Non-intrusive load monitoring method and device, terminal and storage medium

Abstract

The invention discloses a non-intrusive load monitoring method and device, a terminal and a storage medium. The method comprises the following steps: collecting power consumption data of a plurality of residences for training; performing time sequence imaging preprocessing on the acquired power consumption data to obtain a total power image set of each residence and a power image set of each target electric appliance corresponding to the same timestamp; respectively combining the total power image set with the power image set of each target electric appliance, and making a sample set corresponding to each target electric appliance; each constructed TransGIE model is trained respectively, and each TransGIE model is trained; running the trained TransGIE models in parallel to synthesize a GTransGIE model group; and loading a total power image corresponding to the total power time sequence of each electrical device in a to-be-monitored user residence in a predetermined time period into the GTransGIE model group for decomposition to obtain a monitoring result of individual power consumption of each electrical device in a target user residence. According to the invention, the accuracy and efficiency of load monitoring are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power load monitoring, and in particular to a non-intrusive load monitoring method, device, terminal and storage medium. Background Art

[0002] In recent years, energy demand has increased significantly, and appliances such as refrigerators, electric kettles, rice cookers, and toasters have become increasingly indispensable in households, which has led to an increase in energy consumption; therefore, effective monitoring of energy consumption has become very necessary. With the advancement of smart grid technology, household energy consumption can be monitored by using smart meters, that is, a smart meter is installed on each appliance in the home, and the smart meter is used to obtain energy consumption information in real time; but the cost of using smart meters for energy monitoring at the appliance level is quite expensive. For this reason, relevant studies have proposed installing only one meter for each household to monitor the overall energy consumption, and using relevant algorithms to infer the total energy signal into sub-signals at the device level that constitute the signal. This process is also called the energy decomposition task, and the field that studies it is called non-intrusive load monitoring (NILM). The above methods can reduce costs and increase efficiency.

[0003] Some technologies use machine learning methods such as deep neural networks (DNNs) and recurrent neural networks (RNNs) to solve the problems related to NILM. Although load data is time-related, the characteristics of different electrical appliances are complex and changeable. The models of machine learning methods are difficult to accurately capture the characteristics, and they have certain limitations in decomposing energy. Since NILM belongs to a single blind channel conversion problem, the mapping of two sets can be effectively solved by generative models. Therefore, some studies have proposed non-intrusive load monitoring methods based on generative models. Generative models can capture the complex distribution characteristics of power load signals, help achieve more accurate load decomposition, and expand the training set by generating data, especially when the actual load data is insufficient, to improve the robustness of the model; however, generative models may experience mode collapse, resulting in a lack of diversity in the generated load signals, affecting the recognition accuracy, and the black box characteristics of generative models make their decision-making process difficult to explain, which easily affects the user's trust in the results. .

[0004] Although the above-mentioned existing NILM methods can complete basic load decomposition tasks, these methods are more considering the use of sequence-to-sequence decomposition; however, the total aggregate signal in the home corresponds to multiple different electrical appliances without fixed rules, and the equipment operation characteristics are not obvious. Especially when there are too many electrical appliances in the home, there are often multiple appliances turned on at the same time. The traditional single-branch sequence-to-sequence method cannot learn the mapping rules of multiple appliances well, so that the decomposition accuracy cannot meet the application requirements. Summary of the invention

[0005] In order to solve the deficiencies in the prior art, the present invention provides a non-invasive load monitoring method, device, terminal and storage medium, which increases the difference between features by mapping a time series into an image sequence, decomposes a single electrical device in combination with a Transformerbased generative image enhance (TransGIE) structure, and combines multiple TransGIEs into a GTransGIE model group using the idea of ​​grouping. The power characteristics of multiple electrical devices can be learned at the same time, and the load decomposition task is finally completed, which effectively solves the problem that multi-state electrical loads are difficult to decompose, and further improves the accuracy of load monitoring.

[0006] The present invention adopts the following technical solution.

[0007] In a first aspect, the present invention provides a non-intrusive load monitoring method, the method comprising:

[0008] Collecting power consumption data of multiple houses for training;

[0009] Preprocess the collected power consumption data into time series images to obtain a total power image set of each residence and a power image set of each target appliance at the same time stamp;

[0010] The total power image set is combined with the power image set of each target electrical appliance to produce a sample set corresponding to each target electrical appliance;

[0011] Based on the sample sets corresponding to each target appliance, each TransGIE model constructed is trained separately;

[0012] The trained TransGIE models are run in parallel to form a GTransGIE model group;

[0013] The total power image corresponding to the total power time series of each electrical device in the user's residence to be monitored within a predetermined time period is loaded into the GTransGIE model group for decomposition to obtain the monitoring results of the individual power consumption of each electrical device in the target user's residence.

[0014] Optionally, the construction of the TransGIE model includes:

[0015] Transformer Blocks is combined with the generative adversarial network (GAN) to construct the TransGAN architecture, and the Conv convolution layer is used to replace the multi-layer perceptron MLP in the TransGAN architecture, and the DeConv deconvolution layer is used to replace the Linear unflatten linear smoothing layer in the TransGAN architecture to obtain the TransGIE model.

[0016] Optionally, preprocessing the power consumption data into time series images comprises the following steps:

[0017] Align the collected power consumption data according to the sampling time, and supplement the missing values ​​and replace the abnormal values; wherein the power consumption data includes the total power time series of all target electrical appliances and the power time series of each target electrical appliance;

[0018] Normalizing the total power time series and the power time series of each target electrical appliance to linearly change to a range of [0,1];

[0019] According to the preset sliding window and step size, the normalized total power time series and the power time series of each target electrical appliance are respectively segmented by sliding windows to obtain segmented subsequences;

[0020] Each segmented subsequence is converted into an image format of set pixels to obtain a total power image set and a power image set of each target electrical appliance.

[0021] Optionally, the expressions for performing sliding window segmentation on the normalized total power time series and the power time series of each target electrical appliance are as follows:

[0022] I(n)={X s:s+W-1 ,Y s:s+W-1 |s=0,1,……,TW-1}

[0023] Where I(n) represents the result of the nth sliding window segmentation; X represents the total power time series of all electrical appliances at the current time, and Y represents the power time series of a single target electrical appliance at the time aligned with X; X s:s+W-1 and Y s:s+W-1 They represent the subsequences from the sth time point to the s+w-1th time point extracted from sequence X and sequence Y respectively, T represents the total length of the current time series, W represents the size of the sliding window, and s represents the step size of the sliding window.

[0024] Optionally, the preparation of sample sets corresponding to each target electrical appliance includes: taking the total power image as the input sample, taking the power image of each target electrical appliance as the output sample, and pairing them according to the sampling time corresponding to each input sample and each output sample, to form a sample set of each target electrical appliance; wherein the sample set corresponding to each target electrical appliance is divided into a training set and a test set according to a set ratio.

[0025] Optionally, the structure of the TransGIE model includes a generator, a discriminator and a linear filter;

[0026] The generator includes a Conv convolution layer, a Transformerblock, an UpScale upsampling layer and a DeConv deconvolution layer; in the generator, the total power image in the total power image set is used as input data, after the local features are extracted by the Conv convolution layer, the long-term dependency of the local features is processed by the Transformerblock to obtain the global features, and then the global features are expanded by two UpScale upsamplings, and the extended feature sequence is sequentially subjected to multiple Transformerblocks for deep feature extraction to obtain deep features, and the deep features are subjected to the DeConv deconvolution layer to restore the resolution of the power image, and a power sequence diagram of a single target electrical appliance is generated;

[0027] The discriminator includes a Conv convolution layer, a Transformerblock, an average pooling layer and a linear layer; in the discriminator, the power sequence diagram output by the generator is used as input data, and after two Conv convolution layers extract local features, the Transformerblock is used to process the long-term dependency of the local features to obtain global features, and the global features are subjected to two average pooling layers to reduce the time dimension to obtain long sequence features, and the long sequence features are sequentially subjected to multiple Transformerblocks for deep feature extraction to obtain deep features, and the deep features are subjected to feature fusion through two linear layers to obtain a discrimination result; the discrimination result is used to judge whether the power image of a single target electrical appliance generated by the generator is close to the power image of the target electrical appliance in the power image set of each target electrical appliance;

[0028] The linear filter converts the single target electrical appliance power image generated by the generator into a time series, and outputs the power time series corresponding to the single target electrical appliance.

[0029] Optionally, the Transformerblock includes:

[0030] An 8-head multi-head self-attention layer to capture the dependencies between different positions in the sequence input to the multi-head self-attention layer;

[0031] Two normalization layers to accelerate convergence and improve model stability;

[0032] A two-layer feedforward neural network is used to transform and learn features after being processed by the multi-head self-attention layer.

