Cloud edge collaboration based non-intrusive load monitoring data online compressive sensing method

Abstract

The application discloses a kind of non-invasive load monitoring data online compression sensing method based on cloud edge cooperation, the method includes the following steps: step one, data acquisition;Step two, data compression sampling;Step three, sparse representation;Step four, data reconstruction;Step five, load monitoring;Step six, real-time feedback control.The application can learn the dictionary more in line with current signal by online dictionary learning according to the load current signal of real-time acquisition, so as to better adapt to the time-varying nature and nonlinear characteristics of signal.Meanwhile, by using SAMP reconstruction algorithm, signal reconstruction is carried out through sparsity adaptation.The cloud edge cooperation architecture of the application solves the problem of limited hardware computing capacity of local deployment NILM.Meanwhile, the application of online compression sensing method reduces the occupation of communication bandwidth, and improves the ability of signal sparse representation and the accuracy of reconstructed signal.

Description

Technical Field

[0001] This invention relates to a cloud-edge collaborative compressed sensing method, specifically a non-intrusive load monitoring architecture based on cloud-edge collaboration, and an online compressed sensing method through online dictionary learning and sparsity adaptive reconstruction. Background Technology

[0002] Currently, load monitoring has significant application value in power systems, energy management, and smart grids. By collecting and analyzing load information in the power system, load monitoring helps power companies and energy managers understand electricity consumption, optimize energy supply, improve energy efficiency, reduce energy costs, and decrease environmental pollution.

[0003] However, traditional intrusive load monitoring (ILM) methods typically require the installation of numerous sensors or monitoring devices in the power system. This not only increases deployment costs and complexity but may also interfere with the normal operation of equipment within the system. Therefore, using a non-intrusive load monitoring (NILM) method that can achieve accurate monitoring and analysis of load information without the need for additional sensors or equipment has significant practical application value.

[0004] Currently, several load monitoring methods have been proposed and applied in practical systems. For example, the NILM method, based on traditional sampling and signal processing techniques, obtains monitoring information by sampling bus voltage and current signals, and then processing and analyzing the sampled signals. However, to improve monitoring accuracy, high-frequency sampling signals and complex signal processing algorithms are usually required, leading to high computational complexity and hardware costs. Furthermore, because load signals in power systems typically exhibit high dimensionality, non-sparseness, and nonlinearity, traditional lightweight machine learning methods have limitations when processing complex load signals.

[0005] In recent years, distributed NILM system architecture based on cloud-edge collaboration has received considerable attention. Uploading high-frequency sampled signals to a cloud computing center and leveraging its powerful computing and storage capabilities to achieve load monitoring is a current hot research direction in NILM. However, the resulting uploading of high-dimensional data from multiple users to the communication network places enormous pressure on the network. Therefore, how to upload high-dimensional data from the edge to the cloud computing center in real time is a key research issue.

[0006] In summary, the current problems with NILM technology are as follows:

[0007] (1) Traditional NILMs are difficult to process high-frequency sampling signals and implement complex signal processing algorithms on sensor devices (such as electricity meters), while accurate NILMs have high computational complexity and greatly increase the configuration requirements of hardware devices.

[0008] (2) When using the cloud-edge distributed NILM architecture, the communication network is difficult to accommodate the upload requirements of high-dimensional data from multiple users;

[0009] (3) Bus current signals usually have characteristics such as time-varying, high-dimensional, non-sparse and nonlinear. Traditional compressed sensing methods may have certain limitations when processing complex bus current signals. Summary of the Invention

[0010] To address the aforementioned problems in the background technology, this invention provides a non-intrusive online compressed sensing method for load monitoring data based on cloud-edge collaboration. This method reduces the communication bandwidth consumed when uploading non-intrusive load data to the cloud computing center and solves the problem that traditional dictionary learning is difficult to apply to variable bus current signals. This invention can be widely applied in power systems, energy management, smart grids, and other fields for real-time monitoring and feedback control of load status.

[0011] The objective of this invention is achieved through the following technical solution:

[0012] A non-intrusive online compressed sensing method for load monitoring data based on cloud-edge collaboration includes the following steps:

[0013] Step 1: Data Collection

[0014] After the signals collected by the sensor or meter are converted and preprocessed according to the protocol, the original signal x that needs to be uploaded is obtained and stored in the embedded system of the sensor or meter.

[0015] Step 2: Data compression and sampling:

[0016] The original signal x is compressed and sampled by the sensor or meter to obtain the compressed observation signal y. Then, the compressed observation signal is uploaded to the cloud computing center through the communication network. The specific steps are as follows:

[0017] Step 21: Using a Gaussian random matrix as the measurement matrix, the input high-dimensional original signal is projected into a low-dimensional space to obtain the compressed observation signal. This process is described using a matrix operation, namely:

[0018] y = Φx

[0019] Where y is the compressed sampled observation signal, Φ is the measurement matrix, and x is the original signal. Different compression ratios can be achieved by choosing different measurement matrix Φ.

