CN111639586A - Non-invasive load identification model construction method, load identification method and system - Google Patents

Abstract

The invention discloses a non-intrusive load identification model construction method, load identification method and system, belonging to the field of non-intrusive load identification, comprising: establishing a to-be-trained model including a feature extraction module, a classifier module and a similarity measurement module; The training set is composed of electrical samples from multiple families; h families are selected from the training set, k samples are selected from each family, and the feature extraction module is input to obtain the electrical features; the electrical features are input into the classifier module and converted into class labels, and calculated Category loss; input the electrical features into the similarity measurement module, splicing the electrical features of each two families correspondingly, and calculate the similarity loss; take the sum of the two losses as the overall loss, and update the model parameters to minimize the overall loss; repeat training After several rounds, the load identification model is composed of the feature extraction module and the classifier module. The invention can construct a load identification model with domain generalization ability, and improve the accuracy of non-intrusive load identification.

Description

Non-intrusive load identification model construction method, load identification method and system

technical field

The invention belongs to the field of non-invasive load identification, and more particularly, relates to a non-invasive load identification model construction method, load identification method and system.

Background technique

Countries around the world are developing and deploying smart grids and related applications. Using the data collected by smart meters, the advantages of smart grids can be maximized. Using the data collected by the smart meter, the type of electrical appliances in the user's home can be identified, that is, load identification can be realized.

Load identification methods can be divided into two types: intrusive and non-intrusive according to the number of sensors. Intrusive load identification requires the installation of corresponding sensors for each electrical appliance. Additional equipment and higher costs make the intrusive method difficult to promote. Non-intrusive load identification only needs to collect data from a single bus smart meter installed in the home, and through the data analysis of the smart meter, the total household electricity consumption is decomposed into the energy consumption of a single appliance, which helps to realize electricity consumption feedback , help users save energy, and at the same time facilitate accurate billing on the supply side. Compared with invasive load identification, non-invasive load identification has lower cost and is easy to promote, so it has been widely studied.

Non-intrusive load identification methods often need to use models to identify electrical appliances. In practical applications, electrical appliances in different households vary greatly due to user behavior and electrical appliance models. It covers all possible samples of electrical appliances from different families, so different home environments create different application scenarios. In addition, commonly used machine learning algorithms and deep learning models are prone to overfitting to the training set data. Therefore, how to obtain a generalized and generalized load identification model across application scenarios is a difficulty in non-intrusive load identification.

SUMMARY OF THE INVENTION

In view of the defects and improvement requirements of the prior art, the present invention provides a non-intrusive load identification model construction method, load identification method and system, the purpose of which is to construct a generalized and universal load identification model across application scenarios , in order to improve the identification accuracy of non-intrusive load identification.

In order to achieve the above object, according to an aspect of the present invention, a method for constructing a non-invasive load identification model with domain generalization capability is provided, including:

A model to be trained is established based on the deep neural network, and the model to be trained includes: a feature extraction module, a classifier module and a similarity measurement module;

The current signal data including electrical switching events are intercepted from the historical current signal data collected at the bus, and converted into a three-dimensional complex spectrogram. A three-dimensional complex spectrogram and the corresponding electrical appliance category constitute an electrical appliance sample; The electrical samples constitute the training set;

Randomly select h families from the training set, and randomly select k from the electrical samples of each family as the training samples of the current round; input the three-dimensional complex spectrograms in all training samples into the feature extraction module for feature extraction, and obtain Corresponding electrical features; input the obtained electrical features into the classifier module, convert them into corresponding category labels, and calculate the category loss in combination with the real electrical categories; input the obtained electrical features into the similarity measurement module, for each two households The electrical features are spliced, and a splicing feature is composed of two electrical features belonging to different families, and the similarity score between the two electrical features in each splicing feature is calculated, and the similarity loss is calculated in combination with the real electrical category; The sum of the similarity loss and the loss is used as the overall loss of the current round, and the back-propagation algorithm is used to update the parameters of each layer of the model to be trained to minimize the overall loss, thus completing the training of the current training round;

Perform multiple rounds of training on the model to be trained, and after the training is completed, use the trained feature extraction module and classifier module to form a load identification model;

Among them, the similarity score is used to measure the similarity between two electrical features; h≥2, k≥1.

