Non-intrusive load monitoring method applied to limited data scene

Abstract

The invention discloses a non-intrusive load monitoring method applied to a limited data scene, and relates to the field of non-intrusive load monitoring, and the method comprises the steps: dividing an original data set, constructing an initial input and output sample pair, and obtaining an initial sample set; changing the long tail distribution of the initial input and output sample pair based on the improved GAN to obtain a balanced data set, and then performing data enhancement to obtain a training sample set; reconstructing the variational auto-encoder by using a full convolutional network to obtain a reconstructed variational auto-encoder; an IBN-Net sub-network and a CBAM module are introduced into the reconstructed variational auto-encoder, and a load decomposition model is obtained; and training the load decomposition model based on the initial sample set, the balanced data set and the training sample set, and performing load decomposition by using the trained model. According to the method, the adaptability of non-intrusive load monitoring is improved, the precision and efficiency of the balance load decomposition model are improved, and a power grid company and a user can be helped to carry out adjustment according to real power demands.

Description

Technical Field

[0001] The present invention relates to the technical field of non - intrusive load monitoring, and more particularly to a non - intrusive load monitoring method applied to limited - data scenarios. Background Art

[0002] Energy consumption in buildings has accounted for 40% of the total global energy consumption. Optimizing household electricity consumption and reducing wasted electricity consumption have become important ways to save electricity. Real - time and accurate energy consumption monitoring of users' electrical loads can help power grid companies and users adjust power supply according to actual electricity demand.

[0003] Previous energy consumption monitoring methods have often been limited to the total load data of households recorded by traditional electricity meters, far from meeting the refined electricity load data requirements under accurate energy consumption monitoring. Non - intrusive load monitoring (NILM) infers the status and power consumption of appliance - level devices through the total feeder signal in a building. However, in the practical application of demand - side energy management, data - driven NILM methods usually rely on a large amount of label data that is difficult to collect and have poor adaptability in the case of data scarcity. At the same time, most of the load data containing appliance - on segments generated by existing data augmentation (DA) methods have a long - tail distribution similar to the original data, and the data distribution may be highly unbalanced; most existing load decomposition models tend to seek more complex DNN architectures, ignoring the balance between model accuracy and training efficiency.

[0004] Therefore, how to improve the adaptability of non - intrusive load monitoring and balance the accuracy and efficiency of the load decomposition model is an urgent problem for those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a non - intrusive load monitoring method applied to limited - data scenarios. The method balances data through an improved GAN, then performs data augmentation, introduces a reconstruction variational auto - encoder, and introduces IBN - Net and CBAM modules to construct a load decomposition model to achieve load decomposition.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention discloses a non - intrusive load monitoring method applied to limited - data scenarios, and the specific steps are as follows:

[0008] S1, divide the original data set by a sliding window method, construct initial input - output sample pairs, and obtain an initial sample set;

[0009] S2, change the long - tail distribution of the initial input - output sample pairs based on the improved GAN to obtain a balanced data set;

[0010] S3. Perform data augmentation on the balanced data set to obtain a training sample set;

[0011] S4. Use a fully convolutional network to reconstruct the variational autoencoder to obtain a reconstructed variational autoencoder;

[0012] S5. Introduce an IBN-Net sub-network and a CBAM module into the reconstructed variational autoencoder to obtain a load decomposition model;

[0013] S6. Train the load decomposition model based on the initial sample set, the balanced data set, and the training sample set, and use the trained load decomposition model for load decomposition.

[0014] Specifically, the non-intrusive load monitoring method of the present invention is divided into three aspects: data set division, data augmentation, and load decomposition. Step S1 corresponds to the data set division part. Steps S2 and S3 correspond to the data augmentation part, and the two combined constitute the ETEDA data augmentation scheme of the present invention. Since most electrical appliances are only in the on state for a small part of the time, the generated data samples may only have a small part containing the on segment, so the data set may be highly unbalanced. First, the improved GAN model can generate data with translation and medium distribution. Further, data augmentation techniques such as complete replication, horizontal scaling, vertical scaling, and mixed scaling are used to generate training data, and finally, the performance of the model trained based on the generated data can be improved. Steps S4 and S5 correspond to the load decomposition part, and the two combined constitute the FAEwA load decomposition model of the present invention. First, use a fully convolutional network to reconstruct the variational autoencoder to reduce the overall parameter scale of the model. Secondly, a normalization network is introduced into the encoder to enhance the feature coding extraction ability, and the decoder fuses mixed attention to reduce the loss of decoded features, improving the generalization ability of the model facing unknown data and the decomposition accuracy of multi-state devices. Step S6 realizes the combination of the data augmentation part and the load decomposition part.

