Non-intrusive load identification method and device

Abstract

The invention provides a non-intrusive load identification method and device. The method comprises the following steps: acquiring first operation data including different types of electric equipment; extracting features of the multi-modal signals of the first operation data, and fusing the features of the multi-modal signals to generate multi-modal fusion features; generating a feature map based on the multi-modal fusion features, and forming an image load signature by the feature map; performing supervised training on the load identification model based on the image load signature; second operation data including different types of electric equipment are collected in real time, and image load signatures corresponding to the second operation data are obtained; inputting the image load signature into a trained load identification model for classification to obtain a total load label; and decomposing a total load corresponding to the second operation data according to the total load label, and identifying a load type corresponding to each load component in the total load. According to the method, the accuracy of the load labeling data is ensured, and high-precision electric appliance identification can be realized even in a complex load environment.

Description

Technical Field

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

[0002] Non-Intrusive Load Monitoring (NILM) analyzes the total power consumption signal to identify which electrical devices are in the loop and their power consumption. The core of this technology is that there is no need to install a separate sensor on each appliance. Instead, the total power consumption signal of the entire building or facility is intelligently analyzed to infer the usage and energy consumption distribution of each appliance. Compared with traditional invasive methods, NILM technology has the advantages of low cost, simple deployment, and little impact on user use. Therefore, NILM technology has broad application prospects in the fields of smart grid, home energy management, equipment fault diagnosis, etc.

[0003] Although NILM technology has broad application prospects, it still faces many challenges in practical applications. First, existing NILM methods usually rely on manual feature engineering, that is, experts are required to design and select features based on experience and specific application scenarios. This process is not only time-consuming and labor-intensive, but also the effectiveness of the features is highly dependent on the experience of experts, and it is difficult to automatically adapt to different application scenarios. Secondly, under complex load modes, it is difficult for existing NILM methods to accurately distinguish the power consumption of different electrical appliances. For example, when multiple electrical appliances work at the same time, the power signals will be superimposed, resulting in feature ambiguity and a significant decrease in recognition accuracy. In addition, some electrical appliances have similar power consumption patterns, such as different models of electric lamps, whose working modes and power curves are very close, making it difficult for traditional methods to distinguish.

[0004] Traditional NILM methods also have problems in real-time performance. They usually require a large amount of historical data for training and model updates. The training process takes a long time and cannot meet the needs of real-time monitoring. In the power system, the load changes frequently and rapidly. Traditional methods cannot complete model updates in a short period of time, resulting in delayed recognition results. In addition, existing methods perform poorly when faced with dynamically changing load environments. Users' electricity consumption behaviors vary greatly in different time periods, and appliance usage patterns also change with seasons and climate changes. Traditional methods find it difficult to maintain good recognition performance in these changes. NILM technology requires a large amount of labeled data for training, but in actual applications, it is very difficult to obtain accurate load labeling data. Especially in large buildings and industrial scenarios, manual labeling is a huge workload and prone to errors, resulting in insufficient model training data and affecting recognition results. Summary of the invention

[0005] The present invention provides a non-invasive load identification method and device, which are used to solve the technical problem that the existing NILM method usually relies on manual feature engineering and is difficult to obtain accurate load labeling data.

[0006] According to one aspect of the present invention, a non-intrusive load identification method is provided, comprising:

[0007] Collecting first operation data including different types of electrical equipment;

[0008] Extracting features of a multimodal signal of the first operation data, fusing the features of the multimodal signal to generate a multimodal fusion feature;

[0009] generating a feature map based on the multimodal fusion features, and forming an image load signature from the feature map;

[0010] Performing supervised training on a load recognition model based on the image load signature;

[0011] collecting second operation data of different types of electrical equipment in real time, and obtaining the image load signature corresponding to the second operation data;

[0012] The image load signature is input into the trained load identification model for classification to obtain a total load label; the total load corresponding to the second operating data is decomposed according to the total load label to identify the load type corresponding to each load component in the total load.

[0013] Optionally, the collecting includes the first operation data of different types of electrical equipment, and then further includes:

[0014] Preprocessing the first operation data;

[0015] The preprocessing at least includes denoising and standardization.

