A non-intrusive load identification method based on low-frequency sampling and feature fusion

Abstract

The present invention discloses a non-invasive load identification method based on low-frequency sampling and feature fusion. First, event detection is performed on the load data sampled at a low frequency, and the large and small power load events are classified according to the power changes before and after the switching point; the Gram angle field matrix and the power absolute value information matrix are established for the active power signal, and then the power multi-dimensional feature spectrum is obtained by matrix fusion. The load data is divided into single electrical load and combined electrical load according to the steady-state power value before the event is put into use. For single electrical load, a model based on transfer learning is directly used for identification, and the result is used as prior knowledge, and then a total load library is constructed according to the combined electrical load. For the combined electrical load, the prior knowledge is combined with the total load library, and the samples in the sub-load library are selected to form a sample pair, which is input into the twin network together. When a new combination appears, a new sub-load library is created to update the total load library, which can reduce the time for the twin network to identify the combined electrical appliance.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and relates to a load monitoring method for a smart power grid, and specifically to a non-intrusive load identification method based on low-frequency sampling and feature fusion. Background Art

[0002] In Non-Intrusive Load Monitoring (NILM), load identification technology plays a vital role.

[0003] When low-frequency sampling is used for household loads, it is difficult to obtain accurate voltage and current information due to the limitation of sampling frequency, and only the usage rate timing information can be used as the load identification characteristic signal. In the process of using household appliances, there are often multiple appliances running at the same time, and the operating power of these appliances is not constant. Therefore, when a new appliance is put into use, it is impossible to obtain the operating information of the corresponding appliance by performing a simple difference operation on the power signal before and after the input. In particular, when the new appliance is a low-power appliance, it is often masked by the fluctuation signal of the power of the running appliance due to its weak energy consumption characteristics, resulting in inaccurate identification.

[0004] In addition, the generalization performance of the supervised NILM algorithm in the existing technology is poor. The network needs to be retrained for newly emerging combined loads, and the convergence speed is slow and the accuracy is low. This not only affects the overall recognition accuracy, but also limits the applicability of NILM technology in diverse electrical appliance environments. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a non-intrusive load identification method based on low-frequency sampling and feature fusion. The Gram angular field matrix and the power absolute value information matrix are established for the one-dimensional time series signal of active power, and the power multi-dimensional feature spectrum is obtained by feature fusion. The characteristics of a single electrical appliance are introduced into the knowledge base as prior knowledge, and a total load library of combined electrical appliances is constructed; when a new combined electrical appliance appears, the corresponding sub-load library is constructed and added to the total load library of the combined electrical appliance, so as to update the total load library of the combined electrical appliance and enhance generalization and recognition speed.

[0006] A non-intrusive load identification method based on low-frequency sampling and feature fusion, the specific steps are as follows:

[0007] Step 1: Dual-threshold CUSUM event detection and classification

[0008] Set the first and second power thresholds H and h, where h = 1 / 2H, use the CUSUM event algorithm to capture events, and compare the absolute value change of power before and after the i-th event switching point With the size of the power threshold, when When , it is captured as a high power load event; when When it is detected, it is captured as a small power load event.

[0009] Step 2: GAF Matrix Construction

[0010] From the time the appliance is put into operation, after 25 power grid cycles, x sampling points are collected. , is the sampling time, and f is the acquisition frequency. According to the sampling frequency f, the data of x sampling points are converted into a one-dimensional power sequence of length N by decimation or interpolation. . At the same time, calculate the average value of x sampling points before the input time , the power sequence Subtract each value from , perform Gram angle field change and generate GAF matrix.

[0011] Step 3: Construction of power absolute value information matrix

[0012] First, create two The all-zero matrix, and then the one-dimensional power sequence in step 2 Fill in the first column of the matrix in sequence, and the values ​​of each row of the matrix are the same. Set the power value P REF1 <P REF2 <P REF3 , for the power sequence Encode the power absolute value information and obtain two power absolute value information matrices:

[0013] ①For high power load events:

[0014]

[0015]

[0016] Among them, p i represents the value of the i-th row of the matrix, , They represent rounding down and rounding up respectively; min() represents the minimum value, and max() represents the maximum value; L 1 (i) L 2 (i) represent the values ​​of the i-th row in the first and second power absolute value information matrix of the large power load event respectively.