In the model training process of the present invention, in addition to the feature extraction module and the classifier module, the established model to be trained also includes a similarity measurement module that can be used to extract similarity scores between electrical features belonging to different families. The similarity score can be used to further calculate the similarity loss; in the training process, the sum of the similarity loss and the category loss is used as the overall loss. By minimizing the overall loss, the identification result of the load category can be guaranteed to be as close to the basis of the real category as possible. On the other hand, the distance between the characteristics of similar electrical appliances in different user domains is shortened, so that similar electrical appliances belonging to different families are similar or even the same in the feature space, so that the established load identification model can perform load identification for any new sample. The load type corresponding to the new sample can be accurately identified based on the same old samples that have appeared in the training process, which can effectively eliminate the user domain difference caused by different families and realize the generalization of the user domain. Therefore, the load identification model constructed by the present invention has the capability of domain generalization. When the load identification model constructed by the present invention is used for non-intrusive load identification, the influence caused by user domain differences can be eliminated, and the non-intrusive load can be effectively improved. recognition accuracy.

Further, the non-invasive load identification model construction method with domain generalization capability provided by the present invention also includes:

(S1) After establishing the load identification model, select one or more household electrical appliance samples to form a test set, and the test set and the training set do not overlap each other;

(S2) test the load identification model by using the electrical samples in the test set to evaluate whether the identification accuracy of the load identification model meets the application requirements, if not, then go to step (S3); otherwise, go to step (S4);

(S3) after the training parameters are adjusted, the training model is retrained, and after the training, the load identification model is reconstituted by the feature extraction module and the classifier module, and goes to step (S2);

(S4) The test ends.

After constructing the load identification model, the present invention further uses electrical samples in different user domains to form a test set, evaluates the training effect of the load identification model, and re-trains the model when the identification accuracy of the model does not meet the application requirements, thereby It can further ensure that the established load identification model has better generalization ability, thereby ensuring the accuracy of load identification.

Further, h=2.

In the process of model training, the present invention randomly selects electrical samples of two families from the training set as training samples each time, so that the similarity measurement module only needs to simply splicing the samples of two groups of families correspondingly once when performing feature splicing. , without considering other combinations, so that while ensuring the model has strong domain generalization ability, it can avoid the problem that the amount of calculation is too large or even exceeds the computing power of the hardware due to too many training samples in one round. .

Further, when calculating the similarity loss, the similarity score between the two electrical features in the splicing feature and the binary cross entropy of the similarity label of the splicing feature are used as the similarity loss of a single splicing feature, and all splicing in the current round are used as the similarity loss. The average of the similarity loss of features is taken as the similarity loss of the current round;

Among them, the lower the similarity score, the higher the similarity between the two electrical features in the splicing feature; the similarity label of the splicing feature is used to indicate whether the true categories of the two electrical appliances involved in the splicing feature are the same.

In the present invention, the binary cross-entropy of the similarity score and the similarity label is used as the similarity loss of a single splicing feature, and the average value of the similarity loss of the splicing feature is used as the similarity loss of the current round, so that with the training process , the overall loss of training continues to decrease, the similarity loss of each round also continues to decrease, and the similarity loss of a single splicing feature also continues to decrease. Finally, the same kind of electrical samples from different households can obtain a unified feature expression, which can effectively Eliminate user domain differences caused by different families and realize user domain generalization.

Further, the similarity loss of a single splice feature is:

为拼接特征中两个电器特征之间的相似性得分。Among them, l ₁ represents the similarity loss of a single splicing feature; y is the similarity label of the splicing feature. When the true categories of the two electrical appliances involved in the splicing feature are the same, y=0, the real category of the two electrical appliances involved in the splicing feature is the same. When the categories are different, y=1;

is the similarity score between two electrical features in the spliced feature.

