A non-intrusive load identification method and system based on multi-dimensional characteristic matrix

Abstract

The present invention discloses a non-intrusive load identification method and system based on a multidimensional feature matrix, which relates to the field of load identification. The identification method includes: collecting the current circuit data of the main circuit, and after pre-processing the current circuit data, dividing the corresponding current and voltage waveform signal data for each device in the main circuit; extracting the current periodic signal based on the current and voltage waveform signal, generating a known multidimensional feature matrix according to the extraction result, and performing information entropy calculation and weighted processing on the multidimensional feature matrix; inputting the weighted multidimensional feature matrix into the load identification model, calculating the distance between the weighted multidimensional feature matrix and the unknown sample multidimensional matrix, and outputting the load identification result based on the distance result. The present invention uses the multidimensional feature matrix as the input of the load identification model, contains richer information, and is conducive to improving the accuracy and sensitivity of load identification in terms of method.

Description

Technical Field

[0001] The present invention relates to load identification, and in particular to a non-intrusive load identification method and system based on a multi-dimensional characteristic matrix. Background Art

[0002] With the development of smart grid and Internet of Things technology, real-time monitoring and management of power load has become particularly important. Non-intrusive load identification technology, as an emerging power load monitoring technology, has the advantages of low cost, simple maintenance, and easy promotion. The basic principle of non-intrusive load identification technology is to install a data acquisition device at the entrance of residential users to collect data such as the user's terminal voltage and total current, and then bring the processed data into various non-intrusive load identification models to analyze the specific operating status of each load, thereby realizing the monitoring of user loads. This technology does not require invasive modification of the user's power equipment. It only needs to install a data acquisition device on the user side to realize real-time monitoring and management of the user's load.

[0003] In terms of technical implementation, non-intrusive load identification technology mainly includes steps such as data collection and processing, feature extraction and load identification. The user's voltage, current and other data are collected through the data acquisition device, and the data is pre-processed. Statistics reflecting the load power consumption characteristics, namely load characteristics, are extracted from the processed data. The extracted feature quantities are used as the input of the load identification model algorithm, thereby realizing accurate identification of the load through the algorithm.

[0004] Non-intrusive load identification technology (NILM) is a power load monitoring and identification technology. It collects, processes and analyzes power signals such as the voltage and current waveforms of the monitoring equipment to identify and analyze the operating status, category and other information of the equipment. However, for high-frequency voltage and current data, the current and voltage signals of the existing load identification methods will be lost to a large extent during the feature extraction process, making it difficult to achieve satisfactory results in load identification.

[0005] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a non-intrusive load identification method and system based on a multi-dimensional feature matrix to achieve the purpose of improving the accuracy and sensitivity of load identification.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a non-intrusive load identification method based on a multi-dimensional feature matrix, the non-intrusive load identification method comprising:

[0009] Collect the main circuit current circuit data, perform pre-processing on the current circuit data, and divide the corresponding current and voltage waveform signal data for each device in the main circuit;

[0010] Extracting the current periodic signal based on the current and voltage waveform signals, generating a known multi-dimensional feature matrix according to the extraction results, and performing information entropy calculation and weighting processing on the multi-dimensional feature matrix;

[0011] The weighted multidimensional feature matrix is ​​input into the load identification model, the distance between the weighted multidimensional feature matrix and the unknown sample multidimensional matrix is ​​calculated, and the load identification result is output based on the distance result.

[0012] Preferably, extracting the current periodic signal based on the current and voltage waveform signals, generating a known multidimensional feature matrix according to the extraction results, and performing information entropy calculation and weighted processing on the multidimensional feature matrix include:

[0013] The segmented aggregation approximate features of the current and voltage waveform signals are obtained based on the segmented aggregation approximate processing, and the current harmonic features corresponding to the current and voltage waveform signals are extracted using the Fourier transform technology;

[0014] Analyze the active power and total power of the equipment in the power grid based on the current and voltage waveform signals, and calculate the power factor characteristics based on the active power and total power;

[0015] The segmented aggregation approximate features, current harmonic features and power factor features are taken as current periodic signals, and the current periodic signals are normalized and then fused to generate a multi-dimensional feature matrix.

[0016] The information gain of the multidimensional feature matrix is ​​calculated, and the features in the multidimensional feature matrix are weighted according to the gain result.

[0017] Preferably, obtaining the segmented aggregation approximate features of the current and voltage waveform signals based on the segmented aggregation approximate processing, and extracting the current harmonic features corresponding to the current and voltage waveform signals using the Fourier transform technology includes:

[0018] Performing segmented aggregation approximation processing on the current and voltage waveform signal data according to a preset window size set to obtain the current and voltage waveform signal data after dimensionality reduction;

[0019] Calculate the mean square error between the current and voltage waveform signal data after dimension reduction and the current and voltage waveform signal data, and select the window size corresponding to the minimum mean square error as the segmented window size;

[0020] The segmented aggregation approximate features corresponding to the current and voltage waveform signal data are obtained based on the segmented window size, and the current and voltage waveform signal data are converted into frequency domain signals using the fast Fourier transform technology;

[0021] The harmonic component position is found in the frequency domain signal, the harmonic amplitude is extracted according to the position result, and the harmonic amplitude is normalized to the main frequency amplitude to obtain the current harmonic characteristics.

[0022] Preferably, analyzing the active power and total power of the device in the power grid based on the current and voltage waveform signals, and calculating the power factor characteristics based on the active power and the total power include:

[0023] Based on the current and voltage waveform signals, the effective values ​​of current and voltage within the sampling period are obtained, and the effective value of current is multiplied by the effective value of voltage, and the multiplication result is used as the active power of the device in the power grid;

[0024] Analyze the instantaneous current value and instantaneous voltage value corresponding to the current and voltage data points within the sampling period according to the current and voltage waveform signals, and obtain the total power of the device based on the instantaneous current value and the instantaneous voltage value;

[0025] Perform a ratio calculation based on the power and total power, and use the ratio result to describe the power factor characteristics of the device in the power grid.

[0026] Preferably, the calculation formula of the power factor is:

[0027]

[0028] In the formula, λ represents the power factor, P represents the active power, S represents the total power, and I rms Indicates the effective value of current, U rms represents the effective value of voltage, n represents the number of voltage and current data points in the sampling period, i t represents the instantaneous current at the tth data point, u t Represents the instantaneous voltage of the tth data point.