[0015] Further, the specific formula for dividing the original data set by the sliding window method is:

[0016]

[0017] where S represents the data set after sliding window processing, N is the total number of samples in the processed data set, n represents the nth sample; t is the current sampling point, L is the window length, w>0 represents the additional window size of the input, S is the set step length, y t：t+L represents that the output sample window length is L, and x t-w：t+L+w represents that the input sample window length is increased by w before and after the L basis;

[0018] For electrical appliance i, the initial input-output sample pair is expressed as:

[0019]

[0020] Among them, represents the total power consumption of the nth sample, represents the power consumption of a single electrical appliance in the nth sample, and i represents the index of the ith electrical appliance.

[0021] Furthermore, GAN consists of two deep neural networks, a generator G and a discriminator D, and the overall objective is defined as:

[0022]

[0023] Among them, P data (x) represents the real data distribution, and P z (z) represents the noise data distribution, represents the expectation, x represents the real data, and z represents the random noise data; the generator G generates data samples through P z (z), and the discriminator D evaluates these samples and provides feedback to G;

[0024] In the above-defined overall objective, the logarithmic loss is replaced by the squared loss to obtain the loss function of the discriminator of the improved GAN:

[0025]

[0026] Based on the loss function of the GAN generator and introducing a new distance loss function to obtain the loss function of the generator of the improved GAN:

[0027]

[0028] Among them, α is a hyperparameter representing the proportion of the distance loss function, Specifically:

[0029]

[0030] Furthermore, the balanced data set is expressed as Among them respectively represent the total power consumption of the nth sample and the power consumption of the ith electrical appliance in the nth sample, and N is the number of samples; the data augmentation includes one of complete replication, horizontal scaling, vertical scaling, and mixed scaling.

[0031] Furthermore, the complete replication is specifically: Copy and directly to construct new training samples;

[0032] The horizontal scaling is specifically: Scale The length l is scaled by β times to obtain where β ~ N(1, σ 2 ) is a Gaussian random variable with a mean of 1 and a variance of σ 2 ; and the length of is guaranteed to be the same as that of through linear interpolation; finally, is reassigned to and together with a new training sample is constructed;

[0033] The vertical scaling is specifically as follows: The amplitude of is scaled by α times to obtain ; is reassigned to and together with a new training sample is constructed;

[0034] The mixed scaling is specifically as follows: Both horizontal scaling and vertical scaling are performed on , and a new training sample is constructed based on the enhanced and .

[0035] Furthermore, the reconstruction of the variational autoencoder using a fully convolutional network is specifically as follows: Replace the fully connected layers in the variational autoencoder with convolutional layers, so that the encoder, decoder, and latent space layer of the variational autoencoder all use only convolutional neural networks;

[0036] In the variational autoencoder, the process in which the variational parameters φ and the generative model θ participate in learning is expressed as:

[0037]

[0038] where represents the variational lower bound to be optimized, q φ (g|x) represents the encoder, p θ (g|x) represents the decoder, p θ (x) represents the true data distribution, x represents the true data, g represents the generated data, KL represents the Kullback-Leibler divergence, which is used to measure the difference between two probability distributions; is the log-likelihood function.

[0039] Further, introducing the IBN-Net sub-network and the CBAM module into the reconstruction variational auto-encoder specifically includes: introducing the IBN-Net sub-network into the encoder structure of the reconstruction variational auto-encoder; and introducing the CBAM module into the encoder structure to adaptively allocate feature attention weights and enhance the network's attention to important information. The calculation process of the attention mechanism is expressed as:

[0040]

[0041] where F is the input of the CBAM module, F c is the output of the channel attention module, F cs is the output of the CBAM module, M c (F) is the channel attention feature matrix, M s (F) is the spatial attention feature matrix, represents element-wise multiplication between matrices.