Further, the class loss is the cross-entropy loss.

Further, the current signal data including the electrical switching event is converted into a three-dimensional complex spectrogram through short-time Fourier transform.

According to another aspect of the present invention, a non-intrusive load identification method with domain generalization capability is provided, comprising: monitoring the target load to be identified in real time, and when detecting the switching event of the target load, from the bus Intercept the current signal data including the switching event from the collected current signal data;

The intercepted current signal data is converted into a three-dimensional complex spectrogram, and the three-dimensional complex spectrogram is used as input, and the load identification model is obtained by using the non-intrusive load identification model construction method with domain generalization capability provided by the present invention to identify the target load. type.

Since the load identification model constructed by the present invention has the capability of domain generalization, the present invention utilizes the model for non-invasive load identification, which can effectively improve the identification accuracy.

According to another aspect of the present invention, a non-intrusive load identification system with domain generalization capability is provided, comprising: a monitoring module, a conversion module and a load identification module;

The monitoring module is used to monitor the target load that needs to be identified in real time, and when the switching event of the target load is detected, the current signal data containing the switching event is intercepted from the current signal data collected at the bus, and the conversion module is triggered. ;

The conversion module is used to convert the intercepted current signal data into a three-dimensional complex spectrogram, and trigger the load identification module;

The load identification module is used for inputting the three-dimensional complex spectrogram output by the conversion module, and using the non-intrusive load identification model construction method with domain generalization capability provided by the present invention to obtain the load identification model to identify the type of target load.

In general, through the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

(1) In the present invention, a similarity measurement module for calculating similarity scores between electrical features belonging to different families is introduced into the model to be trained, and in the training process, the sum of similarity loss and category loss is used as a whole Loss, can achieve unified feature expression for similar electrical samples from different families, effectively eliminate user domain differences caused by different families, and achieve user domain generalization. Therefore, the load identification model constructed by the present invention has the capability of domain generalization. When the load identification model constructed by the present invention is used for non-intrusive load identification, the influence caused by user domain differences can be eliminated, and the non-intrusive load can be effectively improved. recognition accuracy.

(2) The present invention can monitor the load switching event to be identified in real time, and only needs to intercept a section of current signal data including the electrical switching event during identification, with timely feedback and high accuracy.

(3) The present invention utilizes the idea of domain generalization. The training set and the test set use electrical samples from different families, and the same kind of electrical samples of each family in the training set obtain a unified feature expression, which weakens the user domain characteristics of the electrical samples, and makes the model The generalization ability is enhanced, and when new user scenarios are applied, it can still maintain good accuracy without retraining.

Description of drawings

1 is a flowchart of a method for constructing a non-intrusive load identification model with domain generalization capability provided by an embodiment of the present invention;

2 is a schematic structural diagram of a model to be trained provided by an embodiment of the present invention;

FIG. 3 is a flowchart of a non-intrusive load identification method with domain generalization capability provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

In the present invention, the terms "first", "second" and the like (if present) in the present invention and the accompanying drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

In order to construct a non-intrusive load identification model with domain generalization ability, in an embodiment of the present invention, a method for constructing a non-intrusive load identification model with domain generalization ability is provided, as shown in FIG. 1 , include:

A model to be trained is established based on a deep neural network. As shown in Figure 2, the model to be trained includes: a feature extraction module, a classifier module and a similarity measurement module; wherein, the feature extraction module is used to perform feature extraction on the three-dimensional complex spectrogram to obtain Corresponding electrical features; the classifier module is used to convert the electrical features extracted by the feature extraction module into the category labels of electrical appliances; the similarity measurement module is used to splicing two electrical features belonging to different families to obtain splicing features, and calculate The similarity score between the two electrical features in the spliced feature, the similarity score is used to measure the similarity between the two electrical features;