[0020] Step 22: Transmit the compressed observation signal y through the communication network to reduce network transmission pressure;

[0021] Step 3: Sparse Representation:

[0022] In the cloud computing center, the K-Singular Value Decomposition (KSVD) algorithm is used to obtain a dictionary based on pre-stored bus voltage and current data. This dictionary is then used as the initial dictionary D0 for online dictionary learning. The online dictionary learning algorithm continuously updates the dictionary based on changing signals to better adapt to the varying bus current signals. The specific steps are as follows:

[0023] Step 3: Construct a local dictionary D0 based on the KSVD algorithm, which will serve as the initial dictionary for hot-starting online dictionary learning. The specific steps for constructing the local dictionary D0 based on the KSVD algorithm are as follows:

[0024] (1) In the sparse coding stage, with a fixed learning dictionary D, the Orthogonal Matching Pursuit (OMP) algorithm is used to calculate the sparse coefficient matrix α that satisfies the optimization conditions. The optimization objective is as follows:

[0025]

[0026] Where D is the learning dictionary, α is the sparse coefficient matrix corresponding to the original signal x, and α i Let K be a sparse vector in matrix α, and K be the dictionary sparsity, where the dictionary satisfies the sparsity condition that the number of non-zero elements is less than K.

[0027] (2) During the dictionary update phase, the sparse dictionary is updated column by column. The dictionary update formula is as follows:

[0028]

[0029] Where d t E represents the t-th column of the sparse dictionary D; t Let x be the error matrix, representing the contribution of the t-th column of the sparse dictionary to the original signal x;

[0030] (3) For E t Perform singular value decomposition to obtain d that minimizes the objective function. t With α t The updated value, in order to ensure that the obtained α t To satisfy the sparse constraints of the optimization problem, α t Remove all 0 elements from E, and... t The corresponding column is processed in the same way to obtain E' t Then for E't Perform singular value decomposition:

[0031] E' t =U∑V T

[0032] (4) Update the dictionary column vector d using the first column of matrix U. t Update the sparse vector α using the product of the first column of matrix V and the first singular value. t Then, the previous α t The removed zero elements are added to the appropriate positions to ensure that α is equal to the value before the update. t The sparsity was not reduced;

[0033] (5) Repeat steps (2) to (4) for each column of the sparse dictionary, and combine the updated columns to obtain the optimal sparse dictionary matrix;

[0034] (6) The entire KSVD dictionary learning process will alternately repeat the sparse coding stage and the dictionary update step until the number of iterations reaches the maximum value, thereby obtaining the local dictionary D0.

[0035] Step 3.2: Use an online dictionary learning algorithm to continuously update the dictionary based on the changing bus current signal to better adapt to the changing signal. The specific steps are as follows:

[0036] (1) Original signal It consists of independent and identically distributed samples, with one sample x input at a time. t Input an initial dictionary constructed using KSVD.

[0037] (2) Initialization Reset previous information;

[0038] (3) For t=1 to T, begin iteration:

[0039] (4) Each time a new sample x is added... t ;

[0040] (5) Sparse coding, calculated using LARS Lasso:

[0041]

[0042] (6) Calculate:

[0043] (7) Calculate D using the dictionary update algorithm below. t And using D t-1 As a hot start for the dictionary:

[0044]

[0045] (8) The iteration stops when t = T;

[0046] (9) Output the updated dictionary D T ;

[0047] Step 4: Data Reconstruction

[0048] In the cloud computing center, the Sparse Adaptive Matching Pursuit (SAMP) compressed sensing algorithm is used to reconstruct the data of the observation signal compressed in step two.

[0049] Step 5: Load Monitoring

[0050] Using the large number of reconstructed signals obtained in step four, load monitoring and analysis are performed through some large-scale deep learning algorithms to obtain the real-time working status and changing trends of each load on the user bus.

[0051] Step Six: Real-time Feedback Control

[0052] Based on the load monitoring results from step five, real-time feedback control and adjustment are implemented, such as providing users with energy-saving suggestions and dynamically adjusting electricity prices.

[0053] Compared with the prior art, the present invention has the following advantages:

[0054] This invention proposes a non-intrusive online compressed sensing method for load data based on cloud-edge collaboration. Through online dictionary learning, a dictionary more suitable for the current signal can be learned based on real-time acquired load current signals, thus better adapting to the time-varying and nonlinear characteristics of the signal and achieving better sparse representation. Simultaneously, the SAMP reconstruction algorithm is employed for sparsity adaptive reconstruction of the signal, improving the accuracy of signal reconstruction. The cloud-edge collaborative architecture of this invention solves the problem of limited computing power in locally deployed NILM hardware. Furthermore, the online compressed sensing method of this invention reduces communication bandwidth consumption and improves the ability to represent signals sparsely and the accuracy of signal reconstruction. Attached Figure Description

[0055] Figure 1 For cloud-edge distributed load monitoring;

[0056] Figure 2 The results are based on the local dictionary KSVD dynamic current sparsity.

[0057] Figure 3 The results show the dynamic current sparsity of the online dictionary KSVD-ODL;

[0058] Figure 4 A comparison of the performance of various reconstruction algorithms under different compression lengths;

[0059] Figure 5 The reconstruction results of KSVD-ODL and SAMP algorithms. Detailed Implementation

[0060] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0061] This invention provides a non-intrusive online compressed sensing method for load monitoring data based on cloud-edge collaboration. This method aims to address the problems of high model complexity, low monitoring accuracy, high cloud-edge communication bandwidth requirements, and poor sparse representation of real-time changing signals by local dictionary learning methods when deploying the NILM algorithm at the edge. By leveraging the distributed working mode of cloud-edge collaboration, complex load monitoring can be deployed and executed in the cloud, while the edge only needs to perform simple data acquisition and compressed sampling, effectively avoiding the insufficient computing power problem caused by edge hardware limitations. Considering the transmission bandwidth limitations of cloud-edge communication, compressed sampling of data at the edge can further improve the real-time performance of the overall system and reduce the high bandwidth consumption during cloud-edge communication.