[0029] Preferably, calculating the information gain of the multidimensional feature matrix and performing weighted processing on the features in the multidimensional feature matrix according to the gain result includes:

[0030] Determine the proportion of the preset load category information in the multidimensional feature matrix, and calculate the information entropy of the multidimensional feature matrix based on the proportion result;

[0031] The multidimensional feature matrix is ​​divided into different subsets based on the feature value range information, and the conditional entropy corresponding to the segmented aggregation approximate features, current harmonic features and power factor features in the multidimensional feature matrix is ​​analyzed according to the division results;

[0032] The information gains corresponding to the segmented aggregation approximate features, current harmonic features and power factor features are calculated based on the conditional entropy and information entropy, and the information gains are normalized respectively;

[0033] The normalized processing results are used as feature weighting coefficients to weight the multidimensional feature matrix, and weighted segmented aggregation approximate features, current harmonic features and power factor features are obtained.

[0034] Preferably, the information entropy calculation formula of the multidimensional feature matrix is:

[0035]

[0036] In the formula, H(D) represents the information entropy of the multidimensional feature matrix D, m represents the number of load categories, and P i Represents the proportion of the i-th type of load in the multidimensional feature matrix D.

[0037] Preferably, the weighted multidimensional feature matrix is ​​input into the load identification model, the distance between the weighted multidimensional feature matrix and the unknown sample multidimensional matrix is ​​calculated, and the load identification result is output based on the distance result, including:

[0038] The weighted multi-dimensional feature matrix is ​​used as a training set of known electrical equipment and combined with the proximity algorithm to train the load identification model, and a multi-dimensional feature matrix vector of known category samples is obtained;

[0039] Obtain unknown sample data to generate a multidimensional feature matrix of unknown samples, and input the multidimensional feature matrix of unknown samples into the trained load identification model to obtain a multidimensional feature matrix vector of samples to be identified;

[0040] Calculate the distance between the multidimensional feature matrix vector of samples of known categories and the multidimensional feature matrix vector of samples to be identified, and select the distance between the multidimensional feature matrix vector of samples to be identified and output the load identification result based on the distance result.

[0041] In a second aspect, the present invention further provides a non-intrusive load identification system based on a multi-dimensional characteristic matrix, the non-intrusive load identification system comprising:

[0042] The signal data acquisition unit is used to collect the current circuit data of the main circuit, and after pre-processing the current circuit data, divide the corresponding current and voltage waveform signal data for each device in the main circuit;

[0043] A multi-dimensional matrix generating unit, used for extracting a current periodic signal based on a current and voltage waveform signal, generating a known multi-dimensional feature matrix according to the extraction result, and performing information entropy calculation and weighted processing on the multi-dimensional feature matrix;

[0044] The load result output unit is used to input the weighted multidimensional feature matrix into the load identification model, calculate the distance between the weighted multidimensional feature matrix and the unknown sample multidimensional matrix, and output the load identification result based on the distance result.

[0045] Preferably, the multidimensional matrix generating unit comprises:

[0046] An approximate harmonic feature acquisition module is used to acquire the segmented aggregation approximate features of the current and voltage waveform signals based on the segmented aggregation approximate processing, and to extract the current harmonic features corresponding to the current and voltage waveform signals using Fourier transform technology;

[0047] A power factor characteristic acquisition module is used to analyze the active power and total power of the device in the power grid based on the current and voltage waveform signals, and calculate the power factor characteristic based on the active power and the total power;

[0048] A feature matrix generation module is used to take the segmented aggregation approximate features, current harmonic features and power factor features as current periodic signals, and perform fusion processing on the current periodic signals after normalization to generate a multi-dimensional feature matrix;

[0049] The feature matrix processing module is used to calculate the information gain of the multidimensional feature matrix and perform weighted processing on the features in the multidimensional feature matrix according to the gain result.

[0050] The beneficial effects of the present invention are:

[0051] 1. The present invention collects the user's voltage, current and other data through a data acquisition device, performs preprocessing operations on the data, and extracts statistics reflecting the load power consumption characteristics, namely, load characteristics, from the processed data, and uses the extracted characteristic quantities as the input of the load identification model algorithm, thereby realizing accurate identification of the load through the algorithm. At the same time, the present invention selects three important characteristics of segmented aggregation approximate characteristics, harmonic characteristics and power factor to describe the load signal characteristics, and can flexibly reduce the dimension of the current according to the data sampling frequency, thereby achieving the transformation from high-dimensional data to low-dimensional data, while minimizing the loss of current characteristics.

[0052] 2. Compared with the existing methods of RMS characteristics and current peak-to-peak characteristics, the present invention can reduce the dimension of current by extracting segmented aggregation approximate features, thereby achieving the transformation from high-dimensional data to low-dimensional data, and at the same time minimize the loss of current characteristics and ensure information integrity, so that the results have higher sensitivity and accuracy. The feature extraction of current signals using segmented aggregation approximate algorithm can be dynamically adjusted with the signal sampling frequency, and more detailed information can be obtained without being affected by frequency, thereby improving the accuracy of characterizing current signals. The multidimensional feature matrix composed of segmented aggregation approximate characteristics, current harmonic characteristics and power factor characteristics is used as the input of the load identification model, which contains richer information and is conducive to improving the accuracy and sensitivity of load identification in terms of method. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0054] Figure 1 is a flow chart of a non-intrusive load identification method based on a multi-dimensional feature matrix according to an embodiment of the present invention;

[0055] Figure 2 is a principle block diagram of a non-intrusive load identification system based on a multi-dimensional feature matrix according to an embodiment of the present invention;

[0056] Figure 3 It is a flow chart of a non-intrusive load monitoring method using an improved KNN algorithm and transient and steady-state characteristics in the prior art;

[0057] Figure 4 is a flow chart of a non-intrusive load identification method based on a multi-dimensional feature matrix according to an embodiment of the present invention;

[0058] Figure 5 It is a harmonic characteristic diagram after FFT transformation in a non-intrusive load identification method based on a multi-dimensional characteristic matrix according to an embodiment of the present invention;

[0059] Figure 6 is a load identification warning flow chart in a non-intrusive load identification method based on a multi-dimensional feature matrix according to an embodiment of the present invention;

[0060] Figure 7 is a five-fold cross validation diagram in a non-intrusive load identification method based on a multi-dimensional feature matrix according to an embodiment of the present invention;

[0061] Figure 8 is a cross-validation average accuracy graph in a non-intrusive load identification method based on a multi-dimensional feature matrix according to an embodiment of the present invention;

[0062] Fig. 9 is a confusion matrix in a non-intrusive load identification method based on a multi-dimensional feature matrix according to an embodiment of the present invention;

[0063] Fig.10 It is a confusion matrix of a traditional method in a non-intrusive load identification method based on a multi-dimensional feature matrix according to an embodiment of the present invention.