[0016] Optionally, after collecting the first operation data of different types of electrical equipment, the method further includes:

[0017] The preprocessed first operating data is divided into a training set and a test set, and samples of the training set and the test set both contain the multimodal signal and a load type label corresponding to the first operating data.

[0018] Optionally, the multimodal signal includes at least a current signal, a voltage signal and a power factor signal;

[0019] Then the step of extracting the features of the multimodal signal of the first operation data and fusing the features of the multimodal signal to generate a multimodal fusion feature includes:

[0020] A temporal convolutional network is used to extract the local features of the current signal and the voltage signal respectively;

[0021] Convolutional neural network is used to extract the features of power factor signal:

[0022] The local features of the current signal, the local features of the voltage signal, and the features of the power factor signal are fused through a fully connected layer to generate a multimodal fusion feature.

[0023] Optionally, the feature graph includes at least a learnable cyclic graph model feature graph, a learnable Gram matrix feature graph, and a two-dimensional feature graph;

[0024] Then the generating of a feature map based on the multimodal fusion feature, and forming an image load signature from the feature map, includes:

[0025] Using a fully connected network to learn the relationship between multimodal fusion features, generating the learnable cyclic graph model feature graph;

[0026] Calculating the inner product between the multimodal fusion features, extracting the nonlinear features of the inner product between the multimodal fusion features using a nonlinear activation function, and generating the learnable Gram matrix feature map;

[0027] Flattening the multimodal fusion features into a one-dimensional vector, and generating the two-dimensional feature map through a fully connected network;

[0028] The image load signature is formed by at least one of the cyclic graph model feature map, the learnable Gram matrix feature map, and the two-dimensional feature map.

[0029] Optionally, before the supervised training of the load recognition model based on the image load signature, the method further includes:

[0030] The load identification model is pre-trained using samples without electrical equipment labels to obtain initial weights of the load identification model.

[0031] Optionally, the performing supervised training on the load recognition model based on the image load signature includes:

[0032] Multi-label classification training is performed by minimizing cross entropy as the first loss function of the load recognition model; the input of the load recognition model is the image load signature, and the output is the composition of various types of loads.

[0033] Optionally, also include:

[0034] Using the real-time third operating data to perform online training on the load identification model;

[0035] The adopting the real-time third operating data to perform online training on the load identification model includes:

[0036] collecting the third operation data in real time;

[0037] Acquire the image load signature corresponding to the third operation data;

[0038] An incremental learning algorithm is used to input the image load signature corresponding to the third operating data into the load identification model for classification training; a second loss function of the load identification model adds a regularization term to the first loss function.

[0039] Optionally, the step of inputting the image load signature into the trained load recognition model for classification to obtain a total load label; decomposing the total load corresponding to the second operation data according to the total load label to identify the load type corresponding to each load component in the total load includes:

[0040] Inputting the image load signature into the trained load recognition model to predict the total load label of the image load signature;

[0041] A variational autoencoder is used to decompose the total load corresponding to the second operating data into load components corresponding to each electrical device, thereby identifying the load type corresponding to each load component.

[0042] According to another aspect of the present invention, there is provided a non-intrusive load identification device, comprising:

[0043] A first collection unit, used to collect first operation data including different types of electrical equipment;

[0044] an extraction unit, configured to extract features of a multimodal signal of the first operation data, and fuse the features of the multimodal signal to generate a multimodal fusion feature;

[0045] An imaging unit, used for generating a feature map based on the multimodal fusion feature, wherein the feature map constitutes an image load signature;

[0046] A training unit, configured to perform supervised training on a load recognition model based on the image load signature;

[0047] A second collection unit, configured to collect in real time second operation data of different types of electrical equipment, and obtain the image load signature corresponding to the second operation data;

[0048] The identification unit is used to input the image load signature into the trained load identification model for classification to obtain a total load label; decompose the total load corresponding to the second operating data according to the total load label, and identify the load type corresponding to each load component in the total load.

[0049] The technical solution of the embodiment of the present invention extracts the characteristics of the multimodal signal of the first operating data, fuses the characteristics of the multimodal signal, and generates a multimodal fusion feature; and generates a feature map based on the multimodal fusion feature, and constructs an image load signature from the feature map to obtain the annotation data of the operating data. This increases the expressiveness of the electrical appliance characteristics, and can achieve high-precision electrical appliance identification even in complex load environments, thereby ensuring the accuracy of the load annotation data and improving the accuracy of load identification.