[0062] For cloud-edge collaborative load monitoring applications, edge meters collaborate with cloud computing centers to complete complex load monitoring tasks. This invention designs a cloud-edge coordinated distributed deployment scheme for NILM (Non-Integrated Numerical Load Monitoring), leveraging the high computing power of the cloud to solve the problems of high model complexity and low monitoring accuracy when deploying the NILM algorithm at the edge in traditional methods. By learning from an online dictionary, the most suitable sparse representation of the signal is obtained, minimizing the bandwidth requirements for cloud-edge communication. A sparseness adaptive matching pursuit algorithm is used to improve the accuracy of signal reconstruction. Experiments verify the performance of the proposed method, demonstrating its rationality.

[0063] in, Figure 1 The dashed box on the left represents the edge. In the method of this invention, to ensure the accuracy of the NILM algorithm, the meter with limited computing power at the edge only performs signal acquisition and compressed sampling, transmitting the compressed signal to... Figure 1The dashed box on the right represents the cloud data computing center. At the cloud computing center, sparse representation and compression reconstruction algorithms are used to reconstruct the compressed signal uploaded from the edge. Simultaneously, an online dictionary learning algorithm is employed to continuously update the dictionary based on real-time signal changes, resulting in a more accurate sparse representation of the signal. The reconstructed signal is then input into the NILM algorithm to monitor the operating status and power consumption of each load connected to each meter at the edge. Finally, control feedback adjustments are made based on the load monitoring results. The specific implementation steps are as follows:

[0064] Step 1: Data Collection

[0065] Select appropriate sensors or meters, such as current transformers and voltage transformers, based on actual needs. After protocol conversion and preprocessing, the signals collected by the sensors or meters are used to obtain the raw signal x to be uploaded and stored in the embedded system of the sensor or meter. It is important to note that the sensor's sampling rate and accuracy must meet actual requirements, while also considering factors such as data storage and processing efficiency and cost. Generally, the meter's sampling rate needs to reach at least 15kHz to ensure distortion-free data sampling, providing sufficient effective features for high-precision load monitoring algorithms. This invention selects voltage and current data from the UK-DALE bus (publicly available dataset) at a 16kHz sampling rate as experimental data to verify the compressed sensing performance of this invention.

[0066] Step 2: Data compression and sampling:

[0067] In this sensor or meter, the original signal x is compressed and sampled to obtain the compressed observation signal y, and then the compressed observation signal is uploaded to the cloud computing center through the communication network.

[0068] In compressed sensing, the measurement matrix used in the compressed sampling stage is crucial for projecting high-dimensional signals into a low-dimensional space. The original signal x is projected onto the measurement matrix to obtain the compressed-sampled observation signal y. This process can be described using matrix operations:

[0069] y = Φx

[0070] Here, Φ is the measurement matrix. Different compression ratios can be achieved by choosing different matrices Φ.

[0071] In this invention, a Gaussian random matrix is ​​used as the measurement matrix to project the input high-dimensional signal into a low-dimensional space, resulting in a compressed observation signal. The compressed observation signal is then transmitted through a communication network, reducing network transmission load.

[0072] Different measurement matrices have different characteristics and application scenarios. Commonly used measurement matrices also include: Bernoulli random matrix (each element of the matrix is ​​independently and randomly sampled from the Bernoulli distribution), Fourier matrix (an orthogonal matrix constructed from Fourier transform matrices), permutation matrix (obtained by randomly permuting the rows of the identity matrix, where each element of the matrix is ​​either 1 or 0), and Toeplitz matrix (a matrix with a special structure where each row or column is composed of a set of identical numbers arranged according to a certain pattern, and its construction method is simple).

[0073] According to the application scenario of this invention, in order to maintain the integrity of high-frequency AC signal information during compressed sampling, this invention selects a Gaussian random matrix as the measurement matrix Φ. The elements of the Gaussian random matrix are independent and identically distributed random variables that satisfy a Gaussian distribution, exhibiting good randomness properties. This ensures a high probability that the measurement matrix satisfies the Restricted Isometry Property (RIP). Simultaneously, the Gaussian random matrix has high discreteness, thereby improving the efficiency of bus current signal reconstruction.

[0074] Step 3: Sparse Representation

[0075] The core premise of compressed sensing theory is the sparsity of data. The higher the proportion of non-zero elements in a signal, the more redundant information it contains, and the larger the space that can be compressed. The sparse representation of a signal can be achieved by projecting the non-sparse signal onto a specific mapping space. After transformation, the signal becomes a sparse signal containing only a small number of non-zero elements, and excess redundant information is eliminated. However, in practical applications, most data does not meet the sparsity condition; therefore, it is necessary to design sparse representation methods that conform to the signal to achieve compressed sensing. Common sparse matrix representations for signals include implicit dictionary sparse representation and explicit dictionary sparse representation.

[0076] Implicit dictionary sparse representation methods transform signals by establishing a fixed mapping model, thus quickly constructing a structured sparse dictionary matrix. Discrete Fourier transform matrices and wavelet transform matrices fall under the category of implicit dictionaries. Implicit dictionary construction employs general signal processing methods, which cannot be tailored to specific original signals, and the sparsity effect of the signal cannot be guaranteed.