[0033] Optionally, based on the sample sets corresponding to each target appliance, each constructed TransGIE model is trained separately, including: based on the constructed sample sets, the generator is used to extract the potential power features and information validity in the total power image, and generate a target appliance power image that can deceive the discriminator; when the discrimination result of the discriminator does not meet the preset requirements, the parameters in the generator are adjusted to make the generator gradually good at generating power images that approximate the real single target appliance; the discriminator is used to distinguish between the real single target appliance power image and the image generated by the generator; when the discrimination result of the discriminator meets the preset requirements, the training process reaches a balanced state.

[0034] Optionally, during the training of the TransGIE model, the loss function L of the generator G as follows:

[0035]

[0036] In the formula, represents the adversarial loss of the generator, where Represents the total power image set I xr Total power data extracted from , G() represents the generated result of the generator, and D() represents the output of the discriminator; represents the reconstruction loss of the generator, y ir represents the true value of the i-th power image of the target appliance; y ig Represents the power sequence graph of the generator decomposing the i-th total power image; n represents the number of samples.

[0037] Optionally, when training the TransGIE model, the loss function L of the discriminator D as follows:

[0038]

[0039] In the formula, Represents the power decomposition result of the discriminator D on the generator G output

[0040] judgment; E Iyr～pIyr Represents the target electrical appliance power image set I yr The real electrical power data p extracted from Iyr Calculate the expected value of; G() represents the generation result of the generator, and D() represents the output of the discriminator;

[0041] represents the output expectation of the discriminator for the target electrical appliance power; Represents the total power image set I xr Total power data extracted from Calculate the expected value of ; represents the gradient penalty term, where λ represents the balance coefficient, Denotes the discriminator about I yg The gradient of yg It is the electrical breakdown power generated by the generator; Represents the mixed image set I from the generated electrical power and the target electrical power r Extract the generated electrical power or target electrical power data p Ir Calculate the expected value of .

[0042] In a second aspect, the present invention provides a non-intrusive load monitoring device, which performs the steps of any method described in the first aspect of the present invention, and the device comprises:

[0043] A collection module, used for collecting power consumption data of multiple houses for training;

[0044] A preprocessing module is used to preprocess the collected power consumption data into time series images to obtain a total power image set of each residential building and a power image set of each target electrical appliance corresponding to the same timestamp;

[0045] A production module, used to combine the total power image set with the power image set of each target electrical appliance to produce a sample set corresponding to each target electrical appliance;

[0046] A training module, used for running the trained TransGIE models in parallel to synthesize a GTransGIE model group;

[0047] A synthesis module, used for running the trained TransGIE models in parallel to synthesize a GTransGIE model group;

[0048] The decomposition monitoring module is used to load the total power image corresponding to the total power time series of each electrical device in the user's residence to be monitored within a predetermined time period into the GTransGIE model group for decomposition to obtain the monitoring results of the individual power consumption of each electrical device in the target user's residence.

[0049] Optionally, the construction of the TransGIE model in the training module includes:

[0050] TransformerBlocks is combined with the generative adversarial network GAN to construct the TransGAN architecture, and the Conv convolution layer is used to replace the multi-layer perceptron MLP in the TransGAN architecture, and the DeConv deconvolution layer is used to replace the Linearunflatten linear smoothing layer in the TransGAN architecture to obtain the TransGIE model.

[0051] Optionally, the preprocessing module includes:

[0052] A supplementation and replacement unit, used to align the collected power consumption data according to the sampling time, and to supplement the missing values ​​and replace the abnormal values; wherein the power consumption data includes the total power time series of all target electrical appliances and the power time series of each target electrical appliance;

[0053] A normalization unit, used for normalizing the total power time series and the power time series of each target electrical appliance respectively, so as to linearly change them to a range of [0,1];

[0054] A segmentation unit is used to perform sliding window segmentation on the normalized total power time series and the power time series of each target electrical appliance according to a preset sliding window and step size to obtain segmented subsequences;

[0055] The image conversion unit is used to convert each segmented subsequence into an image format of set pixels to obtain a total power image set and a power image set of each target electrical appliance. 14. The non-intrusive load monitoring device according to claim 13, characterized in that in the segmentation unit, the expression for performing sliding window segmentation on the normalized total power time series and the power time series of each target electrical appliance is as follows:

[0056] I(n)={X s:s+W-1 ,Y s:s+W-1 |s=0,1,……,TW-1}

[0057] Where I(n) represents the result of the nth sliding window segmentation; X represents the total power time series of all electrical appliances at the current time, and Y represents the power time series of a single target electrical appliance at the time aligned with X; X s:s+W-1 and Y s:s+W-1 They represent the subsequences from the sth time point to the s+w-1th time point extracted from sequence X and sequence Y respectively, T represents the total length of the current time series, W represents the size of the sliding window, and s represents the step size of the sliding window.

[0058] Optionally, in the making module, making a sample set corresponding to each target electrical appliance includes:

[0059] The total power image is taken as the input sample, and the power image of each target electrical appliance is taken as the output sample, and they are paired according to the sampling time corresponding to each input sample and each output sample to form a sample set corresponding to each target electrical appliance; wherein the sample set corresponding to each target electrical appliance is divided into a training set and a test set according to a set ratio.

[0060] Optionally, the structure of the TransGIE model includes a generator, a discriminator and a linear filter;

[0061] The generator includes a Conv convolution layer, a Transformer block, an UpScale upsampling layer and a DeConv deconvolution layer; in the generator, the total power image in the total power image set is used as input data, and after the local features are extracted by the Conv convolution layer, the long-term dependency of the local features is processed by the Transformer block to obtain the global features, and then the global features are expanded by two UpScale upsamplings, and the extended feature sequence is sequentially passed through multiple Transformer blocks to extract deep features to obtain deep features, and the deep features are restored to the resolution of the power image by the DeConv deconvolution layer to generate a power sequence diagram of a single target electrical appliance;

[0062] The discriminator includes a Conv convolution layer, a Transformerblock, an average pooling layer and a DeConv deconvolution layer; in the discriminator, the power sequence diagram output by the generator is used as input data, and after two Conv convolution layers extract local features, the Transformerblock is used to process the long-term dependency of the local features to obtain global features, and the global features are subjected to two average pooling layers to reduce the time dimension to obtain long sequence features, and the long sequence features are sequentially subjected to multiple Transformerblocks for deep feature extraction to obtain deep features, and the deep features are subjected to feature fusion through two linear operations to obtain a discrimination result; the discrimination result is used to judge whether the power image of a single target electrical appliance generated by the generator is close to the power image of the target electrical appliance in the power image set of each target electrical appliance;

[0063] The linear filter converts the single target electrical appliance power image generated by the generator into a time series, and outputs the power time series corresponding to the single target electrical appliance.

[0064] Optionally, the Transformer block includes:

[0065] An 8-head multi-head self-attention layer to capture the dependencies between different positions in the sequence input to the multi-head self-attention layer;

[0066] Two normalization layers are used to normalize and converge the features after being processed by the multi-head self-attention layer;

[0067] A two-layer feedforward neural network is used to transform and learn the features after being processed by two normalization layers.

[0068] Optionally, the training module trains each constructed TransGIE model based on the sample set corresponding to each target appliance, including: based on the constructed sample set, the generator is used to extract the potential power features and information validity in the total power image, and generate a target appliance power image that can deceive the discriminator; when the discrimination result of the discriminator does not meet the preset requirements, the parameters in the generator are adjusted to make the generator gradually good at generating power images that approximate the real single target appliance; the discriminator is used to distinguish between the real single target appliance power image and the image generated by the generator; when the discrimination result of the discriminator meets the preset requirements, the training process reaches a balanced state.

[0069] Optionally, during the training of the TransGIE model, the loss function L of the generator G as follows:

[0070]

[0071] In the formula, represents the adversarial loss of the generator, where Represents the total power image set I xr Total power data extracted from , G() represents the generated result of the generator, and D() represents the output of the discriminator; represents the reconstruction loss of the generator, y ir represents the true value of the i-th power image of the target appliance; y ig Represents the power sequence graph of the generator decomposing the i-th total power image; n represents the number of samples.

[0072] Optionally, when training the TransGIE model, the loss function L of the discriminator D as follows:

[0073]

[0074] In the formula, Represents the power decomposition result of the discriminator D on the generator G output

[0075] judgment; E Iyr～pIyr Represents the target electrical appliance power image set I yr The real electrical power data p extracted from IyrCalculate the expected value of; G() represents the generation result of the generator, and D() represents the output of the discriminator;

[0076] represents the output expectation of the discriminator for the target electrical appliance power; Represents the total power image set I xr Total power data extracted from Calculate the expected value of ; represents the gradient penalty term, where λ represents the balance coefficient, Denotes the discriminator about I yg The gradient of yg It is the electrical breakdown power generated by the generator; Represents the mixed image set I from the generated electrical power and the target electrical power r Extract the generated electrical power or target electrical power data p Ir Calculate the expected value of .