[0050] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0052] Figure 1 is a flow chart of an embodiment of a non-intrusive load identification method provided by the present invention;

[0053] Figure 2 is a flow chart of another embodiment of a non-intrusive load identification method provided by the present invention;

[0054] Figure 3 It is a structural schematic diagram of a non-intrusive load identification device for implementing an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0056] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0057] Figure 1 Flow chart of a non-intrusive load identification method provided by an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] S101. Collect first operation data including different types of electrical equipment.

[0059] Different types of electrical equipment may include various electrical appliances in a home or various industrial equipment in the industrial field, for example, including electrical appliances such as electric lights, air conditioners, refrigerators, etc.

[0060] The first operating data may be voltage data, current data, power factor and other data in an environment including multiple types of electrical equipment. The first operating data may be obtained by collecting current signals, voltage signals and power factor signals of multiple types of electrical equipment working simultaneously through high-precision current sensors, voltage sensors and power factor sensors to obtain corresponding current data, voltage data and power factor.

[0061] This embodiment needs to collect the first operating data as training samples, so the types of electrical equipment collected should include all electrical equipment in the test environment as much as possible, and as much first operating data as possible should be collected, for example, 2000 first operating data of dozens of electrical appliances can be included.

[0062] S102: extract features of the multimodal signal of the first operating data, fuse the features of the multimodal signal, and generate multimodal fusion features.

[0063] The multimodal signal in this embodiment refers to a current signal, a voltage signal, a power factor signal, etc. The features of the multimodal signal are extracted at least including the features of the current signal, the features of the voltage signal, and the features of the power factor signal. The method of extracting the features of each signal can be a time convolution network TCN, or a convolutional neural network extraction method, for example, by using a time convolution network TCN or a convolutional neural network to extract the features of the current signal, the voltage signal, and the power factor signal. Of course, other feature extraction methods can also be used to extract the features of each signal.

[0064] After the multimodal signal is feature extracted, the features of each signal can be feature fused. For example, for the extracted features of the current signal, the features of the voltage signal, and the features of the power factor signal, the three features can be feature fused to obtain a multimodal fusion feature. If there are multiple signals, all or part of the features of the multiple signals can also be feature fused. The feature fusion method can be to fuse the features of each signal through a fully connected layer. Of course, other methods can also be used for feature fusion. This embodiment is only an exemplary description.

[0065] S103: Generate a feature map based on the multimodal fusion features, and use the feature map to form an image load signature.

[0066] It should be noted that the types of feature maps in this embodiment may include multiple types, that is, the multimodal fusion features may be used to generate multiple corresponding feature maps, and at least one feature map may be selected from the feature maps to form an image load signature. The feature map generated in this embodiment can well express the load characteristics of each electrical appliance.

[0067] S104: Perform supervised training on a load recognition model based on the image load signature.

[0068] It should be noted that a neural network model can be used as a load identification model. For example, a convolutional neural network can be used as a load identification model to classify and train image load signatures, thereby optimizing model parameters. Specifically, the input of the load identification model is the image load signature, and the output loop contains the composition of various types of loads.

[0069] The loss function of the load identification model can be set as needed. For example, in this embodiment, a cross entropy loss function can be minimized.

[0070] S105 . Collect second operation data of different types of electrical equipment in real time, and obtain the image load signature corresponding to the second operation data.

[0071] For load identification of each electrical device in the current environment, second operation data of different types of electrical devices in the current environment can be collected in real time. The collection method of the second operation data can be the same as that of the first operation data, and the collected second operation data can be used to identify the load type corresponding to each electrical device working in the current environment.

[0072] The image load signature corresponding to the second operating data can also be obtained using the same acquisition method as the image load signature corresponding to the first operating data, that is, first extracting the features of the multimodal signal of the first operating data, fusing the features of the multimodal signal to generate a multimodal fusion feature; generating a feature map based on the multimodal fusion feature, and forming the image load signature from the feature map.

[0073] S106. Input the image load signature into the trained load identification model for classification to obtain a total load label; decompose the total load corresponding to the second operating data according to the total load label to identify the load type corresponding to each load component in the total load.