[0077] Explicit dictionary sparse representation utilizes machine learning and other techniques to explicitly learn the structure of the dictionary matrix. Commonly used methods for explicit dictionary sparse representation include sparse networks, dictionary minimization, K-Singular Value Decomposition (KSVD), and Online Dictionary Learning (ODL) algorithms. This approach can better adapt to different signal sparsity characteristics, thus obtaining a sparse dictionary more suitable for the signal itself and achieving better sparsity performance. Therefore, this invention constructs a local dictionary based on the KSVD algorithm as the initial dictionary D0 for the online dictionary, and then investigates a signal sparse representation method more suitable for online dictionary learning of time-varying alternating current.

[0078] Step 3: 1. Construct a local dictionary D0 based on the KSVD algorithm, and use it as the initial dictionary for hot start of online dictionary learning.

[0079] Given an original signal x, a sparse representation dictionary is trained using the known original signal x to minimize the reconstruction error between the original signal and the sparsely represented signal. The dictionary learning process is represented by the following formula:

[0080]

[0081] Where D is the learning dictionary, α is the sparse coefficient matrix corresponding to the original signal x, and α i Let be the sparse vectors in matrix α, and K be the sparsity. The dictionary satisfies the sparsity condition that the number of non-zero elements is less than K. The specific steps for constructing the local dictionary D0 based on the KSVD algorithm are as follows:

[0082] (1) In the sparse coding stage, with a fixed dictionary D, the Orthogonal Matching Pursuit (OMP) algorithm is used to calculate the sparse matrix α that satisfies the optimization conditions. The optimization objective is as follows:

[0083]

[0084] (2) During the dictionary update phase, the sparse dictionary is updated column by column. The dictionary update formula is as follows:

[0085]

[0086] Where d t E represents the t-th column of the sparse dictionary D; t Let x be the error matrix, representing the contribution of the t-th column of the sparse dictionary to the original signal x;

[0087] (3) For E t Perform singular value decomposition to obtain d that minimizes the objective function. t With αt The updated value, in order to ensure that the obtained α t To satisfy the sparse constraints of the optimization problem, α t Remove all 0 elements from E, and... t The corresponding column is processed in the same way to obtain E' t Then for E' t Perform singular value decomposition:

[0088] E' t =U∑V T

[0089] (4) Update the dictionary column vector d using the first column of matrix U. t Update the sparse vector α using the product of the first column of matrix V and the first singular value. t Then, the previous α t The removed zero elements are added to the appropriate positions to ensure that α is equal to the value before the update. t The sparsity was not reduced;

[0090] (5) Repeat steps (2) to (4) for each column of the sparse dictionary, and combine the updated columns to obtain the optimal sparse dictionary matrix;

[0091] (6) The entire KSVD dictionary learning process will alternately repeat the sparse encoding stage and the dictionary update step until the number of iterations reaches the maximum value, thereby obtaining the local dictionary D0.

[0092] Step 32: Online dictionary learning:

[0093] An online dictionary learning algorithm is employed to analyze and process the preprocessed data, generating a dictionary that is continuously updated based on changing signals to better adapt to dynamic signals. Some classic dictionary learning algorithms, such as KSVD and OMP, have poor adaptability to real-time changing current signals. Therefore, this invention uses an online dictionary and dictionary update algorithm to adapt to the real-time characteristics of bus current signal changes.

[0094] Online dictionary learning methods can be divided into two stages: sparse encoding and dictionary updating. The specific steps are as follows:

[0095] (1) Assume the original signal x consists of independent and identically distributed samples, and given a training set x = [x1, ..., xn] of length n. n Each time a sample x is input... t .

[0096] (2) In the sparse coding stage, the input sample x is calculated using the OMP algorithm. t The sparse coefficient α corresponding to the current dictionary.

[0097] (3) The empirical cost function can be expressed by the following formula:

[0098]

[0099] Where l(x,D) is the loss function:

[0100]

[0101] In l(x,D), D is the sparse dictionary, α is the sparse decomposition coefficient, and f n (D) can be rewritten as a joint optimization problem involving D and α:

[0102]

[0103] in, To verify the convex set of the constraint matrix,

[0104] (4) In the sparse coding stage, the input sample x is calculated using the OMP algorithm. t Dictionary D at the previous moment t-1 The corresponding sparsity coefficient α t .

[0105]

[0106] (5) During the dictionary update phase, the objective function for dictionary update at time t is:

[0107]

[0108] The above equation takes minimizing the reconstruction error of the sparse coefficients at time t as the optimization objective, α satisfies the sparse constraint condition, and λ represents the regularization parameter.

[0109] With a fixed dictionary, problem (5) above is an l1-regularized linear least squares problem. Since the LARS Lasso algorithm has good performance and stability when processing high-dimensional data, it can handle high-frequency AC signals well. Therefore, this invention uses the LARS Lasso algorithm to solve the least squares problem.

[0110] Specifically, the LARS Lasso algorithm can be viewed as an iterative forward stepwise regression algorithm used to progressively learn the dictionary matrix D and the sparse coding matrix α. The iterative process is as follows:

[0111] 1. Initialize the dictionary matrix D as an identity matrix, the sparse coding matrix α as a zero matrix, and the residual r = xt.