[0077] In a third aspect, the present invention provides a terminal, including a processor and a storage medium;

[0078] The storage medium is used to store instructions;

[0079] The processor is used to operate according to the instructions to execute the steps of any one of the methods described in the first aspect of the present invention.

[0080] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods described in the first aspect of the present invention.

[0081] The beneficial effects of the present invention are as follows:

[0082] 1. The present invention aims to solve the problem that the power load sequence often has characteristics such as large jitter, weak feature difference, and outliers, which make it difficult for the model to accurately capture the features. The power consumption data input to the model is visualized in time series, and the time series vector is mapped into a spatial vector representation to expand the feature difference, thereby increasing the model's feature extraction capability for specific electrical appliances, and improving the accuracy of load monitoring by changing the input feature space representation method.

[0083] 2. The present invention uses a multi-layer perceptron MLP to perform preliminary feature extraction on the data in the traditional TransGAN architecture. The MLP is prone to overfitting when the model capacity is large, is not efficient enough when processing data with complex structures, cannot fully capture the dynamic changes in the data, and is prone to training instability. The present invention improves the TransGAN architecture, uses the Conv convolution layer to replace the multi-layer perceptron MLP in the TransGAN architecture, and uses the DeConv deconvolution layer to replace the Linear unflatten linear smoothing layer in the TransGAN architecture. The TransGIE model finally constructed uses the local connectivity and computational efficiency of convolution to improve the accuracy and efficiency of load monitoring.

[0084] 3. The present invention combines multiple TransGIE models into a GTransGIE model group by utilizing the idea of ​​grouping, which can simultaneously learn the power characteristics of multiple electrical equipment and ultimately complete the load decomposition task, effectively solving the problem that multi-state electrical loads are difficult to decompose; in addition, a large number of experiments have been conducted on the present invention and existing methods on public data sets, and the effectiveness of the method proposed in the present invention on multiple data sets has been proved by comparing the same electrical appliances in different households with different aggregated signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 is a flow chart of the non-intrusive load monitoring method of the present invention;

[0086] Figure 2 It is a schematic diagram of the structure of the TranGIE model in the present invention;

[0087] Figure 3 It is a schematic diagram of the overall structure of the GTransGIE model group in the present invention;

[0088] Figure 4 It is a schematic diagram of the sliding window operation in the present invention;

[0089] Figure 5 It is the decomposition result of the refrigerator measured by the REFIT dataset by the GTransGIE model group in the present invention;

[0090] Figure 6 It is the decomposition result of the kettle measured by the REFIT dataset by the GTransGIE model group in the present invention;

[0091] Figure 7 It is the decomposition result of the microwave oven measured by the REFIT dataset by the GTransGIE model group in the present invention;

[0092] Figure 8It is the power decomposition result of the washing machine measured by the GTransGIE model group of the present invention on the REFIT data set;

[0093] Fig. 9 It is the power decomposition result of the dishwasher measured by the GTransGIE model group of the present invention on the REFIT dataset;

[0094] Fig.10 It is a schematic diagram comparing the training time of the GTransGIE model and the MSA-Resnet model on the REFIT dataset and the UKDALE dataset in the present invention;

[0095] Fig.11 It is a structural principle block diagram of the non-intrusive load monitoring device in the present invention. DETAILED DESCRIPTION

[0096] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in the present invention are only part of the embodiments of the present invention, not all of the embodiments. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0097] Embodiment 1:

[0098] Reference Figure 1 The embodiment of the present invention provides a non-intrusive load monitoring method, which specifically includes the following steps:

[0099] Step S1, collecting power consumption data of multiple houses for training;

[0100] Step S2: pre-processing the collected power consumption data into time series images to obtain a total power image set of each house and a power image set of each target electrical appliance at the same time stamp;

[0101] Step S3, combining the total power image set with the power image set of each target electrical appliance to produce a sample set corresponding to each target electrical appliance;

[0102] Step S4: Based on the sample sets corresponding to the target electrical appliances, each constructed TransGIE model is trained respectively;

[0103] Step S5, running the trained TransGIE models in parallel to synthesize the GTransGIE model group;

[0104] Step S6: Load the total power image corresponding to the total power time series of each electrical device in the user's residence to be monitored within a predetermined time period into the GTransGIE model group for decomposition to obtain the monitoring result of the individual power consumption of each electrical device in the target user's residence.

[0105] It should be noted that the problem solved by NILM is to decompose the operating power of a single target device from the aggregate signal of electrical appliances; the composition of the aggregate signal at a certain moment is the sum of the power and operating noise of all devices in the home at the current moment, and the aggregate signal at a certain time is the accumulation of the aggregate signals of a given time period; assuming that there are N devices to be decomposed, then the aggregate signal is X = (x1, x2, .., x T ), represents the total power at T time points, and X(t) represents the total power at time point t. For the i-th electrical equipment, the power consumption at time point T is expressed as Y = (y i1 ,y i2 , ..., y iT ); the relationship between X and Y is:

[0106]

[0107] Wherein, ∈(t) represents the electrical appliances and noise that are not considered; the goal of this embodiment is to decompose X(t) at each moment into Y i (t).

[0108] As an embodiment of the present invention, the preprocessing of the power consumption data into time series images in step S2 includes:

[0109] S2.1. Align the collected power consumption data according to the sampling time, and supplement the missing values ​​and replace the abnormal values; wherein the power consumption data includes the total power time series of all target electrical appliances and the power time series of each target electrical appliance;

[0110] S2.2, normalizing the total power time series and the power time series of each target electrical appliance, and linearly changing all data to the range of [0, 1];

[0111] Specifically, in order to train the deep learning model, the data must be standardized, otherwise it will affect the efficiency of the gradient descent algorithm; read the training set and the test set respectively, calculate the total aggregate signal in the training set and the test set and the maximum power X when the target appliance is turned on max , Y max and minimum power when closed X min , Y min ; According to the maximum power and minimum power normalized to [0, 1], the conversion formula is:

[0112]

[0113]

[0114] Among them, x t represents the original total aggregate signal at time t, y t Represents the original power of the target appliance at time t.

[0115] S2.3, according to the preset sliding window and step size, the normalized total power time series and the power time series of each target electrical appliance are respectively segmented by sliding windows to obtain segmented subsequences;

[0116] S2.4. Convert each segmented subsequence into an image format of set pixels to obtain a total power image set and a power image set of each target electrical appliance.

[0117] Furthermore, in step S3, making a sample set corresponding to each target electrical appliance includes: taking the total power image as an input sample, taking the power image of each target electrical appliance as a respective output sample, and pairing them according to the sampling time corresponding to each input sample and each output sample, to form a sample set corresponding to each target electrical appliance; wherein the sample set corresponding to each target electrical appliance is divided into a training set and a test set according to a set ratio.

[0118] It should be noted that the process of data preprocessing is to vectorize the power consumption data, which can provide preparation for the subsequent image conversion steps; first, the total consumption generated by each power data point of the user and the consumption of a single appliance are converted into multiple vectors of length 64, and these vectors will serve as a prerequisite for converting image data. Assuming that the length of each vector is 64, the sampling frequency of energy load data is about 8 seconds, so a vector of length 64 is equivalent to the energy load data generated in 8 minutes. When selecting data, the time span cannot be too large, otherwise the number of records of appliances that are only in operation for a short time will be reduced within the time span, thereby reducing the prediction accuracy. If the time span is too small, it will increase the complexity of training. When the time series is visualized, the selected sequence is converted into image data of 8×8 pixels in a folding manner using a sliding window, and it is used as the input of an image.

[0119] An example of a sliding window is Figure 4As shown, assuming that the step size is 1 or 4 and the sliding window size is 4, if the step size is smaller than the length of the sliding window, not only the number of samples can be increased, but also two adjacent sets of data can retain stronger temporal correlation; if the step size is equal to the length of the sliding window, the speed of data prediction processing can be increased. In this embodiment, the step size is the same as the sliding window size. For example, 6400 1×n vectors can be converted into 100 8×8 images; the entire data will be cut into multiple 8×8 image data to complete the preprocessing of the overall data.

[0120] Since the aggregate signal and the power signal of a single device need to be converted into an image representation, it is assumed that the window size is W, that is, the corresponding sequence for each time series image is Y s：s+W-1 and X s：s+W-1 , s represents the starting position of the corresponding sequence when converting the image; in this implementation, the window size W = 64; then the expression of the sliding window segmentation in step S2.3 is as follows:

[0121] I(n)={X s：s+W-1 , Y s：s+W-1 |s=0,1,…,TW-1}

[0122] Where I(n) represents the result of the nth sliding window segmentation; X represents the total power time series of all electrical appliances at the current time, and Y represents the power time series of a single target electrical appliance at the time aligned with X; X s：s+W-1 and Y s：s+W-1 They represent the subsequences from the sth time point to the s+w-1th time point extracted from sequence X and sequence Y respectively, T represents the total length of the current time series, W represents the size of the sliding window, and s represents the step size of the sliding window.