[0074] The image load signature corresponding to the second operating data can be input into a trained load identification model to obtain the total load type corresponding to the image load signature. By decomposing the total load, the load components in the total load can be obtained, thereby identifying the load type corresponding to each load component in the total load.

[0075] In this embodiment, the method for decomposing the total load type may be to decompose the total load type through a variational autoencoder VAE.

[0076] Furthermore, after obtaining each load type, the energy consumption of each electrical device can be calculated based on the identification result and the signal size collected in real time.

[0077] The technical solution of the embodiment of the present invention extracts the characteristics of the multimodal signal of the first operating data, fuses the characteristics of the multimodal signal, and generates a multimodal fusion feature; and generates a feature map based on the multimodal fusion feature, and constructs an image load signature from the feature map to obtain the annotation data of the operating data. This increases the expressiveness of the electrical appliance characteristics, and can achieve high-precision electrical appliance identification even in complex load environments, thereby ensuring the accuracy of the load annotation data and improving the accuracy of load identification.

[0078] In a specific implementation manner, the present invention also provides a second embodiment of a non-intrusive load identification method. Figure 2 The flowchart shown, Figure 2 Included:

[0079] S201. Collect first operation data including different types of electrical equipment.

[0080] In this embodiment, a high-precision current sensor, voltage sensor and power factor sensor may be used to respectively collect the current I(t), voltage V(t) and power factor PF(t) signals in the current system.

[0081] When collecting operating data, the sampling frequency can be set to 50kHz to ensure that the characteristics of various electrical signals can be identified with high accuracy. When implementing it, select appropriate sensors and data acquisition systems to ensure that the above requirements can be met.

[0082] S202, preprocessing the first operating data; the preprocessing at least includes denoising and standardization.

[0083] Preprocessing the first operating data may include: performing denoising processing on the first operating data, and extracting a standardized signal within a period from the first operating data.

[0084] Among them, the denoising process for the operating data can remove high-frequency noise in the operating data through a low-pass filter, and smooth the operating data through a mean filter, thereby ensuring the accuracy and smoothness of the operating data and improving the quality of the operating data.

[0085] Specifically, the low-pass filter used can be implemented by the following formula:

[0086] I filtered (t)=I(t)*h(t)

[0087] where h(t) is the impulse response of the low-pass filter.

[0088] The mean filter is used to smooth the running data, specifically:

[0089]

[0090] A suitable filter window size N may be selected, for example, N=5 or N=10.

[0091] Extracting the standardized signal within a cycle may include extracting first operation data within a cycle and performing standardization processing on the first operation data to ensure consistency of the signal.

[0092] The periods of the current signal and the voltage signal may be determined by detecting the zero-crossing point of the current data or the voltage data in the first operation data, thereby determining a period T of the first operation data: T=1 / f.

[0093] Where f is the grid frequency (50 Hz or 60 Hz). When implementing, an algorithm for detecting the zero crossing point can be written to identify the zero crossing point where the voltage signal or current signal changes from negative to positive.

[0094] The voltage, current and power factor signals of one cycle are extracted respectively and standardized:

[0095]

[0096]

[0097]

[0098] Among them, μ I , μ V , μ PF and σ I , σ V , σ PF are the mean and standard deviation of the current signal, voltage signal and power factor signal respectively. When implementing, the mean and standard deviation of the extracted periodic signal are calculated and normalized.

[0099] S203: Divide the preprocessed first operation data into a training set and a test set, wherein samples of the training set and the test set both contain the multimodal signal and a load type label corresponding to the first operation data.

[0100] In this embodiment, the preprocessed data set can be divided into a training set and a test set. Each sample contains a standardized current signal, a voltage signal, a power factor signal, and a load type label corresponding to the sample. The multimodal signal includes at least a current signal, a voltage signal, and a power factor signal. To ensure the diversity of the data set, the data set needs to cover different types of electrical appliances and load characteristics.

[0101] S204: extract features of the multimodal signal of the first operating data, fuse the features of the multimodal signal, and generate multimodal fusion features.