The current signal data including electrical switching events are intercepted from the historical current signal data collected at the bus, and converted into a three-dimensional complex spectrogram. A three-dimensional complex spectrogram and the corresponding electrical appliance category constitute an electrical appliance sample; The electrical samples constitute the training set;

Randomly select h=2 households from the training set, and randomly select k=32 electrical appliances samples from each household, a total of h×k=64 electrical appliances samples, as the training samples of the current round; The three-dimensional complex spectrogram is input into the feature extraction module for feature extraction, and the corresponding electrical features are obtained, and a total of 64 electrical features are obtained; the obtained electrical features are input into the classifier module, converted into corresponding category labels, and combined with real electrical categories Calculate the category loss; input the obtained electrical features into the similarity measurement module, splicing the electrical features of the two selected families, and form a splicing feature from two electrical features belonging to different families. The 32 electrical features of one family and the 32 electrical features from the second family are correspondingly spliced, and a total of 32 splicing features are obtained by splicing, and the similarity score between the two electrical features in each splicing feature is calculated, and a total of 32 similarity scores, combined with the real electrical categories to calculate the similarity loss; the sum of the category loss and similarity loss is used as the overall loss of the current round, and the back-propagation algorithm is used to update the parameters of each layer of the model to be trained. Minimize the overall loss to complete the current training epoch;

The to-be-trained model is trained for 30 rounds. After the training is completed, the trained feature extraction module and the classifier module are used to form a load identification model.

As shown in Figure 2, in this embodiment, the feature extraction module is a convolutional neural network DenseNet-121, and the similarity measurement module includes: three cascaded fully connected layers FC(1)-FC(3), a classifier module It includes one fully connected layer FC(4); it should be noted that this is only an optional implementation manner of the present invention, and should not be construed as the only limitation of the present invention; in other embodiments of the present invention, The feature extraction module can also be replaced by other models capable of feature extraction. The number of cascaded fully connected layers in the similarity measurement module can also be set to other parameters, and even other implementations can be used to implement the similarity measurement module. Similarly, The classifier module can also be implemented by multiple cascaded fully connected layers, or implemented by other implementations;

In this embodiment, the activation function of the last fully connected layer in the similarity measurement module is the sigmoid function, the activation function of the last fully connected layer in the classifier module is the softmax function, and the activation functions of the remaining fully connected layers are ReLU function; also, this is only an optional embodiment of the present invention, and should not be construed as the sole limitation of the present invention.

As shown in FIG. 1 , the method for constructing a non-invasive load identification model with domain generalization capability provided by this embodiment further includes:

(S1) After establishing the load identification model, select one or more household electrical appliance samples to form a test set, and the test set and the training set do not overlap each other;

(S2) test the load identification model by using the electrical samples in the test set to evaluate whether the identification accuracy of the load identification model meets the application requirements, if not, then go to step (S3); otherwise, go to step (S4);

(S3) after the training parameters are adjusted, the training model is retrained, and after the training, the load identification model is reconstituted by the feature extraction module and the classifier module, and goes to step (S2);

(S4) The test ends.

In this embodiment, after the load identification model is constructed, a test set is further formed by using electrical samples in different user domains, the training effect of the load identification model is evaluated, and the model training is re-trained when the identification accuracy of the model does not meet the application requirements. This can further ensure that the established load identification model has good generalization ability, thereby ensuring the accuracy of load identification.

In this embodiment, h=2, that is, the electrical samples of two households are randomly selected from the training set as the training samples of the current round, so that the similarity measurement module only needs to simply combine the two groups of households when performing feature splicing. The samples corresponding to splicing once, compared to other values, for example, when h=3, three households are combined in twos, and a total of 3 corresponding splices are required. For another example, when h=4, four households are combined in twos, a total of Corresponding to 6 times of splicing, this embodiment can effectively reduce the number of training samples in each round, thereby reducing the amount of calculation, and the experimental data shows that only two household electrical samples are selected as training samples in each round, and the final obtained The load identification model already has strong domain generalization ability. Therefore, this embodiment can ensure that the model has a strong domain generalization ability, and at the same time avoid the problem that the amount of calculation is too large due to too many training samples in one round, or even exceeds the computing power of the hardware;

It should be noted that h=2 is only a preferred embodiment of the present invention, and should not be construed as the only limitation of the present invention; in other application scenarios, if there are higher requirements for the domain generalization ability of the model, h can also be set to a larger value accordingly.