[0042] Further, training the load decomposition model based on the initial sample set, the balanced data set, and the training sample set specifically includes:

[0043] First, set the Prob i value. The Prob i value represents the probability that appliance i in the initial input-output sample pair is improved, and set the data augmentation probability where correspond to the probabilities of the four data augmentation methods of full replication, horizontal scaling, vertical scaling, and hybrid scaling respectively;

[0044] Then, select several initial input-output sample pairs from the initial sample set, and execute S2 and S3 according to the Prob i value and the data augmentation probability to obtain the training sample set;

[0045] Finally, use the obtained training sample set to train the load decomposition model.

[0046] Further, the load decomposition model is optimized using the following loss function, and the loss function is:

[0047]

[0048] where T represents the sample length, y t and represent the true power consumption and the model-predicted power consumption at time t respectively.

[0049] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a non-intrusive load monitoring method applied to limited data scenarios.

[0050] It solves the problem of poor adaptability of the NILM method in the case of scarce data. By using methods such as sliding window partitioning, improved GAN, and data augmentation, a balanced and rich training sample set is constructed, enhancing the learning ability of the model in limited data scenarios. Based on the variational autoencoder, a fully convolutional network structure is adopted, and the IBN-Net sub-network and CBAM module are introduced to improve the feature extraction ability of the load decomposition model, enhance the model's attention to important information, and improve the accuracy of load decomposition. A loss function that takes into account both accuracy and efficiency is used to optimize the load decomposition model, achieving high accuracy and high efficiency of non-intrusive load monitoring in limited data scenarios. The present invention can achieve the regression from total power to individual electrical appliance power with only one week of training data. In the case of a small amount of labeled data, it can achieve the same fitting performance as using a complete data set; it can generate samples with balanced and diverse categories. By improving GAN to change the long-tail distribution of existing samples to balance the data set, and combining diverse reinforcement methods, data augmentation is integrated into the model training process to achieve end-to-end learning; by reconstructing the variational autoencoder with a fully convolutional network, the scale of model parameters is reduced. At the same time, the IBN-Net is introduced into the encoder to enhance the feature encoding and extraction ability, and the decoder fuses convolutional attention to reduce the loss of decoded features, achieving the balance between the decomposition accuracy and training efficiency of the model. The present invention improves the adaptability of non-intrusive load monitoring and balances the accuracy and efficiency of the load decomposition model, which is beneficial to helping power grid companies and users make adjustments according to real electricity consumption demands. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to the provided drawings.

[0052] Figure 1 It is a schematic diagram of the overall process of the embodiment of the present invention.

[0053] Figure 2 It is a schematic diagram of the sample distribution generated by each data generation algorithm in the embodiment of the present invention.

[0054] Figure 3(a) is a comparison diagram of the load decomposition effects of each method for a dishwasher and a refrigerator in the embodiment of the present invention.

[0055] Figure 3(b) is a comparison diagram of the load decomposition effects of each method for a microwave oven and a kettle in the embodiment of the present invention.

[0056] Figure 3(c) is a comparison diagram of the load decomposition effects of each method of the washing machine according to the embodiment of the present invention.

[0057] Figure 4(a) is a schematic diagram of the performance comparison of the training model for dishwashers and refrigerators in the complete data set and the limited data scenario with data augmentation according to the embodiment of the present invention.

[0058] Figure 4(b) is a schematic diagram of the performance comparison of the training model for microwave ovens and washing machines in the complete data set and the limited data scenario with data augmentation according to the embodiment of the present invention. Detailed implementation manners

[0059] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0060] The embodiment of the present invention discloses a non-intrusive load monitoring method applied to a limited data scenario, as Figure 1 shown, the specific steps are as follows:

[0061] S1. Divide the original data set by a sliding window method, construct initial input-output sample pairs, and obtain an initial sample set;

[0062] S2. Based on the improved GAN, change the long-tail distribution of the initial input-output sample pairs to obtain a balanced data set;

[0063] S3. Perform data augmentation on the balanced data set to obtain a training sample set;

[0064] S4. Reconstruct the variational autoencoder using a fully convolutional network to obtain a reconstructed variational autoencoder;

[0065] S5. Introduce the IBN-Net sub-network and the CBAM module into the reconstructed variational autoencoder to obtain a load decomposition model;

[0066] S6. Train the load decomposition model based on the initial sample set, the balanced data set, and the training sample set, and use the trained load decomposition model for load decomposition.

