Non-intrusive load decomposition method based on CEEMDAN and fastica

Abstract

The application discloses a non-invasive load decomposition method based on CEEMDAN and FastICA, and belongs to the technical field of load monitoring. The method comprises the following steps: S1, collecting active power of total load and various single loads and preprocessing; S2, constructing a completely adaptive noise ensemble empirical mode decomposition (CEEMDAN) model, and decomposing total load power; S3, estimating the number of sources based on Bayesian information criterion; S4, performing dimension reduction by using maximum information coefficient (MIC); S5, performing load decomposition by using FastICA blind source separation; and S6, evaluating the approximation degree of decomposed signals and source signals. The application realizes load decomposition from the perspective of signal blind source separation, reduces the cumbersome load information feature extraction, and obtains complete load information through decomposition. Compared with deep learning, the model training time is greatly reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of non-intrusive power load monitoring, and more particularly to a non-intrusive load decomposition method based on CEEMDAN and FastICA. BACKGROUND

[0002] Power consumption is an important part of energy consumption, with the rapid economic growth and scientific and technological progress, the power consumption and demand in China is rising. Through load monitoring technology to obtain the details of the electrical load, on the one hand, it promotes the efficient management of power equipment, which is conducive to energy saving and emission reduction; on the other hand, it is helpful for the operation management and optimization of power grid.

[0003] Non-intrusive load monitoring (NILM) technology refers to installing a sensor at the user's inlet, monitoring the power consumption and working state of each or each type of electrical appliance in the house by collecting and analyzing the user's total port data, so as to understand the power consumption and power consumption law of each electrical appliance in the house. Compared with intrusive load monitoring, it does not need to install monitoring devices for each electrical appliance to obtain the state and power consumption of the electrical appliance, which is a low-cost, easy-to-deploy and widely-promoted monitoring method, and at the same time, it strengthens the protection of customer privacy in some aspects.

[0004] Non-intrusive load decomposition realizes the decomposition of total power data into single electrical appliance power consumption. At present, the conventional method is to combine the transient and steady-state characteristics of electrical equipment during operation, and to realize load decomposition by using pattern recognition. However, these methods have poor decomposition effect on electrical loads with similar characteristics. In addition, although the introduction of deep learning improves the accuracy of decomposition, it increases network parameters with the increase of model network layers, and may cause gradient disappearance and network degradation, as well as long model training time. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art, and to provide a non-intrusive load decomposition method based on CEEMDAN and FastICA, which decomposes the load information from the perspective of signal separation, reduces the tedious load information feature extraction, and obtains complete load information.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is: a non-intrusive load decomposition method based on CEEMDAN and FastICA, which comprises the following steps:

[0007] S1: collecting the active power of total load and each type of single load within a period of time, and using a sliding window with a width of w and a step of l to extract power data;

[0008] S2: design a complete adaptive noise set empirical mode decomposition (CEEMDAN) model, decompose the total load active power to obtain a series of modal components IMF;

[0009] S3: calculate the covariance matrix of the signal matrix, and perform singular value decomposition on the covariance matrix, and finally select the source number m using the Bayesian information criterion;

[0010] S4: analyze the correlation degree of each IMF modal component and the original signal by using the maximum information coefficient (MIC) method, and select the IMF component with a higher score;

[0011] S5: take the selected signal component as the input of FastICA, decompose the complete load power signal, and realize load decomposition;

[0012] S6: evaluate the approximation degree of the decomposed signal and the source load power signal.

[0013] Further, the step S2 is specifically as follows:

[0014] S21: add white noise υ μ (t) of standard normal distribution to the total power signal sequence x(t) for μ times (μ=1, 2, …, I), wherein I is the number of noise addition, and the reconstructed power data sequence x μ (t) is:

[0015] x μ (t)=x(t)+υ μ (t)

[0016] S22: perform the first EMD decomposition on x μ (t) to obtain the component IMF1 and the residual r1(t):

[0017]

[0018] r1(t)=x(t)-IMF1

[0019] S23: add υ μ (t) to r1(t) and perform EMD decomposition to obtain IMF2 and r2(t); after i-1 times of decomposition, add υ i-1 (t) to r μ (t), perform the i-th EMD decomposition on , and the obtained IMF i and r i (t) are:

[0020]

[0021] r i(t) = r i-1 (t) = r i

[0022] S24: r K (t) cannot be decomposed by EMD any more, thus the CEEMDAN decomposition process ends, and K intrinsic mode components IMF i and residual term r K (t) is obtained.