In this embodiment, when calculating the similarity loss, the similarity score between two electrical features in the splicing feature and the binary cross entropy of the similarity label of the splicing feature are used as the similarity loss of a single splicing feature, and the current round The average of the similarity losses of all spliced features is taken as the similarity loss of the current round;

Among them, the lower the similarity score, the higher the similarity between the two electrical features in the splicing feature, the similarity score can also be understood as the normalized distance between the two electrical features, and its value ranges from 0 to 1; the similarity label of the splicing feature is used to indicate whether the true categories of the two appliances involved in the splicing feature are the same;

Specifically, the formula for calculating the similarity loss of a single splice feature is:

为拼接特征中两个电器特征之间的相似性得分；Among them, l ₁ represents the similarity loss of a single splicing feature; y is the similarity label of the splicing feature. When the true categories of the two electrical appliances involved in the splicing feature are the same, y=0, the real category of the two electrical appliances involved in the splicing feature is the same. When the categories are different, y=1;

is the similarity score between the two electrical features in the spliced feature;

In some other embodiments of the present invention, the single stitching feature and the overall similarity loss may also be calculated according to other methods such as mean square error;

In this embodiment, the category loss is the cross entropy loss, and the corresponding calculation formula is:

Among them, l ₂ represents the category loss, C is the number of categories, y _i is the ith bit of the category label vector of length C after the one-hot code encoding of the category label, and the vector of the C-bit in the result of the one-hot code encoding only has this category The position is 1, the other bits are 0, zi is the _ith bit of the vector of length C output by the classifier module, and each bit of the vector output by the classifier module represents the probability value belonging to the corresponding category;

In this embodiment, the lengths of the intercepted current signal data including electrical switching events are equal (for example, both are 7s), and in the three-dimensional complex spectrum obtained by conversion, the three dimensions are the frequency dimension, the time dimension, and the real part respectively. and the imaginary part dimension; optionally, in this embodiment, the current signal data including the electrical switching event is converted into a three-dimensional complex spectrogram through short-time Fourier transform.

In general, in this embodiment, a similarity measurement module for calculating similarity scores between electrical features belonging to different households is introduced into the model to be trained, and in the training process, the similarity loss and the category loss are calculated by As the overall loss, it can achieve unified feature expression for similar electrical samples from different families, effectively eliminate user domain differences caused by different families, and achieve user domain generalization.

In another embodiment of the present invention, a non-invasive load identification method with domain generalization capability is provided, as shown in FIG. 3 , including:

Obtain power data samples: monitor the target load to be identified in real time, and when detecting the switching event of the target load, intercept the current signal data containing the switching event from the current signal data collected at the bus;

Convert the sample to a three-dimensional complex spectrogram: convert the intercepted current signal data into a three-dimensional complex spectrogram;

Detecting the target load category to be identified: using the three-dimensional complex spectrogram as an input, using the non-intrusive load identification model construction method with domain generalization capability provided by the above embodiment to obtain the load identification model to identify the target load type.

Since the load identification model constructed in the above embodiment has the capability of domain generalization, this embodiment uses the model to perform non-intrusive load identification, which can effectively improve the identification accuracy.

In yet another embodiment of the present invention, a non-intrusive load identification system with domain generalization capability is provided, comprising: a monitoring module, a conversion module and a load identification module;

The monitoring module is used to monitor the target load that needs to be identified in real time, and when the switching event of the target load is detected, the current signal data containing the switching event is intercepted from the current signal data collected at the bus, and the conversion module is triggered. ;

The conversion module is used to convert the intercepted current signal data into a three-dimensional complex spectrogram, and trigger the load identification module;

The load identification module is configured to take the three-dimensional complex spectrogram output by the conversion module as input, and obtain the load identification model to identify the type of target load by using the non-intrusive load identification model construction method with domain generalization capability provided by the above embodiment.