[0017] ②For small power load events:

[0018]

[0019]

[0020] Among them, l 1 (i) l 2 (i) represent the values ​​of the i-th row in the first and second power absolute value information matrix of the small power load event respectively.

[0021] Step 4: Construction of power multi-dimensional characteristic spectrum

[0022] The GAF matrix and two power absolute value information matrices obtained in steps 2 and 3 are used as the three channels of the RGB color space to synthesize a power multidimensional feature spectrum. The color value of each pixel in the image represents the power absolute value information at the corresponding position. Return to step 2 until the power signals of all event intersection points are converted into power multidimensional feature spectra.

[0023] All event input points are judged. If the average value of x sampling points before the input moment is , marked as a single appliance input event, if the average value of x sampling points before the input time , marked as a combination electrical appliance input event.

[0024] Step 5: Build and train the SE-ResNeXt-50 network based on transfer learning

[0025] The power multi-dimensional feature spectra corresponding to all single electrical appliance input events are collected, and the single electrical appliance feature map data with load labels are divided into training sets and test sets. Then, the pre-trained ResNeXt-50 model is used to initialize the network parameters using the transfer learning method. Then, the pre-trained ResNeXt-50 model is migrated to the electrical appliance type recognition task, and the attention mechanism is introduced to construct the SE-ResNeXt50 model. The attention mechanism SE is adaptively calibrated according to the importance of each feature channel, and the calibrated features are fused with the original features to enhance the generalization ability and classification recognition effect of the model. During the training process, the cross entropy loss function is selected as the optimization target, the gradient is calculated by the back-propagation algorithm, and the SE-ResNeXt50 model parameters are updated using the Adam optimizer. The performance of the model on the training set and the validation set is monitored until the performance on the test set remains stable, and a single electrical appliance load identification network is obtained.

[0026] Step 6: Build a combined electrical appliance total load library and train a twin network

[0027] Step 6.1, collect the power multi-dimensional characteristic spectra corresponding to all combined electrical appliance input events, classify the power multi-dimensional characteristic spectra of the combined loads according to the known electrical appliance tags, build a sub-load library, and integrate all the sub-load libraries together to form a complete combined electrical appliance total load library.

[0028] Step 6.2: Divide all images in the combined electrical appliance total load library into training sets and test sets in a ratio of 4:3. Construct a twin network consisting of two weight-sharing backbone networks. In the training phase, the images in the training set are grouped into sample pairs, which are input into the twin network for feature information extraction. The binary cross entropy loss function is used as the optimization target, the gradient is calculated through the back-propagation algorithm, and the twin network parameters are updated using the Adam optimizer until the performance of the twin network on the test set is stable, and the combined electrical appliance load identification network is obtained.

[0029] Step 7: Test and Verify

[0030] Step 7.1. Collect the low-frequency power signals of the residential users to be identified in the past three days, capture the events according to the method in step 1, and classify the large and small power load events according to the set first and second power thresholds. After eliminating the nearly simultaneous event input points, record the number of remaining events and the time of each event. Initialize the load prior knowledge base.

[0031] Step 7.2: Process the event switching points in sequence according to the order in which the remaining events occur. If it is an event input point, proceed to step 7.3. If it is an event cut-out point, compare the power change with the information of the already input electrical appliances to obtain the information of the shut-down electrical appliances, i.e. the identification result of the cut-out electrical appliances, update the load prior knowledge base, and then process the next event switching point.

[0032] Step 7.3: Obtain the power multi-dimensional characteristic spectrum of the event input point. , marked as a single electrical appliance power multi-dimensional characteristic spectrum, go to step 7.4; if before the event is put into operation, , marked as the multi-dimensional characteristic spectrum of combined electrical power, and go to step 7.5.

[0033] Step 7.4: Input the multi-dimensional characteristic spectrum of the single electrical appliance power into the single electrical appliance load identification network obtained in step 5, output the corresponding identification result, update the load prior knowledge base, and return to step 7.2.

[0034] The updating method of the load prior knowledge base is: when the identification result of the event cut-out point is obtained, the corresponding cut-out electrical appliance information is deleted from the prior knowledge base; when the identification result of the event input point is obtained, the identification result is added to the load prior knowledge base. This method is used to track information in real time, ensuring the real-time performance of the load prior knowledge base.