[0123] As an embodiment of the present invention, the construction of the TransGIE model in step S4 includes: combining TransformerBlocks with a generative adversarial network GAN to construct a TransGAN architecture, replacing the multi-layer perceptron MLP in the TransGAN architecture with a Conv convolutional layer, and replacing the Linear unflatten linear smoothing layer in the TransGAN architecture with a DeConv deconvolutional layer to obtain a TransGIE model.

[0124] It is further explained that the Generative Adversarial Network (GAN) is one of the most attractive ideas in the field of computer science today; in the adversarial process, two models are trained simultaneously: the generator learns to generate realistic images, while the discriminator learns to distinguish between real images and forged images. In addition, unlike traditional neural networks, Transformer uses a self-attention mechanism to process input sequences and output sequences, and extracts the global dependencies between the set and the household appliance signals by relying entirely on the attention mechanism, accurately estimates the power signals of household appliances, and can be calculated in parallel, greatly improving the computational efficiency. The present invention combines Transformer with GAN to construct the TransGAN architecture, aiming to generate high-quality data through the powerful sequence modeling ability of Transformer and the generative adversarial mechanism of GAN; it uses the self-attention mechanism of Transformer to capture long-distance dependencies in the data, and continuously optimizes the output of the generator through adversarial training, so that it can generate samples close to real data. However, the linear operations in the traditional TransGAN architecture cannot fully capture the dynamic changes and complex structures in the data. Therefore, in this embodiment, the difference between features is increased by mapping the time series into an image sequence, the multi-layer perceptron in the TransGAN architecture is replaced by the convolution operation, and the Linear unflatten linear smoothing layer in the TransGAN architecture is replaced by the deconvolution operation, and they are constructed as a model group for application in NILM. The local connectivity and computational efficiency of the convolution are used to improve the accuracy and efficiency of load decomposition, and the operating power of different electrical appliances is monitored simultaneously by grouping, thereby solving the problem of the model decomposing the loads of multiple target electrical appliances at the same time.

[0125] In addition, since the task of the original TransGAN is to generate an image that conforms to a category, its input is random noise; and for the NILM task, the generator needs to complete the mapping of total power to the operating power of a single electrical appliance. Therefore, the present invention changes the input of the generator into an image format; Transformer Blocks are used inside the model, and upsampling and downsampling are combined to convert the output result into an image of the same scale as the input, which is a mapping of the total power image; finally, the image output by the generator is linearly converted into a single electrical appliance power sequence through a filter; the discriminator receives the output of the generator, combines the Transformer Blocks with the linear layer, judges the difference between the input image and the image of the actual single electrical appliance to be decomposed, and thereby optimizes the generator and discriminator of TransGAN.

[0126] Traditional Transformers use deterministic position encoding, while Transformer Blocks use relative position encoding. Compared with deterministic position encoding, the present invention uses relative position encoding to enable the model to learn stronger local position relationships. The formula is as follows:

[0127]

[0128] In the formula, They are represented as query matrix, key matrix and value matrix respectively; H, W, C represent the height, width and embedding dimension of the input feature map; the coordinate difference between each query and the key on the H axis is in the range of -(H-1),H-1, and the coordinate difference on the W axis is similar. By considering both the H and W axes, the relative position can be parameterized by the matrix Indicated.

[0129] In a preferred but non-limiting embodiment, the Transformerblock includes: a multi-head self-attention layer with 8 heads, used to capture the dependencies between different positions in the sequence input to the multi-head self-attention layer; two normalization layers, used to accelerate convergence and improve the stability of the model; and two layers of feedforward neural networks, used to transform and learn the features processed by the multi-head self-attention layer.

[0130] Specifically, the structure of the TransGIE model is as follows Figure 2 As shown, it includes a generator (Generato), a discriminator (Discriminator) and a linear filter (Filter); the generator includes a Conv convolution layer, a Transformer block, an UpScale upsampling layer and a DeConv deconvolution layer; in the generator, the total power image in the total power image set is used as input data, and after the local features are extracted by the Conv convolution layer, the long-term dependency of the local features is processed by the Transformerblock to obtain the global features, and then the global features are expanded by two UpScale upsamplings, and the extended feature sequence is sequentially subjected to multiple Transformerblocks for deep feature extraction to obtain deep features, and the deep features are subjected to the DeConv deconvolution layer to restore the resolution of the power image, and generate a power sequence diagram of a single target electrical appliance;

[0131] The discriminator includes a Conv convolution layer, a Transformerblock, an average pooling layer (AvgPool) and a linear layer (Linea); in the discriminator, the power sequence diagram output by the generator is used as input data, and after two Conv convolution layers extract local features, the Transformerblock is used to process the long-term dependency of the local features to obtain global features. The global features are subjected to two average pooling layers to reduce the time dimension to obtain long sequence features. The long sequence features are sequentially subjected to multiple Transformerblocks for deep feature extraction to obtain deep features. The deep features are subjected to feature fusion through two linear layers to obtain a discrimination result (Validity); the discrimination result is used to judge whether the single target electrical appliance power image generated by the generator is close to the power image of the target electrical appliance in the power image set of each target electrical appliance;

[0132] The linear filter converts the single target electrical appliance power image generated by the generator into a time series, and outputs the power time series corresponding to the single target electrical appliance.

[0133] The linear filter Filter in this embodiment is expressed as follows:

[0134] P i =W f Flatten(I)+b

[0135] Where W f represents the weight matrix of the linear transformation, Flatten(I) represents flattening the image into a one-dimensional vector, b represents the bias term, P i Represents the i-th power time series of the target appliance after linear conversion. After passing through the linear filter, the power image generated by the generator is converted into a one-dimensional power time series; this time series can directly represent the power consumption of a single target appliance in the entire time period.

[0136] The GTransGIE model group constructed in this invention is a model group that can decompose multiple electrical appliances at the same time. Before forming the overall model group, it is necessary to train a single TransGIN separately, and each branch of TransGIN accurately decomposes the power of an electrical appliance. During the training process of the TransGIE model, based on the constructed sample set, the generator extracts the potential power features and information validity in the total power image, and adjusts the network parameters according to the discrimination results fed back by the discriminator to gradually become good at generating power images that approximate the real single target electrical appliance; the discriminator distinguishes between the real single target electrical appliance power image and the generator's generated image. When the discriminator can no longer accurately distinguish between the real image and the generated image, the training process reaches a balanced state. The output of the generator is now the image of the single target electrical appliance that is closest to the real one, and the image is converted into a time series through the linear filter Filter module. In this embodiment, a model group consisting of N TransGAN models is used, where N represents the total number of known electrical appliances in the home. The individual features of the N devices are captured separately through the shared total power features of the devices, so as to achieve the purpose of simultaneous decomposition of multiple devices; the GTransGIE model group is as follows Figure 3 shown.

[0137] Furthermore, in order to enable each branch to accurately capture the characteristics of the target appliance, it is necessary to enable each model to separate and identify the potential power characteristics it wants to identify from the shared device power; therefore, the present invention designs the loss functions of the generator and the discriminator based on the characteristics of the model group.

[0138] The loss function L of the generator G as follows:

[0139]

[0140] In the formula, Represents the adversarial loss of the generator (represents the power decomposition result of the discriminator D on the output of the generator G judgment), among which, Represents the total power image set I xr Total power data extracted from , G() represents the generated result of the generator, and D() represents the output of the discriminator; represents the reconstruction loss of the generator (representing the root mean square error between the electrical appliance decomposition power generated by the generator and the real electrical appliance power, which is used to ensure that the power decomposition result generated by the generator is close to the real data), y ir represents the true value of the i-th power image of the target appliance; y ig Represents the power sequence graph of the generator decomposing the i-th total power image; n represents the number of samples.

[0141] The loss function L of the discriminator D as follows:

[0142]

[0143] In the formula, Represents the power decomposition result of the discriminator D on the generator G output

[0144] judgment; E Iyr～pIyr Represents the target electrical appliance power image set I yr The real electrical power data p extracted from Iyr Calculate the expected value of; G() represents the generation result of the generator, and D() represents the output of the discriminator;

[0145] represents the output expectation of the discriminator for the target electrical appliance power; Represents the total power image set I xr Total power data extracted from Calculate the expected value of ; represents the gradient penalty term, which is used to constrain the gradient of the discriminator to make it smoother and more stable, so as to increase the judgment ability of the discriminator; where λ represents the balance coefficient, Denotes the discriminator about I yg The gradient of yg is the electrical decomposition power generated by the generator; E is the expected value; Represents the mixed image set I from the generated electrical power and the target electrical power r Extract the generated electrical power or target electrical power data p Ir Calculate the expected value of .