Below in conjunction with a specific application example, the technical scheme of the present invention and the beneficial effects that can be obtained are further described:

Using the public household appliance data set UK-DALE (UK domestic appliance-level electricity dataset), the data set contains the electricity load data of 5 households (recorded as household 1 to household 5 in turn), and recorded 1/6Hz bus and minute data. table data. Among them, households 1, 2, and 5 also contain high-frequency bus data of 16 kHz. In order to draw the time spectrum diagram and apply the load identification model construction method and load identification method provided in the above embodiment, the current data of these three households are selected. At the same time, in order to avoid the imbalance of the aspect ratio of the time-spectrogram due to the long frequency axis, the data is downsampled to 2kHz.

In order to reflect the effect of the present invention on the network generalization ability, the data of households 1 and 5 are used as the training set, the data of household 2 is used as the test set, and the 7-second current data including the state change of the electrical appliance is intercepted as a sample. The short-time Fourier transform is processed into a complex time-spectrogram. The size of the obtained time-spectrogram is 224×100×2. The three dimensions represent frequency, time, and real/imaginary parts respectively. The time-spectrogram is used as the model input.

In the dataset UK-DALE, 7 common appliances in 3 households are selected, namely: kettle, refrigerator, dishwasher, microwave oven, washing machine, computer and treadmill.

The load identification model shown in Figure 2 is established. The similarity measurement module includes fully connected layers FC(1), FC(2) and FC(3), and the output dimensions are 1024, 64 and 1, respectively. The classifier module contains a fully connected layer FC(4) with an output dimension of 7. The activation function of FC(1) and FC(2) is ReLU, the activation function of FC(3) is sigmoid, and the activation function of FC(4) is softmax. The Adam optimizer is used in the training process, the learning rate is 0.001, the classification loss is multi-class cross entropy, the similarity loss is binary cross entropy, and the total number of iterations is 1500.

Model training is performed on the samples of family 1 and family 5, and the samples of family 2 are tested after the training is completed. The results are measured by F1-score, which is the harmonic mean of precision and recall, calculated as follows:

Among them, TP indicates the number of events for which the sample is correctly identified, FN indicates the number of events of this category but not identified as this category, and FP indicates the number of events identified as this category but not of this category.

Table 1 The recognition effect of the load recognition model on the UK-DALE dataset

The load identification results are shown in Table 1. According to the results, the above-mentioned non-intrusive load identification method with domain generalization ability can make the model have good generalization ability and perform well in cross-user scenarios, especially washing machines and Dishwashers are two multi-mode and widely different appliances.

Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, etc., All should be included within the protection scope of the present invention.

Claims (9)

1. A non-invasive load identification model construction method with domain generalization capability is characterized by comprising the following steps:

establishing a model to be trained based on a deep neural network, wherein the model to be trained comprises: the device comprises a feature extraction module, a classifier module and a similarity measurement module;

intercepting current signal data containing an electric appliance switching event from historical current signal data collected at a bus, converting the current signal data into a three-dimensional complex spectrogram, and forming an electric appliance sample by the three-dimensional complex spectrogram and a corresponding electric appliance category; forming a training set by electric appliance samples of a plurality of families;

randomly selecting h families from the training set, and randomly selecting k families from the electric appliance samples of each family as training samples of the current round; inputting the three-dimensional complex frequency spectrograms in all the training samples into the feature extraction module for feature extraction to obtain corresponding electrical appliance features; inputting the obtained electric appliance characteristics into the classifier module, converting the electric appliance characteristics into corresponding category labels, and calculating category loss by combining real electric appliance categories; inputting the obtained electrical appliance characteristics into the similarity measurement module, splicing the electrical appliance characteristics of every two families, forming a splicing characteristic by the two electrical appliance characteristics belonging to different families, calculating a similarity score between the two electrical appliance characteristics in each splicing characteristic, and calculating a similarity loss by combining real electrical appliance categories; taking the sum of the category loss and the similarity loss as the overall loss of the current round, and updating parameters of each layer of the model to be trained by using a back propagation algorithm to minimize the overall loss so as to finish the training of the current round;