[0064] In the figure:

[0065] 1. Signal data acquisition unit; 2. Multi-dimensional matrix generation unit; 3. Load result output unit. DETAILED DESCRIPTION

[0066] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0067] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0068] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0069] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0070] See also Figure 3 The non-intrusive load monitoring method proposed in the prior art has a load identification flow chart, and the specific identification steps are as follows:

[0071] The main circuit current and voltage data are collected, and data preprocessing is performed to obtain the current and voltage waveform signals of a single device, and feature extraction is performed on the obtained current and voltage waveform signals.

[0072] The specific feature extraction steps are as follows:

[0073] (1) Root mean square characteristics: The root mean square characteristics of the current cycle are extracted using the root mean square characteristic formula:

[0074]

[0075] In the formula, I rms represents the current RMS characteristic of the current cycle, N represents the number of sampling points in a cycle, t represents any sampling point in the cycle, i t Represents the instantaneous value of the current at sampling point t.

[0076] (2) Current harmonic characteristics: Fourier transform is used to perform harmonic decomposition on the load steady-state process periodic current, and the harmonic order with larger harmonic amplitude is extracted as the load characteristic to reduce noise interference and obtain steady-state characteristics with better load differentiation. Harmonic characteristics are extracted from discrete load periodic currents as follows:

[0077]

[0078] Where i(n) represents the current at the periodic current sampling point n, N k represents the number of sampling points in each cycle, k represents the harmonic order, and D(k) represents the harmonic coefficient in complex form.

[0079] (3) Current peak-to-peak value: The current peak-to-peak value feature is used to extract the RMS feature of the current cycle:

[0080] i pp =i max -i min ;

[0081] In the formula, i pp is the peak-to-peak current characteristic, i max is the maximum current, i min is the minimum current value.

[0082] After the calculation is completed, the eigenvalue is normalized to the maximum value and brought into the identification model for load identification. The k points with the smallest distance from the current point are selected, and the weight factor is added to the selected k points. The category with the largest superposition weight is selected as the load identification result.

[0083] The prior art extracts the RMS feature of the current through the RMS feature formula. According to the RMS feature formula, the process converts the multi-dimensional signal points in the current cycle into one-dimensional features. However, for the current time series signal, the RMS feature does not consider the time series relationship between the data points. Therefore, in the scenario considering the time series characteristics, the feature may not perform well. The RMS feature is the average quantization of the amplitude, which is greatly affected by the overall fluctuation and may not accurately reflect the local characteristics of the data. As the sampling frequency of the current and voltage signals increases, the sampling points in the current and voltage cycle will increase, and the noise signals and burst signals will also increase, which may lead to the loss of some detailed information, thereby affecting the accuracy of load identification.

[0084] In the calculation of harmonic coefficients, Fourier transform is used to decompose the periodic current of the load steady-state process harmonically, and the harmonic order with larger harmonic amplitude is extracted as the load characteristic. The method of only considering the characteristics of larger harmonic amplitude may have limitations, because it may limit the monitoring and identification capabilities of equipment that mainly exhibits its unique electrical characteristics through low-amplitude harmonics. In particular, for low-power electrical appliances, the harmonics generated during operation may not be significant in amplitude. In actual applications, the noise in the power grid and the simultaneous operation of multiple electrical appliances may have a significant impact on the harmonic amplitude, which may in turn affect the accuracy of load identification. If the system only focuses on harmonics with larger amplitudes, some important small-amplitude harmonics may be ignored. These small-amplitude harmonics may actually contain very useful information for load identification.

[0085] This will undoubtedly reduce the adaptability and flexibility of the NILM system. Under different application environments or load combinations, larger harmonics may change, requiring the NILM system to continuously adjust the harmonic frequency and amplitude range it focuses on to maintain a high level of recognition accuracy. This will undoubtedly increase the complexity of the system. When designing a load identification system, if the machine learning model is trained only based on harmonics with larger amplitudes, the generalization ability of the model may be limited. The model may only show good recognition capabilities for loads with specific harmonic characteristics that have been seen during training, while its recognition capabilities for new or unseen load patterns may be greatly reduced. Therefore, when building a load identification system, it is necessary to comprehensively consider various harmonic characteristics, including low-amplitude harmonics, to improve the accuracy and generalization ability of the system.

[0086] Based on the above analysis, the present invention adopts a new feature extraction method, which first ensures that high-dimensional data can be processed, key information can be extracted, dimension reduction can be achieved and the most important features of the current time-frequency domain can be retained. A load identification model is established by training the multi-dimensional feature matrix in the time-frequency domain, making load identification more efficient, thereby improving the accuracy of load identification and achieving the purpose of intelligent power management.

[0087] See also Figure 1 and Figure 4 The present invention provides a non-intrusive load identification method based on a multi-dimensional feature matrix, and the non-intrusive load identification method includes:

[0088] Step S1, collecting the main circuit current circuit data, and after pre-processing the current circuit data, dividing the corresponding current and voltage waveform signal data for each device in the main circuit;

[0089] Step S2, extracting the current periodic signal based on the current and voltage waveform signals, generating a known multidimensional feature matrix according to the extraction results, and performing information entropy calculation and weighting processing on the multidimensional feature matrix;

[0090] Step S3, input the weighted multidimensional feature matrix into the load identification model, calculate the distance between the weighted multidimensional feature matrix and the unknown sample multidimensional matrix, and output the load identification result based on the distance result.