[0146] The GTransGIN model group proposed in the present invention is evaluated through experiments. The power consumption of each electrical appliance is detected through experiments, and the method proposed in the present invention is compared with several existing methods that have adopted different technologies in the past. The REFIT dataset and UKDALE dataset are used in the experiment, which are both open public datasets for load monitoring, so that the experimental results have a certain degree of authenticity.

[0147] The REFIT dataset records the electricity consumption information of 20 households in the UK. Each household contains 9 electrical appliances and corresponding sampling times. The total active power consumption and the consumption of each electrical appliance are sampled once every 8 seconds. There are some missing data and abnormal data in the REFIT dataset. Therefore, the missing data are interpolated during data preprocessing and the abnormal data are deleted. This embodiment selects households No. 2, No. 3, No. 6, and No. 16 from the REFIT dataset to train the model constructed by the present invention, and uses the data of households No. 1 and No. 5 for testing. The selected electrical appliances include refrigerators, microwave ovens, washing machines, kettles and other electrical appliances. Table 1 shows the data characteristics of the REFIT dataset. The UKDALE dataset records the total electricity consumption and device-level electrical appliance consumption of five households in the UK from November 2012 to January 2015, and is sampled every 6 seconds; the five households in the dataset contain 54, 20, 5, 6 and 26 electrical appliances respectively; this embodiment focuses on the decomposition and state recognition of refrigerators, washing machines, kettles and microwave ovens; the training data of the model selects households 1, 3, 4 and 5, and the data of household 2 is tested; Table 2 shows the data characteristics of the UKDALE dataset.

[0148] Table 1. Data characteristics of the REFIT dataset

[0149] Target Electric Maximum Threshold(W) Minimum threshold (W) Minimum opening time (s) Minimum closing time(s) refrigerator 1700 50 70 14 kettle 3000 1000 120 0 Micro-wave oven 1200 300 12 24 washing machine 2500 50 70 182 dishwasher 2400 1000 900 900

[0150] Table 2. Data characteristics of the UKDALE dataset

[0151] Target Electric Maximum Threshold(W) Minimum threshold (W) Minimum opening time (s) Minimum closing time(s) refrigerator 300 50 60 12 kettle 3100 2000 12 0 Micro-wave oven 3000 200 12 30 washing machine 2500 50 1800 160 dishwasher 2500 1000 1800 1800

[0152] In order to compare the model with existing similar methods, this experiment uses three widely used indicators to evaluate the decomposition ability of the model, namely Mean Absolute Error (MAE), Signal Aggregate Error (SAE) and F1 score.

[0153] First, measure the error between the decomposed power and the actual power of the target electrical equipment using MAE and SAE indicators. i and Indicates the real energy of the appliance to be decomposed and the decomposed energy, r and Indicates the total consumption of the target device and breaks down the total consumption of the device:

[0154]

[0155]

[0156] After the power decomposition task is completed, it is necessary to distinguish the switch state of the device; electrical appliances have an on state and an off state. It is often not very accurate to judge the operating state of the device based solely on the data threshold, because the deep learning model has a certain error rate. Therefore, in the experiment, the F1 score classification indicator is used to judge the switch of electrical equipment; specifically, the F1score indicator includes precision and recall indicators, including true positive (TP), false positive (FP), true negative (TN) and false negative (FN). Among them, TP means that the actual state of the device is on and the predicted state is also on. FP means that the predicted state of the device is on, but the actual state is off. TN means that the predicted and actual states of the device are both off. FN means that the actual state of the device is off and the predicted state is on.

[0157] Recall refers to the ratio of the number of samples correctly predicted as positive examples by the classifier to the number of samples of all actual positive examples. Recall indicates the number of true positive examples that are successfully predicted:

[0158]

[0159] The Precision metric indicates the proportion of devices that are actually in the on state when predicted to be in the on state:

[0160]

[0161] Accuracy refers to the ratio of the number of samples for which the device predicts the state correctly to the total number of predictions:

[0162]

[0163] The F1 score metric indicates the model's ability to adequately predict whether an appliance is turned on or off in imbalanced data. This metric maximizes the difference between the correctness of the model's prediction and the likelihood of a false positive:

[0164]

[0165] In this embodiment, the method proposed in the present invention is compared with several typical methods proposed by predecessors. Specifically, S2S is a sequence-to-sequence processing method, which uses a sequence as the input of the model and a sequence of the same length as the output of the model. In the power decomposition task, the input of S2S is a total power sequence, which is subjected to the nonlinear mapping of the model, and finally decomposes the device-level power of the target electrical appliance according to the total power. GRU is an implementation of a recurrent neural network. Like S2S, this model also mainly operates on sequences. The AED model uses the idea of ​​adversarial power to decompose energy by combining a generator and a predictor. MSA-Resnet is an improved Resnet model with a multi-scale attention mechanism. The original task of this model is also to map the total power to the operating power of a single target electrical appliance. Table 3 is the experimental results of the GTransGIE model group proposed in the present invention for the decomposition of a single electrical appliance in the REFIT dataset; the first column in the table is the device predicted in the REFIT dataset, the second column is the method compared with the model proposed in the present invention, and the third to fifth columns are the comparison indicators.

[0166] Table 3 Decomposition and status results of the test households in the REFIT dataset

[0167]

[0168] As can be seen from Table 3, for the experimental electrical appliances: refrigerator, toaster, washing machine, dishwasher, the MAE index of the present invention is lower than that of other compared methods. Although the MAE index of the microwave oven is not as good as that of MSA-Resnet, it is very close to it. In terms of average accuracy, the GTransGIE model group has a lower MAE, which can be explained by the better performance of the model in the decomposition task of load monitoring. In terms of SAE index, although the AED model and MSA-Resnet model exceed the present invention method in the two devices of microwave oven and dishwasher, this is due to the greater power fluctuation of these two appliances when they are in different gears; but there are still three-fifths of the accuracy that exceeds the comparison method, and the present invention method reduces the overall SAE index by 13%. When the experiment measures the F1 score index, the operating status of the device is measured in the real label according to the threshold of the on / off of each electrical device. As can be seen from Table 3, the model proposed by the present invention shows a relatively excellent test result in the F1 score index. Compared with other methods, this index has increased by about 8%, and the F1 index for appliances with more open time is higher.

[0169] Table 4 Decomposition and status results of the test households in the UKDALE dataset

[0170]

[0171] Table 4 is the experimental results of the GTransGIE model proposed in the present invention in the UKDALE data set. It can be seen that the MAE index of refrigerators in all models is relatively high. This is because refrigerators are the appliances with the longest opening time among the five tested devices. Unlike appliances that are normally off, the operating power of refrigerators is always above 0W, with an average maximum value of about 100W. The decomposed data has a large overall gap compared with the real power. Toasters, washing machines, dishwashers and microwave ovens are appliances that are normally off, so their MAE index is lower than that of refrigerators. In the MAE index of the GTransGIE model, only the dishwasher is inferior to the MSA-Resnet model; but overall, the MAE index is reduced by 21% on average. It is worth noting that only one SAE index is higher than the comparison model, and the gap is only 0.3%, indicating that the GTransGIE model proposed in the present invention not only performs better in decomposing normally open appliances, but also has outstanding advantages in decomposing normally closed electrical equipment.

[0172] In addition, the GTransGIE model showed better performance in the F1 score metric; in particular, the F1 score of the refrigerator was 15% higher than the compared model. This is because at a certain time, appliances that are usually turned on have a higher accuracy in judging their operating state, while appliances that are often switched between on and off have a low probability value of being considered to be in operation; this is also reflected in specific electrical devices. The refrigerator, as an electrical device that is usually turned on, is more likely to be correctly identified as being in operation, and the model's prediction is consistent with the true on / off label.

[0173] In order to verify the decomposition performance of the GTransGIE model group when multiple appliances are turned on at the same time, the experiment combined the appliances in the test set in multiple ways; since refrigerators are always on appliances, they are included in the test cases; the combination of appliances also takes into account the power difference between them, so as to show the accuracy of simultaneous decomposition. Table 5 shows the decomposition of multiple appliance combinations by the GTransGIE model group in the REFIT and UKDALE datasets. Since the actual power of microwave ovens and toasters in the two datasets is quite different, the accuracy of each indicator decreases when these two appliances are turned on. At the same time, the refrigerator is almost always on throughout the test time window, which makes the overall indicator in a stable state. Although we generally believe that the more appliances are measured simultaneously, the lower the accuracy will be, it can be seen from Table 5 that this is not the case. This often depends on the user's actual habits and methods of using appliances.