[0035] Step 7.5, extract prior knowledge from the load prior knowledge base, and compare it with the total load base obtained in step 6. If the load prior knowledge base is empty, traverse the total load base of the combined electrical appliances. Determine whether there is a sub-load base corresponding to the multi-dimensional characteristic spectrum of the combined electrical appliance power in step 7.3. If not, create a new sub-load base with the multi-dimensional characteristic spectrum of the combined electrical appliance power as prior knowledge. After manually marking the load information, add the newly created sub-load base to the total load base of the combined electrical appliances; if there is a corresponding sub-load base, directly lock the sub-load base.

[0036] Step 7.6, the multi-dimensional feature spectrum of the combined electrical power and the images in the load library corresponding to the prior knowledge are combined into sample pairs, which are input into the combined electrical load identification network obtained in step 6. Feature information is extracted through two weight-sharing backbone networks, and then the similarity of the sample pairs is predicted.

[0037] Step 7.7: When the similarity is greater than the threshold, the electrical appliances that cause power conversion under the current combined load state and the currently running electrical appliances are output according to the labels of the sub-load library, and the combined electrical appliance load identification network outputs the load identification result. If the similarity is less than the threshold, it is determined that a new combined load has appeared, the unknown load is manually labeled, and a new sub-load library is created with the combined load as prior knowledge, and the sub-load library is added to the combined electrical appliance total load library. Return to step 7.2.

[0038] The present invention has the following beneficial effects:

[0039] 1. Through the RGB feature fusion method, the Gram angular field matrix and the active power feature matrix are fused to obtain the power multi-dimensional feature spectrum, which not only retains the relationship between the load power and time, but also integrates the absolute value information of the load power, further enriching the sample information and describing the characteristics of the load more comprehensively.

[0040] 2. By monitoring the average power of electrical appliances when they are started, the energy consumption characteristics of single appliances and combined appliances are effectively distinguished. The SE-ResNeXt50 network and twin network based on transfer learning are used to process these characteristics separately, which can not only effectively distinguish and process the load patterns of single appliances and combined appliances, but also reduce the complexity of network training, reduce the required training time, and significantly improve the accuracy of load identification due to the targeted processing strategy.

[0041] 3. Construct a priori knowledge base of loads and a total load base of combined electrical appliances. The priori knowledge base of loads provides the network with accurate positioning of electrical appliance categories. The total load base of combined electrical appliances stores multi-dimensional characteristic spectra of known combined electrical appliances, which not only enables the twin network to accurately identify combined electrical appliances, but also effectively reduces the amount of calculation of the twin network. In addition, even when encountering new combined loads, there is no need for additional training of the twin network, which not only improves the flexibility and adaptability of the recognition process, but also effectively improves the accuracy of overall load recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram for constructing a multi-dimensional characteristic spectrum of power;

[0043] Figure 2 Schematic diagram of the SE-ResNeXt50 structure based on transfer learning;

[0044] Figure 3 This is a schematic diagram of the twin network structure in the embodiment;

[0045] Figure 4 Flowchart for model training and testing;

[0046] Figure 5 It is the identification result of transfer learning based on the power multi-dimensional feature map in the embodiment;

[0047] Figure 6 This is the identification result of transfer learning based on the Gram angular field matrix in the embodiment. DETAILED DESCRIPTION

[0048] This embodiment uses LIFTED and EMBED data sets as collected load data, and uses the low-frequency signal sampling information therein to implement a non-intrusive load identification method based on low-frequency sampling and feature fusion. The specific steps are as follows:

[0049] Step 1: Dual-threshold CUSUM event detection and classification

[0050] The low-frequency signal sampling information from the LIFTED and EMBED data sets for seven days is processed by median filtering. Then the CUSUM event algorithm is used to capture the switching events. The CUSUM event algorithm detects possible events by accumulating the difference between the observed value and the expected value, and issues an alarm when the cumulative sum exceeds the preset threshold to capture the occurrence and end time of the event, and records the switching point of each event and the absolute value change of power before and after the i-th event switching point. and the number of times the event occurred.

[0051] Set the first and second power thresholds H and h, where h = 1 / 2H. When the appliance is turned on When the appliance is turned on, it is marked as a high power load event; When , it is marked as a small power load event.