[0091] It should be explained that the non-intrusive load identification technology mainly includes steps such as data collection and processing, feature extraction and load identification. The voltage, current and other data of the user are collected through the data collection device, and the data is pre-processed. The statistics reflecting the power consumption characteristics of the load, that is, the load characteristics, are extracted from the processed data, and the extracted feature quantities are used as the input of the load identification model algorithm, so as to realize the accurate identification of the load through the algorithm. In the non-intrusive load identification method, feature extraction is the key, because these features reflect the similarity or difference of the equipment during operation. Therefore, this embodiment adopts a new feature extraction method to extract features from the current and voltage periodic waveforms, so three important features, namely, the segmented aggregation approximation feature (PAA feature), the harmonic feature and the power factor, are selected to describe the load signal characteristics. The PAA feature can flexibly reduce the dimension of the current according to the data sampling frequency, thereby achieving the transformation from high-dimensional data to low-dimensional data, and at the same time, it can minimize the loss of current features. The harmonic feature reflects the linear and nonlinear characteristics generated by the equipment in the power system. The harmonics generated by the nonlinear load will cause grid instability and other problems. Analyzing the harmonic features is helpful to identify the load.

[0092] Power factor is an important power parameter that describes the ratio of active and reactive power consumed by equipment in the power grid. Different types of loads (resistance, capacitance and inductance) have different power factors, so it can be used for load type identification. By extracting these three features to form a multidimensional feature matrix, the purpose of load identification can be achieved through the KNN algorithm.

[0093] In this embodiment, when extracting the current periodic signal based on the current and voltage waveform signals, generating a known multidimensional feature matrix according to the extraction results, and performing information entropy calculation and weighted processing on the multidimensional feature matrix, the segmented aggregation approximate features of the current and voltage waveform signals can be obtained based on the segmented aggregation approximate processing, and the current harmonic features corresponding to the current and voltage waveform signals can be extracted using Fourier transform technology; the active power and total power of the equipment in the power grid are analyzed based on the current and voltage waveform signals, and the power factor features are calculated based on the active power and the total power; the segmented aggregation approximate features, the current harmonic features, and the power factor features are used as the current periodic signal, and the current periodic signal is normalized and then fused to generate a multidimensional feature matrix; the information gain is calculated for the multidimensional feature matrix, and weighted processing is performed on the features in the multidimensional feature matrix according to the gain results.

[0094] In this embodiment, when obtaining the segmented aggregation approximate features of the current and voltage waveform signals based on the segmented aggregation approximate processing, and extracting the current harmonic features corresponding to the current and voltage waveform signals using the Fourier transform technology, the current and voltage waveform signal data can be subjected to segmented aggregation approximate processing according to a preset window size set to obtain the current and voltage waveform signal data after dimensionality reduction; the mean square error between the current and voltage waveform signal data after dimensionality reduction and the current and voltage waveform signal data is calculated, and the window size corresponding to the minimum mean square error is selected as the segmented window size; the segmented aggregation approximate features corresponding to the current and voltage waveform signal data are obtained based on the segmented window size, and the current and voltage waveform signal data are converted into frequency domain signals using the fast Fourier transform technology; the harmonic component position is found from the frequency domain signal, the harmonic amplitude is extracted according to the position result, and the harmonic amplitude is normalized to the main frequency amplitude to obtain the current harmonic feature.

[0095] Specifically, the specific operation of segmented aggregation approximation is as follows: the steady-state current of the electrical equipment is collected, and a period of current signal is intercepted therefrom. The current signal of one period is segmented according to the signal fluctuation size and the signal sampling frequency, and the average value of each segment is calculated respectively. The average value of each segment is normalized to the maximum value to form a multidimensional feature set for feature fusion of the final multidimensional feature matrix. Signals with large signal fluctuations and high sampling frequencies can choose a smaller segmentation interval length, and signals with large signal fluctuations and low sampling frequencies can choose a larger segmentation interval length.

[0096] It should be explained that when performing segmented aggregation approximate feature extraction, different window sizes need to be selected to perform segmented aggregation approximate processing on all device current data. After processing each window size, the reduced-dimensional time series data is obtained, which is compared with the original sequence to calculate the mean square error. The minimum mean square error is selected as the segmented window size for this type of equipment. The specific calculation formula is:

[0097]

[0098] Where, X i is the value in the original time series of each device, It is the corresponding value in the time series after the segmented aggregation and approximate processing of each device under different window sizes. i represents the i-th data point in the time series, i = 1, 2, 3, ...; j represents the j-th device in the device sequence, and k represents the total number of rows of collected data. The window size corresponding to the minimum mean square error for each device.

[0099] The PAA feature values ​​obtained by dimensionality reduction with different window sizes have different dimensions and contain different original information. The window size with the smallest error is selected to obtain more comprehensive information. The features obtained by dimensionality reduction with a determined window size are the final PAA features.

[0100] Based on the final segmentation window size N, the time series X is represented as X = X1, X2, ..., X n , after being processed by the segmented aggregation approximation algorithm, the time series X can be It is referred to as PAA feature. The calculation formula of the i-th vector of the specific sequence is:

[0101]

[0102] final The PAA feature vector is extracted. The current features extracted by this method reduce the original high-dimensional characteristics and minimize the current feature loss.

[0103] And by adjusting the parameters n and N, the PAA features of current signals with different sampling rates can be extracted. The current features extracted by this method reduce the original high-dimensional characteristics and minimize the current feature loss. At the same time, the vector calculation not only considers the harmonics with larger amplitudes, but also considers the high-order harmonics with small amplitudes. These small-amplitude harmonics may actually contain very useful information for load identification. This approach will undoubtedly improve the load identification accuracy of the NILM system.

[0104] At the same time, the harmonic feature extraction method is to convert the time domain signal into the frequency domain signal through fast Fourier transform (FFT), such as Figure 5 As shown in the figure, the harmonic components are located in the frequency domain signal, and the amplitude of the harmonics is extracted to calculate the amplitude of the first 20 harmonics of the power frequency (x f0 =50HZ,x f1 =100HZ,x f2 =150HZ,x f3 =200HZ, ...), and normalize it to the main frequency amplitude. The calculation formula of current harmonic characteristics is:

[0105]

[0106] Where F represents the harmonic eigenvector consisting of 20 elements.

[0107] In this embodiment, when analyzing the active power and total power of the device in the power grid based on the current and voltage waveform signals, and calculating the power factor characteristics based on the active power and the total power, the effective values ​​of the current and voltage in the sampling period can be obtained based on the current and voltage waveform signals, and the effective value of the current is multiplied by the effective value of the voltage, and the multiplication result is used as the active power of the device in the power grid; the instantaneous current value and the instantaneous voltage value corresponding to the current and voltage data points in the sampling period are analyzed based on the current and voltage waveform signals, and the total power of the device is obtained based on the instantaneous current value and the instantaneous voltage value; the ratio calculation is performed based on the power and the total power, and the ratio result is used to describe the power factor characteristics of the device in the power grid.