[0023]

[0024] Further, the step S3 is specifically as follows:

[0025] S31: combine the power sequence x(t) and the IMF components obtained by CEEMDAN decomposition into a multi-dimensional signal matrix X = [x(t), imfl, imf2, …, imf k , r(t)], solve the covariance matrix R of the multi-dimensional signal matrix X, and perform singular value decomposition on the matrix R.

[0026] The number of sources is selected using Bayesian information criterion (BIC). The objective of BIC is to find an estimated number k = m of signal sources with minimum cost, 1≤k≤L, where L is the number of non-zero eigenvalues of the covariance matrix R, and m is the dimensionality of the power sequence x(t) implied. The calculation formula of BIC is as follows:

[0027]

[0028]

[0029] wherein λ j is the jth eigenvalue of the covariance matrix R; and N is the length of the covariance matrix data. Take the k corresponding to the minimum BIC(k) as the estimated number m of signal sources, i.e. k = m.

[0030] Further, the step S4 is specifically as follows:

[0031] S41: calculate the maximum information coefficient (MIC), and the calculation expression used is as follows:

[0032]

[0033] Wherein, MIC(P;f) is the maximum information coefficient between the total load power P and each IMF component, p(P,f) is the joint probability between the total load power P and the IMF modal component f, a and b are the number of grid division in the x and y axis direction, B is the upper limit value of the number of grid division, the value range of MIC is [0,1], the larger the value is, the higher the correlation degree between the total power and the corresponding IMF classification is.

[0034] S42: Reserve the IMF with larger MIC of the source signal, the number is the same as the estimated number in step S3, which is used as the input of the FastICA model in step S5.

[0035] Further, the step S6 is specifically as follows:

[0036] The similarity coefficient δ is used to evaluate the model decomposition performance, and the expression is as follows:

[0037]

[0038] In the formula, x and y are two time domain signals; M is the sampling point. When δ=0, x and y are independent of each other; when δ=1, x and y are completely similar. The closer |δ| is to 1, the more similar the decomposition signal is to the source signal, and the better the decomposition effect is. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flow chart of the CEEMDAN algorithm Figure 1 Figure 2 is a model diagram of the non-intrusive load decomposition method based on CEEMDAN and FastICA

[0040] Figure 3 is a flow chart of the CEEMDAN algorithm Figure 2 Figure 4 is a specific embodiment of the non-intrusive load decomposition method based on CEEMDAN and FastICA DETAILED DESCRIPTION

[0041] The application will be further described below in combination with the drawings and embodiments. It should be pointed out that the following embodiments are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application.

[0042] The non-intrusive load decomposition method based on CEEMDAN and FastICA of the application has a flow chart as shown in Figure 1 which can be divided into the following steps:

[0043] S1: Collect the active power of the total load and each single load in a building for a period of time, and extract the power data by using a sliding window with a width of w and a step of l.

[0044] The low-frequency data in the energy decomposition data set (Reference Energy Disaggregation Data, REDD) is selected for the embodiment. The sliding window width w is set to 2400, the step l is set to 50, and the power data is extracted.