performing multi-round training on the model to be trained, and after the training is finished, forming a load identification model by using the trained feature extraction module and the classifier module;

wherein, the similarity score is used for measuring the similarity between the two electric appliance characteristics; h is more than or equal to 2, and k is more than or equal to 1.

2. The method for constructing a non-invasive load recognition model with domain generalization capability according to claim 1, further comprising:

(S1) after the load recognition model is established, selecting 1 or more household electrical appliance samples to form a test set, wherein the test set and the training set are not overlapped with each other;

(S2) testing the load identification model by using the electric appliance samples in the test set to evaluate whether the identification precision of the load identification model meets the application requirement, and if not, turning to the step (S3); otherwise, go to step (S4);

(S3) regulating training parameters, then retraining the model to be trained, after training, reconstructing a load recognition model by the feature extraction module and the classifier module, and turning to the step (S2);

(S4) the test is ended.

3. The method for constructing a non-invasive load identification model with domain generalization capability according to claim 1 or 2, wherein h is 2.

4. The method for constructing the non-invasive load identification model with the domain generalization capability according to claim 1 or 2, wherein when the similarity loss is calculated, the similarity loss of a single splicing feature is determined by using the similarity score between two electrical appliance features in the splicing feature and the binary cross entropy of the similarity label of the splicing feature, and the average value of the similarity losses of all the splicing features in the current round is determined as the similarity loss in the current round;

the lower the similarity score is, the higher the similarity between the two electrical appliance characteristics in the splicing characteristics is; the similarity label of the stitching feature is used to indicate whether the genres of the two appliances involved in the stitching feature are the same.

5. The method for non-intrusive load identification with domain generalization capability of claim 4, wherein the similarity loss of a single stitching feature is:

wherein l₁Representing a loss of similarity of individual stitching features; y is a similarity label of the splicing characteristics, when the real categories of the two electric appliances related to the splicing characteristics are the same, y is 0, when the real categories of the two electric appliances related to the splicing characteristics are different, y is 1;

a similarity score between two electrical features in the stitched feature.

6. The method for constructing a non-invasive load recognition model with domain generalization capability according to claim 1 or 2, wherein the class loss is a cross-entropy loss.

7. The method for constructing a non-invasive load identification model with domain generalization capability according to claim 1 or 2, wherein the current signal data containing the appliance switching event is converted into a three-dimensional complex spectrogram by short-time Fourier transform.

8. A method for non-intrusive load identification with domain generalization capability, comprising: monitoring a target load to be identified in real time, and intercepting current signal data containing a switching event from current signal data collected from a bus when the switching event of the target load is detected;

converting the intercepted current signal data into a three-dimensional complex spectrogram, taking the three-dimensional complex spectrogram as an input, and identifying the type of the target load by using the load identification model obtained by the non-invasive load identification model construction method with the domain generalization capability of any one of claims 1 to 7.

9. A non-intrusive load identification system with domain generalization capability, comprising: the system comprises a monitoring module, a conversion module and a load identification module;

the monitoring module is used for monitoring a target load to be identified in real time, intercepting current signal data containing a switching event from current signal data collected from a bus when the switching event of the target load is detected, and triggering the conversion module;

the conversion module is used for converting the intercepted current signal data into a three-dimensional complex frequency spectrogram and triggering the load identification module;

the load identification module is used for obtaining a load identification model by taking the three-dimensional complex spectrogram output by the conversion module as input and utilizing the non-invasive load identification model construction method with the domain generalization capability of any one of claims 1 to 7 to identify the type of the target load.

Images