[0052] Step 2: GAF Matrix Construction

[0053] like Figure 1 As shown in the figure, from the time the appliance is put into operation, after 25 power grid cycles, x sampling points are collected. , is the sampling time, Take more than 10s, f is the acquisition frequency. According to the size of the sampling frequency f, the collected data is converted into a power sequence of length N by extraction or interpolation. . At the same time, calculate the average value of x sampling points before the input time , the power sequence Subtract each value from , and then perform Gram angle field change to generate the GAF matrix. The specific steps are as follows:

[0054] Step 2.1: Subtract The power sequence after Normalize to [ − 1 , 1 ] In scope:

[0055]

[0056] in, Subtract the i-th sampling point in the sequence The power value after represents the normalized power value, is the maximum value in the power series, is the minimum value in the power series.

[0057] Step 2.2: Convert the normalized power value from the rectangular coordinate system to the polar coordinate system:

[0058]

[0059] in, is the polar coordinate angle, is the polar coordinate radius; Representative The time point corresponding to each sampling point; is the size of the two-dimensional image.

[0060] Step 2.3: Define the inner product of the time series Calculation method:

[0061]

[0062] Among them, j represents the jth sampling point, I is the identity matrix. The Gram matrix retains the normalized values ​​of the data points and their temporal positions through a dual mapping function, allowing time to increase as the position moves from the upper left corner to the lower right corner, thereby encoding the time dimension into the geometric structure of the matrix. Converted to GAF matrix:

[0063]

[0064] Step 3: Construction of power absolute value information matrix

[0065] In order to To encode the absolute value of the power, first create two The all-zero matrix, in turn, transforms the power sequence The elements in are filled into the first column of the two matrices, and the values ​​of other columns in the matrix are consistent with the first value of the row. Set the power value =300, =1000, =2000, according to the power sequence The event corresponding to the first sampling point in is encoded:

[0066] ①For high power load events:

[0067]

[0068]

[0069] Among them, p i represents the value of the i-th row of the matrix, , They represent rounding down and rounding up respectively; min() represents the minimum value, and max() represents the maximum value; L 1 (i) L 2 (i) represent the values ​​of the i-th row in the first and second power absolute value information matrix of the large power load event respectively.

[0070] ②For small power load events:

[0071]

[0072]

[0073] Among them, l 1 (i) l 2 (i) represent the values ​​of the i-th row in the first and second power absolute value information matrix of the small power load event respectively.

[0074] In a high-power load event, the power span before and after the event is large. Using two standard values ​​to normalize it can effectively reflect the absolute power difference of different electrical appliances when they are running. Some electrical appliances have similar power sizes when they are running. In order to highlight their differences in the image, their power energy levels are accurately captured. Therefore, for the power absolute value information matrix The line adopts saturation value design, and then The remainder design is used to reflect its small fluctuations.

[0075] Step 4: Construction of power multi-dimensional feature spectrum

[0076] The GAF matrix and two power absolute value information matrices obtained in steps 2 and 3 are used as the three channels of the RGB color space to synthesize a power multidimensional feature spectrum. The color value of each pixel in the image represents the power absolute value information at the corresponding position. Return to step 2 until the power signals of all event intersection points are converted into power multidimensional feature spectra.

[0077] All event input points are judged. If the average value of x sampling points before the input moment is , marked as a single appliance input event, if the average value of x sampling points before the input time , marked as a combination electrical appliance input event.

[0078] Step 5: Build and train the SE-ResNeXt-50 network based on transfer learning

[0079] The power multi-dimensional feature spectra corresponding to all single electrical appliance input events are collected, labeled with the appliance names, and then divided into training and test sets in a ratio of 4:3.

[0080] Using the transfer learning method, we use ResNeXt-50 pre-trained on the ImageNet-1K dataset to initialize the network parameters, and then add the SE attention mechanism to construct Figure 2 The SE-ResNeXt-50 model based on transfer learning is shown.

[0081] During the training process, the cross entropy loss function is selected as the optimization objective, the gradient is calculated through the back propagation algorithm, and the Adam optimizer is used to update the SE-ResNeXt-50 model parameters. The performance of the model on the training set and the test set is monitored until its performance on the test set remains stable, thus obtaining a single electrical load identification network.