[0108] It should be explained that power factor is an important power parameter that describes the ratio of active and reactive power consumed by the equipment in the power grid. Different types of loads (resistance, capacitance and inductance) have different power factors, so it can be used to identify the load type. The power factor characteristic calculation formula is as follows:

[0109]

[0110] P=I rms ×U rms ;

[0111]

[0112] In the formula, λ represents the power factor, P represents the active power, S represents the total power, and I rms Indicates the effective value of current, U rms represents the effective value of voltage, n represents the number of voltage and current data points in the sampling period, i t represents the instantaneous current at the tth data point, u t Represents the instantaneous voltage of the tth data point, t=1, 2, 3,…, n.

[0113] The three extracted eigenvalues ​​are normalized according to the normalization formula to remove the implicit feature weighting caused by significant differences in the feature range, as follows:

[0114]

[0115] In the formula, represents the normalized features, Z represents the original features, and Z min represents the minimum characteristic value, Z maxIt represents the maximum value of the feature. After normalization, the three features can be fused to form a multi-dimensional feature matrix. The multi-dimensional features are put together to form a multi-dimensional feature matrix D, which can be expressed as:

[0116]

[0117] The information gain of the PAA feature, current harmonic feature and power factor feature contained in the matrix D is calculated, and each feature is weighted according to the information gain. In the problem of fusion of the three features of PAA feature, current harmonic feature and power factor feature, the information gain method can effectively filter out the eigenvalues ​​of the three features that have little to do with the target equipment, reduce the interference of the redundant information of the three features, and at the same time, by quantifying the importance of each of the three features, the non-intrusive load identification model can focus more on valuable information, support a more effective and accurate prediction process, and understand the influence of PAA feature, current harmonic feature and power factor feature on the prediction results can help understand the equipment performance and make maintenance decisions.

[0118] In this embodiment, when the information gain of the multidimensional feature matrix is ​​calculated and the features in the multidimensional feature matrix are weighted according to the gain results, the proportion of the preset load category information in the multidimensional feature matrix can be determined, and the information entropy of the multidimensional feature matrix can be calculated based on the proportion result; the multidimensional feature matrix is ​​divided into different subsets based on the feature value range information, and the conditional entropy corresponding to the segmented aggregation approximate features, current harmonic features and power factor features in the multidimensional feature matrix is ​​analyzed according to the division results; the information gain corresponding to the segmented aggregation approximate features, current harmonic features and power factor features is calculated based on the conditional entropy and the information entropy, and the information gain is normalized respectively; the normalization result is used as the feature weighting coefficient to weight the multidimensional feature matrix to obtain the weighted segmented aggregation approximate features, current harmonic features and power factor features.

[0119] Specifically, there is a certain correlation between the PAA characteristics and the current harmonic characteristics. Under normal circumstances, a stable high current average value (the corresponding PAA characteristic value will also be large) may correspond to a more complex harmonic characteristic. For example, some large industrial equipment has a large current average value during operation. At the same time, due to its internal power electronic components or nonlinear characteristics, it may generate more harmonics.

[0120] The PAA feature is also related to the power factor feature. Devices with a high average current will have a specific power factor range. Generally, high-power electrical equipment has a larger power factor variation range, while low-power electrical equipment has a smaller power factor variation range. For example, a high-power air conditioner has a larger current when cooling, and its power factor is relatively low; while some low-power purely resistive loads have a relatively small average current and a power factor close to 1.

[0121] There is also a close relationship between the current harmonic characteristics and the power factor characteristics. Loads with high harmonic content often have low power factors. This is because harmonics will lead to ineffective consumption of electrical energy and increase reactive power, thereby reducing the ratio of active power to apparent power, that is, the power factor.

[0122] Taking these three characteristics into consideration can provide more abundant information for non-invasive load identification, which helps to more accurately identify and distinguish different types of power loads. For example, by monitoring the PAA characteristics of the current, current harmonics and power factor, it can be determined whether the electrical equipment running at a certain moment is a high-power, high-harmonic device (such as a frequency converter) or a low-power, low-harmonic and high-power factor device (such as an incandescent lamp), etc., which can realize real-time monitoring and analysis of power loads, understand the usage of various electrical equipment, and provide important basis for power management, energy-saving optimization and fault diagnosis.

[0123] At the same time, high-order harmonic features provide more dimensional information for load identification. When different loads generate harmonics, the spectral distribution, amplitude, phase and other characteristics of their high-order harmonics are different. For example, electronic devices such as computer power supplies usually generate higher-order harmonics, and their spectral distribution is relatively complex; while traditional resistive loads such as electric ovens have fewer high-order harmonic components. By extracting these high-order harmonic features, they can be combined with other features (such as PAA features, power factor, etc.) to form a rich feature set, thereby more comprehensively describing the characteristics of the load.

[0124] Non-linear loads are the main source of high-order harmonics. By analyzing the characteristics of high-order harmonics, different types of non-linear loads can be distinguished more accurately. For example, although variable-frequency air conditioners and inverters are both non-linear loads, the high-order harmonics they generate may have obvious differences in frequency and amplitude. Based on these differences, specific load types can be identified more accurately.

[0125] Moreover, the basic electrical characteristics of some loads may change under different working conditions (such as voltage fluctuations, load changes, etc.), but the high-order harmonic characteristics are relatively stable and less affected by external conditions. Therefore, the introduction of high-order harmonic characteristics can enhance the robustness of the load identification model, so that it can still maintain a high accuracy under various complex actual working conditions.

[0126] Therefore, the harmonic characteristics reflect the linear and nonlinear characteristics generated by the equipment in the power system. The harmonics generated by nonlinear loads can cause grid instability and other problems. Therefore, analyzing the harmonic characteristics helps to identify the load. The power factor is an important power parameter that describes the ratio of active and reactive power consumed by the equipment in the power grid. Different types of loads (resistance, capacitance and inductance) have different power factors, so they can be used for load type identification. By extracting these three features to form a multidimensional feature matrix, the purpose of load identification is achieved through the KNN algorithm.