[0102] In this embodiment, a time convolution network can be used to extract local features of the current signal and the voltage signal respectively; a convolutional neural network can be used to extract features of the power factor signal: the local features of the current signal, the local features of the voltage signal, and the features of the power factor signal are fused through a fully connected layer to generate multimodal fusion features.

[0103] Specifically, a temporal convolutional network can be used to extract local features of current and voltage signals:

[0104] F I =TCN(I norm )

[0105] F V =TCN(V norm )

[0106] Among them, I norm represents the normalized current signal, F I is the current feature matrix extracted by the temporal convolutional network (TCN), with dimension d I ×n. V norm represents the standardized voltage signal, F V is the voltage feature matrix extracted by the temporal convolutional network (TCN), with dimension d V ×n.

[0107] Use Convolutional Neural Network (CNN) to extract the features of the power factor signal:

[0108] F PF =CNN(PF norm )

[0109] Among them, PF norm Represents the normalized power factor signal, F PF is the power factor feature matrix extracted by convolutional neural network (CNN), with dimension d PF ×n.

[0110] The multimodal features can be fused through the fully connected layer (FC):

[0111] F fused =FC([F I ,F V ,F PF ])

[0112] Among them, F I 、F V and F PF represent the features extracted from the current signal, voltage signal and power factor signal respectively, [F I , F v , F PF ] represents the characteristic matrix obtained by connecting the current, voltage and power factor characteristics in series. The dimension of the characteristic matrix is ​​(d I +d V +d PF )×n. F fused It is the fused feature matrix, which represents the multimodal fusion feature obtained by fusing the multimodal features through the fully connected layer, and its dimension is d fused ×n,d fused is the feature dimension after fusion.

[0113] S205: Generate a feature map based on the multimodal fusion feature, and use the feature map to form an image load signature.

[0114] By learning the relationship between fusion features, graphical features are generated, which can better represent the load characteristics of electrical appliances.

[0115] In one embodiment, a fully connected network may be used to learn the relationship between multimodal fusion features to generate the learnable cyclic graph model feature graph.

[0116] Specifically, a fully connected network is used to learn the relationship between fusion features and generate a cyclic graph model feature graph LRG:

[0117] α i，j =ReLU(W 1 ·(F fused,i -F fused，j )+b 1 )

[0118] LRG=ReLU(W 2 α+b 2 )

[0119] Among them, F fused,i and F fused，j Respectively represent the fusion feature matrix F fused The i-th and j-th columns of i，j Represents the feature relationship after being processed by the ReLU activation function, W 1 and W 2 is the weight matrix of the fully connected network, b 1 and b 2 is the bias vector.

[0120] In one embodiment, the inner product between the multimodal fusion features may be calculated, and a nonlinear activation function may be used to extract the nonlinear features of the inner product between the multimodal fusion features to generate the learnable Gram matrix feature map.

[0121] Specifically, by calculating the inner product between multimodal fusion features, a Gram matrix is ​​generated to further enhance the representation ability of the image features.

[0122] The feature matrix F can be calculated fused Gram inner product of . Assume that F fused The dimension is d×n, where d is the dimension of the feature and n is the number of samples. The Gram matrix G is calculated as follows:

[0123]

[0124] Among them, G is a d×d matrix, which represents the inner product between features.

[0125] The calculated Gram matrix G is processed to generate a Gram matrix feature map LGM. One processing method is to extract nonlinear features through a nonlinear activation function, such as ReLU:

[0126] LGM=ReLU(G)

[0127] This can enhance the representation capability of features, accurately generate graphical features, and ensure that these features can better represent the load characteristics of electrical appliances, providing a high-quality data foundation for subsequent load identification.

[0128] In one embodiment, the multimodal fusion features are flattened into a one-dimensional vector, and the one-dimensional vector is passed through a fully connected network to generate the two-dimensional feature map.

[0129] Specifically, the multimodal fusion features can be expanded into a one-dimensional vector, and the one-dimensional vector can be converted into a two-dimensional feature map through a fully connected network, making the features easier to process by the deep learning model.

[0130] First, flatten the fused features into a one-dimensional vector:

[0131] x flat =Flatten(F fused )

[0132] Among them, the input tensor F fused The shape is d×n. After flattening, the one-dimensional vector x flat The shape of is d×n. Generate a two-dimensional feature map through a fully connected network:

[0133] GG = Reshape(ReLU(W 3 ·x flat +b 3 ))

[0134] Among them, W 3 is the weight matrix of the fully connected network, b 3 is the bias vector.