[0067] In a specific embodiment, the specific formula for dividing the original data set by a sliding window method is:

[0068]

[0069] Among them, S represents the dataset after sliding window processing, N is the total number of samples in the processed dataset, n represents the nth sample; t is the current sampling point, L is the window length, w > 0 represents the additional window size of the input, s is the set step size, and y t：t+L represents that the output sample window length is L, and x t-w：t+L+w represents that the input sample window length increases by w before and after the basis of L.

[0070] Specifically, if the context information outside the window is not considered, only the input x with a time window length of L i is used to estimate the output y with the same time window length i , this non-linear modeling process is quite difficult and complex. Adding an additional window to the training sample input, that is, setting w > 0, means that the input length is longer than the output length, thus providing "extra" context information for the output.

[0071] For electrical appliance i, its initial input-output sample pair is expressed as:

[0072]

[0073] Among them, represents the total power consumption of the nth sample, represents the power consumption of a single electrical appliance of the nth sample, and i represents the index of the ith electrical appliance.

[0074] In a specific embodiment, GAN consists of two deep neural networks, a generator G and a discriminator D, and the overall objective is defined as:

[0075]

[0076] Among them, P data (x) represents the real data distribution, P z (z) represents the noise data distribution, represents the expectation, x represents the real data, and z represents the random noise data; the generator G generates data samples through P z (z), and the discriminator D evaluates these samples and provides feedback to G; in order to solve the potential gradient disappearance problem existing in the discriminator D during learning, in the overall objective definition, the logarithmic loss is replaced by the squared loss, and the loss function of the discriminator of the improved GAN is obtained:

[0077]

[0078] In order to facilitate the generation of a large number of samples with opening time periods and balance the sample imbalance phenomenon of the original dataset, a new distance loss function is introduced Used to evaluate the distance between the opening segment length of the generated samples and the maximum opening segment of the samples contained in the original data, and balance the data types. Generator loss function based on GAN And introduce a new distance loss function Obtain the loss function of the generator of the improved GAN:

[0079]

[0080] Among them, α is a hyperparameter representing the proportion of the distance loss function Specifically:

[0081]

[0082] In a specific embodiment, the balanced data set is represented as Where respectively represent the total power consumption of the nth sample, the power consumption of the i-th electrical appliance in the nth sample, and N is the number of samples. After the data set is balanced in step S3, diverse training samples are synthesized through various sequence data augmentation techniques. After the data set is balanced by the improved GAN, in order to diversify the training data, a time series augmentation method is used, and the data augmentation includes one of complete replication, horizontal scaling, vertical scaling, and hybrid scaling.

[0083] In a specific embodiment, the complete replication is specifically: And are directly replicated to construct new training samples;

[0084] The horizontal scaling is specifically: The length l of is scaled by β times to obtain Among them, β~N(1, σ 2 ) is a Gaussian random variable with a mean of 1 and a variance of σ 2 ; and the length of is ensured to be the same as that of through linear interpolation; finally, is the same as ; finally, is reassigned to and together with to construct new training samples, where represents the input sample The sequence data that does not contain the additional window w in is first subtracted from the original output sample and then added to the enhanced input sample Because the enhanced is the output sample (single electrical appliance power), it is necessary to ensure that the input sample (total power) contains the enhanced output sample (single electrical appliance power), so the original output is first subtracted from the corresponding position of the input and then the current output is added.

[0085] The vertical scaling is specifically: The amplitude of is scaled by α times (i.e., ) to obtain Assign to be re-assigned as and together with to construct a new training sample;

[0086] The hybrid scaling is specifically as follows: simultaneously perform horizontal scaling and vertical scaling on , and based on the enhanced and to construct a new training sample.