[0127] It needs to be explained that the information entropy calculation formula of the multidimensional feature matrix D is:

[0128]

[0129] Where m represents the number of load categories, P i It represents the proportion of the i-th type of load in the characteristic matrix D, which is the number of rows of data for the same type of equipment to the number of rows of data for all equipment.

[0130] The feature matrix D is divided into different subsets according to different features. Since the value range of PAA features, current harmonic features and power factor features is [0, 1], each feature is divided into different feature matrix subsets with values ​​greater than 0.5 and less than 0.5. The three features can be divided into 6 subsets {D1, D2, D3, D4, D5, D6}, where D1 and D2 are subsets of PAA features, D3 and D4 are subsets of current harmonic features, and D5 and D6 are subsets of power factor features. This division method is easy to operate and simplifies subsequent analysis and processing. It can quickly screen out important features in PAA features, current harmonic features and power factor features, reduce the number of information gains that need to be calculated when building a non-intrusive load identification model, and thus improve the prediction efficiency of the non-intrusive load identification model.

[0131] Among them, the conditional entropy calculation formula of PAA feature, current harmonic feature and power factor feature is:

[0132]

[0133] In the formula, |D represents the total number of feature data in the entire feature matrix, |D a | represents the total number of features in each subset, a = 1, 2, 3, ..., 6;

[0134] H(DF) and H(Dλ) represent the conditional entropy of PAA feature, current harmonic feature and power factor feature, respectively.

[0135] At the same time, the information gain calculation formula of PAA feature, current harmonic feature and power factor feature is:

[0136]

[0137] IG(D,F)=H(D)-H(D|F);

[0138] IG(D,λ)=H(D)-H(D|λ);

[0139] In the formula, IG(D, F) and IG(D, λ) represent the information gains of PAA feature, current harmonic feature and power factor feature respectively.

[0140] In the problem of feature fusion, the information gain method can effectively filter out features that are not closely related to the target device and reduce the interference of redundant information. At the same time, by quantifying the importance of features, the model can focus more on valuable information and support a more effective prediction process.

[0141] Normalize the information gain value of each feature, and use the normalized value as the feature weighting coefficient W i The features of the multidimensional feature matrix D are weighted as follows:

[0142]

[0143]

[0144] Among them, W i (i=1, 2, 3) are weighting coefficients of PAA feature, current harmonic feature and power factor feature respectively.

[0145] In this embodiment, when the weighted multidimensional feature matrix is ​​input into the load identification model, the distance between the weighted multidimensional feature matrix and the unknown sample multidimensional matrix is ​​calculated, and the load identification result is output based on the distance result, the weighted multidimensional feature matrix can be used as a known electrical equipment training set and combined with a proximity algorithm to train the load identification model to obtain a multidimensional feature matrix vector of known category samples; unknown sample data is obtained to generate an unknown sample multidimensional feature matrix, and the unknown sample multidimensional feature matrix is ​​input into the trained load identification model to obtain a multidimensional feature matrix vector of samples to be identified; the distance between the multidimensional feature matrix vector of known category samples and the multidimensional feature matrix vector of samples to be identified is calculated, and based on the distance result, the distance to the multidimensional feature matrix vector of samples to be identified is selected to output the load identification result.

[0146] It should be explained that the unknown sample data corresponding to the unknown sample multi-dimensional feature matrix is ​​collected by connecting the electrical equipment through a hardware data acquisition device.

[0147] Specifically, the neighboring algorithm can adopt the K neighboring classification algorithm, perform maximum normalization processing on the PAA features, current harmonic features and power factor features extracted from the known electrical equipment features, perform early feature fusion on the processed data to form a multi-dimensional feature matrix for use as a training data set for the K neighboring algorithm, and perform model training by bringing the training data set into the K neighboring classification algorithm (early feature fusion: also called data layer fusion, in the data input stage, multiple original features are directly connected or combined to form a unified feature vector, and input into the model for learning and prediction).

[0148] The three features extracted from the unknown electrical equipment are normalized to their maximum values, and the processed data are fused with early features to form a multidimensional feature matrix for use as a test data set. The test data set is brought into the training model, and the model compares each set of test data with each set of data in the training data set, and selects the k sample data points closest to the set of test data points. The electrical equipment that appears most frequently among the k sample data points is predicted to be the electrical equipment of the set of test data points.

[0149] It should be noted that the weighted new multidimensional feature matrix is ​​brought into the model as the input of the load recognition model to calculate the distance between the multidimensional feature matrix vector of the known category sample and the multidimensional feature matrix vector of the current sample to be identified. The distance calculation formula is:

[0150]

[0151] Where x represents the test point set, which is (x1, x2, ..., x n ), where x i represents the i-th coordinate of the test point set x, i = 1, 2, ..., n; y represents the sample point in the known sample set in the training model, specifically (y1, y2, ..., y n ), where y i represents the i-th coordinate of the sample set y, i=1, 2, ..., n; D(x, y) represents the distance between the unknown category point x and the known category point y.

[0152] Then, k known categories that are closest to the feature matrix vector of the unknown sample are selected for statistics, and the predicted classification results are output.

[0153] See also Figure 6, calculate the Euclidean distance between the multi-dimensional features of the same category of equipment in the training set, and use the maximum distance as the initial threshold. During the load identification process, bring the test data set into the training model. The model will calculate the Euclidean distance between each set of test data and each set of data in the training data set, and select the k sample data points closest to the test data points. The electrical equipment that appears most frequently among the k sample data points is predicted to be the electrical equipment of the test data points. Based on the specific equipment identified, calculate the distance between the equipment and similar equipment in the training set. If the distance is less than the initial threshold, the prediction is correct; if the distance is greater than the threshold, the staff will be warned to check whether the prediction is correct. If the identification is correct, the model will update the new threshold to the new distance size.

[0154] In order to facilitate understanding of the above technical solutions of the present invention, the working principle or operation mode of the present invention in the actual process is described in detail below.

[0155] This embodiment uses the LILAC dataset. The LILAC dataset contains data collected from 15 types of load equipment, including various types of motors, drilling machines, dryers, coffee machines, electric lamps and other electrical equipment. According to the data sampling frequency of the dataset is 50kHz, one cycle (0.02 seconds) of the current signal contains 1000 data sampling points, and 20 points are used as the segment window size. Finally, the 1000 time series signals are characterized by 50-dimensional PAA features.