[0174] Table 5 Decomposition and status results of multiple appliances turned on at the same time

[0175]

[0176] By summarizing the experiments of the model on the REFIT dataset and the UKDALE dataset, it is shown that the GTransGIE model group constructed by the present invention has a better decomposition effect in mapping the aggregate signal to the target electrical equipment, with an average improvement of 11% and 7% in MAE and SAE indicators, and an improvement of 6% and 4% over the second best model MSA-Resnet model, thus confirming the superiority of our method in energy decomposition. In terms of F1score, the experimental results show that the F1 score is 11% higher than the reference method. The GTransGIE model mainly improves the accuracy index by reducing the number of false alarms. Although a larger window was not set in the experiment, it does not mean that the model provided by the present invention does not support a larger window.

[0177] Figure 5 , Figure 6 , Figure 7 , Figure 8 and Fig. 9 The decomposition results of each electrical appliance measured by the GTransGIE model proposed in the present invention on the REFIT data set are shown; the data of five electrical devices are decomposed in the figure, each row represents an electrical device, the first column is the total aggregate power and the real target electrical power, and the second column is the real target electrical power and the power decomposed by the model. As can be seen from the figure, the GTransGIE model proposed in the present invention performs well in the decomposition of normally-on electrical appliances, and can accurately identify the operating power characteristics of the target device from the aggregated signal. In some cases where the interference is relatively large (such as toasters and washing machines), the GTransGIE model can still eliminate other interference powers and correctly separate the target electrical appliances. This is because the approximate running time of each electrical appliance needs to be manually checked, and a reasonable window size and step size are set based on these running times, so that the converted image contains as much complete power information as possible when it is turned on, and the other image contains as much complete power information as possible when it is turned off. Therefore, the feature differences between the two different images are relatively obvious, making it easier for the GTransGIE model to distinguish the power difference when it is turned on or off, thereby improving the decomposition accuracy.

[0178] Fig.10 The figure shows the comparison of the training time of the GTransGIE model proposed in the present invention and the MSA-Resnet model with the second best test results in the REFIT dataset and the UKDALE dataset. We selected five electrical appliances in each dataset. The blue bar graph in the figure represents the training time of our proposed model, and the orange bar graph represents the training time of the MSA-Resnet model. On average, the time taken by the GTransGIE model to train each electrical appliance is almost 60% of that of the MSA-Resnet model. Therefore, the GTransGIE model also has an advantage in training time.

[0179] The non-invasive load monitoring method based on the GTransGIE model proposed in the embodiment of the present invention introduces the idea of ​​computer vision into solving the problem of energy decomposition. Using TransGIE to train and predict power data is an extension of its model function. In order to better learn the feature connection between sequences and images and capture the complex multi-modal structure of various devices, a new method of processing time series by sequence imaging is used, and a model group method is used to decompose multiple appliances turned on at the same time. A large number of experiments conducted on the REFIT dataset and the UKDALE dataset, and compared with existing methods, verified that GTransGAN can achieve better decomposition results. The MAE index of the GTransGIE model in the two datasets was reduced by an average of 11%, the SAE index was reduced by an average of 7%, and the F1 score index was improved by an average of 11%. Simulation experiments have proved the superiority of the method of the present invention.

[0180] The beneficial effects of the present invention are as follows:

[0181] 1. The present invention aims to solve the problem that the power load sequence often has large jitter, weak feature difference, outliers and other characteristics, which makes it difficult for the model to accurately capture the features. The power consumption data input to the model is visualized in time series, and the time series vector is mapped into a spatial vector representation, which expands the feature difference to increase the model's feature extraction capability for specific electrical appliances. By changing the input feature space representation method, the accuracy of load monitoring is improved.

[0182] 2. The present invention uses a multi-layer perceptron MLP to perform preliminary feature extraction on the data in the traditional TransGAN architecture. The MLP is prone to overfitting when the model capacity is large, is not efficient enough when processing data with complex structures, cannot fully capture the dynamic changes in the data, and is prone to training instability. The present invention improves the TransGAN architecture, uses the Conv convolution layer to replace the multi-layer perceptron MLP in the TransGAN architecture, and uses the DeConv deconvolution layer to replace the Linear unflatten linear smoothing layer in the TransGAN architecture. The TransGIE model finally constructed uses the local connectivity and computational efficiency of convolution to improve the accuracy and efficiency of load monitoring.

[0183] 3. The present invention combines multiple TransGIE models into a GTransGIE model group by utilizing the idea of ​​grouping, which can simultaneously learn the power characteristics of multiple electrical equipment and ultimately complete the load decomposition task, effectively solving the problem that multi-state electrical loads are difficult to decompose; in addition, a large number of experiments have been conducted on the present invention and existing methods on public data sets, and the effectiveness of the method proposed in the present invention on multiple data sets has been proved by comparing the same electrical appliances in different households with different aggregated signals.

[0184] Embodiment 2:

[0185] like Fig.11 As shown, the present invention provides a non-intrusive load monitoring device, which is used to implement the steps of the method in the above embodiment 1, and the device specifically includes:

[0186] A collection module, used for collecting power consumption data of multiple houses for training;

[0187] A preprocessing module is used to preprocess the collected power consumption data into time series images to obtain a total power image set of each residential building and a power image set of each target electrical appliance corresponding to the same timestamp;

[0188] A production module, used to combine the total power image set with the power image set of each target electrical appliance to produce a sample set corresponding to each target electrical appliance;

[0189] A training module, used to train each TransGIE model constructed based on a sample set corresponding to each target appliance;

[0190] A synthesis module, used for running the trained TransGIE models in parallel to synthesize a GTransGIE model group;

[0191] The decomposition monitoring module is used to load the total power image corresponding to the total power time series of each electrical device in the user's residence to be monitored within a predetermined time period into the GTransGIE model group for decomposition to obtain the monitoring results of the individual power consumption of each electrical device in the target user's residence.

[0192] Furthermore, the construction of the TransGIE model in the training module includes:

[0193] TransformerBlocks is combined with the generative adversarial network GAN to construct the TransGAN architecture, and the Conv convolution layer is used to replace the multi-layer perceptron MLP in the TransGAN architecture, and the DeConv deconvolution layer is used to replace the Linearunflatten linear smoothing layer in the TransGAN architecture to obtain the TransGIE model.

[0194] In this embodiment, the preprocessing module includes:

[0195] A supplementation and replacement unit, used to align the collected power consumption data according to the sampling time, and to supplement the missing values ​​and replace the abnormal values; wherein the power consumption data includes the total power time series of all target electrical appliances and the power time series of each target electrical appliance;

[0196] A normalization unit, used for normalizing the total power time series and the power time series of each target electrical appliance respectively, so as to linearly change them to a range of [0,1];

[0197] A segmentation unit is used to perform sliding window segmentation on the normalized total power time series and the power time series of each target electrical appliance according to a preset sliding window and step size to obtain segmented subsequences;

[0198] The image conversion unit is used to convert each segmented subsequence into an image format of set pixels to obtain a total power image set and a power image set of each target electrical appliance.

[0199] In the segmentation unit of this embodiment, the normalized total power time series and the power time series of each target electrical appliance are respectively segmented by sliding windows, and the expressions are as follows:

[0200] I(n)={X s:s+W-1 ,Y s:s+w-1 |s=0,1,……,TW-1}

[0201] Where I(n) represents the result of the nth sliding window segmentation; X represents the total power time series of all electrical appliances at the current time, and Y represents the power time series of a single target electrical appliance at the time aligned with X; X s:s+W-1 and Y s:s+W-1 They represent the subsequences from the sth time point to the s+w-1th time point extracted from sequence X and sequence Y respectively, T represents the total length of the current time series, W represents the size of the sliding window, and s represents the step size of the sliding window.

[0202] Furthermore, the sample set corresponding to each target electrical appliance is prepared in the preparation module, including:

[0203] The total power image is taken as the input sample, and the power image of each target electrical appliance is taken as the output sample, and they are paired according to the sampling time corresponding to each input sample and each output sample to form a sample set corresponding to each target electrical appliance; wherein the sample set corresponding to each target electrical appliance is divided into a training set and a test set according to a set ratio.

[0204] Specifically, the structure of the TransGIE model includes a generator, a discriminator, and a linear filter;

[0205] The discriminator includes a Conv convolution layer, a Transformer block, an average pooling layer and a linear layer; in the discriminator, the power sequence diagram output by the generator is used as input data, and after two Conv convolution layers extract local features, the Transformer block is used to process the long-term dependency of the local features to obtain global features, and the global features are subjected to two average pooling layers to reduce the time dimension to obtain long sequence features, and the long sequence features are sequentially subjected to multiple Transformer blocks for deep feature extraction to obtain deep features, and the deep features are subjected to feature fusion through two linear layers to obtain a discrimination result; the discrimination result is used to judge whether the power image of a single target electrical appliance generated by the generator is close to the power image of the target electrical appliance in the power image set of each target electrical appliance;

[0206] The discriminator includes a Conv convolution layer, a Transformer block, an average pooling layer and a DeConv deconvolution layer; in the discriminator, the power sequence diagram output by the generator is used as input data, and after two Conv convolution layers extract local features, the Transformer block is used to process the long-term dependency of the local features to obtain global features, and the global features are subjected to two average pooling layers to reduce the time dimension to obtain long sequence features, and the long sequence features are sequentially subjected to multiple Transformer blocks for deep feature extraction to obtain deep features, and the deep features are subjected to feature fusion through two linear operations to obtain a discrimination result; the discrimination result is used to judge whether the power image of a single target electrical appliance generated by the generator is close to the power image of the target electrical appliance in the power image set of each target electrical appliance;

[0207] The linear filter converts the single target electrical appliance power image generated by the generator into a time series, and outputs the power time series corresponding to the single target electrical appliance.