[0087] In a specific embodiment, the reconstruction of the variational autoencoder using a fully convolutional network is specifically as follows: replace the fully connected layers in the variational autoencoder with convolutional layers, so that the encoder, decoder, and latent space layer of the variational autoencoder all use only convolutional neural networks. The reason for adopting this network structure is that the output neurons of the convolutional layer are only connected to a small part of the input neurons, greatly reducing the aggregation time in the entire network layer, and the convolutional neural network can better retain temporal information than the fully connected layer neural network.

[0088] In the variational autoencoder, the process of the variational parameters φ and the generative model θ participating in the learning is expressed as:

[0089]

[0090] where represents the variational lower bound to be optimized, q φ (g|x) represents the encoder, p θ (g|x) represents the decoder, p θ (x) represents the true data distribution, x represents the true data, g represents the generated data, KL represents the Kullback-Leibler divergence, which is used to measure the difference between two probability distributions; is the log-likelihood function.

[0091] In a specific embodiment, an IBN-Net sub-network and a CBAM module are introduced into the reconstructed variational autoencoder, specifically as follows: an IBN-Net sub-network is introduced into the encoder structure of the reconstructed variational autoencoder; and a CBAM module is introduced into the encoder structure to adaptively allocate feature attention weights and strengthen the network's attention to important information. The calculation process of the attention mechanism is expressed as:

[0092]

[0093]

[0094] Among them, F is the input of the CBAM module, and F c is the output of the channel attention module, and F cs is the output of the CBAM module, M c (F) is the channel attention feature matrix, M s (F) is the spatial attention feature matrix, represents element-wise multiplication between matrices.

[0095] Specifically, the IBN-Net network combines instance and batch normalization. The batch normalization in the convolutional layer increases the encoder's ability to recognize deep features, thereby allowing the encoder to have more relevant feature maps mapped to the VAE latent space; while the instance normalization in the shallow layer of the network improves the generalization performance and can help the trained model effectively handle unknown data.

[0096] In a specific embodiment, the load decomposition model is trained based on the initial sample set, the balanced data set, and the training sample set. Specifically:

[0097] First, set the Prob i value. The Prob i value represents the probability that the electrical appliance i in the initial input-output sample pair obtains improvement, and set the data augmentation probability where correspond to the probabilities of four data augmentation methods: complete replication, horizontal scaling, vertical scaling, and mixed scaling respectively;

[0098] Then, select several initial input-output sample pairs from the initial sample set, and execute S2 and S3 according to the Prob i value and the data augmentation probability to obtain the training sample set;

[0099] Finally, use the obtained training sample set to train the load decomposition model.

[0100] Specifically, integrate the data augmentation scheme combining steps S2 and S3 into the load decomposition model training process. Once a batch of original samples is selected during the training process, randomly select whether to perform data augmentation and the mode of data augmentation according to the set probability to achieve immediate learning. As the training continues, new training samples will be continuously generated, which will improve the generalization ability of the deep learning model. Once the new training samples cannot contribute to performance improvement or reach the maximum training cycle, stop the training.

[0101] In a specific embodiment, the load decomposition model is optimized using the following loss function. The loss function is:

[0102]

[0103] Among them, T represents the sample length, y tand represent the true power consumption and the model predicted power consumption at time t, respectively. Specifically, y t contains y t：t+L , where t represents the sampling point. The divided y t:t+L forms y t in the order of sampling points. The reason for this representation is that the loss function of the deep learning model is judged based on all samples in a batch, while y t：t+L can only represent one sample.

[0104] In a specific embodiment, the publicly available datasets UK-DALE and REDD are selected as the sources of experimental data. In this embodiment, only five types of appliances including refrigerators, dishwashers, microwave ovens, washing machines, and electric kettles included in UK-DALE and four types of appliances including refrigerators, dishwashers, microwave ovens, and washing machines included in the REDD dataset are considered as experimental objects.