[0112] 2. Calculate the inner product γ of the residual r and the dictionary matrix D. j = <D j, r>, find the atom j with the greatest correlation.

[0113] 3. Calculate the increment γ of feature j. j =sign( <D j ,r>)·min(| <D j r>|, s K ), where s K It is a threshold related to the current regularization parameter λ, used to control the rate at which feature weights increase.

[0114] 4. Add the increment γ to the j-th column of the sparse coding matrix α. j , i.e. α j =D j +γ j The residual r is then updated to r = r - γ. j D j .

[0115] 5. Stop iterating when the number of non-zero elements has reached the preset maximum value or the current correlation can no longer be increased.

[0116] Overall, the LARS Lasso algorithm's iterative process optimizes the model step by step by selecting the feature with the highest correlation to the residual and increasing the feature weight based on the correlation of that feature. It has good interpretability and computational efficiency.

[0117] Combining the LARS Lasso algorithm and comparing the compression and reconstruction effects of different sparse dictionary construction methods, KSVD, which performed best, was selected to construct the initial dictionary. The online dictionary learning model uses minimizing the error at the current time step as its optimization objective to estimate the future dictionary structure. The specific steps are as follows:

[0118] (1) Original signal It consists of independent and identically distributed samples, with one sample x input at a time. t Input an initial dictionary constructed using KSVD.

[0119] (2) Initialization Reset previous information;

[0120] (3) For t=1 to T, begin iteration:

[0121] (4) Each time a new sample x is added... t ;

[0122] (5) Sparse coding, calculated using LARS Lasso:

[0123]

[0124] (6) Calculate:

[0125] (7) Calculate D using the dictionary update algorithm below. t And using D t-1 As a hot start for the dictionary:

[0126]

[0127] (8) The iteration stops when t = T;

[0128] (9) Output the updated dictionary D T .

[0129] In order to speed up the convergence of the algorithm during the dictionary update phase, while ensuring convergence, this invention uses a block coordinate descent algorithm, which selects the direction that minimizes the objective function for updating in each iteration, instead of updating each coordinate axis individually.

[0130] The steps for updating a dictionary are as follows:

[0131] (1) Input dictionary

[0132] (2) Repeat the execution for j = 1 to K:

[0133] (3) Update the j-th column of dictionary D:

[0134]

[0135]

[0136] (4) Until convergence.

[0137] After the dictionary update step is completed, the online dictionary D at time t is obtained. t .

[0138] Step 4: Data Reconstruction

[0139] In the cloud computing center, compressed sensing theory based on the Sparsity Adaptive Matching Pursuit (SAMP) algorithm reconstructs the data from the observed signal compressed in step two, restoring the original compressed data. Several data reconstruction methods based on convex optimization and iterative algorithms can be employed, such as the OMP algorithm based on l1 norm minimization and the ADMM-based iterative algorithm. Comparative experiments demonstrate that the SAMP algorithm performs better in signal reconstruction for the same compression length.

[0140] In this step, after receiving the compressed observation signal y, the signal needs to be reconstructed based on the observation signal y and the learned dictionary D. This can still be transformed into a constrained optimization problem under the l1-norm:

[0141] min ||α||1s.ty=Φx=ΦDα=Θα

[0142] Where Φ is the measurement matrix and Θ is the sensing matrix.

[0143] To adaptively adjust the sparsity of the signal to better adapt to the time-varying bus current signal, this invention uses the SAMP algorithm to solve the problem, with the following steps:

[0144] (1) Input m×n dimensional sensing matrix Θ, n×1 dimensional observation signal y, step size s;

[0145] (2) Initialization: Reconstruct residual r0 = y; Index set Iteration count index t = 1; Sparsity K = s; Block index stage = 1;

[0146] (3) If t≤m, calculate u=|Θ T r t-1 |;

[0147] (4) Select the K largest values ​​in u, and form a column index set by indexing the column indices of Θ corresponding to these values.

[0148] (5) Constructing the candidate set:

[0149] (6) Solving the least squares problem y = Θ t α t :

[0150]

[0151] (7) Calculate:

[0152] (8) Select K elements from the initial support set and denote them as Λ tK The corresponding K column in Θ is denoted as Final episode

[0153] (9) Update residuals:

[0154] (10) If the residual r is less than the specified threshold or t>m, then stop the iteration and reconstruct the... exist If there are non-zero terms, output

[0155] (11) If ||r||2≥||r t-1 ||2, Update block index stage = stage + 1, Update sparsity K = stage × s, Return to step (3);

[0156] (12) If neither of the first two conditions is met, then update the final set. Update residual r t =r, update the iteration index t = t + 1, return to step (3), until the iteration stops;

[0157] The coefficient matrix is ​​obtained through the above steps. Then according to Obtain the reconstructed signal This completes the signal reconstruction in step four.

[0158] Step 5: Load Monitoring

[0159] Load monitoring and analysis are performed based on the reconstructed signal to obtain the real-time status and trend of the load. Classic load monitoring algorithms such as Short-Time Fourier Transform, Wavelet Transform, and Time-Frequency Analysis can be used, as well as deep learning-based algorithms such as Recurrent Neural Networks (RNNs) and Long Short-Time Memory (LSTM) networks. Deploying the load monitoring model in a cloud computing center allows it to learn from diverse data, increasing its robustness, accuracy, and generalization ability.