[0045] S2: design a completely adaptive noise set empirical mode decomposition (CEEMDAN) model, decompose the total load power sequence to obtain a series of modal components IMF;

[0046] The CEEMDAN decomposition model adds white noise υ μ (t) of standard normal distribution to the total power signal sequence x(t) μ times (μ = 1, 2, …, I), where I is the number of noise addition, and the reconstructed power data sequence x μ (t) is:

[0047] x μ (t) = x(t) + υ μ (t)

[0048] The first EMD decomposition of x μ (t) is performed to obtain the component IMF1 and the residual r1(t):

[0049]

[0050] r1(t) = x(t) - IMF1

[0051] Add υ μ (t) to r1(t) and perform EMD decomposition to obtain IMF2 and r2(t); after i-1 times of decomposition, add υ i-1 (t) to r μ (t), and perform the i-th EMD decomposition on to obtain IMF i and r i (t):

[0052]

[0053] r i (t) = r i-1 (t) - IMF i

[0054] After K times of decomposition, r K (t) cannot be further decomposed by EMD, thus the CEEMDAN decomposition process ends, K intrinsic mode components IMF i and residual term r K (t) are obtained, and the total load power signal sequence is:

[0055]

[0056] The standard deviation of Gaussian white noise in CEEMDAN is 0.2, the number of white noise addition I is set to 100, and the maximum iteration number is 500.

[0057] S3: Calculate the covariance matrix of the signal matrix, and perform singular value decomposition on the covariance matrix, and finally select the source number m using the Bayesian information criterion;

[0058] The power sequence x(t) and the IMF components decomposed by CEEMDAN are combined into a multi-dimensional signal matrix X = [x(t), imf1, imf2, …, imf k , r(t)], the covariance matrix R of the multi-dimensional signal matrix X is solved, and the matrix R is singular value decomposed.

[0059] The Bayesian information criterion (BIC) is used to select the source number. The goal of BIC is to find an estimated number k = m of signal sources with the minimum cost, 1 ≤ k ≤ L, where L is the number of non-zero eigenvalues of the covariance matrix R, and m is the dimensionality implied by the power sequence x(t). The calculation formula of BIC is as follows:

[0060]

[0061]

[0062] In the formula: λ j is the jth eigenvalue of the covariance matrix R; N is the length of the covariance matrix data. Take the k corresponding to the minimum BIC(k) as the estimated number m of signal sources, that is, k = m.

[0063] The maximum information coefficient (MIC) method is used to analyze the correlation degree of each IMF modal component and the original signal, and the IMF component with a higher score is selected;

[0064] Compared with the Pearson correlation coefficient, MIC is more intuitive in exploring the correlation between features, and MIC measures more comprehensively, including linear and nonlinear relationships, while the former can only calculate the linear correlation between features.

[0065] The calculation expression of the maximum information coefficient (MIC) is as follows:

[0066]

[0067] Where MIC(P; f) is the maximum information coefficient between the total load power P and each IMF component, p(P, f) is the joint probability between the total load power P and the IMF modal component f, a and b are the number of grid divisions in the x and y axis directions, B is the upper limit of the number of grid divisions, and the value of MIC ranges from 0 to 1. The larger the value, the higher the correlation between the total power and the corresponding IMF classification.

[0068] Reserve the IMF which is more similar to the source signal MIC, the number is the same as the number of step S3.

[0069] The IMF which is more similar to 1 means the correlation is greater. That is, the m IMF which is most similar to 1 is reserved.

[0070] The selected signal component is taken as the input of FastICA, the complete load power signal is decomposed, and the load decomposition is realized.

[0071] S6: The approximation degree of the decomposed signal to the source load power signal is evaluated.

[0072] The similarity coefficient δ is used to evaluate the model decomposition performance, and the expression is as follows:

[0073]

[0074] In the formula, x and y are two time domain signals; M is the sampling point. When δ = 0, x and y are independent of each other; when δ = 1, x and y are completely similar. The more similar |δ| is to 1, the more similar the decomposed signal is to the source signal, and the better the decomposition effect is.

Claims

1. A non-intrusive load decomposition method based on CEEMDAN and FastICA, characterized in that, The method comprises the following steps: S1, collecting total load and active power of various single loads in a period of time, and performing power data extraction by using a sliding window with a width of w and a step of l; S2, designing a completely adaptive noise set empirical mode decomposition (CEEMDAN) model to decompose the total load active power to obtain a series of modal components IMF; S3, calculating a covariance matrix of a signal matrix, performing singular value decomposition on the covariance matrix, and finally selecting a source number m by using a Bayesian information criterion; S4, analyzing the correlation degree of each IMF modal component and the original signal by using a maximum information coefficient (MIC) method, and selecting an IMF component with a higher score; S5, taking the selected signal component as an input of FastICA to decompose complete load power signals and realize load decomposition; S6, evaluating the approximation degree of the decomposed signal and the source load power signal.