[0082] The cross entropy loss function for:

[0083]

[0084] In the formula, is the true label of the i-th training sample, is the model prediction label, and n is the number of samples.

[0085] Step 6: Build a combined electrical appliance total load library and train a twin network

[0086] Step 6.1: Collect the power multi-dimensional characteristic spectra corresponding to all combined electrical appliance input events, classify the power multi-dimensional characteristic spectra of combined electrical appliances according to known electrical appliance tags, and build the corresponding sub-load library. Integrate all sub-load libraries together to form a complete combined electrical appliance total load library.

[0087] Step 6.2: Construct a twin network consisting of two weight-sharing backbone networks, such as Figure 3 As shown in the figure, all the multi-dimensional feature spectra of power in the total load library of combined electrical appliances are divided into training set and test set in a ratio of 4:3. In the training stage, the images in the training set are paired into sample pairs as training samples of the twin network for feature extraction and recognition.

[0088] During the training process, the binary cross entropy loss function is used as the optimization target, the gradient is calculated through the back propagation algorithm, and the Adam optimizer is used to update the twin network parameters until the performance of the network on the test set is stable, and the combined electrical load identification network is obtained.

[0089] The binary cross entropy loss function for:

[0090]

[0091] in, is the true label of the sample pair, is the network output, is the Sigmoid function.

[0092] like Figure 4 As shown, through steps five and six, the training process of the identification model and the construction of the total load library of the combined electrical appliances are completed.

[0093] Step 7: Test and Verify

[0094] Step 7.1. In addition, collect the low-frequency power signals of the residential users to be identified in the past three days as the load data to be identified, and perform event capture according to the method of step 1, and classify the large and small power load events according to the set first and second power thresholds. If there are other event input points within 25 grid cycles after an event input point or among the x points collected, that is, there are nearly simultaneous event input points, it will result in the inability to obtain the power multi-dimensional feature spectrum of the event. Therefore, the input points of nearly simultaneous events should be eliminated, and the number of remaining event occurrences and the time of each event should be recorded. Initialize the load prior knowledge base.

[0095] Step 7.2: Process the event switching points in sequence according to the order in which the remaining events occur. If it is an event input point, proceed to step 7.3. If it is an event cut-out point, compare the power change with the information of the already input electrical appliances to obtain the information of the shut-down electrical appliances, i.e. the identification result of the cut-out electrical appliances, update the load prior knowledge base, and then process the next event switching point.

[0096] Step 7.3: Obtain the power multi-dimensional characteristic spectrum of the event input point. , marked as a single electrical appliance power multi-dimensional characteristic spectrum, go to step 7.4; if before the event is put into operation, , marked as the multi-dimensional characteristic spectrum of combined electrical power, and go to step 7.5.

[0097] In order to demonstrate the advantages of the power multidimensional feature spectrum proposed in this method, the power multidimensional feature spectrum and the ordinary Gram angular field matrix image are used for load identification respectively.

[0098] The results of load identification using the power multidimensional feature spectrum are shown in Table 1. The confusion matrix is ​​shown in Figure 5 As shown:

[0099] Table 1

[0100]

[0101] The results of load identification using the normal Gram angular field matrix image are shown in Table 2. The confusion matrix is ​​shown in Figure 6 As shown:

[0102] Table 2

[0103]

[0104] Among them, Precision Pre represents the ratio of the number of samples correctly predicted as positive examples by the classifier to the number of samples predicted as positive examples by the classifier; Recall R represents the ratio of the number of samples correctly predicted as positive examples by the classifier to the number of actual positive examples; F 1The score is the harmonic mean of precision and recall; accuracy is the ratio of the number of correctly classified samples by the classifier to the total number of samples.

[0105] From the table data, we can see that the accuracy of load identification using the ordinary Gram angular field matrix image is only 0.951, while the accuracy of load identification using the power multidimensional feature spectrum proposed by this method can reach 0.995, and P re , R, F 1 The values ​​are significantly improved, proving that the power multidimensional characteristic spectrum proposed by this method can better reflect the information of the original data.

[0106] Step 7.4: Input the single electrical appliance power multi-dimensional characteristic spectrum obtained in step 7.3 into the single electrical appliance load identification network obtained in step 5 to obtain the corresponding identification result. Update the load prior knowledge base and return to step 7.2.