[0105] In this embodiment, we first verified the effectiveness of the proposed ETEDA scheme, mainly focusing on the effectiveness of the improved GAN algorithm in balancing the dataset. We classified the samples generated by data generation algorithms such as GAN and DDMP and the samples generated by the improved GAN algorithm according to their on-segment lengths, and the sample distribution of each algorithm is as Figure 2 shown. All data generation algorithms use 1000 samples of the UK-DALE dataset to generate 10000 samples. The length of the samples in this data is 232, and we divide this length into four intervals: [0, 58], [59, 116], [117, 174], and [175, 232]. Each sample is classified into the above intervals according to the on-segment it contains.

[0106] Comparing the effects of each algorithm, both GAN and DDMP tend to generate data similar to the distribution of the original dataset, while the improved GAN model proposed by the present invention can generate data with translation and medium distribution, which can ultimately improve the performance of the data-driven model and also prove the effectiveness of the ETEDA scheme.

[0107] In this embodiment, we then verified the effectiveness of the proposed FAEwA model. The experimental data used was the UK-DALE dataset. The proposed FAEwA model was compared with advanced models such as DAE-s2s, S2P, FCN-DAE, and VAE. The specific decomposition effects are shown in Figures 3(a), 3(b), and 3(c). As can be seen from the figures, the start of each device in the decomposition trajectories of the five devices was successfully detected by all methods, but the difference lies in the integrity of the predicted start segments. There are many irrelevant segments in the activations predicted by DAE-s2s, which do not correspond to the true values. Although S2S, FCN-DAE, and VAE can capture each activation, the predicted power levels are still not accurate enough. The prediction effect of the method of the present invention has significant advantages and is close to the true values in terms of both the integrity and accuracy of the predicted start segments.

[0108] In this embodiment, we finally verified the effectiveness of the proposed non-intrusive load monitoring method (LDSNilm framework). The experimental data used was the REDD dataset. First, the dataset was divided into a complete dataset (S1) and a limited data scenario (S2). The former used the complete dataset, and the latter only used one week of labeled data for data augmentation (DA). The training models and parameter configurations used in both were FAEwA. Through the verification of the two scenarios, it was determined whether the LSDNilm framework combining the ETEDA scheme and the FAEwA model could achieve training performance comparable to that using the complete dataset under the condition of limited labeled data. The specific decomposition effects are shown in Figures 4(a) and 4(b).

[0109] By comparing the local decomposition effects of each electrical appliance in the two scenarios, it can be concluded that the model trained with the limited data after data augmentation can achieve competitive performance compared to the model trained on the complete dataset even with only one week of data, and even has better performance than the model trained on the complete dataset for most electrical appliances. This further verifies that the proposed LDSNilm framework can achieve test performance comparable to that of training with the complete dataset in the limited data scenario.

[0110] The various embodiments in this specification are described in a progressive manner. The key point of each embodiment is the difference from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0111] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A non-intrusive load monitoring method applied to limited data scenarios, characterized in that, The specific steps are as follows: S1. Divide the original data set in a sliding window manner, construct initial input-output sample pairs, and obtain an initial sample set; S2. Based on the improved GAN, change the long-tail distribution of the initial input-output sample pairs to obtain a balanced data set; S3. Perform data augmentation on the balanced data set to obtain a training sample set; S4. Use a fully convolutional network to reconstruct the variational autoencoder to obtain a reconstructed variational autoencoder; S5. Introduce an IBN-Net sub-network and a CBAM module into the reconstructed variational autoencoder to obtain a load decomposition model; S6. Train the load decomposition model based on the initial sample set, the balanced data set, and the training sample set, and use the trained load decomposition model for load decomposition.

2. The non-intrusive load monitoring method applied to a limited data scenario according to claim 1, wherein The specific formula for dividing the original data set in a sliding window manner is: Among them, S represents the dataset after sliding window processing, N is the total number of samples in the processed dataset, n represents the nth sample; t is the current sampling point, L is the window length, w > 0 represents the additional window size of the input, s is the set step size, and y t：t+L represents that the output sample window length is L, and x t-w：t+L+w represents that the input sample window length increases by w before and after the basis of L. For electrical appliance i, the initial input-output sample pair is expressed as: Among them, represents the total power consumption of the nth sample, represents the power consumption of a single electrical appliance in the nth sample, and i represents the index of the ith electrical appliance.