[0135] In one embodiment, the image load signature is formed by at least one of the cyclic graph model feature map, the learnable Gram matrix feature map, and the two-dimensional feature map.

[0136] It should be noted that the image load signature may include at least one of a cyclic graph model feature map, a learnable Gram matrix feature map, and a two-dimensional feature map. For example, a two-dimensional feature map GG may be selected as the image load signature, or three feature maps may be selected as the final image load signature (LRG, LGM, GG).

[0137] S206: Pre-train the load identification model using samples without electrical device labels to obtain initial weights of the load identification model.

[0138] When training the load identification model, firstly, the load identification model can be pre-trained using samples without electrical equipment labels to obtain the initial weights of the load identification model so as to improve the generalization ability of the load identification model.

[0139] Specifically, a self-supervised learning task, such as a signal reconstruction task, can be constructed. A self-supervised task is designed, using unlabeled first run data, with the goal of predicting future time steps of a multimodal signal or reconstructing an input signal.

[0140] For example, using a time window with a time step of Δt, given the multimodal fusion features F of the first T-Δt time steps fused , predict the multimodal fusion features of the next Δt time steps

[0141]

[0142] Secondly, use the model weights pre-trained by self-supervised learning as the initial weights to perform supervised learning tasks. During the training process, minimize the prediction error:

[0143]

[0144] in, is the future signal predicted by the model, IF fused (t i +Δt) is the actual current signal. The weights of self-supervised learning are used as the initial weights of supervised learning through transfer learning, thereby improving the training efficiency and the generalization ability of the model.

[0145] S207: Perform supervised training on a load recognition model based on the image load signature.

[0146] In this embodiment, a convolutional neural network (CNN) can be used to perform multi-classification training on the generated image features to optimize model parameters. In one embodiment, multi-label classification training can be performed by minimizing cross entropy as the first loss function of the load recognition model; the input of the load recognition model is the image load signature, and the output is the composition of each type of load.

[0147] Specifically, for the various types of complex features reflected by electrical characteristics, a multi-classification task is constructed, the input is a two-dimensional feature map GG, and the output is the composition of various types of loads contained in the circuit, expressed as:

[0148]

[0149] in, It represents the load power of the i-th electrical appliance. If it is 0, it means there is no such appliance.

[0150] The generated image load signature is trained for multi-label classification using a convolutional neural network (CNN), minimizing the cross entropy loss function:

[0151]

[0152] Among them, L sup is the cross entropy loss function, y i is the true label vector, and the CNN model training can be implemented using the deep learning framework.

[0153] S208: Use the real-time third operating data to perform online training on the load identification model.

[0154] In order to cope with the dynamically changing load environment, this embodiment can use real-time operation data to perform online training on the load identification model, so that the load identification model can adapt to the current power consumption environment to ensure the accuracy of identification.

[0155] In one embodiment, the third operating data may be collected in real time.

[0156] The image load signature corresponding to the third operation data is obtained.

[0157] An incremental learning algorithm is used to input the image load signature corresponding to the third operating data into the load identification model for classification training; a second loss function of the load identification model adds a regularization term to the first loss function.

[0158] The third operation data is the operation data of different types of electrical equipment in the current environment. The method for obtaining the image load signature corresponding to the first operation data is the same as the method for obtaining the image load signature corresponding to the first operation data. That is, the third operation data is first preprocessed, and then the features of the multimodal signal of the third operation data are extracted, and the features of the multimodal signal are fused to generate multimodal fusion features; a feature map is generated based on the multimodal fusion features, and the image load signature is formed by the feature map.

[0159] Using the incremental learning algorithm Elastic Weight Consolidation (EWC), old knowledge is retained when training on new data. EWC prevents forgetting by adding a regularization term to the loss function:

[0160]

[0161] Among them, F k is the Fisher information matrix, θ kis the current parameter, is the optimal parameter for the old task, and λ is the regularization coefficient.