[0160] Step Six: Real-time Feedback Control

[0161] Based on load monitoring results, corresponding measures are taken for real-time feedback control, such as adjusting the generator output power and regulating the transformer turns ratio, to ensure the stability and security of the power grid supply. In practical applications, this invention can be applied to power grids, industrial control, transportation, and other fields to achieve real-time load monitoring and control, improve power grid efficiency and reliability, reduce energy consumption and emissions, thereby achieving environmental protection and energy conservation goals.

[0162] Effect verification:

[0163] 1. Local dictionary selection

[0164] (1) Comparison of signal compression effects

[0165] To evaluate the performance of the compressed sensing algorithm, Hoyer sparsity was selected as the evaluation index for signal sparsity in this invention. The sparsity calculation formula is as follows:

[0166]

[0167] Where n is the length of the original signal before compression. The closer sparseness(x) is to 1, the stronger the sparsity of signal x. When it equals 1, it means that the signal has only one non-zero element. The closer the result is to 0, the more uniform the distribution of signal x and the weaker the sparsity. When the result equals 0, all elements of the signal are of equal size.

[0168] To verify the performance of the local sparse dictionary D0 of this invention in signal compression and reconstruction, different sparse dictionaries and redundant dictionaries jointly constructed with the identity matrix or Biorthogonal basis wavelet transform were selected and compared with the local dictionary learning method of this invention. Table 1 shows the comparison results of the original signal and the Hoyer sparsity of different sparse dictionaries under different sparse dictionary construction methods.

[0169] Table 1 Sparsity under different sparse dictionary construction methods

[0170]

[0171] As shown in Table 1, the original signal has low sparsity. After being sparsified by sparse dictionaries constructed using different methods, the signal sparsity is significantly improved. Among them, the Hoyer sparsity of the voltage and current signals after sparsification by the KSVD dictionary is close to 1, indicating that the KSVD dictionary has a strong sparsification capability for the original signal and is the optimal choice for the initial dictionary D0 for local dictionary learning.

[0172] (2) Comparison of signal reconstruction effects

[0173] This invention uses Mean Squared Error (MSE), Signal-to-Noise Ratio (SNR), and Signal Energy Preservation Ratio (SERP) to measure the signal reconstruction performance of the method. MSE is a crucial indicator for evaluating the reconstruction error of compressed sensing models. It measures the average error between the original and reconstructed signals; a smaller MSE indicates a closer reconstructed signal to the original signal and a better reconstruction effect.

[0174]

[0175] Where n is the original signal length, x t For the input sample, The sample is used for reconstruction. SNR refers to the signal-to-noise ratio, which is used to measure the quality of the signal. In this invention, SNR is used to evaluate the quality of the reconstructed signal, that is, the comparison between the reconstructed signal and the original signal. The higher the SNR, the smaller the difference between the reconstructed signal and the original signal, and the better the reconstruction effect.

[0176]

[0177] Where x is the original signal, The SERP (Symptom Energy Recovery Rate) is used to reconstruct the signal. It is the ratio of the energy of the reconstructed signal to the energy of the original signal, and is used to measure the ability of a compressed sensing algorithm to recover signal energy. If SERP equals 1, it means that the reconstructed signal has the same energy as the original signal, and the reconstruction effect is better.

[0178]

[0179] Next, the signal reconstruction effects of different sparse dictionary construction methods under local dictionary learning are compared. Table 2 shows the comparison results of MSE, SNR, and SEPR of the reconstructed signals of different sparse dictionaries under different local sparse dictionary construction methods.

[0180] Table 2. Signal reconstruction results under different sparse dictionary construction methods.

[0181]

[0182] As shown in Table 2, the compressed sensing method based on KSVD sparse dictionary learning exhibits the lowest MSE and the highest SNR and SEPR in both voltage and current signal compression and reconstruction. Compared to other comparative methods, the KSVD sparse dictionary demonstrates superior reconstruction performance for power load signals. Therefore, the online dictionary learning in this invention utilizes KSVD to construct the initial dictionary D0.

[0183] 2. Online dictionary performance

[0184] In real-world load scenarios, bus voltage signals are relatively stable, and good compression and reconstruction results can be achieved through offline dictionary training. However, current signals dynamically change with the switching of power loads, exhibiting characteristics such as time-varying, high dimensionality, non-sparseness, and nonlinearity. Offline-trained local dictionaries struggle to handle complex current signal scenarios, and sparsity cannot be guaranteed. Therefore, sparsification of dynamic current signals using an offline-trained local sparse dictionary is crucial. Figure 2 As shown. From Figure 2 As can be seen, the sparsity of the current signal after sparsification does not meet the reconstruction requirements, resulting in the inability to accurately reconstruct the signal and thus affecting the load monitoring calculation results. Therefore, it is necessary to update the current signal sparsity dictionary using an online learning method. The signal sparsification result after online dictionary learning in this invention is as follows: Figure 3 As shown. From Figure 3 As can be seen, after the online dictionary learning algorithm completes its learning, the sparsity of the bus current signal is significantly improved compared to the offline trained local dictionary.

[0185] 3. Reconstruction Algorithm Selection

[0186] To better reconstruct the signal, the optimal current signal reconstruction algorithm was selected through experiments, and the maximum compression limit of the compressed sensing method was verified. The accuracy of 50 current signal reconstructions under different compressed signal lengths was compared. Figure 4 As shown.