3. The non-intrusive load monitoring method applied to a limited data scenario according to claim 1, characterized in that, GAN consists of two deep neural networks, a generator G and a discriminator D, and the overall objective is defined as: Among them, P data (x) represents the true data distribution, and P z (z) represents the noise data distribution, represents expectation, x represents the true data, and z represents the random noise data; the generator G generates data samples through P z (z), and the discriminator D evaluates these samples and provides feedback to G; In the definition of the overall objective, replace the logarithmic loss with the squared loss to obtain the loss function of the discriminator of the improved GAN: Generator loss function based on GAN And introduce a new distance loss function Obtain the loss function of the generator of the improved GAN: Among them, α is a hyperparameter representing the proportion of the distance loss function, Specifically:

4. A non-intrusive load monitoring method applied to a limited data scenario according to claim 1, characterized in that The balanced data set is represented as where respectively represent the total power consumption of the nth sample and the power consumption of the ith electrical appliance in the nth sample, and N is the number of samples; the data augmentation includes one of complete replication, horizontal scaling, vertical scaling, and hybrid scaling.

5. The non-intrusive load monitoring method applied to a limited data scenario according to claim 4, characterized in that, The complete replication is specifically as follows: Copy and directly to construct a new training sample; The horizontal scaling specifically is: Multiply the length l of by β to obtain where β ∼ N(1, σ 2 ) is a Gaussian random variable with a mean of 1 and a variance of σ 2 ; and ensure that the length of is the same as that of through linear interpolation; finally, reassign to and construct a new training sample together with ; The vertical scaling specifically refers to: multiplying the amplitude of by α to obtain Reassign to and together with construct a new training sample; The specific hybrid scaling is as follows: simultaneously perform horizontal scaling and vertical scaling on , and based on the enhanced and , construct new training samples.

6. The non-intrusive load monitoring method applied to a limited data scenario according to claim 1, wherein The specific method of using a fully convolutional network to reconstruct the variational autoencoder is: replace the fully connected layers in the variational autoencoder with convolutional layers, so that the encoder, decoder, and latent space layer of the variational autoencoder only use convolutional neural networks; In the variational autoencoder, the process of the variational parameters φ and the generation model θ participating in the learning is expressed as: Among them, denotes the variational lower bound to be optimized, and q φ (g|x) denotes the encoder, and p θ (g|x) denotes the decoder, and p θ (x) represents the true data distribution, x represents the true data, g represents the generated data, and KL represents the Kullback-Leibler divergence, which is used to measure the difference between two probability distributions; is the log-likelihood function (log-likelihood).

7. A non-intrusive load monitoring method applied to a limited data scenario according to claim 1, characterized in that, The method of introducing an IBN-Net sub-network and a CBAM module into the reconstructed variational autoencoder is: introduce an IBN-Net sub-network into the encoder structure of the reconstructed variational autoencoder; and introduce a CBAM module into the encoder structure to adaptively allocate feature attention weights and strengthen the network's attention to important information. The calculation process of the attention mechanism is expressed as: Among them, F is the input of the CBAM module, F c is the output of the channel attention module, F cs is the output of the CBAM module, M c (F) is the channel attention feature matrix, M s (F) is the spatial attention feature matrix, represents element-wise multiplication between matrices.

8. The non-intrusive load monitoring method applied to a limited data scenario according to claim 4, wherein The method of training the load decomposition model based on the initial sample set, the balanced data set, and the training sample set is specifically: First, set the Prob i value, where the Prob i value represents the probability that the electrical appliance i in the initial input-output sample pair obtains improvement, and set the data augmentation probability where correspond to the probabilities of the four data augmentation methods of the complete copy, the horizontal scaling, the vertical scaling, and the hybrid scaling, respectively; Then, several initial input-output sample pairs are selected from the initial sample set, and according to the Prob i value and the data augmentation probability, steps S2 and S3 are executed to obtain a training sample set; Finally, use the obtained training sample set to train the load decomposition model.

9. A non-intrusive load monitoring method applied to a limited data scenario according to claim 1, characterized in that The load decomposition model is optimized using the following loss function, and the loss function is: where T represents the sample length, and y t and represent the true power consumption and the model predicted power consumption at time t, respectively.

Images