, b For any real number, according to the wavelet function Ψ a,b ( t It can handle continuous signals. x ( t Perform Continuous Wavelet Transformation (CWT), as shown in Equation (2): （2） (3) By discretizing the parameters of the scaling transform and displacement according to equation (2), the discrete form of the wavelet transform can be obtained, as shown in equation (3): (3) (4) Arrange the discrete wavelet transform sequence according to certain rules to obtain the DWT sparse basis, denoted as . D 0; Obtain DWT sparse base D After 0, through the formula Y _sample= D 0 α sample Sparsity is achieved for the sample data of each simple power behavior pattern, and computation is performed. α sample The average sparsity, denoted as K MEA and order K up =10* K MEA ; The specific steps of step 3 include: (1) In the sparse coding stage of the K-SVD learning dictionary, before using the coding algorithm to sparsely encode each training sample, the power behavior pattern of the training sample is determined by the Bit_Y calculated by the pattern detection algorithm in step 1. (2) If the sample data belongs to the complex power behavior pattern sample data, then the data obtained in step 2 will be used as the sample data. K up The maximum allowable sparsity of the encoding algorithm is used for solving; if the sample data belongs to a simple power behavior pattern sample data, then the wavelet basis obtained in step 2 is used. D 0. Perform a sparse representation on the sample and obtain the sparsity of the sample under the wavelet basis. KDWT i and will KDWT i As the maximum allowable sparsity K down ; (3) Through the above steps, the learned sparse dictionary D can be obtained through training samples, thereby making the power observation data of power users sufficiently sparse.

2. The power data sparsity compression observation method for non-intrusive load monitoring according to claim 1, characterized in that: The specific steps and methods for step 4 are as follows: Step 3 involves processing the sample dataset. Y Training to obtain a sparse dictionary D Afterwards, order Ψ =D, then the actual collected user power monitoring data can be obtained through formula (4). x real Compressed observations were conducted to obtain observed values ​​of actual power monitoring data. y real This reduces the amount of data transmitted for NILM real-time monitoring. y real = Φx real =ΦΨα (4) In the formula, Φ ∈R M×N The observation matrix needs to satisfy the RIP condition. Ψ ∈R N×N The sparse base obtained during training. α ∈R N×1 The sparsity coefficient is . x real = Ψα ∈R N×1 For real-time monitoring data of power consumption, y real ∈R M×1 The power monitoring data is compressed sensing observations; M << N, therefore the amount of power data observations is much smaller than the amount of actual collected power consumption data.

Images