[0187] Reconstructed current signals with an MSE less than 0.06 are considered to have accurate reconstruction results. Fifty experiments were conducted at 10-bit intervals for compressed signal lengths ranging from 100 to 600 to compare the accurate reconstruction probability of the bus current signal under different compressed signal lengths. Figure 4 As can be seen, when the compressed length of the current signal exceeds 260, the accurate reconstruction probability of the current signal can reach over 90% under the SAMP algorithm. Figure 4 The accurate reconstruction probability of the bus current signal under different compressed signal lengths shown demonstrates that the signal reconstruction algorithm based on the SAMP algorithm of this invention has high reconstruction capability.

[0188] 4. The effectiveness of the KSVD-ODL and SAMP algorithms of this invention

[0189] After sparse representation of the signal using the KSVD-ODL online dictionary, the bus current signal is reconstructed using the SAMP algorithm. The reconstruction result is as follows: Figure 5 As shown.

[0190] from Figure 3 and 5 As can be seen, the compressed sensing algorithm combining KSVD-ODL and SAMP exhibits excellent compression and reconstruction performance for real-time changing bus current signals. This demonstrates the effectiveness of the method described in this invention for compressing and reconstructing load data under different scenarios.

[0191] It should be noted that the above steps are merely an example of the implementation of this invention, and adjustments and optimizations can be made according to specific needs in actual applications. For example, in terms of meter selection and data acquisition, appropriate sensors and acquisition devices can be selected according to specific application scenarios to reduce costs and improve efficiency; in terms of online dictionary learning and SAMP algorithm, different algorithms and parameters can be selected to adapt to different load change characteristics and real-time requirements; in terms of load monitoring and control, appropriate algorithms and control strategies can be selected according to specific application scenarios to improve monitoring accuracy and control effectiveness.

[0192] The method described above by the present invention solves the following technical problems:

[0193] 1. This solution addresses the bandwidth limitations faced by multiple users uploading high-dimensional electrical signals from the edge to the cloud in a distributed NILM system deployed on both the cloud and the edge. It reduces bandwidth consumption during data upload, lowering data transmission costs. Furthermore, it enables collaborative work between the edge and cloud through cloud-edge cooperation, improving the efficiency and reliability of load monitoring.

[0194] 2. Online dictionary learning is an online dictionary learning method that updates and optimizes the dictionary based on time-varying bus electrical signals. This invention uses an online dictionary learning method to overcome the limitations of traditional dictionary learning algorithms in processing bus current signals, which are characterized by time-varying, high-dimensionality, non-sparseness, and nonlinearity. By continuously updating and optimizing the dictionary based on time-varying bus current signals, it adapts to changes in power system load and the time-varying nature of current signals.

[0195] 3. The Sparsity Adaptive Matching Pursuit (SAMP) algorithm is a compressed sensing algorithm used for signal reconstruction, specifically for sparse adaptive recovery and reconstruction of time-varying compressed signals. This invention employs the SAMP algorithm to adaptively adjust the number of iterations and the sparsity threshold based on the signal sparsity, achieving sparse recovery and reconstruction of time-varying compressed signals. This improves the accuracy and efficiency of signal reconstruction, providing precise training data for subsequent load monitoring.

Claims

1. A non-intrusive online compressed sensing method for load monitoring data based on cloud-edge collaboration, characterized in that... The method includes the following steps: Step 1: Data Collection After the signals collected by the sensor or meter are converted and preprocessed according to the protocol, the original signal x that needs to be uploaded is obtained and stored in the embedded system of the sensor or meter. Step 2: Data compression and sampling: The original signal x is compressed and sampled by the sensor or meter to obtain the compressed observation signal y, and then the compressed observation signal is uploaded to the cloud computing center through the communication network. Step 3: Sparse Representation: In the cloud computing center, the KSVD algorithm is used to obtain a dictionary based on the pre-stored bus voltage and current data as the initial dictionary for online dictionary learning. Then, the online dictionary learning algorithm is used to continuously update the dictionary according to the changing signals to better adapt to the changing bus current signals. Step 4: Data Reconstruction In the cloud computing center, a sparse adaptive matching pursuit compressed sensing algorithm is used to reconstruct the data of the observation signal compressed in step two. Step 5: Load Monitoring The large number of reconfiguration signals obtained in step four are used for load monitoring and analysis to obtain the real-time working status and changing trend of each load on the user bus. Step Six: Real-time Feedback Control Real-time feedback control and adjustment are performed based on the load monitoring results from step five.

2. The non-intrusive online compressed sensing method for load monitoring data based on cloud-edge collaboration according to claim 1, characterized in that... The specific steps of step two are as follows: Step 21: Using a Gaussian random matrix as the measurement matrix, the input high-dimensional original signal is projected into a low-dimensional space to obtain the compressed observation signal. This process is described using a matrix operation, namely: y = Φx Where y is the compressed sampled observation signal, Φ is the measurement matrix, and x is the original signal. Different compression ratios can be achieved by choosing different measurement matrix Φ. Step 22: Transmit the compressed observation signal y through the communication network to reduce network transmission pressure.