[0162] Fisher information matrix F k Used for weight update in the EWC algorithm, it reflects the importance of the parameter to the model output. For each parameter θ k , the element of the Fisher information matrix F k The calculation formula is as follows:

[0163]

[0164] Where p(y|θ) represents the probability distribution of the model output.

[0165] S209: Collect second operation data of different types of electrical equipment in real time, and obtain the image load signature corresponding to the second operation data.

[0166] For load identification of each electrical device in the current environment, the second operation data of different types of electrical devices in the current environment can be collected in real time. The collection method of the second operation data can be the same as that of the third operation data, and the collected second operation data can be used to identify the load type corresponding to each electrical device working in the current environment.

[0167] The image load signature corresponding to the second operation data can also be obtained using the same acquisition method as the image load signature corresponding to the third operation data, that is, first preprocessing the third operation data, then extracting the features of the multimodal signal of the third operation data, and fusing the features of the multimodal signal to generate a multimodal fusion feature; generating a feature map based on the multimodal fusion feature, and the image load signature is composed of the feature map.

[0168] S210. Input the image load signature into the trained load identification model for classification to obtain a total load label; decompose the total load corresponding to the second operating data according to the total load label to identify the load type corresponding to each load component in the total load.

[0169] In one embodiment, step S210 includes inputting the image load signature into the trained load recognition model to predict a total load label of the image load signature.

[0170] Specifically, the imaged load signature GG is input into the trained convolutional neural network (CNN) classifier:

[0171]

[0172] Where GG represents the generated graphical load signature, is the label predicted by the model.

[0173] A variational autoencoder is used to decompose the total load corresponding to the second operating data into load components corresponding to each electrical device, thereby identifying the load type corresponding to each load component.

[0174]

[0175] in Represents the load component of the i-th electrical appliance.

[0176] Furthermore, after obtaining each load type, the energy consumption of each electrical device can be calculated according to the identification result and the signal size collected in real time. The calculation method is as follows:

[0177]

[0178] The present invention enhances the expressiveness of electrical appliance features by introducing image load signatures and multimodal feature fusion technology, and can achieve high-precision electrical appliance recognition even in complex load environments. This method can effectively distinguish electrical appliances with similar power curves and improve the accuracy of recognition. Pre-training through self-supervised learning technology reduces the dependence on a large amount of labeled data and improves the generalization ability of the model. The use of incremental learning technology enables the model to be updated quickly to meet real-time monitoring needs and ensure efficient and accurate recognition performance in an environment where power load changes frequently. And through adaptive feature extraction and model update mechanisms, the system can automatically adapt to load changes in different time periods, seasons and application scenarios, ensuring the stability of recognition performance.

[0179] Figure 3 This is an embodiment of a non-intrusive load identification device provided by the present invention. Figure 3 As shown, the device comprises:

[0180] The first collecting unit 301 is used to collect first operation data including different types of electrical equipment;

[0181] An extraction unit 302 is used to extract features of a multimodal signal of the first operation data, and fuse the features of the multimodal signal to generate a multimodal fusion feature;

[0182] An imaging unit 303, configured to generate a feature map based on the multimodal fusion feature, and the feature map constitutes an image load signature;

[0183] A training unit 304, configured to perform supervised training on a load recognition model based on the image load signature;

[0184] The second collection unit 305 is used to collect second operation data of different types of electrical equipment in real time, and obtain the image load signature corresponding to the second operation data;

[0185] The identification unit 306 is used to input the image load signature into the trained load identification model for classification to obtain a total load label; decompose the total load corresponding to the second operating data according to the total load label, and identify the load type corresponding to each load component in the total load.

[0186] The non-intrusive load identification device provided in the embodiment of the present invention can execute the non-intrusive load identification device provided in any embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method.

[0187] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0188] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A non-intrusive load identification method, characterized in that: include: Collecting first operation data including different types of electrical equipment; Extracting features of a multimodal signal of the first operation data, fusing the features of the multimodal signal to generate a multimodal fusion feature; generating a feature map based on the multimodal fusion features, and forming an image load signature from the feature map; Performing supervised training on a load recognition model based on the image load signature; collecting second operation data of different types of electrical equipment in real time, and obtaining the image load signature corresponding to the second operation data; Inputting the image load signature into the trained load recognition model for classification to obtain a total load label; The total load corresponding to the second operating data is decomposed according to the total load label, and the load type corresponding to each load component in the total load is identified.