[0156] The LILAC (Laboratory-measured Industrial appliance Characteristics, LILAC) dataset is a power database specifically used for load identification and transient event detection of industrial equipment. Assume that the LILAC dataset contains data collected from 15 types of load equipment, including various types of motors, drilling machines, dryers, coffee machines, electric lamps and other electrical equipment. The sampling frequency of the data in the dataset is 50kHz, which is suitable for the analysis of steady-state load characteristics. In order to verify the effectiveness of the characteristic matrix load identification method based on the K-nearest neighbor algorithm and its performance in load identification accuracy, this embodiment randomly extracts 380 data from the dataset for experimental verification. The extracted data covers the operating data of 15 types of equipment under various different states.

[0157] In order to verify the effectiveness of the method proposed in this embodiment, a load identification method using the current root mean square (RMS) as a feature was compared. To ensure stable and reliable classification performance, a 5-fold cross validation was adopted in the experiment. Figure 7 and Fig.10As shown, this method ensures that each data point is used in at least one training and testing. In order to more comprehensively evaluate the classification model generated by this embodiment, the experimental results use the confusion matrix tool to analyze the classification results of the two methods. Fig. 9 and Fig.10 is the confusion matrix of the test data.

[0158] According to the results in Table 1, the feature fusion matrix load identification method based on the K nearest neighbor algorithm shows a high level of load identification accuracy for a single electrical appliance, including single-phase capacitor asynchronous motor (1~Asynchronousmachine with capacitor operation), three-phase asynchronous motor cage runner (3~AsynchronousMachineCageRunner), three-phase asynchronous motor slip ring runner (3~AsynchronousMachineSlipRingRunner), and three-phase synchronous machine damping cage (3~SynchronousMachineDamperCage).

[0159] Table 1 Device recognition accuracy

[0160]

[0161]

[0162] See also Figure 2 The present invention also provides a non-intrusive load identification system based on a multi-dimensional characteristic matrix, the non-intrusive load identification system comprising:

[0163] The signal data acquisition unit 1 is used to collect the current circuit data of the main circuit, and after pre-processing the current circuit data, divide the corresponding current and voltage waveform signal data for each device in the main circuit;

[0164] A multi-dimensional matrix generating unit 2 is used to extract the current periodic signal based on the current and voltage waveform signals, generate a known multi-dimensional feature matrix according to the extraction result, and perform information entropy calculation and weighting processing on the multi-dimensional feature matrix;

[0165] The load result output unit 3 is used to input the weighted multidimensional feature matrix into the load identification model, calculate the distance between the weighted multidimensional feature matrix and the unknown sample multidimensional matrix, and output the load identification result based on the distance result.

[0166] In this embodiment, the multi-dimensional matrix generating unit 2 includes:

[0167] An approximate harmonic feature acquisition module is used to acquire the segmented aggregation approximate features of the current and voltage waveform signals based on the segmented aggregation approximate processing, and to extract the current harmonic features corresponding to the current and voltage waveform signals using Fourier transform technology;

[0168] A power factor characteristic acquisition module is used to analyze the active power and total power of the device in the power grid based on the current and voltage waveform signals, and calculate the power factor characteristic based on the active power and the total power;

[0169] A feature matrix generation module is used to take the segmented aggregation approximate features, current harmonic features and power factor features as current periodic signals, and perform fusion processing on the current periodic signals after normalization to generate a multi-dimensional feature matrix;

[0170] The feature matrix processing module is used to calculate the information gain of the multidimensional feature matrix and perform weighted processing on the features in the multidimensional feature matrix according to the gain result.

[0171] To sum up, with the help of the above-mentioned technical scheme of the present invention, the present invention collects the user's voltage, current and other data through a data acquisition device, and pre-processes the data, and extracts statistics reflecting the load power consumption characteristics from the processed data, that is, load characteristics, and uses the extracted characteristics as the input of the load identification model algorithm, so as to realize accurate identification of the load through the algorithm. At the same time, the present invention selects three important characteristics of segmented aggregation approximate characteristics, harmonic characteristics and power factor to describe the load signal characteristics, and can flexibly reduce the dimension of the current according to the data sampling frequency, thereby achieving the transformation from high-dimensional data to low-dimensional data, and at the same time minimizing the loss of current characteristics. Compared with the existing methods of RMS features and current peak-to-peak features, the present invention can reduce the dimension of current by extracting segmented aggregation approximate features, thereby achieving the transformation from high-dimensional data to low-dimensional data, and at the same time minimize the loss of current features and ensure information integrity, so that the results have higher sensitivity and accuracy. The feature extraction of current signals using the segmented aggregation approximation algorithm can be dynamically adjusted with the signal sampling frequency, and more detailed information can be obtained without being affected by the frequency, thereby improving the accuracy of characterizing the current signal. The multidimensional feature matrix composed of PAA features, current harmonic features and power factor features is used as the input of the load identification model, which contains richer information and is conducive to improving the accuracy and sensitivity of load identification in terms of method.

[0172] Those skilled in the art will appreciate that the units, i.e., algorithm steps, of the various examples described in conjunction with this embodiment can be implemented in electronic hardware or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0173] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A non-intrusive load identification method based on a multidimensional feature matrix, characterized in that: The non-intrusive load identification method includes: Collect the main circuit current circuit data, perform pre-processing on the current circuit data, and divide the corresponding current and voltage waveform signal data for each electrical device in the main circuit; Extracting the current periodic signal based on the current and voltage waveform signals, generating a known multi-dimensional feature matrix according to the extraction results, and performing information entropy calculation and weighting processing on the multi-dimensional feature matrix; The weighted multi-dimensional feature matrix is ​​used as a training set of known electrical equipment and combined with the proximity algorithm to train the load identification model, and a multi-dimensional feature matrix vector of known category samples is obtained; Obtain unknown sample data to generate a multidimensional feature matrix of unknown samples, and input the multidimensional feature matrix of unknown samples into the trained load identification model to obtain a multidimensional feature matrix vector of samples to be identified; Calculate the distance between the multidimensional feature matrix vector of the known category sample and the multidimensional feature matrix vector of the sample to be identified, and select the vector closest to the multidimensional feature matrix vector of the sample to be identified based on the distance result. k Perform statistics on the multi-dimensional feature matrix vectors of samples of known categories, and output the load identification results based on the statistical results; The method of extracting the current periodic signal based on the current and voltage waveform signal, generating a known multidimensional feature matrix according to the extraction result, and performing information entropy calculation and weighted processing on the multidimensional feature matrix includes: The segmented aggregation approximate features of the current and voltage waveform signals are obtained based on the segmented aggregation approximate processing, and the current harmonic features corresponding to the current and voltage waveform signals are extracted using the Fourier transform technology; Analyze the active power and total power of the equipment in the power grid based on the current and voltage waveform signals, and calculate the power factor characteristics based on the active power and total power; The segmented aggregation approximate features, current harmonic features and power factor features are taken as current periodic signals, and the current periodic signals are normalized and then fused to generate a multi-dimensional feature matrix. The information gain of the multidimensional feature matrix is ​​calculated, and the features in the multidimensional feature matrix are weighted according to the gain result.