3. The non-intrusive online compressed sensing method for load monitoring data based on cloud-edge collaboration according to claim 1, characterized in that... The specific steps of step three are as follows: Step 3: Construct a local dictionary D0 based on the KSVD algorithm, which will serve as the initial dictionary for hot-starting online dictionary learning. The specific steps for constructing the local dictionary D0 based on the KSVD algorithm are as follows: (1) In the sparse coding stage, the learning dictionary D is fixed, and OMP is used to calculate the sparse coefficient matrix α that satisfies the optimization conditions. The optimization objective is as follows: Where D is the learning dictionary, α is the sparse coefficient matrix corresponding to the original signal x, and α i Let K be a sparse vector in matrix α, and K be the dictionary sparsity, where the dictionary satisfies the sparsity condition that the number of non-zero elements is less than K. (2) During the dictionary update phase, the sparse dictionary is updated column by column. The dictionary update formula is as follows: Where d t E represents the t-th column of the sparse dictionary D; t Let x be the error matrix, representing the contribution of the t-th column of the sparse dictionary to the original signal x; (3) For E t Perform singular value decomposition to obtain d that minimizes the objective function. t With α t The updated value, in order to ensure that the obtained α t To satisfy the sparse constraints of the optimization problem, α t Remove all 0 elements from E, and... t The corresponding column is processed in the same way to obtain E' t Then for E' t Perform singular value decomposition: AND' t =U∑V T (4) Update the dictionary column vector d using the first column of matrix U. t Update the sparse vector α using the product of the first column of matrix V and the first singular value. t Then, the previous α t The removed zero elements are added to the appropriate positions to ensure that α is equal to the value before the update. t The sparsity was not reduced; (5) Repeat steps (2) to (4) for each column of the sparse dictionary, and combine the updated columns to obtain the optimal sparse dictionary matrix; (6) The entire KSVD dictionary learning process will alternately repeat the sparse coding stage and the dictionary update step until the number of iterations reaches the maximum value, thereby obtaining the local dictionary D0. Step 3.2: Use an online dictionary learning algorithm to continuously update the dictionary based on the changing bus current signal to better adapt to the changing signal. The specific steps are as follows: (1) Original signal It consists of independent and identically distributed samples, with one sample x input at a time. t Input an initial dictionary constructed using KSVD. (2) Initialization Reset previous information; (3) For t=1 to T, begin iteration: (4) Each time a new sample x is added... t ; (5) Sparse coding, calculated using LARS Lasso: (6) Calculate: (7) Calculate D using the dictionary update algorithm t And using D t-1 As a hot start for the dictionary: (8) The iteration stops when t = T; (9) Output the updated dictionary D T .

4. The non-intrusive online compressed sensing method for load monitoring data based on cloud-edge collaboration according to claim 3, characterized in that... In step 3.2, to accelerate the convergence speed of the algorithm during the dictionary update phase, while ensuring convergence, a block coordinate descent algorithm is used. In each iteration, the direction that minimizes the objective function is selected for updating. The dictionary update steps are as follows: (1) Input dictionary (2) Repeat the execution for j = 1 to K: (3) Update the j-th column of dictionary D: (4) Continue until convergence; After the dictionary update step is completed, the online dictionary D at time t is obtained. t .

5. The non-intrusive online compressed sensing method for load monitoring data based on cloud-edge collaboration according to claim 3, characterized in that... In step 3.2, the LARS Lasso algorithm is used to learn the dictionary matrix D and the sparse coding matrix α step by step. The iterative process is as follows: 1) Initialize the dictionary matrix D as the identity matrix, the sparse coding matrix α as a zero matrix, and the residual r = x t ; 2) Calculate the inner product γ of the residual r and the dictionary matrix D. j = <D j Find the atom j with the highest correlation, r>. 3) Calculate the increment γ of feature j j =sign( <D j ,r>)·min(| <D j ,r>|,s K ), where s K It is a threshold related to the current regularization parameter λ, used to control the rate at which feature weights increase; 4) Add the increment γ to the j-th column of the sparse coding matrix α. j , i.e. α j =D j +γ j The residual r is then updated to r = r - γ. j D j ; 5) Stop iterating when the number of non-zero elements has reached the preset maximum value or the current correlation can no longer be increased.

6. The non-intrusive online compressed sensing method for load monitoring data based on cloud-edge collaboration according to claim 1, characterized in that... In step four, the SAMP algorithm is used to solve for the sparse coefficients. The specific steps are as follows: (1) Input m×n dimensional sensing matrix Θ, n×1 dimensional observation signal y, step size s; (2) Initialization: Reconstruct residual r0 = y; Index set Iteration count index t = 1; Sparsity K = s; Block index stage = 1; (3) If t≤m, calculate (4) Select the K largest values ​​in u, and form a column index set by indexing the column indices of Θ corresponding to these values. (5) Constructing the candidate set: (6) Solving the least squares problem y = Θ t α t : (7) Calculate: (8) Select K elements from the initial support set and denote them as Λ tK The corresponding K column in Θ is denoted as Final episode (9) Update residuals: (10) If the residual r is less than the specified threshold or t>m, then stop the iteration and reconstruct the... exist If there are non-zero terms, output (11) If ||r||2≥||r t-1 ||2, Update block index stage = stage + 1, Update sparsity K = stage × s, Return to step (3); (12) If neither of the first two conditions is met, then update the final set. Update residual r t =r, update the iteration index t = t + 1, return to step (3), until the iteration stops; The coefficient matrix is ​​obtained through the above steps. Then according to Obtain the reconstructed signal This completes the signal reconstruction in step four.

Images