2. The non-intrusive load decomposition method based on CEEMDAN and FastICA according to claim 1, characterized in that: In step S2, the CEEMDAN decomposes the total load active power, comprising the following steps: S21, add white noise υ of standard normal distribution to the total power signal sequence x(t) for μ (μ = 1, 2, …, I) times μ (t), wherein I is the number of noise addition times, and the reconstructed power data sequence x μ (t) is: x μ (t) = x(t) + v μ (t) S22, to x μ (t) is decomposed into IMF1 and its residual r1(t) as follows: r1(t)=x(t)-IMF1 S23, add υ to r1(t) μ (t) and perform EMD decomposition to obtain IMF2 and r2(t); after decomposition i-1 times, r i-1 (t) and add υ μ (t), perform i-th EMD decomposition on to obtain IMF i and r i (t) is: r i (t) = r i-1 (t) - IMF i S24, r K (t) cannot be decomposed by EMD any more, and the CEEMDAN decomposition process ends, obtaining K intrinsic mode components IMF i and residual term r K (t) is the total load power signal sequence 3. The non-intrusive load decomposition method based on CEEMDAN and FastICA according to claim 1, characterized in that, In step S3, the source number is selected based on the Bayesian information criterion, comprising the following steps: S31, combine the power sequence x(t) and the IMF components decomposed by CEEMDAN into a multi-dimensional signal matrix X=[x(t), imfl, imf2, …, imfn, r(t)], solve the covariance matrix R of the multi-dimensional signal matrix X, and singular value decompose the matrix R; k S31, combine the power sequence x(t) and the IMF components decomposed by CEEMDAN into a multi-dimensional signal matrix X=[x(t), imfl, imf2, …, imfn, r(t)], solve the covariance matrix R of the multi-dimensional signal matrix X, and singular value decompose the matrix R; S32, the Bayesian information criterion (BIC) is used to select the source number, and the target of the BIC is to find an estimated number k=m of signal sources with a minimum cost, 1≤k≤L, wherein L is the number of non-zero eigenvalues of the covariance matrix R, and m is the dimension of the power sequence x(t) implied; the calculation formula of the BIC is as follows: In the formula: λ j Let be the j-th eigenvalue of the covariance matrix R; N is the length of the covariance matrix data; take the k corresponding to the smallest BIC(k) as the estimated number m of the signal sources, i.e., k = m.

4. The non-intrusive load decomposition method based on CEEMDAN and FastICA according to claim 1, characterized in that, In step S4, the maximum information coefficient (MIC) is used for dimension reduction, comprising the following steps: S41, the maximum information coefficient (MIC) is calculated, and the calculation expression is as follows: Wherein, MIC(P;f) is the maximum information coefficient between the total load power P and each IMF component, p(P,f) is the joint probability between the total load power P and the IMF modal component f, a and b are the number of grid division in the x and y axis directions, B is the upper limit value of the grid division number, the value range of MIC is [0, 1], the greater the value, the higher the correlation degree between the total power and the corresponding IMF classification; S42, the IMF with a larger MIC with the source signal is retained, the number is the same as the estimated number in step S3, and is used as the input of the FastICA model in step S5.

5. The non-intrusive load decomposition method based on CEEMDAN and FastICA according to claim 1, characterized in that, In step S6, the approximation degree of the decomposed signal and the source load power signal is evaluated, and the expression is as follows: Wherein, x and y are two time domain signals; M is a sampling point; when δ=0, x and y are independent of each other; when δ=1, x and y are completely similar; the closer δ is to 1, the more similar the decomposed signal and the source signal are, and the better the decomposition effect is.

Images