[0107] The updating method of the load prior knowledge base is: when the identification result of the event cut-out point is obtained, the corresponding cut-out electrical appliance information is deleted from the prior knowledge base; when the identification result of the event input point is obtained, the identification result is added to the prior knowledge base.

[0108] Step 7.5, extract prior knowledge from the load prior knowledge base, and compare it with the total load base obtained in step 6. If the load prior knowledge base is empty, traverse the total load base of the combined electrical appliances. Determine whether there is a sub-load base corresponding to the multi-dimensional characteristic spectrum of the combined electrical appliance power in step 7.3. If not, create a new sub-load base with the multi-dimensional characteristic spectrum of the combined electrical appliance power as prior knowledge. After manually marking the load information, add the newly created sub-load base to the total load base of the combined electrical appliances; if there is a corresponding sub-load base, directly lock the sub-load base.

[0109] Step 7.6, the multi-dimensional feature spectrum of the combined electrical power and the images in the load library corresponding to the prior knowledge are combined into sample pairs, which are input into the combined electrical load identification network obtained in step 6. Feature information is extracted through two weight-sharing backbone networks, and then the similarity of the sample pairs is predicted.

[0110] Step 7.7: When the similarity is greater than the threshold, the electrical appliances that cause power conversion under the current combined load state and the currently running electrical appliances are output according to the labels of the sub-load library, and the combined electrical appliance load identification network outputs the load identification result. If the similarity is less than the threshold, it is determined that a new combined load has appeared, the unknown load is manually labeled, and a new sub-load library is created with the combined load as prior knowledge, and the sub-load library is added to the combined electrical appliance total load library. Return to step 7.2.

[0111] The monitor and refrigerator are used as the load distribution library to identify other switched electrical appliances. The specific results are shown in Table 3:

[0112] Table 3

[0113]

[0114] It can be seen that the recognition accuracy of this method reached 0.9733, and the precision P re and F 1 The scores are all above 0.9, and the recall rate R Keep it at 1. It is proved that this method can effectively realize the accurate identification of combined electrical loads.

Claims

1. A non-intrusive load identification method based on low-frequency sampling and feature fusion, which sets a power threshold, performs event identification on the low-frequency sampling data of the power load, and obtains a series of event switching points, characterized in that: The specific steps are as follows: Step 1: For an input event, the sampling point data after the event occurs is converted into a one-dimensional power sequence to generate a Gram angular field matrix; at the same time, the power absolute value information in the one-dimensional power sequence is encoded to obtain two power absolute value information matrices; Step 2: The Gram angle field matrix and the power absolute value information matrix of an input event are respectively used as three channels of the RGB color space to obtain a power multi-dimensional feature spectrum corresponding to the input event; Step 3: Calculate the average value of multiple sampling points before the input event occurs. If the average value is less than the threshold, the input event is determined to be a single appliance input, otherwise it is a combination of appliances input; Step 4: Use the power multi-dimensional feature spectrum of a single electrical appliance input event to train a classification network as a single electrical appliance load identification network; classify the power multi-dimensional feature spectrum of the combined load according to the known electrical appliance labels, build a sub-load library, and integrate all the sub-load libraries together to form a complete combined electrical appliance total load library; All images in the combined electrical appliance total load library are grouped into sample pairs, which are input into the twin network. The twin network composed of two weight-shared backbone networks is trained to obtain a combined electrical appliance load identification network. Step 5: Initialize a load prior knowledge base. For the power load data to be identified, firstly capture the event, then process each event switching point in sequence according to the order of event occurrence. When the identification result of the event input point is obtained, the identification result is added to the load prior knowledge base; when the identification result of the event cut-out point is obtained, the corresponding cut-out appliance information is deleted from the prior knowledge base: ① For the event cut-out point, compare the power change before and after the event with the information of the electrical appliances that have been put into use, and output the identification result of the cut-out electrical appliances; ② For a single electrical appliance input event, the power multi-dimensional characteristic spectrum corresponding to the event is input into the single electrical appliance load identification network to obtain the identification result of the single electrical appliance input; ③ For the event of combined electrical appliances being put into operation, firstly, the prior knowledge in the load prior knowledge base is compared with the total load base of the combined electrical appliances to find the corresponding sub-load base, and the samples in the sub-load base are selected to form a sample pair with the power multi-dimensional feature spectrum of the combined electrical appliances being put into operation. The sample pairs are input into the combined electrical appliances load identification network to calculate the similarity. When the similarity is greater than the threshold, the combined electrical appliances identification result is output, otherwise it is deemed that a new combined load has appeared. When the corresponding sub-load base cannot be found or a new combined load has appeared, a new sub-load base is created after manual marking, and the total load base of the combined electrical appliances and the load prior knowledge base are updated.