2. The non-intrusive load identification method according to claim 1, characterized in that: The collecting includes the first operation data of different types of electrical equipment, and then further includes: Preprocessing the first operation data; The preprocessing at least includes denoising and standardization.

3. The non-intrusive load identification method according to claim 2, characterized in that: After collecting the first operation data of different types of electrical equipment, the method further includes: The preprocessed first operating data is divided into a training set and a test set, and samples of the training set and the test set both contain the multimodal signal and a load type label corresponding to the first operating data.

4. The non-intrusive load identification method according to claim 1, characterized in that: The multimodal signal includes at least a current signal, a voltage signal and a power factor signal; Then the step of extracting the features of the multimodal signal of the first operation data and fusing the features of the multimodal signal to generate a multimodal fusion feature includes: A temporal convolutional network is used to extract the local features of the current signal and the voltage signal respectively; Convolutional neural network is used to extract the features of power factor signal: The local features of the current signal, the local features of the voltage signal, and the features of the power factor signal are fused through a fully connected layer to generate a multimodal fusion feature.

5. The non-intrusive load identification method according to claim 4, characterized in that: The feature graph includes at least a learnable cyclic graph model feature graph, a learnable Gram matrix feature graph, and a two-dimensional feature graph; Then the generating of a feature map based on the multimodal fusion feature, and forming an image load signature from the feature map, includes: Using a fully connected network to learn the relationship between multimodal fusion features, generating the learnable cyclic graph model feature graph; Calculating the inner product between the multimodal fusion features, extracting the nonlinear features of the inner product between the multimodal fusion features using a nonlinear activation function, and generating the learnable Gram matrix feature map; Flattening the multimodal fusion features into a one-dimensional vector, and generating the two-dimensional feature map through a fully connected network; The image load signature is formed by at least one of the cyclic graph model feature map, the learnable Gram matrix feature map, and the two-dimensional feature map.

6. The non-intrusive load identification method according to claim 3, characterized in that: Before the supervised training of the load recognition model based on the image load signature, the method further includes: The load identification model is pre-trained using samples without electrical equipment labels to obtain initial weights of the load identification model.

7. The non-intrusive load identification method according to claim 1, characterized in that: The supervised training of the load recognition model based on the image load signature includes: Multi-label classification training is performed by minimizing cross entropy as the first loss function of the load recognition model; the input of the load recognition model is the image load signature, and the output is the composition of various types of loads.

8. The non-intrusive load identification method according to claim 7, characterized in that: Also includes: Using the real-time third operating data to perform online training on the load identification model; The adopting the real-time third operating data to perform online training on the load identification model includes: collecting the third operation data in real time; Acquire the image load signature corresponding to the third operation data; An incremental learning algorithm is used to input the image load signature corresponding to the third operating data into the load identification model for classification training; a second loss function of the load identification model adds a regularization term to the first loss function.

9. The non-intrusive load identification method according to claim 1, characterized in that: The image load signature is input into the trained load recognition model for classification to obtain a total load label; Decomposing the total load corresponding to the second operation data according to the total load tag, and identifying the load type corresponding to each load component in the total load, including: Inputting the image load signature into the trained load recognition model to predict the total load label of the image load signature; A variational autoencoder is used to decompose the total load corresponding to the second operating data into load components corresponding to each electrical device, thereby identifying the load type corresponding to each load component.

10. A non-intrusive load identification device, characterized in that: include: A first collection unit, used to collect first operation data including different types of electrical equipment; an extraction unit, configured to extract features of a multimodal signal of the first operation data, and fuse the features of the multimodal signal to generate a multimodal fusion feature; An imaging unit, used for generating a feature map based on the multimodal fusion feature, wherein the feature map constitutes an image load signature; A training unit, configured to perform supervised training on a load recognition model based on the image load signature; A second collection unit, configured to collect in real time second operation data of different types of electrical equipment, and obtain the image load signature corresponding to the second operation data; An identification unit, used for inputting the image load signature into the trained load identification model for classification to obtain a total load label; The total load corresponding to the second operating data is decomposed according to the total load label, and the load type corresponding to each load component in the total load is identified.

Images