2. The non-intrusive load identification method based on a multidimensional feature matrix according to claim 1 is characterized in that: The method of obtaining the segmented aggregation approximate features of the current and voltage waveform signals based on the segmented aggregation approximate processing and extracting the current harmonic features corresponding to the current and voltage waveform signals using the Fourier transform technology includes: Performing segmented aggregation approximation processing on the current and voltage waveform signal data according to a preset window size set to obtain the current and voltage waveform signal data after dimensionality reduction; Calculate the mean square error between the current and voltage waveform signal data after dimension reduction and the current and voltage waveform signal data, and select the window size corresponding to the minimum mean square error as the segmented window size; The segmented aggregation approximate features corresponding to the current and voltage waveform signal data are obtained based on the segmented window size, and the current and voltage waveform signal data are converted into frequency domain signals using the fast Fourier transform technology; The harmonic component position is found from the frequency domain signal, the harmonic amplitude is extracted according to the position result, and the harmonic amplitude is normalized to the main frequency amplitude to obtain the current harmonic characteristics.

3. The non-intrusive load identification method based on a multidimensional feature matrix according to claim 2 is characterized in that: The active power and total power of the equipment in the power grid are analyzed based on the current and voltage waveform signals, and the power factor characteristics are calculated based on the active power and the total power, including: Based on the current and voltage waveform signals, the effective values ​​of current and voltage within the sampling period are obtained, and the effective value of current is multiplied by the effective value of voltage, and the multiplication result is used as the active power of the device in the power grid; Analyze the instantaneous current value and instantaneous voltage value corresponding to the current and voltage data points within the sampling period according to the current and voltage waveform signals, and obtain the total power of the device based on the instantaneous current value and the instantaneous voltage value; The ratio calculation is performed based on the active power and the total power, and the ratio result is used to describe the power factor characteristics of the electrical equipment in the power grid.

4. The non-intrusive load identification method based on a multidimensional feature matrix according to claim 3 is characterized in that: The calculation formula of the power factor characteristic is: ； ； In the formula, λ Indicates the power factor characteristics, P Indicates active power, S Indicates the total power, I rms Indicates the effective value of current, U rms Indicates the effective value of voltage. n Indicates the number of voltage and current data points within the sampling period, i t Indicates t The instantaneous current of each data point, u t Indicates t The instantaneous voltage of a data point.

5. The non-intrusive load identification method based on a multidimensional feature matrix according to claim 4 is characterized in that: The information gain calculation is performed on the multidimensional feature matrix, and weighted processing is performed on the features in the multidimensional feature matrix according to the gain result, including: Determine the proportion of the preset load category information in the multidimensional feature matrix, and calculate the information entropy of the multidimensional feature matrix based on the proportion result; The multidimensional feature matrix is ​​divided into different subsets based on the feature value range information, and the conditional entropy corresponding to the segmented aggregation approximate features, current harmonic features and power factor features in the multidimensional feature matrix is ​​analyzed according to the division results; The information gains corresponding to the segmented aggregation approximate features, current harmonic features and power factor features are calculated based on the conditional entropy and information entropy, and the information gains are normalized respectively; The normalized processing results are used as feature weighting coefficients to weight the multidimensional feature matrix, and weighted segmented aggregation approximate features, current harmonic features and power factor features are obtained.

6. The non-intrusive load identification method based on a multidimensional feature matrix according to claim 5 is characterized in that: The information entropy calculation formula of the multidimensional feature matrix is: ； In the formula, H ( D ) represents a multidimensional feature matrix D The information entropy of m represents the number of load categories, P i Indicates i Class load in multidimensional feature matrix D The proportion in .

7. A non-intrusive load identification system based on a multi-dimensional characteristic matrix, used to implement the non-intrusive load identification method based on a multi-dimensional characteristic matrix as claimed in any one of claims 1 to 6, characterized in that: The non-intrusive load identification system includes: The signal data acquisition unit is used to collect the current circuit data of the main circuit, and after pre-processing the current circuit data, divide the corresponding current and voltage waveform signal data for each electrical device in the main circuit; A multi-dimensional matrix generating unit, used for extracting a current periodic signal based on a current and voltage waveform signal, generating a known multi-dimensional feature matrix according to the extraction result, and performing information entropy calculation and weighted processing on the multi-dimensional feature matrix; The load result output unit is used to input the weighted multidimensional feature matrix into the load identification model, calculate the distance between the weighted multidimensional feature matrix and the unknown sample multidimensional matrix, and output the load identification result based on the distance result.

8. The non-intrusive load identification system based on a multi-dimensional characteristic matrix according to claim 7 is characterized in that: The multidimensional matrix generating unit comprises: An approximate harmonic feature acquisition module is used to acquire the segmented aggregation approximate features of the current and voltage waveform signals based on the segmented aggregation approximate processing, and to extract the current harmonic features corresponding to the current and voltage waveform signals using Fourier transform technology; A power factor characteristic acquisition module is used to analyze the active power and total power of the device in the power grid based on the current and voltage waveform signals, and calculate the power factor characteristic based on the active power and the total power; A feature matrix generation module is used to take the segmented aggregation approximate features, current harmonic features and power factor features as current periodic signals, and perform fusion processing on the current periodic signals after normalization to generate a multi-dimensional feature matrix; The feature matrix processing module is used to calculate the information gain of the multidimensional feature matrix and perform weighted processing on the features in the multidimensional feature matrix according to the gain result.

Images