2. A non-intrusive load identification method based on low-frequency sampling and feature fusion as claimed in claim 1, characterized in that: In step 1, the specific method of generating the Gram angular field matrix is ​​as follows: starting from the input event, after 25 power grid cycles, collect x sampling points, x = τ × f, τ is the sampling time, f is the acquisition frequency; convert the data of the x sampling points into a one-dimensional power sequence P(t) of length N by decimation or interpolation; calculate the average value of the x sampling points before the input moment Subtract each value in the power sequence P(t) from Then perform Gram angle field transformation to generate Gram angle field matrix.

3. The non-intrusive load identification method based on low-frequency sampling and feature fusion as claimed in claim 1, characterized in that: In step one, the specific method for obtaining the power absolute value information matrix is as follows: Starting from the input event, after 25 power grid cycles, x sampling points are collected, where x = τ × f, τ is the sampling time, and f is the sampling frequency; the data of the x sampling points is converted into a one-dimensional power sequence P(t) of length N through decimation or interpolation. Two all-zero matrices of size N×N are created, and the values of the one-dimensional power sequence P(t) are filled into the first column of the matrix in sequence, and the values of each row of the matrix are the same. Set the first-class and second-class power thresholds H and h, where h = 1 / 2H. When |P| ≥ H, it is captured as a high-power load event; when h ≤ |P| < H, it is captured as a low-power load event, and |P| represents the change in the absolute value of the power before and after the event injection point. ① For high-power load events: Among them, p i represents the value of the i-th row of the matrix, They represent rounding down and rounding up respectively; min() represents taking the minimum value, and max() represents taking the maximum value; L1(i) and L2(i) represent the values ​​of the i-th row in the first and second power absolute value information matrix of the large power load event respectively; ② For low-power load events: Wherein, l1(i) and l2(i) represent the values ​​of the first and second absolute power information matrices of the small power load event, respectively; P REF1 , P REF2 , P REF3 are all power absolute value thresholds, and P REF1 <P REF2 <P REF3 .

4. The non-intrusive load identification method based on low-frequency sampling and feature fusion as claimed in claim 3, characterized in that: P REF1 =300、P REF2 =1000、P REF3 =2000。 5. The non-intrusive load identification method based on low-frequency sampling and feature fusion as claimed in claim 1, characterized in that: In step four, the specific method for training a single electrical appliance load identification network is as follows: Construct a classification model SE-ResNeXt50 based on the attention mechanism SE. The attention mechanism SE adaptively calibrates according to the importance of each feature channel and fuses the calibrated features with the original features; using the method of transfer learning, the parameters of SE-ResNeXt50 are initialized with the pre-trained ResNeXt-50 model; the power multi-dimensional feature spectrogram of a single electrical appliance input event is used to train the classification model SE-ResNeXt50. During the training process, the cross-entropy loss function is selected as the optimization objective until the performance of the model remains stable.

6. A non-intrusive load identification method based on low-frequency sampling and feature fusion as claimed in claim 5, characterized in that: The ResNeXt-50 model is pre-trained using the ImageNet-1K dataset.

7. The non-intrusive load identification method based on low-frequency sampling and feature fusion as claimed in claim 1, characterized in that: In step four, the binary cross-entropy loss function is used as the optimization objective, the gradient is calculated through the backpropagation algorithm, and the parameters of the siamese network are updated using the Adam optimizer until the performance of the siamese network is stable.

8. The non-intrusive load identification method based on low-frequency sampling and feature fusion as claimed in claim 7, characterized in that: The binary cross-entropy loss function f2(x) is: Among them, y is the true label of the sample pair, is the network output result, and σ is the Sigmoid function.

Images