[0208] The Transformer block includes: an 8-head multi-head self-attention layer, which is used to capture the dependencies between different positions in the sequence input to the multi-head self-attention layer; two normalization layers, which are used to normalize and converge the features processed by the multi-head self-attention layer; and two layers of feedforward neural networks, which are used to transform and learn the features processed by the two normalization layers.

[0209] As an embodiment of the present invention, in the training module, based on the sample sets corresponding to the target electrical appliances, the constructed TransGIE models are trained separately, including: based on the constructed sample sets, the generator is used to extract the potential power features and information validity in the total power image, and generate a target electrical appliance power image that can deceive the discriminator; when the discrimination result of the discriminator does not meet the preset requirements, the parameters in the generator are adjusted to make the generator gradually good at generating power images that approximate the real single target electrical appliance; the discriminator is used to distinguish between the real single target electrical appliance power image and the generated image of the generator; when the discrimination result of the discriminator meets the preset requirements, the training process reaches a balanced state.

[0210] Furthermore, during the training of the TransGIE model, the generator loss function L G as follows:

[0211]

[0212] In the formula, represents the adversarial loss of the generator, where Represents the total power image set I xr Total power data extracted from , G() represents the result of the generator, and D() represents the output of the discriminator; represents the reconstruction loss of the generator, y ir represents the true value of the i-th power image of the target appliance; y ig Represents the power sequence graph of the generator decomposing the i-th total power image; n represents the number of samples.

[0213] The loss function L of the discriminator D as follows:

[0214]

[0215] In the formula, Represents the power decomposition result of the discriminator D on the generator G output

[0216] judgment; E Iyr～pIyr Represents the target electrical appliance power image set I yr The real electrical power data p extracted from Iyr Calculate the expected value of; G() represents the generation result of the generator, and D() represents the output of the discriminator;

[0217] represents the output expectation of the discriminator for the target electrical appliance power; Represents the total power image set I xr Total power data extracted from Calculate the expected value of ; represents the gradient penalty term, where λ represents the balance coefficient, Denotes the discriminator about I yg The gradient of yg It is the electrical breakdown power generated by the generator; Represents the mixed image set I from the generated electrical power and the target electrical power r Extract the generated electrical power or target electrical power data p Ir Calculate the expected value of .

[0218] The non-intrusive load monitoring device provided in the embodiment of the present invention and the non-intrusive load monitoring method provided in Example 1 are based on the same technical concept and can produce the beneficial effects as described in Example 1. For the contents not described in detail in this embodiment, please refer to Example 1.

[0219] Embodiment three:

[0220] A terminal provided by an embodiment of the present invention includes a processor and a storage medium;

[0221] The storage medium is used to store instructions;

[0222] The processor is used to operate according to the instruction to execute the steps of the method according to any one of the embodiments.

[0223] Embodiment 4:

[0224] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of any method described in Embodiment 1 are implemented.

[0225] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0226] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0227] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0228] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0229] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A non-intrusive load monitoring method, characterized in that the method include: Collecting power consumption data of multiple houses for training; Preprocess the collected power consumption data into time series images to obtain a total power image set of each residence and a power image set of each target appliance at the same time stamp; The total power image set is combined with the power image set of each target electrical appliance to produce a sample set corresponding to each target electrical appliance; Based on the sample sets corresponding to each target appliance, each TransGIE model constructed is trained separately; The trained TransGIE models are run in parallel to form a GTransGIE model group; The total power image corresponding to the total power time series of each electrical device in the user's residence to be monitored within a predetermined time period is loaded into the GTransGIE model group for decomposition to obtain the monitoring results of the individual power consumption of each electrical device in the target user's residence.

2. The non-intrusive load monitoring method according to claim 1, characterized in that: The construction of the TransGIE model includes: Transformer Blocks is combined with the generative adversarial network (GAN) to construct the TransGAN architecture, and the Conv convolution layer is used to replace the multi-layer perceptron MLP in the TransGAN architecture, and the DeConv deconvolution layer is used to replace the Linear unflatten linear smoothing layer in the TransGAN architecture to obtain the TransGIE model.

3. The non-intrusive load monitoring method according to claim 1, characterized in that: The preprocessing of the power consumption data into time series images comprises the following steps: Align the collected power consumption data according to the sampling time, and supplement the missing values ​​and replace the abnormal values; wherein the power consumption data includes the total power time series of all target electrical appliances and the power time series of each target electrical appliance; Normalizing the total power time series and the power time series of each target electrical appliance to linearly change to a range of [0,1]; According to the preset sliding window and step size, the normalized total power time series and the power time series of each target electrical appliance are respectively segmented by sliding windows to obtain segmented subsequences; Each segmented subsequence is converted into an image format of set pixels to obtain a total power image set and a power image set of each target electrical appliance.

4. The non-intrusive load monitoring method according to claim 3, characterized in that: The expressions for performing sliding window segmentation on the normalized total power time series and the power time series of each target electrical appliance are as follows: I(n)={X s:s+W-1 ,Y s:s+W-1 |s=0,1,……,T-W-1} Where I(n) represents the result of the nth sliding window segmentation; X represents the total power time series of all electrical appliances at the current time, and Y represents the power time series of a single target electrical appliance at the time aligned with X; X s:s+W-1 and Y s:s+W-1 They represent the subsequences from the sth time point to the s+w-1th time point extracted from sequence X and sequence Y respectively, T represents the total length of the current time series, W represents the size of the sliding window, and s represents the step size of the sliding window.

5. The non-intrusive load monitoring method according to claim 3, characterized in that: The step of preparing a sample set corresponding to each target electrical appliance comprises: The total power image is taken as the input sample, and the power image of each target electrical appliance is taken as the output sample, and they are paired according to the sampling time corresponding to each input sample and each output sample to form a sample set corresponding to each target electrical appliance; wherein the sample set corresponding to each target electrical appliance is divided into a training set and a test set according to a set ratio.

6. The non-intrusive load monitoring method according to claim 1 or 2, characterized in that: The structure of the TransGIE model includes a generator, a discriminator and a linear filter; The generator includes a Conv convolution layer, a Transformer block, an UpScale upsampling layer and a DeConv deconvolution layer; in the generator, the total power image in the total power image set is used as input data, and after the local features are extracted by the Conv convolution layer, the long-term dependency of the local features is processed by the Transformer block to obtain the global features, and then the global features are expanded by two UpScale upsamplings, and the extended feature sequence is sequentially subjected to multiple Transformer blocks for deep feature extraction to obtain deep features, and the deep features are subjected to the DeConv deconvolution layer to restore the resolution of the power image, and a power sequence diagram of a single target electrical appliance is generated; The discriminator includes a Conv convolution layer, a Transformer block, an average pooling layer and a linear layer; in the discriminator, the power sequence diagram output by the generator is used as input data, and after two Conv convolution layers extract local features, the Transformer block is used to process the long-term dependency of the local features to obtain global features, and the global features are subjected to two average pooling layers to reduce the time dimension to obtain long sequence features, and the long sequence features are sequentially subjected to multiple Transformer blocks for deep feature extraction to obtain deep features, and the deep features are subjected to feature fusion through two linear layers to obtain a discrimination result; the discrimination result is used to judge whether the power image of a single target electrical appliance generated by the generator is close to the power image of the target electrical appliance in the power image set of each target electrical appliance; The linear filter converts the single target electrical appliance power image generated by the generator into a time series, and outputs the power time series corresponding to the single target electrical appliance.

7. The non-intrusive load monitoring method according to claim 6, characterized in that: The Transformerblock includes: An 8-head multi-head self-attention layer to capture the dependencies between different positions in the sequence input to the multi-head self-attention layer; Two normalization layers are used to normalize and converge the features after being processed by the multi-head self-attention layer; A two-layer feedforward neural network is used to transform and learn the features after being processed by two normalization layers.

8. The non-intrusive load monitoring method according to claim 6, characterized in that: Based on the sample sets corresponding to each target appliance, each constructed TransGIE model is trained separately, including: based on the constructed sample sets, the generator is used to extract the potential power features and information validity in the total power image, and generate a target appliance power image that can deceive the discriminator. When the discrimination result of the discriminator does not meet the preset requirements, the parameters in the generator are adjusted to make the generator gradually good at generating power images that approximate the real single target appliance; the discriminator is used to distinguish between the real single target appliance power image and the image generated by the generator. When the discrimination result of the discriminator meets the preset requirements, the training process reaches a balanced state.

9. The non-intrusive load monitoring method according to claim 8, characterized in that: In the process of training the TransGIE model, the loss function L of the generator G as follows: In the formula, represents the adversarial loss of the generator, where Represents the total power image set I xr Total power data extracted from , G() represents the result of the generator, and D() represents the output of the discriminator; represents the reconstruction loss of the generator, y ir represents the true value of the i-th power image of the target appliance; y ig Represents the power sequence graph of the generator decomposing the i-th total power image; n represents the number of samples.

10. The non-intrusive load monitoring method according to claim 8, characterized in that: When training the TransGIE model, the loss function L of the discriminator is D as follows: In the formula, Represents the power decomposition result of the discriminator D on the generator G output judgment; E Iyr～pIyr Represents the target electrical appliance power image set I yr The real electrical power data p extracted from Iyr Calculate the expected value of; G() represents the generation result of the generator, and D() represents the output of the discriminator; represents the output expectation of the discriminator for the target electrical appliance power; Represents the total power image set I xr Total power data extracted from Calculate the expected value of ; represents the gradient penalty term, where λ represents the balance coefficient, ▽ Iyg D(G(I yg )) represents the discriminator about I yg The gradient of yg It is the electrical breakdown power generated by the generator; Represents the mixed image set I from the generated electrical power and the target electrical power r Extract the generated electrical power or target electrical power data p Ir Calculate the expected value of .

11. A non-intrusive load monitoring device, running the non-intrusive load monitoring method according to any one of claims 1 to 10, characterized in that: The device includes: A collection module, used for collecting power consumption data of multiple houses for training; A preprocessing module is used to preprocess the collected power consumption data into time series images to obtain a total power image set of each residential building and a power image set of each target electrical appliance corresponding to the same timestamp; A production module, used to combine the total power image set with the power image set of each target electrical appliance to produce a sample set corresponding to each target electrical appliance; A training module, used for running the trained TransGIE models in parallel to synthesize a GTransGIE model group; A synthesis module, used for running the trained TransGIE models in parallel to synthesize a GTransGIE model group; The decomposition monitoring module is used to load the total power image corresponding to the total power time series of each electrical device in the user's residence to be monitored within a predetermined time period into the GTransGIE model group for decomposition to obtain the monitoring results of the individual power consumption of each electrical device in the target user's residence.

12. The non-intrusive load monitoring device according to claim 11, characterized in that: The construction of the TransGIE model in the training module includes: Transformer Blocks is combined with the generative adversarial network (GAN) to construct the TransGAN architecture, and the Conv convolution layer is used to replace the multi-layer perceptron MLP in the TransGAN architecture, and the DeConv deconvolution layer is used to replace the Linear unflatten linear smoothing layer in the TransGAN architecture to obtain the TransGIE model.

13. The non-intrusive load monitoring device according to claim 11, characterized in that: The preprocessing module comprises: A supplementation and replacement unit, used to align the collected power consumption data according to the sampling time, and to supplement the missing values ​​and replace the abnormal values; wherein the power consumption data includes the total power time series of all target electrical appliances and the power time series of each target electrical appliance; A normalization unit, used for normalizing the total power time series and the power time series of each target electrical appliance respectively, so as to linearly change them to a range of [0,1]; A segmentation unit is used to perform sliding window segmentation on the normalized total power time series and the power time series of each target electrical appliance according to a preset sliding window and step size to obtain segmented subsequences; The image conversion unit is used to convert each segmented subsequence into an image format of set pixels to obtain a total power image set and a power image set of each target electrical appliance.

14. The non-intrusive load monitoring device according to claim 13, characterized in that: In the segmentation unit, the normalized total power time series and the power time series of each target electrical appliance are segmented by sliding windows respectively, as follows: I(n)={X s:s+W-1 ,Y s:s+W-1 |s=0,1,……,T-W-1} Where I(n) represents the result of the nth sliding window segmentation; X represents the total power time series of all electrical appliances at the current time, and Y represents the power time series of a single target electrical appliance at the time aligned with X; X s:s+W-1 and Y s:s+W-1 They represent the subsequences from the sth time point to the s+w-1th time point extracted from sequence X and sequence Y respectively, T represents the total length of the current time series, W represents the size of the sliding window, and s represents the step size of the sliding window.

15. The non-intrusive load monitoring device according to claim 13, characterized in that: In the production module, producing a sample set corresponding to each target electrical appliance includes: The total power image is taken as the input sample, and the power image of each target electrical appliance is taken as the output sample, and they are paired according to the sampling time corresponding to each input sample and each output sample to form a sample set corresponding to each target electrical appliance; wherein the sample set corresponding to each target electrical appliance is divided into a training set and a test set according to a set ratio.

16. The non-intrusive load monitoring device according to claim 12, characterized in that: The structure of the TransGIE model includes a generator, a discriminator and a linear filter; The generator includes a Conv convolution layer, a Transformer block, an UpScale upsampling layer and a DeConv deconvolution layer; in the generator, the total power image in the total power image set is used as input data, and after the local features are extracted by the Conv convolution layer, the long-term dependency of the local features is processed by the Transformer block to obtain the global features, and then the global features are expanded by two UpScale upsamplings, and the extended feature sequence is sequentially subjected to multiple Transformer blocks for deep feature extraction to obtain deep features, and the deep features are subjected to the DeConv deconvolution layer to restore the resolution of the power image, and a power sequence diagram of a single target electrical appliance is generated; The discriminator includes a Conv convolution layer, a Transformer block, an average pooling layer and a linear layer; in the discriminator, the power sequence diagram output by the generator is used as input data, and after two Conv convolution layers extract local features, the Transformer block is used to process the long-term dependency of the local features to obtain global features, and the global features are subjected to two average pooling layers to reduce the time dimension to obtain long sequence features, and the long sequence features are sequentially subjected to multiple Transformer blocks for deep feature extraction to obtain deep features, and the deep features are subjected to feature fusion through two linear layers to obtain a discrimination result; the discrimination result is used to judge whether the power image of a single target electrical appliance generated by the generator is close to the power image of the target electrical appliance in the power image set of each target electrical appliance; The linear filter converts the single target electrical appliance power image generated by the generator into a time series, and outputs the power time series corresponding to the single target electrical appliance.

17. The non-intrusive load monitoring device according to claim 16, characterized in that: The Transformerblock includes: An 8-head multi-head self-attention layer to capture the dependencies between different positions in the sequence input to the multi-head self-attention layer; Two normalization layers are used to normalize and converge the features after being processed by the multi-head self-attention layer; A two-layer feedforward neural network is used to transform and learn the features after being processed by two normalization layers.

18. The non-intrusive load monitoring device according to claim 16, characterized in that: In the training module, based on the sample sets corresponding to the target electrical appliances, the constructed TransGIE models are trained separately, including: based on the constructed sample sets, the generator is used to extract the potential power features and information validity in the total power image, and generate a target electrical appliance power image that can deceive the discriminator. When the discrimination result of the discriminator does not meet the preset requirements, the parameters in the generator are adjusted to make the generator gradually good at generating power images that approximate the real single target electrical appliance; the discriminator is used to distinguish between the real single target electrical appliance power image and the image generated by the generator. When the discrimination result of the discriminator meets the preset requirements, the training process reaches a balanced state.

19. The non-intrusive load monitoring device according to claim 18, characterized in that: In the process of training the TransGIE model, the loss function L of the generator G as follows: In the formula, represents the adversarial loss of the generator, where Represents the total power image set I xr Total power data extracted from , G() represents the result of the generator, and D() represents the output of the discriminator; represents the reconstruction loss of the generator, y ir represents the true value of the i-th power image of the target appliance; y ig Represents the power sequence graph of the generator decomposing the i-th total power image; n represents the number of samples.

20. The non-intrusive load monitoring device according to claim 18, characterized in that: When training the TransGIE model, the loss function L of the discriminator is D as follows: In the formula, Represents the power decomposition result of the discriminator D on the generator G output judgment; E Iyr～pIyr Represents the target electrical appliance power image set I yr The real electrical power data p extracted from Iyr Calculate the expected value of; G() represents the generation result of the generator, and D() represents the output of the discriminator; represents the output expectation of the discriminator for the target electrical appliance power; Represents the total power image set I xr Total power data extracted from Calculate the expected value of ; represents the gradient penalty term, where λ represents the balance coefficient, ▽ Iyg D(G(I yg )) represents the discriminator about I yg The gradient of yg It is the electrical breakdown power generated by the generator; Represents the mixed image set I from the generated electrical power and the target electrical power r Extract the generated electrical power or target electrical power data p Ir Calculate the expected value of .

21. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-10.

22. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images